137
THE DEVELOPMENT AND EVALUATION OF A MODEL OF TIME-OF-ARRIVAL UNCERTAINTY by BECKY L. HOOEY A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Graduate Department of Mechanical and Industrial Engineering University of Toronto © Copyright by Becky Hooey 2009

The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

THE DEVELOPMENT AND EVALUATION OF A MODEL OF

TIME-OF-ARRIVAL UNCERTAINTY

by

BECKY L. HOOEY

A thesis submitted in conformity with the requirements

for the degree of Doctor of Philosophy

Graduate Department of Mechanical and Industrial Engineering

University of Toronto

© Copyright by Becky Hooey 2009

Page 2: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

ii

The Development and Evaluation of a Model of Time-of-Arrival Uncertainty

Becky L. Hooey

Doctor of Philosophy, 2009 Department of Mechanical and Industrial Engineering, University of Toronto

Abstract

Uncertainty is inherent in complex socio-technical systems such as in aviation, military, and surface

transportation domains. An improved understanding of how operators comprehend this uncertainty is

critical to the development of operations and technology. Towards the development of a model of time of

arrival (TOA) uncertainty, Experiment 1 was conducted to determine how air traffic controllers estimate

TOA uncertainty and to identify sources of TOA uncertainty. The resulting model proposed that operators

first develop a library of speed and TOA profiles through experience. As they encounter subsequent

aircraft, they compare each vehicle!s speed profile to their personal library and apply the associated

estimate of TOA uncertainty.

To test this model, a normative model was adopted to compare inferences made by human observers to

the corresponding inferences that would be made by an optimal observer who had knowledge of the

underlying distribution. An experimental platform was developed and implemented in which subjects

observed vehicles with variable speeds and then estimated the mean and interval that captured 95% of

the speeds and TOAs.

Experiments 2 and 3 were then conducted and revealed that subjects overestimated TOA intervals for

fast stimuli and underestimated TOA intervals for slow stimuli, particularly when speed variability was

high. Subjects underestimated the amount of positive skew of the TOA distribution, particularly in

slow/high variability conditions. Experiment 3 also demonstrated that subjects overestimated TOA

uncertainty for short distances and underestimated TOA uncertainty for long distances. It was shown that

subjects applied a representative heuristic by selecting the trained speed profile that was most similar to

the observed vehicle!s profile, and applying the TOA uncertainty estimate of that trained profile.

Page 3: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

iii

Multiple regression analyses revealed that the task of TOA uncertainty estimation contributed the most to

TOA uncertainty estimation error as compared to the tasks of building accurate speed models and

identifying the appropriate speed model to apply to a stimulus. Two systematic biases that account for

the observed TOA uncertainty estimation errors were revealed: Assumption of symmetry and aversion to

extremes. Operational implications in terms of safety and efficiency for the aviation domain are

discussed.

Page 4: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

iv

Acknowledgements

I would like to express my gratitude to the members of my thesis committee (Professors Paul

Milgram, John Senders, and Greg Jamieson) for their guidance and insightful comments throughout

the entire PhD process. I am particularly grateful to Paul who was willing to support a long-distance

mentoring relationship as we attempted to close the chasm between engineering and psychology,

and between theoretical and applied research. I am honoured to have had the opportunity to share

many discussions with John Senders who brought interesting insights to the TOA uncertainty

problem. I am grateful to Greg for inspiring my PhD studies, as it was because of Greg!s seminar at

NASA Ames Research Center in 2003 that I pursued graduate studies at the University of Toronto. I

also extend my appreciation to my external committee members, Professors Ian Spence and Esa

Rantanen, for their time and effort.

I have been very fortunate to have the support of my colleague and friend, Dr. David Foyle of NASA

Ames Research Center, who supported this endeavour in every way possible including: endless

scientific discussions about time of arrival uncertainty, experimental methods, and data analyses;

facilitating the development and testing of the experiments; and his continual moral support and

encouragement. I would also like to thank Ron Miller for his programming expertise and support in

developing the experimental platform.

I am very grateful for the financial support that I received from the following sources: NSERC (Canada

Graduate Scholarship), Queen!s University (Marty Memorial Scholarship), University of Toronto

(Dissertation Completion Grant), and NASA Ames Research Center (support for the human-in-the-

loop experiments).

Finally, I would like to thank Brian Gore for sharing in all of the trials and tribulations along the way.

This could not have been completed without Brian!s support and encouragement.

Page 5: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

v

The Development and Evaluation of a Model of Time-of-Arrival Uncertainty

Table of Contents

!"#$%&'$((((((((((((((((((((((((((((((((((((((((((((((((((((((( ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((())!

!'*+,-./01/2/+$# (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((()3!

45!6789:;<::7=>8?@A?!99=B!C:D7@!E:FG4897!=G7H:=G:4@>6C8I::J@4=@?7845G=4!C:JHJ78>J(((((((((((((((;!

;(;:=G79@KF47=@G((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((;!

;(;(;:7)2/:,L:!%%)3&.:F+'/%$&)+$M:)+:$N/:!)%:7%&LL)':>&+&1/2/+$:JM#$/2 (((((((((((((((((((((((((((((((((((((((((((((; !

;(;(O::7)2/:,L:!%%)3&.:K/')#),+:JPQQ,%$:JM#$/2#((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((R!

;(;(R::K/')#),+:>&*)+1:&+0:J)$P&$),+:!-&%/+/## (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((S!

;(;(S::K/L)+)+1:F+'/%$&)+$M(((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((T!

;(O::98J8!945:@UV847=B8 ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((W!

45!6789:O<::8I689=>8G7:;:X:!G:=GB8J7=Y!7=@G:!GK:698C=>=G!9H:>@K8C:@A:7@!:

FG4897!=G7H:=G:!B=!7=@G:JF9A!48:D7!I=E:@689!7=@GJ ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((Z !

O(;::!B=!7=@G:JF9A!48:D7!I=E:@689!7=@GJ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((Z!

O(O:8I689=>8G7:[=75:!=9:79!AA=4:4@G79@C:D!74E:JFUV847:>!7789:8I6897J((((((((((((((((((((((((((( ;\!

O(O(;::>/$N,0(((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( ;;!

"#"#$#$!%&'()*)+&,(-################################################################################################################################################################## $$!

"#"#$#"!!.++&'&(/- #################################################################################################################################################################### $$!

O(O(O::6%,'/0P%/ (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( ;;!

"#"#"#$!!0,('12/*()1,################################################################################################################################################################ $$!

"#"#"#"!!3)-(&,*456+442!7&8)9'&()1,################################################################################################################################# $"!

"#"#"#:!!;<+4')=4,(&8!>')&8-################################################################################################################################################# $:!

"#"#"#?!!%1-(@-(/2A!349')4B ################################################################################################################################################### $?!

O(O(R::9/#P.$# (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( ;S!

"#"#:#$!!>C.!74'(&),(A!D&(),E-############################################################################################################################################ $F!

"#"#:#"!!0,(4'G)4H-#################################################################################################################################################################### $I!

O(R:!:>@K8C:@A:7@!:FG4897!=G7H:8J7=>!7=@G(((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( ;Z!

O(S::JF>>!9H((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( O;!

45!6789:R<::>875@KJ:!GK:>879=4J:7@:8B!CF!78::7@!:FG4897!=G7H:8J7=>!7=@G(((((((((((((((((((((((((((((((OO!

R(;::!UJ79!47=GY:K=J79=UF7=@G!C:69@6897=8J (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( OO!

R(O::8I689=>8G7!C:>875@KJ:!GK:>879=4J(((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( O]!

R(O(;::8^Q/%)2/+$&.:7&#*(((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( OZ!

R(O(O::>/$%)'#:$,:83&.P&$/:F+'/%$&)+$M:4&.)"%&$),+ ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( R\!

45!6789:S<::8I689=>8G7:O:X:758:8AA847:@A:J688K:!GK:B!9=!U=C=7H:@G::7@!:FG4897!=G7H:

8J7=>!7=@G((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((R_!

S(;::J684=A=4:98J8!945:`F8J7=@GJ:!GK:5H6@758J8J (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( R_!

S(O::8I689=>8G7!C:K8J=YG(((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( RT!

S(R::>875@K ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( RW!

S(R(;::JP"a/'$#(((((((((((((((((((((((((((((((((((((((((((((((( ((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( RW!

S(R(O::>&$/%)&.# (((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((((( R]!

?#:#"#$!!.++&'&(/- #################################################################################################################################################################### :J!

?#:#"#"!!6()=/8) ########################################################################################################################################################################### :J!

Page 6: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

vi

!"#"#$$%&'()*+&) """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" #,!

"#$#$#%!!&'()*+,-(.*'!/'+!0)/.'.'1#################################################################################################################################### $2!

"#$#$#$!!345).46!/'+!7*88*9:;<########################################################################################################################################### "%!

!"!$$-./012/ """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" !3!

!"!"4$$.567896)*$/:))* """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" !#!

!"!"3$$.567896)*$/:))*$;7<*'=5""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" !>!

!"!"#$$.567896)*$2?@"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" !A!

!"!"!$$.567896)*$2?@$;7<*'=5 """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" BC!

!"!"B$$.567896)*$2?@$;7<*'=$/D88)6&D""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" B3!

!"B$$EF/G0//F?H """""""""""""""""""""""""""""""""""""""""""""""" """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" BB!

!"B"4$$/+889&D$'I$-)5+J65"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" BB!

!"B"3$$17876967'<5$'I$6K)$G+&&)<6$.L:)&78)<6$9<*$/+MM)567'<5$I'&$N+6+&)$/6+*7)5"""""""""""""""""""""""""" B>!

!"B"#$$F8:J7(967'<5$I'&$6K)$H)L6$O)<)&967'<$@P7967'<$/D56)8"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" BA!

GQ@%2.-$BR$$.S%.-FT.H2$#$U$2?@$0HG.-2@FH2V$./2FT@2F?H$@HE$.S2-@%?1@2F?H"""""""""""""""""""""""""""""""B,!

B"4$$FH2-?E0G2F?H"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" B,!

B"4"4$$@**&)557<M$17876967'<5$'I$.L:)&78)<6$3 """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" B,!

B"3$$-./.@-GQ$W0./2F?H/ """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" >C!

B"#$$T.2Q?E """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" >4!

B"#"4$$/+XY)(65"""""""""""""""""""""""""""""""""""""""""""""""" """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" >4!

B"#"3$$.L:)&78)<69J$E)57M<"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" >4!

=#$#>#%!?@/A4!$/B!!34C48*<.'1!/!D.5)/)E!*6!F<44+!?)*6.84A###################################################################################### G%!

=#$#>#>#!!?@/A4!$5B!!HA(.I/(.'1!0JK!;'-4)(/.'(E!6*)!0)/.'4+!?)*6.84A############################################################### G"!

=#$#>#$#!!?@/A4!$-B!!HA(.I/(.'1!0JK!;'-4)(/.'(E!6*)!L49M!;'()/.'4+!?)*6.84A ############################################### GN!

B"!$$-./012/ """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" >,!

B"!"4$$%K95)$#9R$E)P)J':7<M$9$17X&9&D$'I$/:))*$%&'I7J)5"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" >,!

=#"#%#%!!HA(.I/(4+!O4/'!F<44+########################################################################################################################################### G2!

=#"#%#>!HA(.I/(4+!F<44+!P.'+*9A#################################################################################################################################### NQ!

B"!"3$$%K95)$#XR$.5678967<M$2?@$+<()&697<6D$I'&$6&97<)*$:&'I7J)5 """"""""""""""""""""""""""""""""""""""""""""""""""""""""" A3 !

=#"#>#%#!!HA(.I/(4+!0JK!;'-4)(/.'(E ################################################################################################################################ N>!

=#"#>#>#!!0JK!;'-4)(/.'(E!HA(.I/(.*'!H))*) ################################################################################################################## NG!

=#"#>#$!!HA(.I/(4+!0JK!&'(4)C/8!FEII4()E################################################################################################################### NN!

B"!"#$$%K95)$#(R$.5678967<M$2?@$0<()&697<6D$I'&$0<6&97<)*$%&'I7J)5"""""""""""""""""""""""""""""""""""""""""""""""""""" A, !

=#"#$#%!!&+4'(.6.-/(.*'!*6!0)/.'4+!/'+!;'()/.'4+!?)*6.84A ####################################################################################### RQ!

=#"#$#>!!S*))4-(!&+4'(.6.-/(.*'!*6!F<44+!?)*6.84############################################################################################################ RQ!

=#"#$#$!!F.I.8/).(E!T/(.'1A ##################################################################################################################################################### R%!

=#"#$#"!!HA(.I/(4+!0JK!P.'+*9######################################################################################################################################## R>!

B"B$$G?HG10/F?H$@HE$?%.-@2F?H@1$FT%1FG@2F?H/ """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ZA!

GQ@%2.-$>R$$T?E.1$-.NFH.T.H2 """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""Z,!

>"4$2FT.$?N$@--F[@1$0HG.-2@FH2V$T?E.1$-./[email protected]$@HE$-.NFH.E """""""""""""""""""""""""""""""""""""""""""""""" Z,!

>"3$$/?0-G./$?N$2FT.$?N$@--F[@1$0HG.-2@FH2V$./2FT@2F?H$.--?-""""""""""""""""""""""""""""""""""""""""""""""" ,C!

>"#$$G@0/./$?N$2FT.$?N$@--F[@1$0HG.-2@FH2V$./2FT@2F?H$.--?-"""""""""""""""""""""""""""""""""""""""""""""""""" ,4!

>"#"4$$@55+8:67'<$'I$/D88)6&D """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ,4!

>"#"3$$@P)&57'<$6'$.L6&)8)5""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ,!!

GQ@%2.-$AR$$/0TT@-V\$1FTF2@2F?H/$@HE$FT%1FG@2F?H/""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""",>!

A"4$$T?E.1$@//0T%2F?H/$@HE$1FTF2@2F?H/ """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ,A!

A"4"4$$/:))*$75$H'&89JJD$E756&7X+6)*"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ,A!

A"4"3$$@P97J9X7J76D$'I$@((+&96)$/:))*$-)9*'+65 """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ,Z!

Page 7: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

vii

!"#"$%&'()%*+%,--'./0%123)-4/'245%674'(/4'*2%,89)-)7%4*%49)%:-*;*7)8%&<*=74);%:-*3)77%>*8)0

""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ??!

!"@%%A:6B,&CAD,E%C>:ECF,&CADG%HAB%,IC,&CAD""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#JJ!

!"@"#"%%C(;0'3/4'*27%+*-%KL(/2%6--*- """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#JJ!

!"@"@%%C(;0'3/4'*27%+*-%KL(/2=F)24-)8%,L4*(/4'*2%4*%GL;;*-4%,'-%&-/++'3%F*24-*0 """""""""""""""""""""#J@!

!"$%%H1&1B6%MCB6F&CADG """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#JN!

FK,:&6B%OP%%B6H6B6DF6G """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#J!!

,::6DMCF6G""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""##$!

,::6DMCQ%,#P%%6Q:6BC>6D&%#%:BA&AFAE """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""##$!

,::6DMCQ%R@P%%6Q:6BC>6D&%@%M6>ASB,:KCF%T16G&CADD,CB6""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""##O!

,::6DMCQ%R$P%%6Q:6BC>6D&%@%M6RBC6H%T16G&CADG"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""##?!

,::6DMCQ%RUP%>6&KAM%HAB%M6&6B>CDCDS%,F&1,E%G:66M%,DM%&A,%GFAB6G"""""""""""""""""""""""""""""""""#@J!

,::6DMCQ%F#P%%6Q:6BC>6D&%$%:BA&AFAE """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#@#!

,::6DMCQ%F@P%%6Q:6BC>6D&%$%M6>ASB,:KCF%T16G&CADD,CB6""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#@O!

,::6DMCQ%F$P%%6Q:6BC>6D&%$%M6RBC6H%T16G&CADG """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""#@?!

Page 8: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

1

CHAPTER 1: TIME-OF-ARRIVAL (TOA) UNCERTAINTY IN COMPLEX

SOCIO-TECHNICAL SYSTEMS

1.1 INTRODUCTION

The need to predict time of arrival (TOA) is becoming increasingly prevalent in complex

socio-technical systems. For example, pilots and air traffic controllers predict TOA in order

to identify and prevent potential collisions, taxi dispatchers predict driver arrival times, and

weather forecasters use complex meteorological and oceanographic models to predict the

time at which a hurricane is expected to make landfall. These diverse domains all share the

common problem of TOA prediction uncertainty. In these cases, even when the data used

to make TOA predictions are essentially accurate at the specific point in time that they are

collected, there is some degree of uncertainty inherent in the prediction, since the world is

dynamic and continues to change. That is, even if the current location, course, and speed

are known, an object or event can change course and speed in the future, and this

introduces uncertainty into the predictions.

1.1.1 Time of Arrival Uncertainty in the Air Traffic Management System

The aviation domain has been selected for investigation in the current research in order to

ground it in a real-world problem. In the Air Traffic Management (ATM) system it is

necessary for air traffic controllers (ATC) to have an accurate view of possible future events

to prevent dangerous or undesirable situations and to optimize traffic flow. ATM is greatly

vulnerable to uncertainty, and controllers and pilots are frequently required to make

important decisions based on uncertain or incomplete information, especially when they

involve predictions of future system states. The sources of TOA uncertainty include factors

such as wind shifts, weather, route complexity, turbulence, pilots! intentions, pilot variability,

and prediction time span. Erroneous interpretation of this uncertain information carries with

it a potential for serious consequences.

Many attempts have been made to minimize uncertainty in the aviation system, such as

reliance on procedures and standardized routes intended to minimize variability and

increase predictability of aircraft movements. Also, great efforts have been taken to improve

the reliability of automation algorithms to minimize uncertainty (Kuchar, 2001; van Doorn,

Page 9: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

2

Bakker & Meckiff, 2001). However, it is not possible to eliminate uncertainty from human-

automation systems altogether, particularly because of the latent variability in human and

environmental factors. Furthermore, this approach may not be sustainable, as new ATM

concepts that require greater precision and greater accuracy, such as reduced separation

minima and coordinated runway crossings, are introduced. These concepts require pilots

and controllers to anticipate trajectories of traffic to facilitate detection and avoidance of

potential conflicts and to make more efficient use of airspace and runways. However, all

these tasks inherently involve uncertainty, as future states within such complex system

cannot be known in advance.

Much research has been conducted to explore planning tasks that require anticipation and

prediction in uncertain, dynamic environments. Boudes and Cellier (2000) studied human

anticipation when controlling dynamic environments (such as ATC or power plants) in which

the operator does not have complete control over the evolution of the system. As per their

definition, !anticipating" involves evaluating the future state of a process and deciding,

according to the subsequent representations developed, which action is to be undertaken

and when. It is, of course, necessary to have an accurate view of possible future events to

prevent dangerous or undesirable situations and to better prepare for favourable conditions

for traffic evolution and optimization.

However, prediction and anticipation are not functions that people perform with high

precision, given known perceptual and cognitive limitations. For example, human operators

are limited in their ability to integrate information over time. Such tasks normally require

mental computation to extrapolate the future from the present and past, and may require that

future estimates be stored in working memory (Wickens & Hollands, 2000). Due to limits on

memory, people tend to give undue weight to early cues (primacy), those that support the

initial hypothesis (anchoring), and very recent cues (recency) (Kahneman, Slovic & Tversky,

1982; Wickens & Hollands, 2000). Furthermore, humans are not good estimators of

variability and are limited in their ability to perceive trends emerging in data that require

extrapolation of exponential or accelerating growth trends.

Non-specialized subjects (Schiff & Oldak, 1990) and trained experts, such as air traffic

controllers (Boudes & Cellier, 2000), also tend to exhibit a conservative bias in that they

make decisions that are risk averse. For example, the conservative bias is demonstrated by

Page 10: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

3

an air traffic controller who is unaware of the exact location or speed of an aircraft and

leaves very large !space envelopes" around each aircraft. In essence, this decision bias is a

way of preparing optimal solutions to absorb unforeseen events, because it allows for a time

margin for solving problems. With large space envelopes, the controller has more time to

issue commands to the aircraft to resolve and avoid potential conflicts. However, it is

important to note that, although functional, the bias may also produce temporal errors and

introduce risk into the ATM system. For example, the temporal errors caused by the

conservative bias may lead controllers to ignore potential conflicts because he/she may

have a false sense that the aircraft are safely separated despite his/her lack of actual

knowledge of the aircraft"s precise location or speeds. The controller"s !safer" space

envelopes may also increase system risk in some cases, for example when the pilot must

carry out a go-around manoeuvre rather than land the aircraft to maintain the controller"s

conservative space envelope. Furthermore, the conservative decision bias may also limit

system efficiency and throughput by creating the need for holding patterns and departure

queues, both of which may be particularly undesirable as we face increasingly congested

airspace and runways.

1.1.2 Time of Arrival Decision Support Systems

To support TOA prediction in dynamic, uncertain environments, Decision Support Systems

(DSSs) are being developed and introduced into many complex domains, in an attempt to

reduce operator workload, improve decision-making, and reduce errors. For example,

automated conflict detection systems have been developed using predictor algorithms and

dynamic models to propagate aircraft state forward in time and predict conflicts over time

and space (Kuchar, 2001; Wickens & Morphew, 1997; Johnson, Battiste & Holland, 1999).

This requires some assumptions regarding the future intentions of each aircraft, such as

assuming that each aircraft will continue to fly at its current heading, speed, and altitude.

When such assumptions are violated, however, the inaccuracies may result in unpredicted

conflicts or unnecessary manoeuvres. In reality, in other words, such predictive conflict

detection systems are rarely 100% reliable. A recent estimate suggested that 85% reliability

is typical of a conflict detection system that predicts conflicts with long look-ahead times (i.e.,

20 minutes) in airspace subject to uncertainties arising from turbulence and future pilot

control actions between the time of alert and the predicted conflict (Xu, Wickens &

Rantanen, 2004).

Page 11: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

4

Much of the research on automation reliability (see Wickens, 2000) has assumed a cueing

system that notifies the operator of impending events. In these systems, automation is often

considered to be either correct or incorrect – that is, the automation correctly cues the

human operator to an object that exists in the world, or does not. However, a DSS that

provides TOA predictions is more likely to have varying degrees of error, rather than error

versus no error. Yet, despite the known lack of perfect reliability and the inherent uncertainty

due to factors that make perfect prediction impossible (i.e., wind shifts, turbulence, pilot

variability, prediction time span, etc.), information about uncertainty associated with

prediction errors is not normally presented to the human decision maker (For one exception,

however, see Lee & Milgram, 2008).

1.1.3 Decision Making and Situation Awareness

In ill-structured, uncertain, and dynamic environments, automation has been less-than-

successful, and has even created serious problems, as the automation becomes "brittle"

and cannot adapt to unknown or unforeseen situations (Cohen, 1993). It fact, the very

features of real-world environments, such as their ill-structuredness, uncertainty, shifting

goals, dynamic evolution, time stress, multiple players, etc., typically defeat the static,

bounded models typically used by decision support systems (Cohen, 1993). Cohen

suggests that each decision involves a unique and complex combination of factors, which

seldom fits easily into a standard decision analytic template, a previously collected body of

expert knowledge, or a predefined set of linear constraints, and that users are sometimes

better than the aids in their ability to recognize and adapt to such complex patterns.

Hansman and Davison (2000) suggest that when decisions must be made based on

incomplete data [or uncertain information], humans are better at decision making than is

automation that follows rule-based strategies, because human operators can formulate a

more comprehensive picture and detect patterns indicating that a system is straying outside

normal limits. It is proposed that, even though humans are better at managing their tasks in

ill-structured, uncertain, dynamic environments, their performance in these environments

may be improved if provided with visualization tools that enable them to forecast and predict

future spatial-temporal states and their associated uncertainty.

Page 12: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

5

Another rationale for placing the human in the role of estimating TOA uncertainty associated

with a TOA prediction is evident from Wickens! (2000) research on automation reliability. It

has been shown that, when the operator is unaware that an automated system is less-than-

perfectly reliable, performance decrements, such as slowed failure detection time and lower

situational awareness (SA), are observed. This occurs because human operators tend to

become complacent and overly trusting of the automated systems when they should be

monitoring operations, considering alternative hypotheses, or preparing for alternative

courses of action. When users are aware of automation unreliability, however, they tend to

calibrate their cognitive strategies with the actual level of imperfection (Wickens, 2000) and

exhibit improved allocation of visual attention to raw data, as manifested by their visual scan

patterns, ultimately yielding improved performance. Paradoxically, because an operator

may attend more to the raw data underlying an automated cue, or implement a wider scan

pattern when they are uncertain of the outcome, they may actually perform better than when

they are certain that the automation is perfectly reliable.

Endsley, Bolte & Jones (2003) proposed that asking operators to manage TOA uncertainty

explicitly may help operators maintain operator attention and situation awareness, and

actually alleviate some of the negative consequences frequently attributed to increasing

levels of automation. In essence, it may not be desirable to totally eliminate uncertainty from

the environment; rather, it is important for operators to maintain an accurate awareness of

the state of the world in which they are operating, and that includes an awareness of the

level of certainty of the world or system. Endsley et al. proposed that the relationship

between confidence and situation awareness (SA) plays a significant role in the quality of

decisions. If confidence is high, and SA is high, an operator is more likely to achieve a good

outcome. However, given the same level of SA, but lower levels of confidence, the operator

will be less likely to act, choosing to gather more information, and thus be ineffective. Lack

of confidence therefore can contribute to indecisiveness and delays and the operator in this

case may be less likely to revise a plan when appropriate. On the other hand, if an operator

with poor SA recognizes that it is poor (i.e., has low confidence) he will correctly choose not

to act or will continue to gather more information to improve SA, and avert a potentially bad

outcome. The worst situation is the operator who has poor SA, but high confidence in his

erroneous picture of the world. This operator will likely act boldly but inappropriately. A

critical issue, according to Endsley et al., is ensuring not only that operators have a good

Page 13: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

6

picture of the situation, but also that they are able to maintain an appropriate amount of

confidence with respect to that picture. As such, it is very important that operators

understand the degree and nature of uncertainties in the system.

Endsley et al. discussed the relationship between situation awareness (SA) and decision

making (Endsley, Bolte & Jones, 2003), together with the task/system factors and individual

factors that influence both of them. They also showed that system capability, interface

design, operator stress, workload, task complexity, and degree of automation can influence

an operator!s SA, decisions, and actions. According to Endsley et al., SA comprises three

levels. Level 1 refers to perception of the elements of the current situation. In an aviation-

related TOA scenario, an operator would be considered as having good level 1 SA if he/she

knows the location of each aircraft, its speed, and distance to go. Level 2 refers to

comprehension of the current situation, or the ability of the operator to determine whether

the aircraft are ahead, behind, or on schedule. Level 3 refers to projection of future states,

or the ability for the operator to survey the airport surface and predict whether all of the

aircraft under his control will arrive at the final destination on time, early, or late, based on

potential traffic and/or route complexities.

1.1.4 Defining Uncertainty

Before embarking on research to determine how operators comprehend TOA uncertainty, it

is imperative to first define TOA uncertainty. There is no clear consensus in the literature as

to how to define or quantify uncertainty associated with a TOA prediction, or what factors

should be included in an uncertainty metric. As a starting point, Schunn, Kirschenbaum and

Trafton (2005) have produced a potentially useful taxonomy of informational uncertainty that

is arguably the most comprehensive to date. Informational uncertainty is defined as

objective uncertainty in the information a person uses to make decisions. Drawing on

expertise in three domains – weather forecasting, submarine / sonar operations, and

medical imaging – they defined four classes of informational uncertainty that are prevalent

and generalize across many complex domains and labelled them as Physics,

Computational, Visualization, and Cognitive uncertainty. The focus of this research is

cognitive uncertainty; however at the same time it is acknowledged that this interacts with

several other forms of uncertainty, such as future prediction uncertainty, and physical

Page 14: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

7

uncertainty associated with aircraft dynamics and environmental factors.

Physics Uncertainty refers to uncertainty in the raw, measured data. This is further

subdivided into three sub-types: Absence of signal being measured, noise or bias in the

signal, and noise or bias introduced as the signal is transduced (or converted into data).

Computational Uncertainty refers to the computational procedures that are applied to data

before they are presented to the human operator. There are three ways that computational

procedures can add new sources of uncertainty. The first (and the focus of the present

research) is future prediction uncertainty. Data are collected and may be accurate at a

specific point in time; however, the world is dynamic and continues to change.

Computational procedures either make no corrections for these changes or make only

partially accurate corrections, and this introduces a potentially large source of uncertainty. A

second form of computational uncertainty is statistical artefact uncertainty, in which

statistical algorithms such as data smoothing introduce artefacts in the displayed data and

distort reality. A third type of computational uncertainty is that related to the use of fast and

cheap algorithms that present the current best guess so that the problem solver can make

quick (though approximate) judgements if necessary. The third category of uncertainty

information within the taxonomy is Visualization Uncertainty. Here, the presentation of data

in the form of a map, table, or graph creates uncertainty, either by failing to represent

relevant information or by presenting it in a misleading way. The final category is Cognitive

Uncertainty. Inclusion of this category in the taxonomy acknowledges that humans, in acting

as an encoding device, information storage/retrieval device, and procedure enactors, are a

key, but error-prone, part of the information system. As possible errors are introduced into

these tasks, sources of information uncertainty are introduced. Cognitive uncertainty can be

created because of perceptual errors, memory encoding errors, information overload,

retrieval errors, background knowledge errors, and skill errors1.

1.2 RESEARCH OBJECTIVE

Given that uncertainty is inherent in systems, it is important to consider how operators in

general comprehend such uncertainty, and to identify conditions under which an operator

1 Because skilled behaviour is conventionally modelled as being automatic, "skill errors" should arguably not be

included as a type of cognitive error.

Page 15: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

8

believes there to be more or less uncertainty than actually exists in the world. With regards

to TOA uncertainty in particular, an improved understanding of how operators comprehend

this uncertainty and the biases inherent in the process is critical to the development of

operations and technology in complex, uncertain, dynamic environments that involve

estimating time of arrival. A model of TOA uncertainty estimation is a critical first step

toward the development of user-centred DSSs. It is expected that the findings of this

research will extend to operators who make decisions in a variety of complex, dynamic, and

uncertain environments – i.e., not only aviation, but also process control, surface

transportation, and military command and control domains.

Page 16: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

9

CHAPTER 2: EXPERIMENT 1 – AN INVESTIGATION AND PRELIMINARY MODEL OF

TOA UNCERTAINTY IN AVIATION SURFACE (TAXI) OPERATIONS

2.1 AVIATION SURFACE (TAXI) OPERATIONS

The concept of TOA uncertainty is applicable to many domains; however, it is helpful to work

in a single problem domain in order to ground the problem in reality. The chosen domain is

that of airport surface (taxi) operations. In current day airport surface operations, a

controller is responsible for both issuing taxi clearances and monitoring aircraft for

conformance to the issued taxi route. Controllers monitor for potential conflicts and act in a

tactical / reactive manner by issuing a hold command to aircraft when needed to prevent

conflicts. The ability of a controller to predict the time of arrival of an aircraft is limited, mostly

because the controller lacks the tools and information needed to make such predictions.

This leads to ultra-conservative, risk averse behaviour, as exhibited by the tendency for

controllers to require larger than necessary separations between aircraft. For example, when

clearing taxiing aircraft to cross an active runway, the controllers tend to build a queue of

several aircraft to cross the runway all at once during a single large break in arriving traffic,

rather than attempting to sequence the taxiing aircraft in a coordinated fashion between

arriving aircraft. This ultimately reduces airport capacity and increases the delay

experienced by passengers on the airport surface, with the first aircraft in the queue often

waiting 20 minutes or more. Furthermore, an aircraft that is stopped requires twice as long to

!spool up" and cross a taxiway/runway than an aircraft that arrives !just in time" without

stopping. This effectively doubles runway occupancy time and reduces arrival and

departure throughput. Upon arrival at its destination (departure runway or gate), an aircraft

will often have to stop and wait (either for take-off or for the gate to become available). This

creates inefficiencies, as a stopped aircraft may block a gateway, a taxiway, or runway exit

and restrict movements.

The future aviation system is moving to 4-D operations2, which will involve the issuance of a

time- or speed-based clearance by the controller, with the additional requirement to monitor

2 The term "4-D operations" is currently used in the aviation industry in reference to the time/speed

element; however, in airport surface (taxi) operations, the third dimension, altitude, is of course constant.

Page 17: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

10

for TOA conformance. It is expected that pilots will be required to arrive at the runway (for

crossing or take-off) within a specified TOA window. The acceptable duration of this

window, and therefore the precision requirements, have not yet been determined.

Currently, automated systems are being developed to assist the controller in issuing

commanded TOA, and also to predict aircraft TOA, but these predictions will be inherently

uncertain, given the dynamic nature of the environment. Pilots! taxiing ability, weather,

traffic, and route complexity are among the factors that may cause an aircraft to alter its

speed and thus may influence its probability of arriving on time. The degree to which

controllers can effectively estimate TOA uncertainty will impact their situational assessment

and decision-making abilities, yielding safer, more efficient operations.

As a first step in creating a model of how ground controllers actually deal with TOA

uncertainty in their environment, experiment 1 was carried out with expert ground

controllers. The goal of this experiment was to understand factors that are relevant to

estimating TOA uncertainty in real-world settings and which should be considered for further

evaluation in future studies. A second goal was to gain an understanding of the task of TOA

uncertainty estimation, as is characterized in a preliminary process model of TOA

uncertainty estimation later in this chapter, and evaluated in more rigorous experimental

studies presented in Chapters 4 and 5.

2.2 EXPERIMENT WITH AIR TRAFFIC CONTROL (ATC) SUBJECT MATTER EXPERTS

This experiment explored factors that contribute to air traffic controllers! certainty that a time-

based taxi plan issued to pilots would be carried out as scheduled. The experiment was

conducted in two parts: First a desktop simulation experiment was conducted to collect

subjective ratings, given a number of factors believed to influence TOA uncertainty. Second,

the same controllers participated in a debrief interview to further understand the factors that

contribute to TOA uncertainty in actual current-day and potential future time-based taxi

operations.

Page 18: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

11

2.2.1 Method

2.2.1.1 Participants

One current and three recently retired air traffic controllers with extensive major airport tower

experience were recruited. Their mean age was 56.75 years and their mean years of

experience at major international and domestic airports was 23.5. All participants had

worked in local controller and ground controller positions, while two of the participants had

additional experience in TRACON (tower control). They were all familiar with tower

operations and procedures and common tower radar and computer systems.

2.2.1.2 Apparatus

Static pictures of typical taxi routes (as shown in Figure 2-1) at a complex airport, Dallas

Forth Worth (DFW), were presented using SuperLab software on a desktop Macintosh G3

computer with a 23” Apple cinema display. Each picture depicted a taxi route (highlighted in

magenta) and an aircraft icon (yellow triangle) that showed the aircraft!s position along the

route. The time that had elapsed since the aircraft started the route was shown in a white

box beside the aircraft icon (shown as 2 minutes in Figure 2-1). Each taxi route ended at a

runway that was highlighted in yellow. An assigned runway crossing time window (a window

of time within which the aircraft had to reach the runway threshold) was presented in a

yellow box beside the runway crossing point (shown as 5:41 to 6:11 in Figure 2-1).

2.2.2 Procedure

2.2.2.1 Introduction

The instructional protocol for this experiment is presented in Appendix A1. The concepts of

coordinated runway crossings and time-based taxi clearances were explained to each

participant, as well as the rationale in terms of both system-wide safety and efficiency

benefits. Participants were told that, if an aircraft arrives early at the runway-to-be-crossed,

the aircraft will have to stop and wait for the assigned crossing time, which doubles the time

required for an aircraft to cross a runway due to engine spool-up times. Furthermore,

arriving late will mean that the aircraft has missed the crossing window and this will

inevitably cause a delay.

Page 19: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

12

Figure 2-1. Experiment 1 stimulus: Dallas Fort Worth (DFW) Airport layout, shown with an aircraft (yellow triangle) taxiing along a taxi route (magenta path) en route to a to-be-crossed runway (yellow highlight). Elapsed time is shown in the white box and assigned runway-crossing window is presented in the yellow box.

2.2.2.2 Distance/Speed Calibration

In order to calibrate the air traffic controllers to the approximate times required to taxi along

different routes at DFW, participants were shown four different taxi routes overlaid on the

DFW airport (as in Figure 2-1 above). Two routes (6000! and 12,000!) running east/west and

two running north/south across the airport were displayed. The time to complete each route

at three different speeds (16 kts, 20 kts, and 24 kts) were also provided, assuming constant

speed, as shown in Table 2-1, along with the graphical route depictions. The controllers

were instructed that typical taxi speeds ranged from 16 to 24 kts, with an average speed of

20 kts.

Page 20: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

13

Table 2-1. Distance and TOA for each route distance and speed combination.

Time to complete route per route distance (in Minutes:Seconds)

Taxi Speed

6,000! 12,000!

16 kts 3:42 7:24

20 kts 2:58 5:56

24 kts 2:28 4:56

2.2.2.3 Experimental Trials

Subjects were shown 162 static pictures of taxi plans in progress as in Figure 2-1. Each

picture showed the taxi plan and an aircraft!s progress in carrying out the plan – both its

current location along the route and the elapsed time since the aircraft started taxiing. Each

picture was a combination of the five independent variables presented in Table 2-2: route

distance, route complexity, runway crossing window size, position along the route, and

speed. Route Distance was expected to influence TOA uncertainty estimations, with longer

routes producing lower TOA certainty than shorter routes. Route complexity was

represented as the number of turns required to complete the route. Turns add variability to

the taxi speed, as pilots must reduce speed to safely manoeuvre the turn and accelerate

after the turn to resume taxi speed. It was expected that the increased speed variability

would increase TOA uncertainty. The duration of the runway-crossing window was a

manipulation of the degree of precision required by the taxiing pilot. It was expected that

controllers! TOA uncertainty would be higher when the aircraft was required to make a very

small crossing window than for a larger crossing window. The aircraft!s position along the

route was also manipulated. It was expected that TOA uncertainty would be larger for

aircraft at the start of their route than for those that had progressed appropriately along their

route towards their destination. The final independent variable was the nominal taxi speed

of the aircraft.

The subjects! task was to assess the information available in the depicted taxi plan and

provide a rating of how certain he/she was that the aircraft would arrive at its destination

within the allotted runway crossing window – neither too early, nor too late. Following the

sizeable literature (e.g., see Soll & Klayman, 2004; Block & Harper, 1991) that use subjects!

Page 21: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

14

ratings of certainty as a proxy for estimating probability, subjects in this experiment provided

ratings on a seven-point scale from 1 (Very Low Certainty) to 7 (Very High Certainty) by

entering a number on a standard keyboard, and were encouraged to use the entire range of

the seven-point scale. Subjects were told to assume that it was a clear day at DFW airport

and that no traffic conflicts were anticipated. Also, they were told that the aircraft represent

a Boeing 757 that taxis at 20 kts on average but that any number of factors could influence

their ability to maintain a constant 20 kts.

Table 2-2. Independent Variables

Independent Variable Levels

Route Distance 6000!, 12000!

Route Complexity 1, 2, 3 turns per 6000!

Runway Crossing Window Size 15, 30, 45 seconds

Position along route Start, 1/3rd, 2/3rd

Taxi Speed 16, 20, 24 kts

Subjects first completed 18 practice trials that exposed them to each level of the five factors.

Then, after a short break, subjects completed 162 experimental trials comprising each

combination of the five factors. The order of the 162 trials was randomized, and they were

presented in 4 blocks, each separated by a short break. The task was self-paced.

2.2.2.4 Post-study Debrief

Upon completion of the experimental trials, a semi-structured interview was conducted with

each of the ground controllers to better understand their current tasks, how these might

change with the introduction of 4-D clearances, and their perceptions of TOA uncertainty on

the airport surface. Specifically, the purpose of the interviews was to explore factors that

contributed to their perceptions of certainty that a time-based taxi plan once issued to a pilot

would be carried out as scheduled. The interview questions are presented in Appendix A2.

2.2.3 Results

Prior to presenting the results of this experiment, is it noteworthy to recall that the main goal

of this experiment was to determine which factors are relevant to the task of estimating TOA

Page 22: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

15

uncertainty, for the purposes of developing a model of the TOA uncertainty estimation task

for future empirical evaluation.

2.2.3.1 TOA Certainty Ratings

Recall that the dependent measure was a subjective rating of TOA certainty, with lower

numbers representing low certainty, and higher numbers representing higher certainty. The

mean and standard error of the ratings for each level of the five independent variables are

presented in Table 2-3. A Distance (3) x Complexity (3) x Runway window crossing size (3)

x Position (3) x Speed (3) repeated measures analysis of variance (ANOVA) was conducted.

Given the exploratory nature of this experiment, an alpha level of 0.1 was used in

interpreting these results.

There was some evidence that, as route complexity (number of turns) increased, the

subjects! TOA certainty ratings decreased, F(2,4)=5.00, p=.08. In particular, the high

complexity condition (3 turns) yielded lower certainty scores than both low (1 turn) and

medium complexity (2 turns; p<.05). Recall that route complexity represents speed

variability, thus this suggests that factors that contribute to speed variability on the airport

surface may reduce TOA certainty. (These factors were explored further in the debrief

interviews, as discussed in section 2.2.2).

Table 2-3. Certainty Ratings Mean and Standard Error

Independent Variable Condition Mean Rating

Standard Error

Route Distance 6000!

12000! 4.9 5.2

.627

.415

Route Complexity 1 turn 2 turn 3 turn

5.3 5.2 4.7

.367

.602

.566

Runway Crossing Window 15 sec 30 sec 45 sec

4.9 5.1 5.1

.496

.534

.502

Aircraft Position Start 1/3rd 2/3rd

5.3 4.6 5.3

.617

.406

.555

Speed 16 20 24

5.5 5.1 4.5

.388

.424

.730

Page 23: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

16

There was a small, but significant, increase in subjectively rated TOA certainty as a function

of runway crossing window size, F(2,4) = 15.45, p=.013, with lower certainty ratings for the

15 sec window than both the 30 and 45 second window. This finding suggests a possible

threshold level associated with the required TOA precision at which the subjects! certainty

associated with an aircrafts! TOA may drop.

Speed yielded interesting results, with subjects! certainty scores decreasing as speed

increases, F(2,4)=5.829, p=.065. Comments made by the subjects during the debrief

interview suggested that certainty scores were lower under high-speed conditions because

they were unsure if the aircraft could maintain the high speed throughout the entire route,

especially given the required turns in the route.

Although there was no significant effect of route distance, there was a significant effect of

the aircraft position along the route on the subjects! certainty ratings, F (2,4)=4.84, p=.086.

As can be seen in Table 2-3, when the aircraft was 1/3rd of the way along the route, certainty

ratings were lower than when the aircraft was either at the start or 2/3rd along the route.

This suggests that the most uncertainty was perceived when the aircraft had not travelled

very far (thus the ability to assess mean speed was limited) and when there was a longer

distance remaining (thus the potential for increased variability in the speed on the route

remaining).

In summary, turn complexity, or factors expected to increase speed variability reduced the

subjects! certainty ratings, as did aircraft speed, short runway window crossing sizes (or high

TOA precision requirements), and unknowns regarding aircrafts! speed variability when only

a short distance has been travelled and larger distance remained. These results indicate

that future empirical studies should manipulate speed variability and distance remaining.

Related, feedback from the SMEs interviewed after this experiment revealed the need for

dynamic simulations so that subjects could better understand the amount of speed variability

for a given aircraft. Showing the aircraft position along the route with elapsed time (see

Figure 2-1) provided no indication as to whether the aircraft!s speed was constant or

variable.

Page 24: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

17

Also, while the precision required for an aircraft to arrive at its destination (runway crossing

window) affected the subjects! ratings of certainty, results from the debrief interviews

suggested that this was a coarse level of granularity. Future studies should consider using

TOA estimation as a dependent variable rather than an independent variable – that is, rather

than asking subjects to rate their certainty associated with the aircraft arriving within a

certain time window, future studies should ask subjects to estimate the TOA window in

which they are certain the aircraft will arrive.

2.2.3.2 Interviews

The post-study interviews with the air traffic controllers were useful to provide insights into

the ground controller!s task. First, it is clear that, at any given time, a controller is

responsible for monitoring several aircraft simultaneously, often more than 20 aircraft, each

with its own starting point, destination, and route. Second, it was noted that, in time-metered

situations, it is generally beneficial for a controller to detect an aircraft that will arrive earlier

or later than the commanded time, so that the controller can issue a new speed command,

reroute the aircraft, or reroute another aircraft. Third, the uncertainty associated with on-

time arrival is due primarily to the variability of speed of the aircraft, which is determined by

the pilot in control of the aircraft. An important assumption for future operations is that each

pilot will be made aware of the optimal speed to travel and uncertainty with regard to the

TOA will be based on the pilot!s ability to maintain that speed.

The debrief interviews from the four air traffic controllers were synthesized into three

relevant topic areas, as discussed next.

1. How controllers determine TOA and TOA uncertainty

At the outset of this experiment, it was assumed that air traffic controllers calculate time of

arrival, albeit with rough approximations, by considering approximate route distances and

aircraft speeds. However, results of the debrief clearly suggest otherwise. Instead, it was

revealed that controllers develop heuristics to determine how long a taxiing aircraft should

take to complete a given route. These heuristics are built over years of on-the-job

experience. For example, the controllers stated:

Page 25: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

18

• “I never thought about how long a 12000! route at SFO takes. I really don!t

know. I just know by experience how long it takes to get from one side of the

airport to the other”

• “Today, I found myself using math calculations [to do the experimental task]. I

would never do this in the real world.”

• “With experience at an airport, you just know how long it takes to travel a taxi

route”

These TOA heuristics developed by the controllers take a number of factors into

consideration, including the airport layout, aircraft type, airline culture, and traffic flow. For

example, controllers had developed expectations that pilots of a certain airline could

consistently be relied on to make an expedited runway crossing, whereas for others they

would take a different, more conservative approach, because the controller was uncertain

whether those pilots would comply. They cited issues such as corporate culture, experience

at the airport, and aircraft type as factors in their decision. The following are direct quotes

from the air traffic controllers selected to shed light on this process:

• “Pilots flying out of their own base [airport] are very familiar with airport

operations. They know the system and I can be more certain of their TOA.

They are easier than the "once a month! pilots, who tend to be more cautious

and apprehensive”.

• “We [ATC] have running jokes about airlines and which ones will comply to

expedited clearances. Some we know and others … maybe they will, maybe

they won!t”

• “I never told [Airline X] to do anything [with a time element to it] but [Airline Y]

would always be able to do it”

2. Sources of TOA uncertainty

The experiment reported above considered four factors that were expected to contribute to a

controller!s uncertainty of an aircraft!s TOA: route distance, route complexity, runway-

crossing window size, and position along the route. One of the findings of that experiment

was that speed variability, as manipulated by route complexity, was a large contributor to

TOA uncertainty. This finding was echoed by the controllers in the debrief interviews. One

Page 26: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

19

pilot stated “It is the human element that is the big problem. The pilot has to be comfortable

with the visibility, the equipment, and the taxi route/ airport layout. The pilot has the final

say”. Other factors, in addition to those tested as independent variables in experiment 1,

that contribute to speed variability, and that controllers consider in estimating TOA

uncertainty include:

• Traffic patterns/ volume

• Aircraft type

• Weather

• Visibility

• Airline (familiarity with airport, past experience)

• Arrival traffic, separation on final

• Departure queue

• Flow control

• Aircraft equipage (with speed/time displays)

3. Effect of TOA uncertainty on ATC tasks

As aviation operations move from current-day operations to future time-based operations, it

is only reasonable to expect a change to ATC tasks and the way controllers operate. When

asked during the debrief interviews, each controller expressed concern about a need for

increased monitoring. For example, one controller stated that “[in cases of high TOA

uncertainty] controllers would have to monitor the aircraft more closely, and more frequently.

It may require re-routing to ensure conflict-free routing”. Similarly, another controller stated

that “it will change the way controllers monitor”, and also added that this change in

operations would mandate new displays and technologies to support the task. “We would

need more information, such as expected TOA and speed readouts, to know who!s not

going to make their commanded arrival time. A speed readout would be important, so that

we can see if the speed is reasonable and how much it fluctuates”

2.3 A MODEL OF TOA UNCERTAINTY ESTIMATION

Based on the results of this first experiment and its debrief interviews, a preliminary model of

the TOA uncertainty estimation task is proposed. This model will be used to guide

subsequent experiments. It is proposed that the process of estimating TOA uncertainty is a

Page 27: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

20

two-step process, as depicted in Figure 2-2 and described next. Controllers, through years

of experience, develop a library of speed profiles that describe vehicles or situations that

they encounter, based on particular airlines, traffic patterns, or weather conditions. Each

speed profile consists of both an average speed and amount of speed variability (for

example, “this aircraft tends to be fast, but very variable – or consistently slow”). This speed

profile may be generated based on previous experience and expectations developed over

time and exposure, real-time observation, or training. At one extreme, if the variability is

zero, the aircraft will travel at a constant speed, there is no uncertainty in TOA, and the

operator should be able to predict TOA perfectly. At the other extreme, if the aircraft speed

is very variable, the uncertainty will be high, and the controller cannot easily predict TOA.

Figure 2-2. A model of the TOA uncertainty estimation process

During operations, controllers determine if each vehicle matches a known speed profile in

their personal library. For example, debrief comments suggested that controllers considered

all aircraft from a particular airline as sharing the same profile. Generally, airlines that flew

into an airport very infrequently were considered to have a speed profile that was slow and

highly variable, whereas those airlines that were based out of that airport were faster and

more consistent.

Page 28: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

21

Operators then use knowledge of the speed profile (e.g., fast and variable) combined with

an estimate of the distance to travel, and an associated estimate of TOA and TOA

uncertainty to make decisions and act on those decisions, depending on factors such as

personal risk tolerances and consequences of error. These decision-making processes,

however, are beyond the scope of this research effort.

The model must also anticipate the situation in which the operator is faced with a new

vehicle profile with which he/she is not familiar. In this case, it is expected that the controller

would apply a heuristic to estimate TOA uncertainty for unknown profiles, based on their

similarity to known profiles. This was suggested by the controllers! comments, that

suggested a tendency to lump all aircraft from a single airline together – even though each is

piloted by a different pilot, whom they may or may not have encountered before.

2.4 SUMMARY

Experiment 1 resulted in a simple model of TOA uncertainty based on insights gathered

from skilled professionals who engage in TOA uncertainty tasks. The next step in the

process was to evaluate the model using empirical human-in-the-loop experiments. To

accomplish this, first an appropriate experimental method and platform was developed, as is

discussed next in Chapter 3. Two additional experiments, presented in Chapters 4 and 5,

were conducted to evaluate the above TOA uncertainty model, and a refined model is

presented and discussed in Chapter 6.

Page 29: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

22

CHAPTER 3: METHODS AND METRICS TO EVALUATE

TOA UNCERTAINTY ESTIMATION

No literature was identified to date that has attempted to develop experimental methods and

metrics to assess an operator!s comprehension of TOA uncertainty. Indeed, there is a

small body of human factors literature that has investigated uncertainty displays. However,

much of that work was conducted to evaluate the effectiveness of a specific uncertainty

display for a domain-specific need (e.g. Andre & Cutler, 1998; Banbury et al., 1998) and

thus the generalizability of those findings, and methods, is limited. Other researchers, such

as Schaefer, Gizdavu & Nicholls (2004) have evaluated the usability of uncertainty displays

and assessed the impact of the uncertainty display characteristics on operator workload and

situation awareness. Just as with the subjective certainty ratings used in Experiment 1 of

the current research effort (reported in Chapter 2), those studies provided valuable insights,

but they are not sufficient to further our knowledge of the understanding and management of

uncertainty.

To objectively evaluate TOA uncertainty, methods and metrics that compare objective

uncertainty to subjective uncertainty are necessary. Objective uncertainty is a property of

the system under study, in this case derived from the dynamics of the aircraft and other

environmental factors. As such, objective uncertainty can be objectively computed, given

knowledge of the variability in the speeds of the aircraft. On the other hand, subjective

uncertainty refers to estimates by an observer of the system given his/her state of

information at the time. When perfectly calibrated, subjective uncertainty equals objective

uncertainty. On the other hand, if the observer is not well calibrated, he/she will either

overestimate or underestimate TOA uncertainty. Thus, the first challenge of the present

research is to develop an experimental task and metric that will enable evaluation of

uncertainty calibration.

3.1 ABSTRACTING DISTRIBUTIONAL PROPERTIES

The ability to learn the distributional properties of probabilistic information is critical to the

task of estimating TOA uncertainty. In daily experience with probabilistic processes, it is

very rare that we are told explicitly what an underlying distribution is – i.e., whether it is

normal or skewed, or what are the mean and standard deviation that characterize the

Page 30: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

23

distribution (Pitz, 1980). In these cases, it is necessary for an observer to infer properties of

the population from information that is contained in the samples of objects that they observe.

Such tasks require a person to engage in !abstraction," or the process of incorporating into

the representational structure information that has not been presented directly (Pitz, 1980).

Pitz used the term !prototype" to refer to the structure that describes the representation of

information about a population that is inferred from sample information. A prototype is a

theoretical concept that has been employed to describe the end product of abstraction.

Although it is still largely an open question what properties of the population might be

abstracted from a sample, it is known that the process of inferring an average value for

uncertain quantities is a very general process, and one that humans tend to do quite well

(Pitz, 1980; Peterson and Beach, 1967). On the other hand, humans have more difficulty

estimating statistical variance, while skewness and bimodality of a distribution were

recognized and used by the subjects in making their predictions. Pitz concluded that

information about central tendency is abstracted as a prototype; however it appeared that

certain critical features of the sample information are stored directly – such as the extreme

values, the smallest and the largest that have been observed.

Peterson and Beach (1967) found that estimates of variance tend to be influenced by the

mean value of the stimuli. They provide the analogy of comparing the smoothness of the

surface of a forest to that of a desktop. The treetops seem to form a fairly smooth surface,

however, the items on the surface of a desktop would seem very bumpy and variable. This

forest top is far more variable than the surface of the desk, but not relative to the sizes of the

objects being considered. They concluded that, instead of estimating variance, subjects tend

to estimate the coefficient of the variation (the standard deviation divided by the mean).

The issue of how a person represents knowledge about variability and skewness is less

clear. Pitz (1980) proposed that it is possible that measures of variability are inferred and

stored as part of the prototype, along with information about the central tendency. Estimates

of variability can be generated by finding the (absolute) difference between each value and

the currently estimated average value, and taking the average of these differences.

However, such a procedure would place a fairly heavy load on short-term memory and is

unlikely to occur as a spontaneous strategy. Alternatively, it is possible that people use

Page 31: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

24

information about the central tendency together with their knowledge of the extreme values.

The difference between the two extremes provides information about variability – while the

location of the central tendency relative to the extremes provides information about

skewness.

Related to this is the issue of how human operators factor in extreme values and outliers.

Alpert and Raiffa (1982) found that subjects were unwilling to weight large deviations

heavily. They gave five groups of subjects almanac questions. All subjects were asked to

report 25th, 50th, and 75th percentiles. In addition, Group 1 was asked to report 1st and 99th

percentiles; Group 2 was asked to report 0.1th and 99.9th percentiles; Group 3 was asked to

report “minimum" and “maximum" values; and Group 4 was asked to report “astonishingly

low" and “astonishingly high" values. In every case, the estimated spread of the tails of the

distributions was too small, regardless of the definition of the extremes and, although

feedback did improve the spread, it did not completely eliminate the overconfidence bias.3

O!Connor and Lawrence (1989) utilized a task that involved time series predictions and they

found that calibration of subjects! estimated confidence intervals was influenced by the

degree of forecasting difficulty. For simple series, subjects were underconfident (that is, the

subjects! estimated intervals were too large), but for medium to high difficulty series, the

subjects were overconfident (the estimated intervals were too small).

The research discussed above evaluated operators! ability to abstract mean and variance

information of a population from samples of numbers. However, the ability to abstract the

mean and variance of a series of times of arrival (TOAs) has evidently not been considered

in the literature, in spite of the fact that understanding human capability to abstract TOA

information from observing a sample of multiple objects with variable speed is particularly

important to understanding how human operators comprehend TOA uncertainty. This

becomes a particularly interesting problem when one considers that TOA is a non-intuitive

metric, due to the fact that, because TOA varies inversely with speed for a given distance

3 In statistics, a confidence interval is an interval estimate of a particular population parameter. For example, a

95% confidence interval means that it is estimated that 95% of samples from the population will lie within this

interval. If a person’s subjective 95% confidence interval were to be narrower than the actual interval, that is

they underestimated the confidence interval, then it can be said that the subject is overconfident with respect to

the actual corresponding interval.

Page 32: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

25

(T=D/V), TOA, by definition, is not a linear function of speed. Rather it exhibits a function as

depicted in Figure 3-1, which plots TOA as a function of speed for variables that are along

the order of those encountered in the ground controller!s world.

As can be seen in Figure 3-1, because TOA is the reciprocal of speed, slower speeds have

a bigger effect on TOA. That is, a difference of 1 mph at slow speeds (for example, an

increase from 1 to 2 mph) has a large impact on TOA (in this example, assuming a distance

of 1 mile, TOA at 1 mph = 1 hour, while TOA at 2 mph = .5 hour.) However a difference of 1

mph at a higher speed (for example, from 60 to 61 mph) changes the TOA by only .003

hours (TOA at 60 mph = .0167, while TOA at 61 mph = .0164 for a distance of 1 mile).

Related to this, if one considers a range of speeds that spans 5 to 15 mph, one can see in

Figure 3-1 that the resultant range of TOAs is much larger than if the same 10 mph range

were centred around a speed of 85 mph. It follows then that speeds selected from a normal

distribution of slower speeds would yield a more positively skewed4 distribution when

converted to TOAs than would speeds selected from a normal distribution of faster speeds.

Figure 3-1. The relationship between speed (mph) and TOA (hours) assuming distance = 1 mile

4 A positively skewed distribution is one whose tail is elongated to the right; that is, more data are encountered in

the right tail than would be expected in a symmetrical distribution.

!"

!#$"

!#%"

!#&"

!#'"

!#("

!#)"

!#*"

!#+"

!#,"

$"

!" $!" %!" &!" '!" (!" )!" *!" +!" ,!" $!!"

Tim

e o

f A

rriv

al (h

ours

)"

!"#$%&'()*+,-.)Speed (mph)

Page 33: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

26

Furthermore, as the standard deviation (SD) associated with the speed distribution

increases, the positive skew of the resulting TOA distribution would be exaggerated. This

becomes clear if one considers that numbers drawn randomly from a distribution with a

larger SD will consist of more extreme speeds than if drawn from a distribution with a

smaller SD. This is demonstrated in Figure 3-2 that shows in the left panel, histograms

derived from three normal distributions of speed with the same mean (20 mph) and different

standard deviations (2, 4, and 8). The right panel shows histograms of the corresponding

times of arrival (hours) assuming a distance of 1 mile. As can be seen, the positive skew of

the TOA distribution increases considerably as the standard deviation increases from 2 to 8.

Also, consider Table 3-1, which shows the Mean, SD and skewness for distributions of TOA

that are defined by various combinations of means and standard deviations of speed.

Clearly the distribution is much more skewed (skew = 20.1) at lower speeds (M = 20) with a

high standard deviation (SD = 8), than for higher speeds (M = 60) with the same standard

deviation, in which case the skew was 1.2.

Table 3-1. Effect of mean and standard deviation of speed on TOA

Speed Distributions

TOA M 20 SD 2

M 20 SD 4

M 20 SD 8

M 30 SD 2

M 30 SD 4

M 30 SD 8

M 60 SD 2

M 60 SD 4

M 60 SD 8

M .05 .05 .07 .03 .03 .04 .02 .02 .02

SD .005 .01 .23 .002 .005 .013 .0006 .0011 .0025

SKEW .9 2.5 20.1 .5 1.2 3.9 .4 .6 1.2

Notes: For each speed distribution (defined by a Mean and SD of Speed), 1000 samples were drawn using the Regress+ software package, which possesses a normal distribution random number generator. Each speed was converted to a TOA assuming a distance of 1 (i.e., TOA = 1 / speed) yielding 1000 TOAs for each combination of Mean and SD of speed. Descriptive statistics (mean, SD, and skew) were calculated for the 1000 TOAs of each speed distribution condition.

Page 34: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

27

Figure 3-2. Left panel shows three histograms of speeds drawn from a normal distribution with mean speed 20 and (top to bottom) SD=2, 4, or 8. Right panel shows corresponding TOA distributions with increasingly positive skew as standard deviation increases.

!"#$%&%'(%

)*%&%'%

)+,-%&%(%

!"#$%&%.(/0%

)*%&%.((/%

)+,-%&%.1%

!"#$%&%'(%

)*%&%2%

)+,-%&%(%

!"#$%&%.(/'%

)*%&%.(0'%

)+,-%&%'./%

!"#$%&%'(%

)*%&%3%

)+,-%&%(%

!"#$%&%.(40%

)*%&%.(23%

)+,-%&%'(.0%

Speed Time of Arrival

%%%%%.('%%%%%%%%%%%%%%%%%%%.(2%%%%%%%%%%%%%%%%%%.(4%%%%%%%%%%%%%%%%%%.(3%%%%%%%%%%%%%%%%%%.0%%%%%%%%%%%%%%%%%%%.0'%

.(2%%%%%%%%%%%%%%%%%%%%%%%.(/%%%%%%%%%%%%%%%%%%%%%%%%.(4%%%%%%%%%%%%%%%%%%%%%%%%.(5%

%%%%%.((% .'%%%%%%%%%%%%%%%%%%%%%%%%%%%.2%%%%%%%%%%%%%%%%%%%%%%%%%%.4% %%%%%%%%%%%%%%%%%%%%%%%.3%%%%%%%%%%%%%%%

%%%%%%%%Velocity (mph) Time of Arrival (hours) %

Speed Time of Arrival

Speed (mph) Time of Arrival (hours)

Fre

qu

en

cy

Fre

qu

en

cy

Fre

qu

en

cy

Fre

qu

en

cy

Fre

qu

en

cy

Fre

qu

en

cy

Page 35: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

28

3.2 EXPERIMENTAL METHODS AND METRICS

To evaluate the model of TOA uncertainty, experimental task and metrics are required that

capture the essence of the real-world task faced by air traffic controllers in time-based

environments. Based on the findings of the experiment reported in Chapter 2, the task

should possess the following characteristics.

1) Provide an opportunity for subjects to build speed profile models based on

experience. The air traffic controller interviews clearly revealed that, rather than

real-time calculations, controllers estimate TOA and TOA uncertainty based on

expectations, gained through on-the-job experience. To replicate this in a controlled

experiment requires dynamic simulations (as opposed to the static route plans used

in experiment 1 in Chapter 2) that allow subjects to build the experience needed to

generate estimation heuristics. Providing speed read-outs will also facilitate the

subjects! understanding of the vehicle!s speed profile and remove issues

associated with speed perception.5

2) Avoid confounds associated with subjects! previous experiences. Since each

controller has already built expectations based on their unique individual

experiences, using actual controllers and real-world ATC tasks will likely introduce

confounds with previous experience that cannot be experimentally controlled.

Thus, there is a need for experimental controls that dictate the subjects! experience,

rather than relying on the individual experiences that each subject may bring to the

experiment from their own real-world experience, as these are widely variable

across individuals. This will require that speed variability be modelled in a generic

way not tied to expectations of airport operations, airlines, traffic etc.

3) Consider the interaction among speed, distance, and time. To estimate TOA

and TOA uncertainty, air traffic controllers consider both speed and distance.

5 The implications of this premise are discussed in Chapter 7. While issues related to velocity perception are an

interesting aspect of the TOA estimation problem, relying on velocity perception rather than velocity read-outs

would have the potential of confounding the main experimental findings related to how subjects estimate TOA

uncertainty. Furthermore, it can be safely assumed that accurate ground velocity readouts will be available to

both pilots and ATC in future aviation operations.

Page 36: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

29

Furthermore, there are two factors associated with vehicle speed that must be

considered: average speed and speed variability. Inappropriate estimates of TOA

or TOA uncertainty can result from misestimating either speed or distance. This

must be addressed, either by controlling (holding constant) one or more factors or

experimentally manipulating each factor.

4) Approximate ATC workload. While it is not believed to be necessary to replicate

the entire ATC experience by including all task elements in high fidelity, it was

determined that approximating the visual monitoring workload was important. The

interviews with the ATC subjects suggested that ground controllers at major airports

are regularly required to monitor approximately 20 aircraft at a time.

5) Include known and unknown profiles. Controllers must work with both pilots and

airlines that fly into their airports regularly as their base operations and those that

fly in very infrequently. This means that matching the observed speed profile to

those in their library of profiles will be more difficult for those vehicles with

infrequent exposure. The experimental method should address both cases.

3.2.1 Experimental Task

An experiment was designed to simulate the critical elements of the task of an air traffic

ground controller (or other operator), who is responsible for monitoring and predicting the

ability of an aircraft (or object) to arrive at a defined location at a specific time, or within an

allowable time window around a specific time. The task involved several objects, each

travelling along a separate path (with different lengths). All objects started at the same time

and shared a common commanded TOA. Each was assigned a mean speed (which was

computed based on the speed required for on-time arrival) and a standard deviation. Each

object changed speed at pre-determined intervals, with the new speed selected from a

normal distribution defined by the assigned mean and standard deviation. Based on this

distribution, the TOA uncertainty could be computed objectively. Given a symmetrical

distribution of speeds about the mean speed, the expected TOA of each aircraft will be

correct and will not change (that is, if each is given a mean speed at the outset, based on a

proper estimated TOA). However, the probability of arriving within the permitted time

window will change, depending on the variability of the speed.

Page 37: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

30

3.2.2 Metrics to Evaluate Uncertainty Calibration

A metric is required that can be used to determine how well human operators can estimate

measures of central tendency and uncertainty for both speed and TOA of aircraft that move

with variable speed. The approach adopted within the current research was to use a

normative model, in order to identify factors relevant to the inference process for speed and

TOA. In particular, the approach was to examine the relationship between inferences made

by human observers and corresponding inferences that would be made by an optimal

observer who had knowledge of the underlying distribution. This approach is analogous to

the use of the ideal observer used in signal detection theory and !economic man" in

economics (Peterson and Beach, 1967). In this sense, Peterson and Beach argue,

probability theory and statistics fulfil a role similar to that of optics and acoustics in the study

of vision and hearing. Statistics can be used as a theory of the uncertain environment and

provide a basis for a descriptive theory of imperfect human inference (Peterson and Beach,

1967).

Peterson and Beach refer to two measures: accuracy and optimality. Accuracy is assessed

when subjects" estimates are compared to descriptive statistics of an observed sample. For

example, if subjects observe a sample of 20 aircraft, each with a speed drawn from a

population with a known distribution, accuracy would be assessed by comparing the

subjects" estimate of the mean speed of the 20 aircraft to the objective mean of the 20

aircraft speed samples. Optimality, on the other hand, is assessed when subjects"

estimates are compared to parameters associated with the underlying population from which

the sample is drawn (rather than with the computed sample estimates). That is, in

assessments of optimality, the subjects" estimates are compared to the actual mean of the

population distribution. The current experiments set out to assess the optimality of subjects"

estimates. Recall that the simple TOA model presented in chapter 2, based on input from

ATC subjects, posits that operators build experience over many exposures and develop a

library of TOA uncertainty expectations that they then apply, rather than assessing the

variability of each object individually. Thus, it is argued that assessing optimality is the more

appropriate variable to evaluate for the current research effort.

Page 38: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

31

The optimality of subjects! estimates can be explored using any of a number of population

parameters, including measures of central tendency, variance, skewness, and confidence

intervals. Given the fairly consistent finding from existing literature that subjects perform

quite well at the task of inferring means, this was chosen as one measure of optimality for

this experiment, in order to determine if these past findings extend to dynamic stimuli and for

the variables of speed and TOA. Subjects! estimates of speed and expected TOA were

compared to the mean speed and TOA of the population distribution.

More relevant to understanding TOA uncertainty, however, is to assess the optimality of

subjects! estimates of TOA variability. No literature was identified that has attempted to

develop experimental methods or metrics to assess an operator!s understanding or

inference of TOA uncertainty. Nevertheless, we can draw upon the literature on interval

estimation for insights into this problem.

The fractile method (Block and Harper, 1991) is a widely used method (see for example:

Juslin, Wennerholm, Olsson, 1999; Soll & Klayman, 2004; McKenzie, Liersch, & Yaniv,

2008; Teigen, Halberg, & Fostervold, 2007) that requires participants to provide the smallest

intervals that include an existing true value with a pre-stated probability. The participant is

considered calibrated if the proportion of values that fall within the interval equals the given

probability. An example of a question that would be provided to a subject using the fractile

method is: “Give the boundaries of the interval for which you are 90% confident that it

encloses the number of chickens consumed by American households daily” (Juslin,

Winman, & Olsson, 2003). Although there are exceptions, the literature suggests that

subjects tend to be over-confident in their ratings, in that the proportion of true values

included by the subjects! estimated intervals is typically lower than the pre-stated probability

– that is, the intervals tend to be too small, such that less than the pre-stated probability of

the true values falls within the subjects! intervals.

However, subsequent studies have shown two factors that reduce this overconfidence: 1)

the format of the question and 2) the type of stimuli. Each is discussed next.

Page 39: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

32

Format of Question.

Soll and Klayman (2004) conducted a series of three studies with the goal of determining

how to best elicit numerical estimates of intervals from subjects. They compared subjects!

estimates across three methods of interval elicitation:

1) Range interval; (e.g. “I am 80% sure Charles Dickens was born between 1750 and

1860”),

2) Two-point method, involving lower and upper boundary; (e.g. “I am 90% sure that

Dickens was born after 1750” and “I am 90% sure that Charles Dickens was born

before 1860”), and

3) Three-point method, involving a measure of central tendency (mean or median),

lower and upper boundary; (e.g. “I think that it is equally likely that Charles Dickens

was born before or after the year 1805” and “I am 90% sure that Dickens was born

after 1750” and “I am 90% sure that Charles Dickens was born before 1860”). 6

They found that the three-point method – asking for an estimate of central tendency in

addition to both a lower and upper boundary – yielded the lowest overconfidence and the

highest degree of calibration. Similarly, Block and Harper (1991) found that, when

participants were required to produce an explicit point-estimate first, before the lower and

upper boundary, overconfidence decreased. This was also the case when an external point-

estimate was provided by the experimenter. Paese and Sniezek (1991) concurred with this

finding by suggesting that, when the point-estimate is required first, the subject is required to

spend more time processing the task and the stimuli before assessing the probability

interval, and thus produces better-calibrated intervals.

Type of Stimuli.

The fractile method, when used with general knowledge content questions such as the

above example of estimating the interval that contains the daily consumption of chicken with

90% certainty, has been shown to generally produce over-confident ratings (Lichtenstein,

Fischhoff, & Phillips, 1982; Juslin, Wennerholm, & Olsson, 1999). Typically in those studies,

questions are asked of undergraduate students and are selected to be values that the

students are unlikely to know much about (McKenzie, Liersch, & Yaniv, 2008). Thus over-

6 Charles Dickens was born on Feb. 7, 1812.

Page 40: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

33

confidence has been attributed to the use of such !general knowledge" questions for which

the subjects" judgment likely involves elaboration of general knowledge through inference

(Juslin, Winman, & Olsson, 2003). In contrast, tasks that require an assessment of sensory

input (e.g., which of two vertical lines is longer) do not show the same over-confidence bias

(Juslin, Winman, & Olsson, 2003). Erev, Wallsten, and Budescu (1994) also report that the

typical over-confidence bias is present when the uncertainty is internal to the subject (i.e.,

related to their knowledge), but not when the uncertainty is external (i.e., a stochastic

property of the environment). Furthermore, the overconfidence bias has been shown to

decrease or disappear when the objects of judgement are randomly sampled from a natural

environment (Gigerenzer, Hoffrage, & Kleinbolting, 1991; Juslin, 1994). Thus, it is believed

that this method is an appropriate method for use when subjects are asked to create

estimates based on an assessment of sensory input.

In the studies that follow, the three-point fractile method was used to elicit estimates of the

TOA interval. Subjects were asked to first provide an estimate of the TOA, and then provide

both a lower boundary and an upper boundary that would produce the smallest window that

they were 95% sure would contain the actual TOA. This measure was chosen because it

reflects both the standard deviation of the distribution and the skewness of the distribution.

The size of the subjects" TOA window was then compared to the objectively computed

intervals to assess the degree to which the human operators were calibrated to the objective

level of TOA uncertainty in the environment. Objective uncertainty was operationalized in the

studies that follow as the 95% confidence intervals around the mean speed or TOA. These

were compared to the subjective TOA window estimates to determine the degree of

calibration. As the variability (standard deviation) of the stimuli to be observed was

increased, so too should the uncertainty associated with it, as reflected by an expected

increase in the size of the estimated window.

The metric chosen to assess TOA uncertainty required that the subject estimate the lower

bound and the upper bound of the smallest time window within which they are 95%

confident the aircraft will arrive at its destination with a specified probability – for example, “I

am 95% confident that the aircraft will arrive between 1:55 and 2:05”. The goal was to

assign the smallest time window possible, while ensuring that the aircraft does not arrive

Page 41: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

34

outside the window. An operator who is “optimally calibrated” will estimate time windows of

size equal to the objectively computed TOA windows, computed as the 95% CI of the normal

distribution. If the subject is not perfectly calibrated, he/she will either overestimate or

underestimate the duration of the subject!s TOA window. If the time window is overestimated

(that is, larger than would ordinarily be needed to be 95% certain), that suggests that the

operator is assuming a variance in TOA arrival times that is larger than the actual objective

variance, which therefore suggests that he/she is in fact underestimating the probability that

the aircraft will arrive within the objectively correct (95%) time window.

Page 42: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

35

CHAPTER 4: EXPERIMENT 2 – THE EFFECT OF SPEED AND VARIABILITY ON

TOA UNCERTAINTY ESTIMATION

The goal of the present experiment was to determine how well human operators estimate

means and uncertainty for both speed and TOA of vehicles that move with variable speed.

Specifically, the goal was to determine how well subjects estimate the mean and 95%

Confidence Intervals of a population distribution based on observing a sample of 20 vehicles

drawn randomly from the population distribution. This research makes the assumption that

speeds are normally distributed.

4.1 SPECIFIC RESEARCH QUESTIONS AND HYPOTHESES

The specific research questions of interest, and their associated hypotheses, are:

1) How well do subjects estimate means of both speed and TOA (computed as the

difference between the objectively computed mean of the population distributions

and the subjects! estimate of the mean speed and TOA)?

Hypothesis 1: It was expected, based on the findings of Pitz (1980), that

subjects will estimate means with little estimation error; that is, no systematic

biases are expected.

2) How well do subjects estimate uncertainty associated with speed and TOA –

computed as the difference between the objectively computed 95% Confidence

Interval (CI) of the population mean and the subjects! estimate of the smallest

window that they are 95% sure will contain the average speed or actual TOA?

Hypothesis 2: Across all conditions, subjects! estimates of both speed and

TOA intervals will be smaller than objective 95% CIs. This is based on the

previously discussed research (e.g., Alpert & Raiffa, 1982; Juslin,

Wennerholm, & Olsson, 1999; Soll & Klayman, 2004; McKenzie, Liersch, &

Yaniv, 2008; Teigen, Halberg, & Fostervold, 2007) that suggests that human

decision-makers tend to be over-confident, and thus estimate less uncertainty

than actually exists.

3) Do the subjects! estimates differ as a function of the mean speed of the observed

stimuli?

Page 43: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

36

Hypothesis 3: It is expected (based on Peterson and Beach, 1967) that the

difference between actual 95% CIs and subjects! estimated speed and TOA

windows will be larger for objects with slower speeds than for those with

faster speeds, due to the non-linear relationship between speed and time-of-

arrival (as shown in figure 3-1).

4) Do the subjects! estimates differ as a function of the standard deviation of the speed

of the observed stimuli?

Hypothesis 4: The difference between actual 95% CIs and subjects!

estimated speed and TOA windows will be larger for speeds with high

variability than for those with low variability. This is based on Alpert and

Raiffa!s (1982) findings that subjects inadequately account for "extreme!

values, of which there will be more in higher variability conditions.

5) Are subjects able to infer skewness in the TOA distribution and apply this

appropriately in their estimation of the TOA intervals?

Hypothesis 5: It was expected, based on Pitz (1980), that subjects will

assume a linear transformation from speed to TOA and therefore will

estimate TOA intervals that are symmetrical as opposed to the objectively

generated positively skewed CIs.

4.2 EXPERIMENTAL DESIGN

This experiment employed a fully-within-subjects experimental design with three

independent variables, summarised in Table 4-1: speed, variability, and repetition. The three

levels of speed7 (slow, moderate, and fast) map to 35 units/time, 45 units/time, and 75

units/time respectively. These speed levels were chosen based on three constraints:

1) Given the relationship between speed and TOA shown in Figure 3-1, an attempt was

made to span a wide range of speeds, in order to capture subjects! ability to estimate

TOA and TOA uncertainty at different points along the speed/TOA curve;

2) The speeds were chosen to yield three expected TOAs that were approximately

equally spaced;

7 While it may be more correct to refer to low and high speeds rather than slow and fast speeds, the latter term

were nevertheless used in the remainder of the thesis to differentiate the levels of speed (slow, moderate, and

fast) from the levels of variability (low, medium, and high).

Page 44: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

37

3) Given concerns over the salience of extreme numbers, an attempt was made to

choose !like numbers"; specifically a decision was made to choose three speeds that

ended in !5". It was believed that the psychological difference between (for example) 35

and 37 is less than between 39 and 41.

The three levels of variability (low, medium, and high) map onto standard deviations of the

speed distributions. Specifically, the standard deviations chosen were 3, 6, and 9

respectively. These were chosen to ensure a wide range of speeds without producing

negative values, given the mean speed values chosen.

Each subject completed six repetitions of each speed x variability combination. This was

deemed a sufficient amount based on pilot testing, which revealed very little difference

among estimations across repetitions. Subjects completed six blocks of trials. Each block

consisted of one trial, representing each of the nine unique experimental conditions –

comprising each factorial combination of three levels of speed and three levels of variability.

The presentation order of the nine conditions within each block was randomized.

Table 4-1. Independent variables

Independent Variables Levels

Speed Slow, Moderate, Fast

Variability Low, Medium, High

Repetition 1, 2, 3, 4, 5, 6

4.3 METHOD

4.3.1 Subjects

Fifteen subjects were recruited from colleges and the community local to NASA Ames

Research Center, where the experiment was conducted. Subjects" ages ranged from 18 to

53, with the mean age being 35. The range of education levels possessed by the

participants included high school (3 participants), two-year college diploma (3 participants),

four-year university (7 participants), and graduate school (2 participants). Eight of the

subjects had taken at least one course in statistics. Prior to participation, all subjects

successfully completed a vision test to ensure that they could read text in the same font

style, size, and colour used for the experimental stimuli under lighting conditions that were

Page 45: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

38

identical to those of the actual experimental trials. The experiment lasted approximately 2.5

hours and subjects were paid $10 an hour for their participation.

4.3.2 Materials

4.3.2.1 Apparatus

The experiment was conducted in a noise-proof, human-subject testing booth. Subjects

viewed a large screen monitor (32 in. x 18 in.) that subtended a visual angle of 60º from the

average viewing distance of 32 in. Subjects input their responses using a standard

keyboard and mouse. They were also provided a pen and paper and were told they could

use them to take notes during the experiment, if they wished. The monitor brightness and

resolution settings were held constant across all subjects.

4.3.2.2 Stimuli

Experimental stimuli, shown in Figure 4-1, consisted of 20 vehicles (shown as round circles

with a diameter of .5 in) travelling simultaneously from left to right along identical routes, with

one vehicle on each route. All vehicles started at the same time, at trial start. Each route

was 18 in. long and was divided into 10 equal length segments (but the segments were not

marked or visible to the participant). Each vehicle!s speed was constant during each

segment but changed from segment to segment according to pre-determined numbers that

were drawn from a normal distribution with a designated mean and standard deviation

according to the experimental condition. Data tags (1.5 in. x .25 in. in size) displayed the

current ground speed (GS) on a digital readout that updated at the beginning of every

segment. A digital clock, centred vertically, immediately to the right of the 10th vehicle,

displayed elapsed time, with a precision of one decimal point.

A numerical simulation was first conducted to create a population or parent distribution,

which was randomly sampled to select the speeds for any given vehicle. This allowed for

later determination of the 95% intervals, by selecting the 2.5 percentile and 97.5 percentile

from the parent distribution. The stimuli were created following the steps outlined below:

1. Using a random number generator (Regress+ software package) 250,000 speeds

were drawn randomly (with replacement) from a normal distribution with a given

mean and standard deviation.

Page 46: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

39

2. The 250,000 speeds were then randomly assigned to 25,000 vehicles providing 10

speeds - one for each segment the vehicle would travel.

3. Assuming a segment distance of .05 miles, the elapsed time for each segment was

computed.

4. The 10 segment elapsed times were summed to provide a total route TOA for each

of the 25,000 vehicles. The 25,000 TOAs were sorted in ascending order, and then

binned into 20 equal groups each representing 5% of the TOAs.

5. For each sample of 20 vehicles, one TOA was randomly selected from each bin.

This ensured that the resulting sample of 20 vehicles was representative of the

distribution. This was necessary because of the relatively small sample size of 20

vehicles.

4.3.3 Procedure

4.3.3.1 Introduction and Training

Participants were welcomed to the test facility at NASA Ames Research Center. First, they

were asked to read the Instruction Protocol (Appendix B1) and then read and sign the

informed consent form. After completing a short demographic questionnaire (Appendix B2),

each participant completed a brief vision test using a standard Snellen eye chart (read from

a distance of 10 feet), and a computer-based vision test which presented ten three-digit

numbers on the experimental apparatus with the same font style, size, and colour used in

the experiment. Participants were asked to read the ten numbers out loud to verify that they

could read them without difficulty.

Next participants were shown the experimental apparatus, as described above, and the

experimenter verbally explained the purpose of the experiment, the experimental stimuli, and

specific details of the task to be performed, including how to input responses using the

keyboard and mouse, as described in the written instructional protocol. The participants

completed three practice trials while the experimenter watched. The three training trials

comprised the following Speed/Variability combinations:

Practice Trial 1: Slow/Low

Practice Trial 2: Moderate/Medium

Practice Trial 3: Fast/High

Participants were permitted to ask questions throughout the three practice trials.

Page 47: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

40

Figure 4-1. Experimental stimuli shows 20 vehicle icons with current ground speed data tags. Actual route length was 18 in. Data tags showing ground speed (GS) were 1.5 in. X .25 in. Elapsed time was shown in a box on the right (shown as 21.1 above).

Page 48: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

41

4.3.3.2 Experimental Trials Each trial consisted of twenty vehicles traversing routes of the same length. All twenty

vehicles shared the same speed profile – that is, their speeds were drawn from the same

population distribution. All vehicles started at the same time at trial start. After the last of the

20 vehicles completed the route, a question screen appeared, with the following question:

A vehicle!s speed profile is similar in nature to the vehicle you just observed and it is

about to traverse the same length route.

Please estimate this vehicle!s average (mean) speed.

Please provide the smallest window that you are 95% sure will contain the

vehicle!s actual average speed

Participants used a standard keypad to enter the average speed and then used the mouse

to manipulate window sliders to set the size of the window by setting the lower and upper

bounds. Once satisfied with these answers, the subject pressed !continue" and the second

query screen appeared with the following question:

A vehicle!s speed profile is similar in nature to the vehicle you just observed and it is

about to traverse the same length route.

Please estimate this vehicle!s time of arrival (TOA).

Please provide the smallest TOA window that you are 95% sure will contain

the vehicle!s actual TOA

Each participant responded in the same manner as above, and once completed, the

participant pressed !continue" to begin the next trial.

Participants completed six blocks of experimental trials, each consisting of 9 trials – one for

each combination of the three mean speeds and three speed standard deviations.

Participants took a short break after the second and the fourth block. The time to complete

each trial varied from 1 minute (in the shortest condition) to 5 minutes (in the longest

condition). The total duration of the experiment, including training, ranged from 2.5 hours to

3 hours.

4.3.3.3 Debrief and Follow-Up

Upon completion of the last trial, a verbal debrief session was conducted (see Appendix

B3). Participants were asked questions to ascertain the strategies that they developed,

Page 49: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

42

whether they used memory aids, and whether they found one task (i.e. estimating speed or

time) easier than the other. The general purpose and objective of the experiment was

explained, the experimenter answered any questions, and the participant was thanked for

his/her time.

4.4 RESULTS

For each trial, subjects estimated the average speed, speed windows, TOA, and TOA

windows of a new !target" vehicle that shared the speed profile of the 20 vehicles that they

had just viewed. The subjects" estimate of the target vehicle"s speed and TOA was

compared to the objectively computed mean speed and TOA of the population distribution

from which the samples of 20 vehicles were drawn. The subjects" estimates of the smallest

window that they were 95% sure contained the vehicles" average speed (estimated speed

window) and TOA (estimated TOA window) were compared to the objectively computed

95% CIs (referred to as speed intervals and TOA intervals) of the population distribution

from which the samples of the 20 vehicle were drawn. The !actual" mean, intervals, and

skew for both speed and TOA are presented below in Table 4-2, and the method by which

the actual scores were calculated is presented in Appendix B4.

The following analyses were conducted:

1) Four separate analyses (3 x 3 x 6 Repeated Measures Analysis of Variance, ANOVA)

were conducted to determine the degree to which the subjects" estimates of speed,

speed windows, TOA, and TOA windows, differed as a function of the independent

variables: speed (3), variability (3), and repetition (6). These analyses were conducted

to assess whether subjects could differentiate among the chosen levels of speed and

variability. The inclusion of repetition in these analyses allowed for the determination of

the effect of learning, fatigue, or scenario differences across trials. Repetition was not

statistically significant in any analysis, suggesting that these effects had minimal

impact, if any, on the data and therefore will not be discussed further.

2) To determine if the subjects" estimates differed from the actual values, the 95% CI

around the mean of the subjects" estimates – of speed, speed windows, TOA, and TOA

windows – were analyzed to determine if the actual score was contained within the CI.

Page 50: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

43

If the actual score was not contained within the CI, the subjects! estimates were

deemed to be significantly different than the actual score.

3) To determine if subjects were more likely to produce sub-optimal estimates as a

function of the independent variables, Estimation Error scores were calculated, by

subtracting the subjects! estimated value from the objectively computed, or actual,

value. A separate 3 x 3 x 6 repeated measures ANOVA was conducted on the absolute

values of the Estimation Error scores, to determine if subjects tend to make larger

estimation errors as a function of the independent variables speed, variability, and

repetition. Absolute values were used as these allow an assessment of estimation

errors regardless of direction – i.e. too fast and too slow, or too early and too late – in

light of the fact that, in the ATC operational setting under consideration, errors of both

kinds are detrimental to system performance. Again, there were no significant effects of

repetition and these results will not be discussed further.

4) To determine the degree to which subjects produce symmetrical or skewed estimated

TOA windows, a symmetry score was developed (defined below) and applied to both

the actual and estimated TOA intervals. The 95% CIs around the subjects! TOA

symmetry score was analyzed to determine if the actual symmetry score was contained

within the CI. Instances in which the actual symmetry score was not contained within

the 95% CI suggested that the subjects! estimated intervals were different in symmetry

than the actual intervals.

4.4.1 Estimated Speed

Table 4-3 presents the objectively computed (actual) mean speed and the size of the 95%

confidence intervals for each combination of speed and variability, the subjects! mean

estimated values for speed, and the mean size of their estimated speed windows. It also

includes the mean estimation error scores, calculated as the mean of the differences

between the actual and estimated values, for both mean and window size.

Page 51: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

44

Table 4-2 Actual mean values, interval sizes, and skew for speed and TOA

Speed TOA Independent

Variables

Speed/Variability Mean

Lower Bound

Upper Bound

Interval Size

Skew Mean Lower Bound

Upper Bound

Interval Size

Skew

Slow/Low 35 33.2 36.9 3.7 .01 51.8 49.1 54.8 5.7 .14

Slow/Medium 35 31.3 38.7 7.4 .00 53.1 47.4 60.1 12.7 .42

Slow/High 35 29.4 40.6 11.2 .00 56.2 46.3 71.5 25.2 34.32

Moderate/Low 45 43.1 46.9 3.7 .00 40.2 38.6 41.9 3.4 .12

Moderate/Medium 45 41.3 48.7 7.4 .01 40.8 37.5 44.6 7.1 .28

Moderate/High 45 39.4 50.5 11.1 .00 41.9 36.6 48.8 12.3 .73

Fast /Low 75 73.1 76.9 3.8 .01 24.1 23.5 24.7 1.2 .08

Fast /Medium 75 71.3 78.7 7.4 .00 24.2 23 25.4 2.4 .17

Fast /High 75 69.4 80.5 11.1 .00 24.4 22.6 26.4 3.8 .47

First, an analysis was conducted to determine if the subjects! estimated mean speed differed

as a function of the experimental conditions. This was to verify that the task conditions

(speed and variability) had been perceived as being significantly different. Subjects!

estimated mean speed, shown as a dashed line in Figure 4-2, did indeed differ as a function

of speed (F (2,28) = 7306.75, p < .001). Post-hoc paired t-tests revealed significant

differences among all three levels of Speed (p < .001 for all pairwise tests), providing

evidence that subjects were able to differentiate among the three levels of speed. There

were no differences in the subjects! estimates of mean speed as a function of variability or

repetition.

Next, the subjects! estimated mean speed was compared to the actual mean speed by

examining the 95% CI around subjects! estimated mean speed to determine if the 95% CI

contained the actual mean speed. Figure 4-2 presents both the actual mean speed and the

Page 52: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

45

subjects! estimated mean speed, the latter plotted with error bars which reflect the 95% CI

around the mean. In conditions in which the 95% CI did not contain the actual speed, it was

concluded that the subjects! estimates were significantly different than the actual speed.

This was the case for three conditions: moderate speed/medium variability, high

speed/medium variability, and high speed/high variability. In each of these three cases, the

actual speed was greater than the estimated values, revealing that subjects underestimated

speed in these three conditions.

Table 4-3. Actual and estimated speed means, intervals, and estimation errors

Actual Speed Subjects! Estimated

Speed

Estimation Error (Actual –

Estimated) Independent Variables (Speed/Variability)

Mean 95% CI Mean Window

Size Mean

Window Size

Slow/Low 35 3.7 34.7 6.0 .3 -2.3

Slow/Medium 35 7.4 35.1 9.0 -.1 -1.6

Slow/High 35 11.2 34.2 12.6 .8 -1.4

Moderate/Low 45 3.7 44.4 6.0 .6 -2.3

Moderate/Medium 45 7.4 43.4 8.8 1.6 -1.4

Moderate/High 45 11.1 44.4 11.9 .6 -.8

Fast /Low 75 3.8 74.7 5.1 .3 -1.4

Fast /Medium 75 7.4 73.9 8.4 1.1 -1.0

Fast /High 75 11.1 73.7 11.4 1.3 -.3

Speed estimation error scores were computed by subtracting the subjects! estimate of the

average speed from the objectively calculated average speed of the original distribution

(presented in Table 4-2). An ANOVA was conducted on the absolute values of the speed

estimation error scores in order to determine if the size of error (regardless of direction)

differed as a function of the experimental variables of interest. The amount of speed

estimation error increased as variability increased (F(2,28) = 119.889, p < .001). All pair-

wise comparisons were significant (p < .001), suggesting that subjects made larger errors in

Page 53: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

46

estimating speed in the high variability condition (M=3.5) than the medium variability

condition (M=2.4), and both the high and medium conditions were larger than the low

variability condition (M=1.2).

Figure 4-2. Actual and estimated mean speed as a function of speed and variability.

4.4.2 Estimated Speed Windows

An analysis was conducted to determine if the size of the subjects! estimated speed

windows differed as a function of the experimental conditions. Recall that subjects were

asked to provide the smallest window that they were 95% sure would contain the actual

average speed of a target vehicle that shared a speed profile with the 20 vehicles just

observed. This measure was intended to assess how well subjects were able to

differentiate among the three levels of variability and provide speed windows that reflected

this variability. The width of the subjects! estimated speed window differed as a function of

variability (F(2,28)=11.12, p < .001), shown by the dashed lines in Figure 4-3. The size of

the estimated speed window was smallest when variability was low (M=5.7), followed by

medium variability (M=8.7), and then by the high variability condition (M=12.0). Paired t-

Page 54: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

47

tests revealed significant differences among all three levels of variability (p < .001). This

provides evidence that subjects! estimated speed windows were sensitive to the variability in

the stimuli and supports the use of 95% CI as a measure of uncertainty. In other words, this

finding is evidence that subjects can indeed create intervals that reflect the amount of

variability inherent in the stimuli. No other main effects or interactions were observed.

To determine if the subjects! estimated speed window size differed significantly from the

actual speed interval, the 95% CIs around the mean subjects! estimated speed window

widths were examined. As can be seen in Figure 4-3, there were only two conditions for

which the 95% CI did not contain the actual speed interval width – the slow speed/low

variability and moderate speed/low variability conditions. In both of these conditions, the

subjects! estimated speed windows were wider than the actual speed intervals. This shows

that, for the most part, the subjects were capable of developing calibrated windows for the

speed estimation task.

Figure 4-3. Estimated speed windows relative to actual speed interval.

4.4.3 Estimated TOA

Table 4-4 presents the actual mean TOA and TOA interval size, the subjects! estimated

mean TOA and TOA window size, and the difference between the actual and estimated

scores. A 3 x 3 x 6 repeated measures ANOVA was conducted to determine if the subjects!

estimated TOA differed significantly as a function of the experimental conditions. Again, this

was conducted to determine whether the participants could distinguish among the

Page 55: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

48

experimental conditions (speed and variability), and to evaluate whether performance

differed across the six repetitions due to learning, fatigue, or scenario effects.

A significant speed x variability interaction (F(4,56) = 96.66, p < .001) provided evidence that

subjects! estimated TOAs differed as a function of the experimental independent variables.

As can be seen by the dashed lines in Figure 4-4, in the slow speed condition subjects

estimated longer TOAs in the high variability condition than for the medium condition (p <

.001) and both were higher than the low variability condition (p < .001 for both). In the

moderate speed condition, estimated TOAs in both the high and medium variability condition

were significantly higher than the low variability condition (p < .05), but there was no

difference between the medium and high variability conditions (p < .05). The same pattern

was observed for the fast speed condition – estimated TOAs in the high and medium

variability conditions were significantly higher than in the low variability condition (p < .05 for

both), but the high and medium variability conditions were not significantly different. These

results demonstrate that the subjects! TOA estimates were sensitive to changes in the

speed and variability of the stimuli.

Table 4-4. Actual TOA scores, estimated TOA scores & TOA estimation error scores

Actual TOAs Subjects! Estimated

TOAs

TOA Estimation Error (Actual –

Estimated)

Independent Variables (Speed / Variability) Mean

Interval Size

Mean Window

Size Mean

Window Size

Slow/Low 51.8 3.7 55.1 4.9 -3.3 -1.2

Slow/Medium 53.1 12.7 56.3 7.2 -3.2 5.5

Slow/High 56.2 25.2 59.5 11.2 -3.3 14.0

Moderate/Low 40.2 3.4 42.7 3.7 -2.5 -0.4

Moderate/Medium 40.8 7.1 43.7 5.2 -2.9 1.9

Moderate/High 41.9 12.3 44.4 6.9 -2.5 5.4

Fast /Low 24.1 1.2 24.6 2.7 -0.5 -1.5

Fast /Medium 24.2 2.4 25.1 3.5 -0.9 -1.1

Fast /High 24.4 3.8 25.3 4.1 -0.9 -0.4

Page 56: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

49

Figure 4-4. Actual and Estimated TOA as a Function of Speed and Variability Graph shows 95% Confidence Intervals around the Estimated TOA.

Next, the subjects! estimated TOAs were compared to the actual mean TOAs by examining

the 95% CI around the mean of the subjects! estimated TOA to determine if the CI contained

the actual mean TOA (as shown in Figure 4-4). In conditions in which the CI does not

contain the actual TOA, it was concluded that the subjects! estimate was significantly

different than the actual TOA. In stark contrast to speed, the actual TOA was not contained

within any of the 95% CIs. This indicates that subjects! TOA estimates were significantly

different than the actual TOA for every condition. This suggests that subjects may perform

less well at the task of estimating TOA than they do at estimating speed. Subjects

overestimated TOA in all conditions.

TOA estimation error scores, given in Table 4-3, were computed by subtracting the subjects!

estimated TOA from the actual TOA of the population distribution. A repeated measures

ANOVA was conducted on the absolute values of the TOA estimation error scores to

Page 57: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

50

determine if the amount of estimation error (regardless of direction) differed as a function of

the independent variables. The ANOVA yielded a speed x variability interaction (F(4, 56) =

3.941, p < .01). As seen in Figure 4-5, at slow speeds, absolute TOA estimation error was

greater in the high variability condition than the medium variability condition (p < .01) and

low variability condition (p < .05), but there was no difference between the low and medium

variability conditions (p > .05). At moderate speed, the same pattern was found – that is,

TOA estimation error was greater when variability was high than medium (p < .01) and low

(p < .05), but there was no difference between the low and medium variability conditions (p >

.05). (However, this last finding approached significance (p = .053).) At fast speeds, there

were no significant differences in the amount of TOA estimation error as a function of

variability.

Figure 4-5. Absolute TOA estimation error as a function of speed and variability.

(Absolute values are plotted, with +/- 1 SE).

4.4.4 Estimated TOA Windows

As shown in Figure 4-6, subjects! estimated TOA windows differed as a function of the

interaction between speed and variability (F(4,56) = 7.241, p < .001). Subjects! estimated

TOA windows were sensitive to the independent variables tested and reflected the relative

amount of variability (uncertainty) in the TOA of the stimuli. Of most interest was whether

Page 58: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

51

subjects! TOA windows increased as a function of variability at each level of speed. Simple

effect tests revealed that this indeed was the case. Subjects! estimated TOA windows were

larger for the high variability condition than the medium variability condition, which in turn

were larger than the low variability condition (p < .05), for each level of speed.

Figure 4-6. Estimated TOA windows as a function of speed and variability. Graph shows 95% CIs around the estimated TOA window.

In order to determine whether the widths of the subjects! estimated TOA windows were

significantly different than actual TOA intervals, the 95% CIs for the three levels of variability

in each level of speed, shown in Figure 4-6, were examined to determine if they contained

the actual TOA interval width. In the slow speed condition, the width of the subjects!

estimated TOA window was significantly smaller than the actual TOA interval width in both

the medium and high variability conditions, but not the low variability conditions. Similarly, in

the moderate speed condition, the width of the subjects! estimated TOA window was again

significantly smaller than the actual TOA interval width in both the medium and high

variability conditions, but not the low variability conditions. In the fast speed condition,

however, the width of the subjects! estimated TOA window was significantly larger than the

actual TOA interval, in the low variability condition only.

A TOA window estimation error score (presented in Table 4-3) was calculated by subtracting

the width of the subjects! TOA window from the width of the actual TOA interval – in other

words, such that positive values represent overestimates of TOA window width. A 3 x 3 x 6

repeated measures ANOVA conducted on the absolute value of the TOA window estimation

Slow Moderate Fast

TO

A I

nte

rva

l W

idth

Low Medium High Low Medium High Low Medium High Variability Variability Variability

TO

A I

nte

rva

l W

idth

TO

A I

nte

rva

l W

idth

Page 59: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

52

error score revealed a speed x variability interaction (F(4,56) = 98.503, p < .001), as seen in

Figure 4-7. Follow-up paired t-tests revealed that, at slow speeds, TOA window estimation

error was significantly higher when variability was high than medium (p < .001) or low (p <

.001) and error for the medium variability condition was significantly higher than low

variability condition, p < .001. At moderate speeds, the same pattern was observed, with

TOA window estimation error being significantly higher when variability was high than

medium (p < .001) or low (p < .001) and medium variability being significantly higher than

low variability (p < .001). However, for fast speeds, there was no significant difference in

TOA window estimation error as a function of variability.

Figure 4-7. TOA interval estimation error as a function of speed and variability (Absolute values, plotted with +/- 1 SE)

4.4.5 Estimated TOA Window Symmetry

In addition to the size of the subjects! estimated TOA window, the symmetry of the estimated

TOA window was also of interest. As was discussed previously, given a normal distribution

of speeds, the resulting TOA distribution is necessarily positively skewed – and furthermore

the amount of positive skew changes as a function of speed and variability – with greater

Page 60: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

53

positive skew for stimuli with low speeds and high variability. To assess the degree to which

the subjects! estimated TOA window appropriately reflected this positive skew, a TOA

symmetry score was calculated for the actual distribution and for each subject!s TOA

window estimate, using the following equation:

TOA Symmetry Score = |UB –TOA| - |LB-TOA|;

where TOA = subjects! estimates of TOA,

LB = TOA Interval Lower Bound,

UB = TOA Interval Upper Bound (UB)

If the TOA symmetry score is equal to zero, this means that the window created by the

subject was perfectly symmetrical; if it is positive, then the window was positively skewed,

with the upper portion larger than the lower portion; and if negative, then the window was

negatively skewed with the lower portion larger than the upper portion.

Symmetry scores were first computed analytically from the underlying distribution for each

combination of speed and variability, as shown by the solid lines in Figure 4-8. The mean

TOA value was used as the measure of central tendency of TOA. As can be seen, the

actual TOA intervals were either symmetrical or positively skewed. The skew was highest in

the slow speed/high variability condition. This is consistent with the inverse relationship

between speed and TOA, as described previously.

Next, TOA symmetry scores were calculated for each subject!s estimated windows around

the subject!s own estimate of TOA. As can be seen by the dashed lines in Figure 4-8, the

subjects! symmetry scores also tended to be positive (overall mean = 0.2), therefore

suggesting that TOA windows provided by subjects were slightly positively skewed. Note

that for fast speeds the amount of skew was low and quite comparable to the actual

symmetry score; however, for slow speeds with high variability, symmetry scores were

positively skewed, but substantially less skewed than the actual symmetry scores.

Page 61: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

54

Figure 4-8. Actual and estimated TOA symmetry scores. Estimated scores are plotted with 95% CIs.

The 95% CIs around the subjects! symmetry scores were examined to determine if they

contained the actual symmetry score. If the 95% CI did not contain the actual symmetry

score, the subjects! symmetry score was considered significantly different than the actual

symmetry score. This was the case for four speed/variability combinations: slow/medium,

slow/high, moderate/medium, and moderate/high. As can be seen in Figure 4-8, in these

four conditions for which the 95% CI did not contain the actual symmetry score, the actual

symmetry score was higher than the upper-bound of the subjects! 95% CI, reflecting that

subjects underestimated the amount of positive skew of the TOA interval. The graph shows

that subjects! estimated windows increased in positive skew as variability increased,

particularly in the slow condition, just as for the actual distributions; however, they

underestimated the magnitude of the skew.

Based on subjective comments made during the debrief session, most subjects detected

that the distribution was skewed. For example, many referred to "laggers! or "stragglers! –

the one or two or three vehicles that seemed substantially later than the rest of the sample.

However, most subjects suggested that they did not know how to factor these extreme

values into their interval estimations. Two strategies were prevalent. Some subjects took the

strategy of excluding the outliers altogether and considered them statistical anomalies.

Others chose to factor the outliers in by adjusting the window boundary "a little bit!, but again

expressed uncertainty as to how much to adjust the window. A hybrid strategy was also

Page 62: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

55

implemented by other subjects, in that they would disregard a single outlier, but adjust the

interval boundary if there were 2 or 3 extremes.

4.5 DISCUSSION

4.5.1 Summary of Results

One objective of this research was to determine how well subjects estimated speed and

TOA. As expected, subjects! estimates of both speed and TOA differed among the levels of

speed, suggesting that the mean values chosen for the three distributions tested in this

experiment were perceived by the subjects as coming from different distributions, even for

the highly variable conditions. As expected, based on research of estimates of central

tendency, as summarized by Pitz (1980), subjects performed quite well at the task of

estimating mean speed, with only two conditions in which subjects significantly

underestimated speed as compared to the actual, or objectively computed, speed. In

estimating TOA, on the other hand, subjects over-estimated TOA compared to the actual

TOA for all conditions. This is a marked deviation from Pitz!s research, which would have

predicted that, when asked to estimate central tendency for other tasks, estimation error is

minimal.

It could be that TOA is inherently different from the types of stimuli used in other studies.

For example, it is likely that the positively skewed nature of the TOA distribution could have

increased TOA estimation error, as will be described later. Alternatively, it may be the case

that subjects weren!t attempting to estimate the mean of the distribution, but instead were

estimating some other measure of central tendency, such as the mode or median. During

the debrief, some subjects stated that they selected the time when the "middle group! of

vehicles finished the route – which supports the possibility that subjects were estimating the

median.

A second, and more important, goal was to determine how well subjects estimated the

amount of uncertainty associated with both speed and TOA. Estimated window size was

used as a measure of the subjects! estimate of uncertainty, in that larger windows reflected

larger estimates of uncertainty. For both speed and TOA, subjects! estimated window sizes

increased as a function of stimulus variability, providing support for the use of 95% intervals

Page 63: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

56

as a measure of uncertainty. It was hypothesized that the size of the subjects! estimated

windows would be smaller than the actual intervals, since other subjects in the literature

tended to be overconfident in their estimates of uncertainty and not factor in extreme values

appropriately. This was not the case for speed windows. In fact, subjects produced well-

calibrated speed intervals and in only two conditions was there an indication of error in the

size of the estimated speed window – and in both cases the subjects overestimated, rather

than underestimated, the size of the window. TOA window estimation error, however,

depended on the speed of the stimuli. In general, for fast stimuli, subjects slightly

overestimated window size; however, for moderate and low speeds, subjects tended to

underestimate window size, as hypothesized. The amount of this underestimation was

greater for higher variability conditions than medium and low variability. The fact that

subjects did produce well-calibrated speed intervals also provides support for the

experimental technique. Any mis-calibration of TOA intervals can thus more likely be

attributed to real bias in TOA uncertainty estimation, as opposed to a general over- or under-

confidence bias in interval estimation due to the experimental technique.

Finally, it was hypothesized that subjects would create TOA windows that were symmetrical,

neglecting to account for the positive skew of the distribution. Although subjects! windows

were positively skewed, they still underestimated the amount of positive skew, particularly in

the slow/high variability condition.

4.5.2 Limitations of the Current Experiment and Suggestions for Future Studies

Being the first in a series of studies to explore how subjects estimate TOA and TOA

uncertainty, the stimuli and the task were quite simple. For example, the routes were

straight lines of equal length and all vehicles started at the same time. This simplicity may

have produced experimental artefacts that the subjects used in their estimates. During the

debrief, some subjects talked about "watching time! or using "visual patterns! of the vehicles

as they finished their routes. This allowed them to "see! the distribution of TOAs as it

unfolded. In essence then, subjects may have been simply describing the visible profile of

TOAs of the 20 vehicles, rather than applying their knowledge of speed variability to

estimate TOA. Future studies with added stimulus complexity, such as staggering the start

times or positions of the vehicles, should add to the robustness of these findings. In addition,

Page 64: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

57

a different experimental paradigm in which subjects undergo a substantial training period to

learn to characterize speed profiles, which are later !applied" in experimental trials, may also

be more informative. This experimental paradigm will be used in the following experiment.

One particularly interesting finding in this experiment was how subjects factored in !extreme"

vehicles – that is those that are much faster or much slower than the others. Both the

objective data and the subjective comments made during the debrief session revealed that

subjects were uncertain how to treat extreme values and how to determine their effect on

TOA or TOA uncertainty. Indeed, this is likely the most difficult task associated with

estimating TOA and TOA uncertainty. With samples of 20 vehicles, there was usually only

one extreme vehicle, and many subjects (but not all) noted that they tended to ignore this

vehicle as a statistical anomaly. The strategy of dealing with extremes, however, was less

clear when there were two or three vehicles that were much slower or faster than the others.

Here, subjects expressed some uncertainty as to how to account for these vehicles.

Therefore, subsequent studies that show larger samples of vehicles to the subjects (i.e., 100

or more) may provide a more realistic opportunity to learn how subjects factor small

numbers of extremes into their estimation process. Having said this, however, it is believed

that 20 vehicles may be at or above the limit of the number of vehicles that a human subject

can process at one time. As such, the follow-on experiment will display a set of 100 vehicles

across five screens, rather than asking subjects to process 100 vehicles at the same time.

4.5.3 Implications for the Next Generation Aviation System

The airspace system is advancing to a !4-D" navigation environment, in which aircraft are

required to not only follow the conventional spatial routes, but also conform to a commanded

arrival time. Of course, for overall system efficiency, the notion of 4-D control has to be

applied to airport surface (taxi) operations as well, to ensure that aircraft are able to depart

on time and meet their next 4-D waypoint. (Note that current convention in the aviation

community utilizes the term “4-D navigation” also on the surface, even though altitude is

fixed.) On the surface, the goal is to minimize wait time by sequencing aircraft so they arrive

at each active intersection !just-in-time" to be able to cross it without waiting for other aircraft

to pass. In this environment, arriving early is as detrimental to system traffic flow as is

arriving late. For airport surface operations, this move to 4-D navigation will mean that pilots

will not only have to navigate the taxiways in order to get from gate to runway (and vice

Page 65: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

58

versa), but also manipulate the aircraft speed (without the advantage of auto-pilot) so as not

to arrive too early or too late. Also, controllers must be able to estimate TOA and assign an

appropriate degree of uncertainty to the TOA estimate, based on the speed profile of the

aircraft and adjust traffic sequencing, or re-route traffic accordingly.

The results of this experiment, if they hold true after further, more robust, investigations,

suggest that, if an air traffic controller knows that a particular aircraft adheres to a certain

speed profile, such as slow and variable, or fast and consistent, they will tend to

overestimate the aircraft!s TOA – that is, a controller will estimate that the aircraft will arrive

later than it actually will – and more so for slower aircraft than faster aircraft. Furthermore,

for slow and moderate speed conditions, controllers will have a tendency to underestimate

uncertainty associated with the estimated TOA, which could result in sequencing aircraft too

close together, or not leaving a wide enough time window (or space envelope) around each

aircraft. Finally, the results of the symmetry analysis suggest that, particularly at slow and

moderate speeds with high variability, controllers will underestimate the likelihood that an

aircraft will arrive late to its destination.

Page 66: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

59

CHAPTER 5: EXPERIMENT 3 – TOA UNCERTAINTY ESTIMATION AND

EXTRAPOLATION

5.1 INTRODUCTION

This experiment explored participants! ability to develop speed profile expectations by

learning a speed profile through training with feedback, and then to apply this speed profile

expectation to estimate TOA uncertainty. In addition, the ability to identify and apply the

appropriate speed profile based on observation of an individual vehicle was examined.

Thirdly, the ability to extrapolate TOA uncertainty estimates from known to unknown speed

profiles was examined.

5.1.1 Addressing Limitations of Experiment 2

This experiment addressed two limitations that were identified in experiment 2 (Chapter 4):

• Subjects appeared to describe the characteristics of TOA data displayed for each of

the 20 vehicles just observed, rather than apply their global knowledge of each

speed profile to estimate TOA.

• With only 20 vehicles per sample in experiment 2, subjects may not have had

enough data to properly understand and characterize the TOA distribution.

Separating the process of learning speed profiles in the training session from applying them

to estimate TOA uncertainty in the experimental trials was expected to minimize the

limitations of experiment 2, in which subjects were simply describing the TOA distribution of

20 vehicles at a time. In the present experiment, subjects were exposed to a training

program in which they learned speed profiles by exposure to a larger number of vehicles

(100 vehicles per speed profile) than in experiment 2. After the training session, subjects

were asked to apply their knowledge of the speed profiles by estimating TOA uncertainty for

subsequent vehicles – which either followed a trained speed profile or a new speed profile.

A third change in the experimental protocol was made in an attempt to improve the subjects!

understanding of the task by clarifying the concept of the intervals. In the previous

experiment, subjects were asked to provide “the smallest TOA window that they were 95%

sure would contain the vehicle!s actual TOA.” In this follow-on experiment, these

instructions were augmented to better define “95% sure”. Specifically, subjects were

instructed to select a window that would capture 95 of 100 vehicles. Because subjects saw

Page 67: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

60

100 vehicles of each speed profile it was expected that this would make the task more

concrete and meaningful.

5.2 RESEARCH QUESTIONS

This experiment set out to address 5 research questions as listed below, and shown

schematically in Figure 5-1, in relation to the model of TOA Uncertainty developed in

Chapter 2. These research questions were addressed across three phases of a single

within-subjects experiment. The experiment phase is listed in parentheses after each

question.

1. Can subjects develop an internal model of a speed profile based on the average speed and speed variability? (Phase 3a)

2. Can subjects apply their internal model of a speed profile to estimate TOA uncertainty – and does this differ as a function of average speed or speed variability? (Phase 3b)

3. Can subjects extrapolate from a trained distance to estimate TOA uncertainty for a new, untrained distance? (Phase 3b)

4. Based on a period of observation of a single vehicle, can subjects identify the vehicle!s speed profile from their repertoire of known speed profiles and apply knowledge of this profile to accurately estimate TOA uncertainty? (Phase 3c)

5. Can subjects extrapolate from trained speed profiles to accurately estimate TOA uncertainty for route distances and speed profiles on which they were not trained. (Phase 3c)

Figure 5-1. Research questions mapped to the preliminary model of TOA uncertainty (Numbers in parentheses refer to phases of the experiment)

Page 68: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

61

5.3 METHOD

5.3.1 Subjects

Sixteen subjects were recruited from colleges and the community local to NASA Ames

Research Center, where the experiment was conducted. Subjects were pre-screened for

colour blindness and instructed to wear glasses/contact lenses normally used for computer

work. Subjects! ages ranged from 21 to 54, with a mean age of 35. The range of education

levels possessed by the participants included high school (3 participants), two-year college

diploma (2 participants), four-year university (7 participants), and graduate school (4

participants). Seven of the subjects had taken at least one university-level course in

statistics. Prior to participation, all subjects successfully completed a vision test to ensure

that they could read text in the same font style, size, and colour used for the experimental

stimuli under lighting conditions that were identical to that of the actual experimental trials.

The experiment lasted approximately 3.5 hours and subjects were paid $10 per hour for

their participation.

5.3.2 Experimental Design

This experiment consisted of three phases, as described below. All subjects completed

each of the three phases in a single session, and in the same order. Prior to embarking on

the first phase, subjects first read a set of instructions (see Appendix C1) that outlined the

purpose and tasks required for each phase, signed an informed consent form, and

completed a short demographic questionnaire (see Appendix C2). Upon completion of the

third phase, subjects participated in a post-study debriefing session consisting of a semi-

structured interview (see Appendix C3).

5.3.2.1 Phase 3a: Developing a Library of Speed Profiles

The goal of this first phase of the experiment was to train subjects to accurately characterize

four different speed profiles through exposure to a large number of vehicles and by providing

feedback.

Task Description. As shown in Figure 5-2, subjects observed on a single screen 20 vehicle

icons that shared a common speed profile (i.e., all speeds were randomly drawn from the

same normal distribution), travelling simultaneously along the same length route (medium

Page 69: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

62

length: 500 pixels). Throughout this phase, all vehicles that shared the same speed profile

also shared the same colour-coded icon. Each route was divided into 10 segments with

visible vertical hatch marks, and the speed changed for each segment. This was similar to

the stimuli used in experiment 2, except the segment hatch marks were not shown in

experiment 2 and the elapsed time readout shown in experiment 2 was not shown in this

current experiment.

Figure 5-2. Phase 3a Stimuli. Each red dot represents a vehicle. Each three-digit label (e.g., 36.4) represents current ground speed. The twenty routes were identified sequentially, with 1 at the top and 20 at the bottom.

Page 70: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

63

After the last vehicle had completed the route, subjects were asked to report the number of

the vehicle that arrived first and last. (This task was aimed at ensuring that subjects actually

attended to the distribution of arrival times.) Next, subjects were asked to characterize the

speed profile by estimating the average speed of the profile, and providing a lower bound

and upper bound for the smallest range of speeds that would capture 95% of the individual

average speeds, in the same manner as experiment 2. Subjects were given feedback after

every trial that indicated whether their answers were “Correct”, “Too High” or “Too Low”

relative to the mean speed and 95% CIs of the population distribution. The feedback was

provided on the subject!s estimate of average speed as well as both the lower bound and

the upper bound individually. The feedback algorithm had a buffer of 1 unit, such that, if a

subject!s answer was within plus or minus 1 of the actual answer, the algorithm reported

“Correct”. In total, subjects experienced five consecutive trials of each colour-coded speed

profile – for a total of 100 vehicles per speed profile. Upon completion of the fifth trial for

each speed profile, subjects also received the actual mean, lower bound and upper bound,

as computed objectively from the speed profile distributions. Subjects then answered the

three questions below pertaining to the speed profile, using their full knowledge of all of the

vehicles that shared that profile – that is, all 100 vehicles observed.

Q1. Based on their average speed, are the vehicles within this profile more likely travelling on a six-lane freeway (i.e., fast) or a two-lane rural road (i.e., slow)? Please circle one:

1 2 3 4 5 6 7 Definitely Definitely two-lane rural six-lane freeway

Q2. Based on their speed fluctuations, are the vehicles within this profile more likely

travelling with cruise control ON or OFF? Please circle one: 1 2 3 4 5 6 7 Definitely Definitely cruise control ON cruise control OFF

Q3. Please characterize the time of arrival of vehicles that share this speed profile.

(use words, symbols, numbers, or pictures).

The purpose of the first two questions was to assess how subjects categorized each profile

on the two dimensions: speed (low or high) and variability (low or high) and to determine if

Page 71: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

64

their categorization matched the stimuli. With regards to Question 3, subjects were

encouraged to make notes to characterize the TOA distribution using words, symbols,

numbers, or pictures. For example, subjects could note that the TOAs were !spread out" or

!clumped together", estimate a number of seconds from the first to the last vehicle, or use

symbols or pictures to depict the spread graphically. They were instructed that these notes

would be important for completing Phase 2 and 3 of the experiment, in which they would be

required to estimate TOA, and that these were the only notes that they could carry through

to the subsequent experimental phases. These notes also served to emphasize the

importance of utilizing their knowledge of the speed profile of the full set of 100 vehicles, not

just of a single vehicle.

Subjects completed this task for four different speed profiles made up of speed/variability

pairings adopted from Experiment 2: slow/low, slow/high, fast/low, fast/high, summarised in

Table 5-1. Each speed profile was labelled by colour-coded icons, allowing subjects to

differentiate among the four speed profiles – slow/low = red; slow/high = yellow; fast/low =

green; fast/high = blue. The five trials for each speed profile were blocked, and the order in

which the speed profile blocks was presented to each subject was determined using a Latin

Square design.

Table 5-1. Phase 3a Experimental Design

Speed Profile

(Speed/Variability)

Profile

Labelled

Number

of Trials

Number of

Vehicles per

Trial

Route

Distances

(in pixels)

Slow/Low Yes 5 20 Medium (500)

Slow/High Yes 5 20 Medium (500)

Fast/Low Yes 5 20 Medium (500)

Fast/High Yes 5 20 Medium (500)

5.3.2.2. Phase 3b: Estimating TOA Uncertainty for Trained Profiles

The goals of this phase were to:

1) Assess whether subjects could apply their knowledge of the four learned speed

profiles to accurately estimate TOA uncertainty; and

Page 72: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

65

2) Assess subjects! ability to extrapolate to distances other than that distance (medium)

for which they were trained.

Task Description. In this block of trials, subjects were shown a dynamic simulation of a

single vehicle with a digital speed readout as it traversed a medium length route (500 pixels,

as used in Phase 3a; see Figure 5-3 top). All vehicles were colour coded to match one of

the four speed profiles learned in Phase 3a. Subjects were instructed that this was a

prototypical vehicle that was representative of all of the vehicles that shared that profile, and

when answering the questions they should consider all vehicles that shared that same

colour-coded speed profile. The stimuli in this phase, shown in Figure 5-3, were the same

as in Phase 3a, except that only one vehicle was shown at a time. The speed readout was

displayed, but the elapsed time was not displayed. The speed profile always matched one of

the four trained profiles, and was identified using the same colour-coding scheme used in

Phase 3a. The vehicle!s speed changed with each segment of the route, and speeds were

drawn from the normal distribution that characterized the speed profile.

Figure 5-3. Phase 3b experimental stimuli. The top panel shows the dynamic simulation of one vehicle traversing a medium-length route adhering to the "red! speed profile (Slow/Low Variability). The icon changes speed at each segment (vertical hatch mark). The bottom panel shows the static probe screen presented after the dynamic simulation (top) was completed, with the same speed profile (red colour-code) and a visual representation of the route!s distance remaining (in this case a short, 250 pixel route). The top dynamic simulation was shown first, followed by the bottom static screen.

Static Probe Screen

Dynamic Simulation

Page 73: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

66

After the vehicle had completed 10 speed change segments and reached the end of the

route, the vehicle looped to the beginning of the route (see Figure 5-3, bottom) and its

movement was paused with a new distance remaining that was either short (250 pixels),

medium (500 pixels), or long (1000 pixels). An !expected TOA" was presented to the

subjects that reflected the time (in seconds) that the vehicle was expected to complete the

route. (The time shown was the mean of the TOA distribution for the distance remaining.)

Subjects were asked to produce TOA intervals that reflected the smallest window that they

were 95% sure would contain the actual TOA, in the same manner as experiment 2. It was

emphasized in the written and on-screen instructions that the subjects should aim to

produce intervals that capture 95 of 100 vehicles in the profile.

As shown in Table 5-2, stimuli consisted of four speed profile conditions at each of three

distances (250, 500, and 1000 pixels). Recall that routes in the training block were 500

pixels long – thereby requiring subjects to apply their knowledge of speed to estimate TOA

windows for the same distance (500 pixels) and also to extrapolate to a shorter route (250

pixels) or a longer route (1000 pixels).

Subjects completed two familiarization trials with the experimenter present: the slow/high-

short route and the fast/low–long route. Then subjects completed one block of 12 training

trials (4 speed profiles x 3 distances), followed by four blocks of 12 experimental trials. The

presentation order of the twelve trials within each block was determined randomly.

Table 5-2. Phase 3b experimental design

Speed Profile (Speed/Variability)

Profile Trained

Profile Labelled

Number of Trials Route Distance

(pixels)

Slow/Low Yes Yes 1 training +

4 experimental 250, 500, 1000

Slow/High Yes Yes 1 training +

4 experimental 250, 500, 1000

Fast/Low Yes Yes 1 training +

4 experimental 250, 500, 1000

Fast/High Yes Yes 1 training +

4 experimental 250, 500, 1000

Page 74: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

67

Hypotheses. If the findings from Experiment 2 hold, it was expected that subjects would

overestimate TOA uncertainty in the fast conditions, particularly the low variability condition

(that is, they were expected to produce TOA windows that were larger than the actual TOA

intervals). For slow speeds, it was expected that estimated TOA windows would be smaller

than actual TOA intervals, and more so for the high variability condition than low variability

condition. The magnitude of TOA estimation error was expected to be smaller than for

experiment 2 because the increased training protocol should yield improved estimates.

Further, if subjects were able to extrapolate from the trained route distance (500 pixels) to

new, untrained distances, a main effect of route distance would be expected, with smaller

estimated TOA windows for shorter distances, and longer TOA windows for longer

distances. Absence of a significant main effect, assuming sufficient power, would suggest

that subjects could not extrapolate from trained distances to untrained distances.

5.3.2.3. Phase 3c: Estimating TOA Uncertainty for New, Untrained Profiles

The goal of this phase was to determine if subjects could identify a known speed profile

based on a short period of observation and apply knowledge of this profile to estimate TOA

uncertainty. A second goal was to assess how well subjects could extrapolate from the

trained speed profiles to assess TOA uncertainty for speed profiles on which they were not

trained.

Task Description. These experimental trials were identical to the trials in Phase 3b, with

three exceptions:

1) The speed profile was not labelled (through colour coding) in any trials;

2) Some trials matched the four trained speed profiles, but some did not, as shown in Table

5-3;

3) All route distances were medium length (500 pixels, the same as the trained route

distance).

This enabled an evaluation of the subjects! ability to extrapolate from trained profiles to new

untrained profiles, presented in Table 5-3 as shaded rows. Subjects completed three trials

for each speed profile. The nine speed profiles were presented in random order in three

blocks. Subjects were informed that not all vehicles shared one of the four trained speed

profiles.

Page 75: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

68

Prior to estimating the TOA windows, subjects were asked to 1) determine if the speed

profile was one of the trained profiles; 2) identify the speed profile (by colour) that it most

closely resembled (or state the colour of the trained profile they believed it matched), and 3)

rate the similarity of the vehicle!s profile to the selected speed profile on a scale from 1 (very

low) to 5 (very high).

Table 5-3. Phase 3c experimental design. (Shaded rows are new, untrained, speed profiles)

Hypotheses. For trained profiles, it was expected that subjects would overestimate TOA

uncertainty in the fast conditions (that is, estimate TOA windows that are larger than the

actual TOA intervals) and underestimate TOA uncertainty (estimate TOA intervals that are

smaller than actual TOA intervals) in the slow conditions, and more so for the high variability

condition than low variability condition. TOA estimation errors that are larger for Phase 3c

trials than 3b trials would indicate that subjects were not able to identify the speed profiles

solely from the observed stimuli.

Speed Profile

(Speed/Variability)

Profile

Trained

Profile

Labelled

Number of

Trials

Route

Distance

(in pixels)

Slow/Low Yes No 3 500

Slow/Medium No No 3 500

Slow/High Yes No 3 500

Moderate/Low No No 3 500

Moderate/Medium No No 3 500

Moderate/High No No 3 500

Fast/Low No No 3 500

Fast/Medium Yes No 3 500

Fast/High Yes No 3 500

Page 76: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

69

5.4 RESULTS

5.4.1 Phase 3a: Developing a Library of Speed Profiles

Recall that the goal of phase 3a was to expose subjects to 100 vehicles travelling with a set

of shared speed profiles to enable them to develop a mental model of the profile in terms of

average speed and speed variability. Analyses were conducted to evaluate how well

subjects were able to estimate the mean speed and speed windows for each speed profile.

Table 5-4 shows actual and estimated mean speed and speed windows for the fifth (final)

trial of each speed profile, and a difference score representing actual minus estimated

values.

Table 5-4. Actual speed scores, estimated speed scores, and speed estimation error scores

Actual Subjects! Estimates Estimation error

(Actual – Estimate) Speed

Profile Mean Interval Mean Window Mean Window

Slow/Low 35 3.75 34.7 4.13 0.3 -0.38

Slow/High 35 11.15 34.7 9.76 0.3 1.39

Fast/Low 75 3.75 75.5 4.26 -0.5 -0.51

Fast/High 75 11.15 73.8 9.38 1.2 1.77

*Data shown in table are for last trial of each speed profile

5.4.1.1 Estimated Mean Speed

The first element required to develop a model of the speed profile was to estimate mean

speed. Figure 5-4 shows the subjects! estimated mean speed (dashed lines) on the fifth

(final) trial for each speed profile, as compared to the actual mean speed (solid lines). A 2 x

2 x 5 ANOVA revealed no significant three-way interaction. But, subjects were sensitive to

the difference in speed, estimating the mean speed for fast stimuli (M = 74.7) as faster than

for slow stimuli (M=34.7) (F(1,15)=5630.17, p<.001). The difference between subjects!

estimated mean speed and the actual mean speed was negligible (p>.05) in all conditions.

The subjects! estimates of mean speed did not differ as a function of variability

(F(1,15)=1.124, p>.05). Although not shown in Figure 5-4, mean speed estimates did differ

Page 77: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

70

as a function of trial (F(4,60)=3.212, p<.05), with subjects underestimating mean speed on

the first trial, but quickly calibrating toward the actual mean for the remaining trials.

Figure 5-4. Estimated mean speed on 5th and final trial (shown with 95% CI)

5.4.1.2 Estimated Speed Windows

The second element required to develop an internal model of a speed profile was the

subjects! ability to estimate the speed windows (or the window that captures 95% of the

mean speeds). A 2 x 2 x 5 ANOVA was conducted. Neither a three-way interaction nor a

main effect of average speed on speed window estimation was found (p>.05). This

supports the premise that the subjects understood that actual speed windows do not differ

as a function of speed, but only as a function of variability, as can be seen in Table 5-4. A

main effect of variability (F(1,15) = 47.63, p<.001), on the other hand, revealed that subjects

did estimate larger speed windows for the high variability condition (M = 8.85) than for the

low variability condition (M = 4.82), reflecting an understanding of the relationship between

interval size and variability. A significant trial x variability interaction (F (4,60)=12.61,

p<.001) revealed that estimated window sizes increased as a function of trial repetition in the

high variability condition, but decreased in the low variability condition. Similarly, a Trial x

Speed interaction (F (4,60)=2.87, p<.05) revealed that subjects! window size estimates

increased as trials progressed for fast speeds and decreased for slow speeds. These

Page 78: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

71

interactions reflect an improvement in performance due to feedback, as will be shown in the

following section.

Figure 5-5 presents the speed window difference scores (actual minus estimated speed

window size) as a function of speed, and trial, where a perfectly calibrated speed window

estimate yields a difference score of zero. A 2 x 2 x 5 ANOVA revealed a significant speed x

trial interaction (F(4,60)=2.874, p<.05). (No significant three way interaction was found.) As

can be seen in Figure 5-5, initially subjects tended to overestimate the size of the speed

windows for the slow speed trials and underestimate the speed windows for the fast speed

trials – however, with feedback after each trial, both converged towards the actual speed

interval size, with a low difference score by the fifth trial. Analyses of just the fifth trial

revealed no significant difference in speed window difference scores as a function of speed

(p>.05).

Figure 5-5. Speed window difference scores as a function of speed (slow, fast) and trial number (shown plotted with ±1 Std Error)

A variability x trial interaction was also observed (F(4,60)=12.607, p<.001), as seen below in

Figure 5-6. Subjects overestimated speed windows for low variability stimuli and

underestimated speed windows for high variability stimuli. A subsequent analysis of the final

trial showed that the speed window difference score was significantly higher for low

variability conditions than high variability conditions (F(1,15)=9.218, p<.05).

Page 79: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

72

Figure 5-6. Speed window population difference scores as a function of variability (low, high) and trial (shown plotted with ± 1 std error)

In summary, in a small number of trials with performance feedback, subjects were able to

effectively develop an internal model of each speed profile and characterize both mean and

speed intervals with little error.

5.4.2 Phase 3b: Estimating TOA uncertainty for trained profiles

Recall that the goal of this second phase of the experiment was to determine if subjects

could apply their internal model of the four speed profiles to accurately estimate TOA

uncertainty for trained distances. A second goal was to determine if subjects could use their

knowledge of trained speed profiles to estimate new (untrained) distances. Each subject

completed five trials for each speed profile x distance combination, in a random order. The

first trial of each condition was considered practice and the analyses that follow were

conducted on the remaining four trials.

5.4.2.1. Estimated TOA Uncertainty

Table 5-5 presents the actual TOA interval size, the subjects! estimated TOA window size,

and the difference between the actual and estimated window size. Recall that the expected

TOA was provided to the subjects to eliminate potential confounds associated with

misestimating the TOA itself.

Trial

Page 80: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

73

Table 5-5. Actual TOA scores, estimated TOA scores, and TOA estimation error scores

Independent Variables

Speed Variability Distance

Actual TOA Interval

Estimated TOA Window

TOA Estimation Error (Actual –

Estimated)

Slow Low Short 2.8 4.6 -1.8

Slow Low Medium 5.6 5.5 0.1

Slow Low Long 11.2 6.6 4.6

Slow High Short 12.5 9.3 3.2

Slow High Medium 25.0 10.3 14.7

Slow High Long 50.0 12.4 37.6

Fast Low Short 0.6 2.7 -2.1

Fast Low Medium 1.2 3.4 -2.2

Fast Low Long 2.4 3.8 -1.4

Fast High Short 1.9 4.6 -2.7

Fast High Medium 3.8 5.4 -1.6

Fast High Long 7.6 6.6 -1.0

A 2 x 2 x 3 x 4 repeated measures ANOVA was conducted to determine if the subjects!

estimated TOA window differed as a function of the experimental conditions (speed,

variability, distance, repetition). Again, this was conducted to investigate whether the

subjects could distinguish among the experimental conditions, and to evaluate whether

performance differed across the four repetitions due to learning, fatigue, or scenario

differences. The analysis revealed a significant main effect of distance (F(2,30)=13.979,

p<.001), and post-hoc comparisons revealed that subjects! estimated TOA window size

increased as a function of distance. TOA windows were smallest for short routes (M=5.29),

followed by medium routes (M=6.15), and long routes (M=7.34). All pairwise comparisons

were significantly different (p<.01). This suggests that subjects did factor the distance of the

route into their TOA uncertainty estimation.

A significant speed x variability interaction (F(1,15)=8.95, p=.009) provided evidence that

subjects! estimated TOA uncertainty differed as a function of the two main experimental

independent variables. As can be seen by the dashed lines in Figure 5-7, at slow speeds

Page 81: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

74

the difference in estimated window size between low (M=5.55) and high (M=10.66)

variability was greater than the difference between low (M=3.3) and high (M=5.5) variability

at fast speeds.

Figure 5-7. Estimated TOA window size as a function of speed (slow, fast) and

variability (low, high). Error bars represent ±1 std error

Next, the subjects! estimated TOA window was compared to the actual TOA interval by

examining the 95% CI around the subjects! estimated TOA window size, to determine if the

actual TOA window size was contained within the 95% CI. Figure 5-8 shows the subjects!

estimated TOA window size (dashed lines) and the actual TOA interval size (solid lines) as a

function of distance for each combination of speed and variability. As indicated by the circles

in the figure, in all but two conditions (slow/low variability/medium distance and fast/high

variability/long distance) the 95% CI around the subjects! estimated window size did not

contain the actual TOA interval size, indicating that the subjects! estimate was significantly

different than the actual interval size. As can be seen in the upper left graph in Figure 5-8,

for the slow/low variability condition, the subjects overestimated TOA window size for short

distances and underestimated TOA window size for long distances. For the slow/high

variability condition (upper right), the subjects underestimated TOA window size for all

distances. The opposite was true for the fast/low variability condition (lower left) and

Page 82: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

75

fast/high variability conditions, in which the subjects overestimated TOA window size for all

distances (with the exception of fast/high variability/long distance).

Figure 5-8. Actual (solid lines) and estimated (dashed lines) TOA windows as a function of speed (slow, fast), variability (low, high), and distance (short, medium, long). Graph shows 95% CIs around the estimated TOA window. Circles show data where actual interval widths fall outside of the 95% CI of the subjects! estimates.

The pattern of overestimated TOA window size for short distances and underestimated TOA

window size for long distances can be seen more readily in Figure 5-9, which adjusts the

subjects! estimates to correct for the initial estimation error seen in the trained (medium)

distance. That is, when the subjects! TOA uncertainty estimate for the medium distance is

shifted to match the actual TOA uncertainty, one can clearly see that the subjects

systematically overestimate TOA uncertainty for short distances, and underestimate TOA

uncertainty for long distances.

Related to this finding, in the debrief session, all subjects were shown three scenarios of a

vehicle with the same speed profile travelling routes of different lengths (see Appendix C3),

to assess their understanding of the effect of distance on TOA uncertainty. In the first

scenario, the route distance was 4 miles, and the TOA window around an expected TOA of

320 seconds was 80 seconds. In the second scenario, subjects were asked to estimate the

TOA window for a vehicle with the same profile travelling a route that was half as long – 2

miles. The third scenario asked subjects to estimate the TOA window for an 8 mile route.

For the second (2 mile route) scenario, all but two subjects correctly estimated a TOA

Page 83: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

76

window of 40 seconds (which was half the TOA window provided in scenario 1). One

subject estimated a shorter window (20 seconds) and one estimated a longer window (60

seconds). For scenario 3 (8 mile route, which was twice the distance of scenario 1), all but

four subjects correctly estimated the TOA window to be 160 seconds (or twice as long as

provided in Scenario 1). One subject estimated a shorter TOA window (90 seconds), while

three estimated longer TOA windows (180, 200, 320 seconds). While there were some

exceptions, the majority of the subjects demonstrated an accurate understanding of the

relationship between TOA uncertainty and route distance. Yet in the experiment, the

subjects systematically misestimated TOA uncertainty as a function of distance.

Figure 5-9. Actual (solid lines) and estimated (dashed lines) TOA windows as a function of speed (slow, fast), variability (low, high), and distance (short, medium, long), corrected for

initial estimation error at the trained (medium) distance.

5.4.2.2. TOA Uncertainty Estimation Error

TOA estimation error scores, shown in Table 5-5, were computed by subtracting the

subjects! estimated TOA window size from the actual TOA interval size generated from the

population distribution. A TOA window estimation error score of zero reflects a perfectly

calibrated TOA window estimate. A repeated measures ANOVA was conducted on the

absolute values of the TOA window estimation error scores to determine if the amount of

estimation error (regardless of direction) differed as a function of the independent variables.

As evident in Figure 5-10, the ANOVA yielded a speed x variability x distance interaction

(F(2,30) = 202.37, p<.01). Subsequent simple effect tests revealed a significant variability x

Page 84: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

77

distance interaction for the slow speed condition (F(2,30)=258.364, p<.001). In the slow/low

variability condition, TOA window estimation error was not significantly different for short

(M=2.02) and medium (M=1.68) distances, p>.05), but TOA window estimation error was

larger for long distances (M=5.18) than both medium and short distances, (p<.001). In the

slow/high variability condition, TOA window estimation error was highest for long distances

(M=37.99), followed by medium distances (M=14.87) and short distances (M=5.27). All

pairwise comparisons were significantly different (p<.001). For fast speeds, no interaction

was found for variability x distance (p>.05). That TOA uncertainty estimation error was

greatest in the slow/high profile for long distances suggests that subjects treated this

condition as a symmetrical distribution, and failed to recognize the substantial increase in

TOA interval size associated with this condition.

Figure 5-10. TOA window estimation error (absolute values), as a function of speed (slow,

fast) and variability (low, high). (Plotted with ±1 SE).

5.4.2.3 Estimated TOA Interval Symmetry

In addition to the size of the subjects! estimated TOA window, the symmetry of the estimated

TOA window was also of interest. As was discussed previously, given a normal distribution

of speeds, the resulting objectively determined distribution of TOAs is necessarily positively

skewed – and furthermore the amount of positive skew changes as a function of speed and

variability – with the greater amount of positive skew for stimuli with slow speeds and high

Page 85: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

78

variability. To assess the degree to which the subjects! estimated TOA windows

appropriately reflected this positive skew, a TOA symmetry score was calculated using the

following equation (and also used in experiment 2):

TOA Symmetry Score = |UB –TOA| - |LB-TOA|;

where TOA = subjects! estimates of TOA,

LB = TOA Interval Lower Bound,

UB = TOA Interval Upper Bound (UB)

If the TOA symmetry score is equal to zero, the interval created by the subject was perfectly

symmetrical; if positive, then the window was positively skewed with the upper portion larger

than the lower portion; and if negative, then the window was negatively skewed with the

lower portion larger than the upper portion.

Symmetry scores were first computed on the population data for each combination of speed

and variability as shown in Figure 5-11. As can be seen from the solid lines, the intervals

were either symmetrical or positively skewed. The skew was highest in the slow speed/high

variability condition. This is consistent with the inverse relationship between speed and

TOA, as discussed previously. Next, TOA symmetry scores were calculated for each

subject!s estimated TOA windows. As can be seen from the dashed lines in Figure 5-11, the

subjects! symmetry scores also tended to be neutral or slightly positive (Overall Mean = .05),

therefore suggesting that subjects provided TOA windows that were slightly positively

skewed. Note that for fast speeds the amount of skew was low and quite comparable to the

actual symmetry score; however, for slow speeds with high variability, symmetry scores

were substantially less skewed than the actual symmetry scores.

The 95% CIs around the subjects! symmetry scores were examined to determine if the CI

contained the actual symmetry score. If the 95% CI did not contain the actual symmetry

score, the subjects symmetry score was considered significantly different than the actual

symmetry score. This was the case for all combinations of variability and distance for the

slow condition, in which the actual TOA windows were more positively skewed than the

subjects! estimates, with the exception of slow/low variability/short distance, which was not

different than the actual symmetry score. In contrast, subjects! estimates of TOA symmetry

for the fast stimuli were not significantly different than the actual symmetry score for any

variability or distance combination.

Page 86: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

79

Figure 5-11. Actual and estimated TOA symmetry scores. Estimated scores are plotted with 95% CIs. Top graph: slow speed; bottom graph: fast speed.

5.4.3 Phase 3c: Estimating TOA Uncertainty for Untrained Profiles

The goal of this third phase was to determine if, after observing an unlabelled vehicle

travelling a medium length route with 10 speed changes, subjects could determine whether it

was a speed profile for which they had been trained – and apply TOA windows accordingly.

A second goal was to determine whether subjects could extrapolate from the trained speed

profiles to assess TOA uncertainty for speed profiles on which they were not trained. In

contrast to Phase 3b, only one distance (medium route, 500 pixels) was used for all trials in

this phase. Recall that after viewing the vehicles, subjects were asked to first determine

Page 87: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

80

whether the vehicle!s speed profile was new or old (i.e., one of the four they had learned in

training). Subjects were then asked to select which of the four trained profiles it was most

similar to – or state the profile they believed it matched – and provide a similarity rating from

1 (very low) to 5 (very high). Then, subjects estimated the TOA windows and entered them

in the same manner as Phase 3b.

5.4.3.1 Identification of Trained and Untrained Profiles

Subjects correctly identified 63% of the profiles as trained or untrained, with a mean of 17

correct identifications out of 27 trials (with a range of 12 – 21). As shown in Figure 5-12, the

correct identification rate for trained profiles (76%) was higher than for new, untrained

profiles (52%) (X2 = 26.27, p<.05).

Figure 5-12. Identification of "trained! and "untrained! profiles

5.4.3.2 Correct Identification of Speed Profile

Next, subjects were asked to identify which of the original four speed profiles the vehicle

was most similar to – or, if the subject stated that it was a trained profile, they were asked to

identify which one of the four profiles it was. Figure 5-13 shows the percent of trials that

were identified correctly for each of the four trained profiles. (Note that this analysis uses

only the four trained profiles, since there was no "correct! closest profile for the untrained

profiles). Identification rates were highest for the slow/low (88%), followed by slow/high

(73%), fast/high (67%) and fast/low (65%) profile. Chi square residual analyses revealed

Trained Untrained Speed Profile

Pe

rce

nt

Co

rre

ct

Page 88: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

81

that the identification rate for the slow/low condition was above the mean percent correct

value and the fast/low and fast/high rates were below the mean percent correct value (X2=

7.8, p<.05). The higher identification rate for the slow conditions is likely attributed to the

skewed distribution yielding a range of TOAs that was much larger, and thus arguably more

salient, than for all other conditions.

Figure 5-13. Correct Identification of Speed Profile

5.4.3.3 Similarity Ratings

As summarised in Table 5-6, when asked to rate the similarity of the observed profile to the

trained profile that the subjects identified as being the closest, subjects rated the trained

profiles higher (more similar) than the untrained profiles. That is, when subjects reported that

it was one of the four trained profiles and it actually was one of the four trained profiles,

subjects rated the similarity rating higher (M = 4.2) than if subjects reported that it was a new

profile that they had not seen before and it was indeed a new profile (M = 2.6)

Table 5-6. Mean similarity ratings

Actual profile was: Subject Response

Trained Untrained

Trained 4.2 3.8

Untrained 3.0 2.6

Pe

rce

nt

Co

rre

ct

Slow/Low Slow/High Fast/Low Fast/High Speed Profile

Page 89: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

82

5.4.3.4 Estimated TOA Window

Next, subjects were asked to estimate the TOA uncertainty window in the same manner as

in Phase 3b. Figure 5-14 shows the subjects! estimated TOA windows (dashed lines)

plotted with 95% CIs, and the actual TOA intervals (solid lines). A 3 x 3 x 3 ANOVA was

conducted using all trials (trained and untrained speed profiles) to determine if the subjects!

estimated TOA window size differed as a function of speed, variability, and trial. (Recall that

for this phase, a new speed (moderate/45) and a new level of variability (medium/6) were

introduced, creating three levels of each variable. Distance was held constant at medium, as

used for Phase 3a.)

Figure 5-14. Estimated and Actual TOA windows for both trained and untrained profiles

Results revealed a significant main effect of speed (F(2,30) = 17.24, p<.05). (No three-way

or two-way interactions were found.) Follow-up post-hoc tests did not reveal any significant

difference between the subjects! estimated TOA window for the slow speed condition

(M=6.7) and the estimated TOA window for the newly added moderate speed condition

(M=7.0), but it did reveal that both were significantly larger than the TOA window for high

speeds (M=4.3) (p<.001). This result suggests that subjects may have simply applied the

TOA window for slow speeds to the new moderate speed condition, rather than interpolating

between the slow and fast conditions.

Page 90: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

83

Subjects! estimated TOA windows differed as a function of variability (F(2,30)=10.84,

p<.001). TOA window estimates for the low variability condition (M=4.8) were lower than for

the new medium variability condition (M=6.353) (p<.05), and for the high variability condition

(M=6.9) (p<.05). Subjects' TOA window estimates for medium and high variability were not

significantly different (p>.05). This result suggests that subjects may have applied the TOA

window for high variability to the new medium variability condition.

There was a significant effect of repetition (F(2,30)=4.33, p<.05). Results revealed that

estimated TOA windows were shorter on Trial 1 (M=5.7) than Trial 2 (M=6.4) (p<.05), but

Trial 3 (M=6.0) was not significantly different than either Trial 1 or 2 (p >.05).

Next, subjects! estimated TOA windows (dashed lines in Figure 5-14) were compared to the

actual TOA intervals (solid lines in Figure 5-14). In the slow speed condition, subjects

underestimated the TOA window size for medium and high variability conditions. In the

moderate speed condition, subjects overestimated the TOA window for low variability and

underestimated the TOA window for the high variability condition. In the fast speed

condition, subjects overestimated the TOA window for both low and medium variability.

An analysis was conducted to determine if the subjects! estimated TOA window differed as a

function of whether the subject thought the speed profile was new or old. For each profile,

the mean TOA window was calculated for those that answered “old/trained” and for those

that answered “new/untrained”. Due to unequal numbers of samples across conditions,

statistical analyses were not conducted, but mean TOA windows are presented in Figure 5-

15 below. Averaged across all speed profiles, the mean estimated TOA window for trials in

which the subject thought the speed profile was "new! was 6.9 (170 trials, SD = 3.9), as

compared to an average of 5.44 (261 trials, SD = 3.7) for trials in which the subject thought

that the speed profile was old. This finding reflects presumably greater uncertainty

associated with an unknown profile. As can be seen in Figure 5-15, the TOA window

estimates for each profile were very similar for those that the subjects thought had a new

speed profile and those which they thought were old, with the exception of two new speed

profile conditions, 45/6 and 45/9. In both cases when subjects correctly identified these as

new profiles, the estimated TOA uncertainty was much larger than when they incorrectly

identified them as old profiles.

Page 91: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

84

Figure 5-15. TOA window estimates for each speed profile plotted by subjects! identification of old/trained or new/untrained profile (shown with +/- 1 Std. Error). Parentheses at the bottom indicate whether the speed profile was actually trained or untrained.

Lastly, subjects! estimated TOA windows from Phase 3c were compared to those from

Phase 3b. The only difference between the two conditions was that the speed profiles were

labelled (by colour code) in 3b, but was not labelled in 3c. Figure 5-16 compares the TOA

window estimation error scores for Phase 3b and 3c, for each of the four trained speed

profiles, revealing a phase x profile interaction (F(3,45)=4.92, p=.005). The difference in

TOA window estimation error scores between Phase 3b (labelled) and Phase 3c (not

labelled) was significant only in the slow/high profile condition (F(1,15)=7.24, p=.017), with

estimation error higher when the profile was not labelled (Phase 3c). This result suggests

that there may be an operational benefit to providing display technology or otherwise

identifying the speed profile to operators, since TOA uncertainty estimation error was larger

in the absence of the profile identification. There were only small differences (p<.05)

between the two studies for the other three profiles, suggesting that, for these profiles,

subjects! TOA window estimates were not altered if the speed profile was labelled or

unlabelled.

Page 92: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

85

Figure 5-16. TOA window error scores when the profile was colour-coded (phase 3b) and not colour-coded (phase 3c) (shown plotted with +/- 1 Std. Error)

Understanding the process of extrapolating to untrained speed profiles. Further analyses

were conducted to explore how subjects extrapolated from trained profiles to assess TOA

uncertainty for a new speed profile. In particular, given that subjects were trained on low

and high variability conditions, this analysis set out to determine how subjects estimated

TOA windows for the new, untrained, medium variability conditions. Nisbett, Krantz, Jepson

& Fong!s (1982) research on inductive reasoning suggests two mechanisms that could

explain how subjects went about estimating TOA uncertainty for untrained conditions,

yielding two competing hypotheses for consideration. First, subjects could have estimated

the TOA window for low and high variability based on their model of uncertainty developed

in Phase 3a, and averaged the two numbers to determine the TOA window size for the

medium variability condition. This is the simplest form of a mathematical model, or statistical

heuristic (Nisbett, Krantz, Jepson & Fong, 1982), that could reasonably be expected to be

employed by the subjects during the experiment. Alternatively, subjects could have selected

the trained condition that they felt was closest to the new condition, and simply assume the

same TOA window. This is commonly referred to in the literature on judgment under

uncertainty as the representativeness heuristic (Tversky & Kahneman, 1982).

Speed Profile

TO

A W

ind

ow

Err

or

Sco

re

Page 93: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

86

To test these two hypotheses, a multiple regression analysis was conducted to determine

which hypothesis better predicted the outcome measure of estimated TOA window for the

new, untrained profiles. The first hypothesis was modelled as a predictor variable that was

computed as the average of each subject!s estimated TOA window for the low and high

variability conditions. The second hypothesis was modelled as a predictor variable that

represented the subject!s estimated TOA window for the speed profile that he/she stated

most closely matched the stimuli. With both predictor variables in the equation, the model

accounted for 80% of the variance in estimated TOA window times.

Referring to Table 5-7, examination of each predictor variable revealed that only the

predictor variable that represented the subject!s estimate for the closest profile was

significant, with a standardized beta score of .878. This suggests that subjects

overwhelmingly employed the representativeness heuristic – that is, they applied the TOA

window of the trained profile that was most similar rather than the average of the low and

high variability. However, the reader must be aware of one potential limitation of this finding.

Because subjects were first asked to identify the closest trained profile and then determine

the TOA window, it is possible that this experimental method cued the subjects to use the

representativeness heuristic.

Table 5-7. Multiple regression analysis to determine which model best explains how subjects estimated TOA uncertainty for untrained profiles.

Model B Beta t sig

Closest Profile .833 .878 11.012 .000

Average of Low and High 0.29 .22 .278 .782

R2 = .804

A subsequent multiple regression analysis was conducted that included the subjects!

similarity rating scores, in addition to the two heuristics as predictor variables. This was

conducted to determine if the degree of rated similarity to a trained profile accounted for any

more variance in estimated TOA window size. Results revealed that it did not – producing a

R2 of .805, as compared to the previously obtained .804. The similarity score predictor

variable was not significant (t=.681, p>.05) with a standardized beta weight of -0.032.

Page 94: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

87

5.5 CONCLUSION AND OPERATIONAL IMPLICATIONS

This experiment addressed two methodological limitations identified in experiment 2. The

inclusion of Phase 3a, an extensive training session that exposed subjects to over 100

vehicles for each of four profiles, served to increase the sample size from which subjects

could understand and characterize the TOA distribution, and also changed the task from one

of (potentially) simply observing and describing a single sample of displayed speed

readouts, to one of developing an internal model for application in subsequent phases of the

experiment. Additionally, changes were made to the experimental protocol to improve

subjects! ability to understand the task.

Not only did this experiment replicate the results of experiment 2 (reported in Chapter 4), it

also revealed further important insights into the TOA uncertainty estimation problem. The

results of Phase 3a revealed that using a training-with-feedback paradigm, subjects can

develop accurate models of speed profiles based on relatively short exposure to only 100

vehicles. Phase 3b then asked subjects to apply their models of each speed profile to

estimate TOA uncertainty, for both the trained distance and two new, untrained distances

(one shorter and one longer). For the trained distance, subjects estimated the TOA

Uncertainty for the slow/low profile very accurately, but underestimated TOA Uncertainty for

slow/high while overestimating for the fast/low and fast/high profiles. The untrained distance

conditions revealed systematic biases – overestimating TOA uncertainty for short distances

and underestimating TOA uncertainty for long distances. The TOA uncertainty estimation

error was greatest for the slow/high profile, and this was attributed to subjects! assumption

that the TOA distribution was symmetrical when in fact it is highly positively skewed.

Phase 3c set out to determine how well subjects could identify an unlabeled speed profile by

observing speed fluctuations and then apply the appropriate TOA uncertainty estimate.

Results revealed that subjects could indeed identify the correct profile about 73% of the

time, and consistently applied TOA uncertainty estimates that were similar to estimates of

the labelled vehicles in Phase 3b; however TOA estimation error was higher when the

vehicle was unlabelled in the slow/high profile. Further, Phase 3c examined how subjects

Page 95: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

88

estimated TOA uncertainty for new, untrained speed profiles. A series of analyses

suggested that subjects applied a representative heuristic by selecting the trained speed

profile that was most similar to the observed vehicle!s profile, and applying the TOA

uncertainty estimate of that trained profile.

Implications of these results for the proposed model of TOA uncertainty are discussed next

in Chapter 6 and their implications for the operational environment of future 4-D air traffic

control operations are then discussed in Chapter 7.

Page 96: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

89

CHAPTER 6: MODEL REFINEMENT

6.1 TIME OF ARRIVAL UNCERTAINTY MODEL RESTATED AND REFINED

The preliminary model of TOA uncertainty, first presented in chapter 2, and reproduced

below in Figure 6-1, was generated based on input from subject matter experts (ATC), after

the first experiment.

Figure 6-1. Original Model of TOA Uncertainty

After exploring the model systematically in two tightly controlled human-in-the-loop

experiments, a proposed modification to the model is presented in Figure 6-2. Results of

experiment 3c revealed that subjects did not change the way they estimated TOA

uncertainty as a function of whether they had seen the profile before or not, as was

previously hypothesised. Rather in all cases, after observing the vehicles, subjects appeared

to determine which of the profiles in their personal library that it most closely matched, and

applied that estimate of TOA uncertainty directly.

Figure 6-2. Refined Model of TOA Uncertainty

Page 97: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

90

6.2 SOURCES OF TIME OF ARRIVAL UNCERTAINTY ESTIMATION ERROR

Based on the results of experiments 2 and 3, it is now possible to explain this model further,

by characterizing the sources of TOA uncertainty error and determining how each of the

three processes represented in the refined model of Figure 6-2 (i.e., Build library of speed

profiles; Identify closest speed profile, and Apply estimate of TOA uncertainty) contribute to

TOA uncertainty estimation error observed in study 3c. A multiple regression analysis was

conducted to determine the relative contribution of each of the model processes in

explaining the observed TOA estimation error. In the multiple regression analysis, the

subjects! ability to build a library of speed profiles was modelled using two variables from

study 3a: 1) speed mean estimation error and 2) speed window estimation error. Subjects!

ability to identify the closest speed profile was modelled using two variables from study 3c:

1) error associated with identifying whether the profile was a trained one, or a new one, and

2) error associated with identifying the speed profile for trained profiles only. Finally,

subjects! ability to estimate TOA uncertainty was modelled as TOA uncertainty estimation

errors from study 3b.

The multiple regression revealed that, collectively, the five predictor variables accounted for

75% of the variance. However, as can be seen in Table 6-1, the variable that contributed

most significantly to this was the subjects! TOA uncertainty estimation error score (t=4.65,

p=.001). This suggests that, based on the experimental studies, TOA uncertainty estimation

error can be attributed primarily to the actual TOA estimation process itself, and not to the

process associated with developing the library of speed profiles, or comparing observed

vehicles to the library and identifying the closest profile.

It is prudent to raise one caveat associated with this conclusion. That is, given the

experimental method employed, it was necessary to provide feedback to subjects in study

3a, during which they built their library of speed profiles. While it is believed that this

feedback would be obtained implicitly over time by operators in the actual environment, it is

unknown how long it would take to achieve the same level of performance without the

explicit feedback. Actual operations may yield more error in the first two processes (Building

a library of speed profiles and Identifying the closest profile from the library) than was

observed in these studies.

Page 98: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

91

Table 6-1. Multiple Regression to determine source of TOA Uncertainty Error

Variable B Beta t Sig

Speed Mean Estimation Error 0.082 0.032 0.187 0.855

Speed Window Estimation Error 0.203 0.192 1.127 0.286

Old/New Identification 0.034 0.037 0.218 0.832

Speed Profile Identification -0.192 -0.163 -0.914 0.38

TOA Uncertainty Estimation Error for labelled profiles

0.845 0.793 4.65 0.001

R2 = .75

6.3 CAUSES OF TIME OF ARRIVAL UNCERTAINTY ESTIMATION ERROR

Having identified the process of TOA uncertainty estimation as the potentially largest

contributor to error, the next step is to understand the causes of these errors. Further

examination of the results from experiments 2, 3b, and 3c reveal two systematic biases that

contribute to TOA uncertainty estimation errors.

6.3.1 Assumption of Symmetry

First, data in both experiments 2 and 3 revealed that subjects typically failed to recognize

the asymmetrical shape of the TOA distribution and thus underestimated the amount of

actual TOA uncertainty. (Recall, that approximately half of the subjects had completed at

least one university-level course in statistics). As discussed previously, if speed is normally

distributed, then by definition TOA is positively skewed, since TOA is the reciprocal of

speed. Furthermore, this amount of positive skew is more pronounced for vehicles travelling

at slower speeds and with higher variability. However, subjects appeared to treat all

distributions as symmetrical, or in any case to underestimate the amount of positive skew.

This can be seen clearly in the symmetry metric calculated in both experiment 2 (shown

previously in chapter 4) and experiment 3b, shown previously in chapter 5 and reproduced

here in Figure 6-3.

A second piece of evidence that points to the assumption of normality explanation as a

cause of TOA uncertainty estimation error is the speed x variability interaction observed in

Page 99: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

92

experiment 3b, shown previously in chapter 5 and reproduced here in Figure 6-4. As can be

seen, subjects estimated the TOA window to be larger for slower vehicles than faster

vehicles, but failed to recognize the sharp increase in TOA window size associated with the

slow/high variability condition due to the positively skewed TOA distribution.

Figure 6-3. TOA window symmetry scores from experiment 3b show that subjects

underestimated the magnitude of positive skew for slow speeds.

Page 100: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

93

Figure 6-4. Estimated and actual TOA window size from experiment 3b shows that subjects underestimated the window size for the slow/high condition, which is highly skewed.

(Error bars represent ±95% CI)

Lastly, subjective reports elicited during the debrief interview also support the assumption of

normality hypothesis. Most subjects expressed a strong bias towards symmetrical TOA

windows, as evidenced by the following comments:

• “My TOA windows were always symmetrical, because speed was always

symmetrical.”

• “TOA windows were symmetrical. There is no point in making asymmetrical

windows. Asymmetrical windows would have been strange.”

• “TOA windows were symmetrical, I have an obsession with symmetry.”

• “I thought all TOA windows should be symmetrical.”

• “I made nice even windows around the TOA.”

• “Symmetrical windows just seemed natural.”

In some cases, subjects expressed confusion or uncertainty as to whether the windows

should be symmetrical, positively skewed, or negatively skewed. The following quotes

demonstrated that these subjects did not clearly understand that TOA distributions are

positively skewed:

Page 101: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

94

• “As variability increased, the TOA windows tended to be more asymmetrical, but not

always in the same direction. The red profile [slow/low] had more early vehicles, and

the blue profile [fast/high] had more later vehicles.”

• “My TOA windows tended to be asymmetrical – sometimes bigger on the left [more

earlier than later] and sometimes bigger on the right [more later than earlier]. I don!t

know why.”

However two subjects did recognize that TOA windows were asymmetrical, as evidenced by

the following quotes:

• “ The TOA windows were symmetrical, except for the yellow profile [slow / high

variability] because they lagged a bit.”

• “TOA windows were asymmetrical. In actuality, there is no way that arrivals would

be symmetrical – there is more of a chance of being slow and taking longer.”

6.3.2 Aversion to Extremes

A second potential source of TOA Uncertainty estimation error was an apparent aversion to

extremes. In both experiment 2 and 3, subjects overestimated conditions of small

uncertainty and underestimated conditions of large uncertainty. This systematic bias can be

seen in the speed x variability interaction observed in both experiment 2 and 3b shown

above in Figure 6-4. Subjects overestimated TOA uncertainty for the fast speed profiles that

possessed low actual TOA uncertainty (fast / low variability and fast / high variability) and

underestimated TOA uncertainty for the slow profiles which possessed higher actual TOA

uncertainty. Finally this can also be seen in the adjusted distance extrapolation results of

experiment 3b, reproduced here in Figure 6-5. Subjects systematically overestimated the

low actual TOA uncertainty of short distances and underestimated the higher actual TOA

uncertainty of longer distances. This was seen for each of the four speed profiles. This

systematic aversion to extremes is compatible with previous research on the estimation of

probabilistic events conducted by Tversky and Simonson (2000), which suggests that most

people tend to avoid extremes and choose an intermediate option if available.

Page 102: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

95

Figure 6-5. TOA window estimates as a function of speed profiles for trained and untrained distances, showing subjects! systematic bias for overestimating low uncertainty and underestimating high uncertainty.

Page 103: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

96

CHAPTER 7: SUMMARY, LIMITATIONS AND IMPLICATIONS

The ability to estimate TOA uncertainty is a critical cognitive process in supervisory control

systems such as air traffic management. It is expected to become a critical issue for the next

generation aviation system, which will rely on strict conformance and monitoring of 4-D

trajectories both in the air and on the ground. Chapter 1 discussed the introduction of new

decision support system technologies to support 4-D operations. There the need for

controllers to develop effective estimates of TOA uncertainty rather than relying on

automation for this purpose was identified.

In chapter 2, experiment 1 solicited input from ATC subject matter experts to shed light on

how TOA is estimated in the real world, together with the uncertainty associated with those

estimates. Several factors believed to impact TOA uncertainty estimates were identified.

This experiment led to the development of a preliminary model of TOA uncertainty

estimation.

In chapter 3, literature from domains such as decision making under uncertainty and

probabilistic reasoning was applied to develop a generic experimental method and metrics

for evaluating TOA uncertainty, extracted from relevant aspects of the ATC TOA estimation

task. This experimental method was then applied in experiment 2 in chapter 4, which

identified that subjects tended to over-estimate TOA uncertainty for fast speeds, and

underestimate it for slow speeds. The latter was attributed to underestimating the amount of

positive skew associated with the TOA distribution.

In chapter 5, the results of experiment 3 addressed two limitations of the experimental

paradigm and replicated the results of experiment 2. Experiment 3 also extended the

findings to better understand how subjects estimate TOA uncertainty for untrained distances

and profiles. From experiment 3, it was concluded that subjects appeared to apply the

representativeness heuristic when estimating TOA uncertainty for new speed profiles that

they have not previously encountered.

Page 104: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

97

7.1 MODEL ASSUMPTIONS AND LIMITATIONS

The refined model of TOA uncertainty estimation presented in chapter 6, along with the

sources and causes of estimation error, are based on certain assumptions as follows.

7.1.1 Speed is Normally Distributed

First, the model and results are predicated on the assumption that speeds are normally

distributed, which was the case in the two experimental studies. Although this is a

simplification of the operational environment, data from a recent airport human-in-the-loop

taxiing simulation (Williams, Hooey & Foyle, 2006) nevertheless support that this is the case

for nominal taxi conditions. Figure 7-1 presents a histogram of a short sample of taxi speed

data taken from that study, in which the pilot was taxiing in a medium-fidelity aircraft

simulator while trying to maintain a commanded average taxi speed, unencumbered by

traffic delays. As can be seen, those taxi speeds do approximate a normal distribution in this

simple taxi condition. Of course, this would not necessarily hold in more complex scenarios

with traffic and adverse weather. In addition, it is unknown whether the subjects understood,

or assumed, that the distribution of speeds was normally distributed. It is indeed possible,

for example, that subjects envisioned a uniform distribution in which case they would not

necessarily expect the resulting TOA distribution to be positively skewed. This could have

accounted for the subjects! underestimates of positive skew.

Figure 7-1 Taxi speed data from a medium fidelity pilot-in-the-loop study, showing that

nominal taxi speed maintenance data are approximately normally distributed.

Page 105: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

98

7.1.2 Availability of Accurate Speed Readouts

Second, the model assumes that operators have access to speed data, such as in the form

of a digital readout, and are not required to perceive changes in speed visually. In many

ways this simplifies the task and the model, but in a manner that is consistent with future

aviation operations. It is a safe assumption that both pilots and controllers will be equipped

with accurate speed-readouts to assist them in advanced 4-D trajectory environments.

However, there are potential implications for environments where this assumption may not

hold true. Estimates of TOA uncertainty may be systematically influenced by subjects!

misperception of vehicle speed that may occur as a function of the stimulus field, object size,

and the background texture.

As summarized by Ryan and Zanker (2001) and Goldstein (1989), perception of speed is

affected by the size of both the moving object being observed and the framework through

which it moves. The perceived speed of regularly spaced objects moving within apertures

decreases with larger aperture size (Brown, 1931). In particular, Brown asked observers

sitting in a dimly lit room to adjust the speed of a large dot moving across a large rectangle

to make it appear equal to the speed of a small dot moving across a small rectangle. Brown

found that, if the large rectangle was ten times larger than the small one, the large dot had to

move seven times faster than the small one for them to appear to move at the same speed.

This effect, known as speed transposition, shows that two images moving across the visual

field at different speeds can be perceived to be moving at the same speed. This could be

due to a size constancy effect by which, in the absence of veridical depth information, a

smaller stimulus is interpreted as being further away from the observer. This could have

important implications in some domains, particularly if the user is permitted to resize

windows on their display, thus making objects appear to be moving faster or slower. This

effect would be particularly problematic given the results found here, which show that

subjects tend to overestimate TOA uncertainty for faster objects and underestimate TOA

uncertainty for slower objects.

Another factor that may affect the perception of speed is the background on which the object

is displayed. This may range from objects in the real world, to icons on a display with

backgrounds that consist of simple, solid colour backgrounds, to complex grid backgrounds.

Page 106: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

99

The threshold for perceiving movement in a homogeneous field is a speed of about 1/6 to

1/3 of a degree of visual angle per second (Goldstein, 1989). However, if vertical lines are

present in the background, the object!s movement can be perceived even at speeds as low

as one-sixtieth of a degree of visual angle per second. Thus, perception of motion is

determined by the background, which affects not only the threshold for movement but also

the perception of speed. Indeed, the grid marks placed on the vehicle!s route in experiment

3 (see chapter 5) could have caused subjects to perceive the speed to be greater than it

really was. Related to this, it is also unknown if the perception of velocity was uniform

across the display (as opposed to the possibility that higher speeds might have been

perceived at the edges of the display), and precisely how this may have affected the results.

However, it is important to note that the factors outlined above are unlikely to have

confounded the results in the current experiments, for two reasons: 1) Subjects in these

experiments relied primarily on the digital speed readouts to estimate speed and thus did so

with negligible perceptual error; and 2) Speed estimation performance in experiment 3 was

very similar to that in experiment 2, in which there were no vertical hatch marks along the

route. Nevertheless, these factors could substantially influence speed perception in

operational settings in which operators are provided displays that show vehicles moving

against a background that changes from visually simple to complex as a function of the

environment such as a road network or airport layout.

7.1.3 Time of Arrival Uncertainty Estimation Adheres to the Proposed Two-step

Process Model

The model and experiments were also predicated on the assumption that subjects first learn

speed profiles and for each speed profile they estimate an expected TOA uncertainty

window. Indeed, this assumption was generated based on results of the first study in which

air traffic controllers described how they estimated TOA and TOA uncertainty. Controllers

reported that they had mentally assigned certain airlines with a speed profile (e.g.,

“Southwest Airlines taxied fast and consistently”) and estimated TOA based on that profile.

It is for this reason that in experiment 3 subjects were trained (with feedback) on speed and

then tested on their ability to estimate TOA. If, instead, subjects had been trained to

estimate TOA by means of a training protocol that provided feedback on TOA instead of

speed, the results most likely would have yielded improved TOA uncertainty estimates.

Page 107: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

100

Nonetheless, it is believed that operators, such as air traffic controllers, do indeed develop

models of speed profiles from which they estimate TOA, rather than developing models of

TOA directly.

Nonetheless, this is an important assumption to consider and raises implications for real-

world operational training and display needs. In the operational setting, there are four

important differences from the studies presented in this research: 1) controllers don!t always

get immediate feedback in terms of their speed and TOA estimates; 2) the on-the-job

training would not provide feedback that is as precise and unambiguous as in experiment 3;

3) controllers won!t likely experience 100 aircraft in a row of the same profile, and thus may

experience degradations of their estimates due to interruptions and integration of multiple

profiles; and 4) controllers would need to generate many more speed profiles to capture the

range of operational conditions and the differences among these profiles may be more

subtle than those tested in experiment 3. Exactly how these differences help or hinder

controllers! ability to build a library of speed profiles and apply them to estimate TOA

uncertainty is a topic for further research using operational settings and experienced

controllers.

7.2 OPERATIONAL IMPLICATIONS FOR AVIATION

A number of important implications for the operational environment have been revealed.

These implications are reported below, with the caveat that, without further operational or

field research, the degree to which these laboratory-based studies generalize to expert air

traffic controllers in the operational environment is unknown.

7.2.1. Implications for Human Error

First, it was shown that subjects underestimate TOA uncertainty in conditions where TOA

objective uncertainty is high, such as aircraft travelling long distances and aircraft that travel

at slow average speeds that are highly variable. This may have implications for actual

aviation operations, in that aircraft in these operating conditions may be sequenced too

closely because controllers will think the aircraft need less separation than they actually do.

In the example of runway sequencing, it is likely that controllers will produce smaller-than-

Page 108: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

101

necessary runway-crossing windows (a gap between two arriving aircraft in which a third

aircraft can cross the active runway). That is, if a controller creates a runway-crossing

window that is based on an underestimated TOA uncertainty window, the resulting runway-

crossing window will be too small. It is likely that the second arriving aircraft will be required

to conduct a go-around manoeuvre because the crossing aircraft will not have cleared the

runway in time. This could be a safety hazard, leading to potential loss-of-separation, or

create system inefficiencies if aircraft need to be stopped and/or re-routed.

Second, the subjects underestimated the amount of positive skew of TOA distributions for

slow and highly variable aircraft, thus a higher number of aircraft arrived later than expected.

This implies that if actual controllers also exhibit this tendency, they will underestimate the

likelihood that an aircraft will arrive late at a destination (e.g., waypoint, runway, or gate).

This can have both safety and efficiency implications in the case of paired departure

operations (Lunsford, 2009), in which two aircraft must depart parallel runways within 30

seconds of each other to ensure the following aircraft is not subject to excessive wake

turbulence from the lead aircraft. If controllers underestimate the likelihood that the following

aircraft may arrive late at the departure runway, the following aircraft may either experience

a wake vortex incident or experience substantial delays, as it would be required to stop and

wait for the next available departure slot, resulting in system inefficiencies.

Third, subjects overestimated TOA uncertainty when the distance was short, and when

actual TOA uncertainty was low, as with aircraft that travel at a consistently fast speed. If

actual controllers were also to exhibit this tendency, the result would be larger-than-

necessary runway crossing windows. That is, for example, a controller might leave a gap of

two minutes between arriving aircraft to ensure that a third aircraft can cross the active

runway, when in fact a gap of only one minute is required. With accurate knowledge of TOA

uncertainty, the controller could have doubled runway-crossing throughput by crossing two

aircraft in the gap, instead of one. In these cases, overestimates of TOA uncertainty reduce

airport efficiency. Also, in this case, a controller may unnecessarily require an arriving

aircraft to conduct a go-around manoeuvre if he/she believes the runway will not be clear,

which creates additional risks for the arriving aircraft.

Page 109: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

102

Finally, subjects tended to generalize unknown profiles to the closest known profile. If

controllers do this in the operational environment, and there is evidence from experiment 1

that they do, this could lead to over-generalizations, resulting in TOA windows that are either

too small or too large as described above, again resulting in either unsafe or inefficient

operations.

7.2.2 Implications for Human-Centred Automation to Support Air Traffic Control

Given the prevalence of the TOA uncertainty estimation errors noted above, it is next

reasonable to consider how the performance of air traffic controllers, particularly in next-

generation time-based environments, could be supported by automation. The results of this

research are interpreted using Parasuraman, Sheridan, and Wickens! (2000) "stages and

levels! of automation framework. The framework proposes that automation can support four

broad classes of functions: 1) Information Acquisition; 2) Information Analysis; 3) Decision

and Action Selection; and 4) Action Implementation. Within each of these functions,

automation can be applied across a continuum of levels from low to high (i.e., from fully

manual to fully automatic). Suggestions for how automation can be used to support air

traffic controllers in each stage are discussed next.

Information Acquisition refers to the process by which raw data are selectively attended or

filtered (Wickens, 2000). In the context of TOA uncertainty estimation, this includes the

controllers! ability to accurately perceive the aircraft!s current state (location, speed, distance

to travel), identify the factors that contribute to a particular aircraft!s TOA uncertainty, and to

accurately characterize the aircraft!s speed profile. The present research demonstrated

human limitations associated with the task of accurately characterizing an aircraft!s speed

profile. Specifically, it was noted that subjects were able to accurately identify a speed

profile after a short period of observation only 65 to 88% of the time for trained profiles.

Automation to support this information acquisition task may take quite different forms

ranging from simple displays of raw data (particularly speed; see section 7.1.2 regarding the

assumption of this minimum level of automation) to target cueing (Yeh, Wickens, & Seagull,

1999), intelligent information management (Hammer, 1999), and highlighting (Wickens,

Kroft, & Yeh, 2000), each of which explicitly directs the controller!s attention to certain

Page 110: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

103

information sources at the expense of others. The automation may guide attention to

sources of environmental information that the controller may use to determine the aircraft!s

speed profile (such as information about traffic, weather, aircraft type, and pilot familiarity

with the airport) or may explicitly provide an indication of an aircraft!s speed profile, possibly

through color-coding on the ATC display.

In the second stage (Information Acquisition), data must be integrated and interpreted so as

to draw some intelligent inference regarding the state of the world (Wickens, 2000). In ATC

operations, this refers to the controllers! estimation of TOA uncertainty based on their library

of speed and TOA profiles. As was seen in the results of experiment 2 and 3, this was the

task that was the most prone to bias, and thus presumably may benefit the most from

automation. To support this task, automation could compute and display TOA uncertainty

and prioritize or rank order the aircraft in terms of TOA uncertainty. A controller could use

the display to choose the aircraft that is most likely to arrive at the departure runway on time,

or to identify aircraft that are unlikely to make their takeoff time.

The third phase (Decision and Action Selection), in the context of ATC, refers to the

controllers! task of deciding whether or not the estimated TOA uncertainty is acceptable, or if

action is required to reroute aircraft. Although not addressed in the current research, there

is evidence that controllers may typically apply a conservative bias in this decision-making

phase (see Boudes & Cellier, 2000) and thus may be more likely to determine that the TOA

uncertainty is unacceptable and reroute aircraft. Although this conservative bias is actually

quite adaptive in today!s ATC environment (as described previously in section 1.1.1), it may

be detrimental in NextGen 4-D environments, which will require aircraft to arrive "just-in-time!

(that is, neither early nor late). Automation in the form of visual or auditory alerts that

indicate when objective TOA uncertainty exceeds a pre-determined threshold could be

applied to support this stage.

Finally, the Action Execution stage in this ATC context refers to the task of issuing a revised

clearance to one or more aircraft as a result of excessively high TOA uncertainty. A lower

level of automation to support this stage could take the form of automation that generates

time-based clearances that the controller would approve or modify and then communicate to

Page 111: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

104

the aircraft. This approach would take advantage of the automation!s ability to quickly

compute accurate TOA!s and distributions, but at the same time take advantage of the

human!s ability to improvise and use flexible procedures (see Fitts, 1951). A higher level of

automation could operate completely autonomously and issue time-based clearances

directly to the aircraft and modify them when pilots fail to meet their time requirements. The

controller could over-ride the automation to provide certain aircraft with more time and a

larger uncertainty buffer where mandated. However, it must be acknowledged that a

weakness of this approach is the operator "out-of-the-loop! problem (Endsley & Kris, 1995).

As discussed previously in Section 1.1.3, when placed in a monitoring role, it is often difficult

for the operator to maintain sufficient situation awareness to make the necessary

interventions.

While automation to support the Action-Execution stage is currently being researched (for

example, Balakrishnan & Jung, 2007; Rathinam, Montoya & Jug, 2008), many of these

prototype systems typically address TOA uncertainty in an overly simplistic manner by either

applying a constant uncertainty value for each aircraft or assuming perfect on-time

performance from each aircraft. The current research showed that the TOA uncertainty

values incorporated into the automation should consider (at least) the mean speed,

variability, and distance to travel. Experiment 1 also uncovered several other factors that

human controllers consider in estimating TOA uncertainty, such as airline, weather

conditions, traffic levels, and even whether or not the pilot spoke with an accent (suggesting

to the controllers that the pilot might not be familiar with the airport surface). While some of

these could presumably be incorporated into the automation algorithms (with varying

degrees of success) others, such as the pilot!s accent, are not easily considered by

automation. Failure to account for these important variables may result in inaccurate

uncertainty buffers around each aircraft. Uncertainty buffers that are too large will limit

system efficiency and uncertainty buffers that are too small may yield excessively high error

rates.

In summary, automation that supports controllers in the Information Acquisition task may be

the most fruitful for several reasons. First, as noted above, subjects exhibited systematic

biases in their abilities to estimate TOA uncertainty. Second, applying automation at the

earlier stages, rather than at the decision and action stages, may indeed be the best option

Page 112: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

105

if the airport surface environment proves to be too dynamic, thus rendering the automation

"brittle" (Cohen, 1993; as discussed previously in section 1.13). This approach would

minimize the operator out-of-the-loop concern because the controller would be actively

controlling the airport traffic, albeit in a more informed manner. Third, a further potential

advantage of automation at the earlier stages that were shown to exhibit systematic bias, is

that it is likely that controllers! TOA uncertainty estimation errors would gradually correct

over time with use of the automation and their estimates would become more accurate.

Indeed, such an improvement in estimation error was observed in Experiment 3 for the task

of velocity and velocity window estimation errors. It is plausible that displaying the TOA

uncertainty information to the controller would serve as adequate feedback to support this

learning process in the actual operational environment. If this is the case, it is possible that

the controllers would rely less on the automation over time, and use it as an aid only when

presented with an aircraft with an unfamiliar profile.

The results of the current research suggest that automation, particularly when applied to

Stages 1 and 2, could help compensate for the systematic biases observed in the tasks

associated with identifying an aircraft!s speed profile and applying the TOA estimate. It is

still unclear, however, exactly how a controller would, or should, use such TOA uncertainty

information, and this remains a topic for future research.

7.3 FUTURE DIRECTIONS

This research culminated in the development of a model that explains how subjects

estimated TOA uncertainty, and the sources and causes of errors in the TOA uncertainty

estimation process. Armed with this model of the systematic biases inherent in human

operators! TOA uncertainty estimation, three future research directions have been identified.

First, the results of the current research were based on a simplification of a real-world task.

Although it is believed that the important elements of the task were represented in the

experimental tasks reported in chapters 4 and 5, future research is required to increase the

realism and operational robustness of the experimental task, to ensure that the findings

generalize to real world ATC operations. This could include modifications to the current

experimental platform, such as including variable start times and route lengths,

Page 113: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

106

representations of traffic sequencing tasks and other environmental cues that affect

variability, and implementing variable profiles. Future analyses of actual operational data in

an ATC setting would be of value, as would modelling profiles with realistic, potentially non-

normal distributions.

Second, various forms of ATC automation were presented and discussed in section 7.2.2.

Each of these approaches requires extensive research including human-in-the-loop

simulations to fully define and evaluate. Only with relatively mature automation solutions (or

at least emulations) can the relative merits of each automation level for a particular stage be

evaluated.

Third, given that TOA uncertainty exists in many complex environments and humans

systematically misestimate this uncertainty, there is a need for further research and

development efforts geared toward the design and development of DSSs and other

advanced visual displays to support users in the TOA uncertainty estimation task (such as

Lee & Milgram, 2008). It is expected that elements of the experimental method and metrics

developed within the current research effort will be useful for this purpose.

Fourth, the current research has examined how subjects comprehend TOA uncertainty

relative to actual TOA uncertainty; however, there is still much to be learned regarding what

users will do with this information. Future research is required to better understand how this

comprehension of TOA uncertainty will affect operator situation awareness and the actual

decisions that operators make.

Page 114: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

107

CHAPTER 8: REFERENCES

Alpert, M. & Raiffa, H. (1982). A progress report on the training of probability assessors. In

D. Kahneman, P. Slovic and A. Tversky (Eds.). Judgment Under Uncertainty:

Heuristics and Biases. Cambridge: Cambridge University Press.

Andre, A. D. & Cutler, H. A. (1998). Displaying uncertainty in advanced navigation systems.

Proceedings of the Human Factors and Ergonomics Society 42nd Annual Meeting. (pp.

31-35). Santa Monica, CA: HFES.

Balakrishnan, H., & Jung, Y. (2007). A framework for coordinated surface operations

planning at Dallas-Fort Worth International Airport (Paper # AIAA 2007-6553).

Proceedings of the AIAA Guidance, Navigation, and Control Conference and Exhibit.

Alexandria, VA: AIAA

Banbury, S., Selcon, S., Endsley, M. Gorton, T. & Tatlock, K. (1998). Being certain about

Banbury, S., Selcon, S., Endsley, M. Gorton, T. & Tatlock, K. (1998). Being certain

about uncertainty: How the representation of system reliability affects pilot decision

making. Proceedings of the Human Factors and Ergonomics Society 42nd Annual

Meeting. (pp. 36-39). Santa Monica, CA: HFES

Block, R. A. & Harper, D. R. (1991). Overconfidence in estimation: Testing the anchoring

and adjustment hypothesis. Organizational Behavior and Human Decision Processes,

49, 188-207.

Boudes, N. & Cellier, J-M. (2000). Accuracy estimations made by air traffic controllers. The

International Journal of Aviation Psychology, 10(2), 207-225.

Brown, J. F. (1931). The visual perception of velocity. Psychologische Forschung, 14, 199-

232.

Cohen, M. S. (1993). The bottom line: Naturalistic decision aiding. In G. A. Klein, J.

Orasanu, R. Calderwood, & C. E. Zsambok (Eds.). Decision Making in Action: Models

and Methods. Norwood, NJ: Ablex Publishing Corporation.

Page 115: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

108

Endsley, M. R., Bolte, B. & Jones, D. G. (2003). Designing for Situational Awareness: An

approach to user-centered design. New York: Taylor & Francis.

Endsley, M. R., & Kris, E. O. (1995). The out-of-the-loop performance problem and level of

control. Human Factors, 37 (2), 381-394.

Erev, I., Wallsten, T.S., & Budescu, D. V. (1994). Simultaneous over- and underconfidence:

The role of error in judgment processes. Psychological Review, 101, 519-527

Fitts, P. M. (1951). Human Engineering for an Effective Air Navigation and Traffic Control

System. Washington DC: National Research Council.

Gigerenzer, G., Hoffrage, U., & Kleinbolting, H. (1991). Probabilistic mental models: A

Brunswikian theory of confidence. Psychological Review, 98, 506-528.

Goldstein, E. B. (1989). Sensation and Perception. Belmont, California: Wadsworth

Publishing Company.

Hammer, J. M. (1999). Human factors of functionality and intelligent avionics. In D. J.

Garland, J. A. Wise, & V. D. Hopkin (Eds.), Handbook of Human Factors in Aviation. (pp.

549-565). Mahwah, NJ: Lawrence Erlbaum.

Hansman, R. J. & Davison, H. J. (2000). The effect of shared information on pilot / controller

and controller / controller interactions. 3rd USA/Europe ATM R&D Seminar, Napoli, Italy.

Johnson, W.W., Battiste, V. & Holland, S. (1999) A cockpit display designed to enable

limited flight deck separation responsibility. Proceedings of the 1999 World Aviation

Conference, Anaheim, CA.

Juslin, P. (1994). The overconfidence phenomenon as a consequence of informal

experimenter-guided selection of times. Organizational Behavior and Human Decision

Processes, 57, 226-246.

Juslin, P., Wennerholm, P., & Olsson, H. (1999). Format dependence in subjective

probability calibration. Journal of Experimental Psychology: Learning Memory and

Page 116: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

109

Cognition, 25(4), 1038-1052.

Juslin, P., Winman, A., & Olsson, H. (2003). Calibration, additivity, and source

independence of probability judgments in general knowledge and sensory discrimination

tasks. Organizational Behavior and Human Decision Processes. 92, 34-51.

Kahneman, D., Slovic, P. & Tversky, A. (1982). Judgment under Uncertainty: Heuristics and

Biases. NY: Cambridge University Press.

Kuchar, J. K. (2001). Managing uncertainty in decision-aiding and alerting systems.

Proceedings of the 6th CNS/ATM Conference. Taipei, Taiwan.

Lee, H. & Milgram. P (2008). An Integrated Air Traffic Control Display Concept for Conveying

Temporal and Probabilistic Conflict Information. Proceedings of the Human Factors and

Ergonomics Society (pp. 84-84). Santa Monica, HFES.

Lichtenstein, S., Fischhoff, B., & Phillips, L. D. (1982). Calibration of subjective probabilities:

The state of the art up to 1980. In D. Kahneman, P. Slovic, & A. Tversky (Eds.),

Judgment under Uncertainty: Heuristics and Biases (pp. 306-334). New York:

Cambridge University Press

Lunsford. C. (2009). Feasibility and availability of pairing departures from closely spaced

parallel runways for wake avoidance. Proceedings of the 2009 IEEE/AIAA Integrated

Communications, Navigation and Surveillance Conference (ICNS) conference.

Arlington, VA, May 2009.

McKenzie, C. R. M., Liersch, M. J., & Yaniv, I. (2008). Overconfidence in interval estimates:

What does expertise buy you? Organizational Behavior and Human Decision

Processes, 107, 179-191.

Nisbett, R. E., Krantz, D. H., Jepson, C. & Fong, G. T. (1982). Improving inductive

inference. In D. Kahneman, P. Slovic, & A. Tversky (Eds.). Judgment under

uncertainty: Heuristics and biases. Cambridge: Cambridge University Press

O!Connor, M., & Lawrence, M. (1989). An examination of the accuracy of judgemental

confidence intervals in time series forecasting. Journal of Forecasting, 8(2), 141-155.

Page 117: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

110

Paese, P. W., & Sniezek, J. A. (1991). Influences on the appropriateness of confidence in

judgment: Practice, effort, information, and decision making. Organizational Behavior

and Decision Making, 48, 100-130.

Peterson, C. R. & Beach, L. R. (1967). Man as an intuitive statistician. Psychological

Bulletin, 68(1), 29-4.

Pitz, G. F. (1980). The very guide of life: The use of probabilistic information for making

decisions. In T. S. Wallsten (Ed.) Cognitive Processes in Choice and Decision.

Hillsdale, N.J: Erlbaum and Associates.

Rathinam, S., Montoya, J., & Jung, Y. (2008). An optimization model for reducing aircraft

taxi times at the Dallas Fort Worth International Airport. Proceedings of the 26th

International Congress of the Aeronautical Sciences (ICAS). Sept 14-19, 2008,

Anchorage, Alaska.

Ryan, J. & Zanker, J. M. (2001). What determines the perceived speed of dots moving

within apertures? Exp Brain Res. 141, 79-87.

Schaefer, D., Gizdavu, A., & Nicholls, D. (2004). The display of uncertainty information on

the controller working position. Proceedings of the Digital Avionics Systems Conference.

St. Louis, Mo: IEEE.

Schiff, W. & Oldak, R. (1990). Accuracy of judging time to arrival: effects of modality,

trajectory, and gender. J Exp Psychol Hum Percept Perform. 16(2). 303-16.

Schunn, C. D., Kirscehnbaum, S. S., and Trafton, J. G. (2005). The ecology of uncertainty:

Sources, indicators, and strategies for informational uncertainty. Technical Report

(Available at www.au.af.mil/au/awc/awcgate/navy/nrl_uncertainty_taxonomy.pdf)

Sheridan, T. B. (2002). Humans and Automation: System Design and Research Issues.

Santa Monica, CA: HFES / John Wiley& Sons.

Soll, J. B. & Klayman, J. (2004). Overconfidence in Interval Estimates. Journal of

Experimental Psychology: Learning, Memory, and Cognition. 30(2), 29-314.

Page 118: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

111

Teigan, K. H., & Halberg, A., & Fostervold, K. I. (2007). More than, less than, or minimum,

maximum: How upper and lower bounds determine subjective interval estimates.

Journal of Behavioral Decision Making, 20, 179-207

Tversky, A. & Kahneman, D. (1982). Judgment under uncertainty: Heuristics and biases.

In D. Kahneman, P. Slovic, & A. Tversky (Eds.). Judgment under uncertainty:

Heuristics and biases. Cambridge: Cambridge University Press

Tversky, A. & Simonson, I. (2000). Context-dependent preferences. In D. Kahneman and

A. Tversky (Eds). Choices, Values, and Frames. Cambridge: Cambridge University

Press

van Doorn B., Bakker, B., & Meckiff, C. (2001). Evaluation of advanced conflict modelling in

the highly interactive problem solver. Proceedings of the 3rd USA/ Europe ATM R&D

Seminar, Napoli.

Xu, A., Wickens, C. D., & Rantan, E. (2004). Imperfect conflict alerting systems for the

cockpit display of traffic information. Technical Report AHFD-04-8/NASA-03-2. Savoy,

Illinois: University of Illinois at Urbana-Champaign, Aviation Human Factors Department.

Wickens, C. D. (2000). Imperfect and unreliable automation and its implications for attention

allocations, information access, and situation awareness. Technical Report ARL-00-

10/NASA-00-02. Savoy, Illinois: ARL, University of Illinois at Urbana-Champaign.

Wickens, C. D., & Hollands, J. G. (2000). Engineering Psychology and Human

Performance. London: Prentice Hall.

Wickens, C. D., Kroft, P., & Yeh, M. (2000). Database overlay in electronic map design:

Testing a computational model. Proceedings of the IEA 2000/HFES 2000 Congress.

Santa Monica, CA: HFES.

Wickens, C. D., & Morphew, E. (1997). Predictive features of a cockpit traffic display: A

workload assessment (Technical Report ARL-97-6/NASA-97-3). Savoy, IL: University

of Illinois, Institute of Aviation, Aviation Research Lab.

Page 119: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

112

Williams, J. L., Hooey, B. L., & Foyle, D. C. (2006). 4-D taxi clearances: Pilots' usage of

time- and speed-based formats. Proceedings of the AIAA Modelling and Simulation

Technologies Conference. Paper AIAA-2006-6611.

Yeh, M., Wickens, C. D., & Seagull, F. J. (1999). Target cueing in visual search: The effects

of conformality and display location on the allocation of visual attention). Human Factors,

41, 524-542.

Page 120: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

113

APPENDICES

APPENDIX A1: EXPERIMENT 1 PROTOCOL

As we look ahead to the future and the potential implementation of time-based taxi clearances, we acknowledge that the current ground control system is very dynamic and possesses many sources of uncertainty. The purpose of this study is to explore factors that contribute to your perceptions of certainty (or confidence) that a time-based taxi plan that you issue to pilots will be carried out as scheduled. In particular, I am interested in how a number of factors influence your perception of certainty. 1) Distance / Speed Calibration

Before we begin, I will start by providing you with some information to calibrate you to the time required to taxi different routes at two different speeds. You can safely assume that typical taxi speeds range from 16 to 24 knots, with an average speed being 20 knots. This shows two routes running east/west across the airport. One is 6,000! and the other is 12,000!. These are intended to give you a rough idea of how long each taxi route would take to complete at average speeds of 16 kts, 20 kts, and 24 kts. You can see that the distance from one inside runway to the other is about 6,000! and a 12,000! route stretches across most of the airport. Here are the same length routes, this time running north/south on the airport. 2) Scenario Descriptions

In this study, you will see several static pictures of taxi plans in progress. Each picture will show you the taxi plan and an aircraft!s progress in carrying out the plan. Your task will be to assess the information and provide a rating of how certain you are that the plan will succeed (that is, that the aircraft will arrive at the runway threshold within the provided window of time – neither too early, nor too late). Each scenario will contain the following features:

Magenta taxi route: In all trials, you see a map of DFW airport with one taxi route shown in magenta.

Aircraft Icon: The aircraft icon indicates the position of the taxiing aircraft along the route of the aircraft. The aircraft will either be positioned at the beginning of the route or part way through. If it is at the beginning, you can assume that the aircraft has spooled up to a nominal taxi speed.

Data tag: Associated with each aircraft is a data tag with a TIME. This refers to the elapsed time since the pilot started taxiing.

Runway Arrival Time Window: All taxi routes will end at a runway. A runway arrival time window will be provided in the yellow box beside the runway. This time window indicates the acceptable window of time in which the aircraft may arrive at the runway. It DOES NOT include time required to cross the runway.

Page 121: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

114

Block 1

You have seen the full range of the factors for this experiment. When you are ready,

we!ll continue with more trials.

Remember to consider all of the factors together and to use the entire scale from 1 (if you have LOW certainty that the plan will succeed) to 7 (if you have HIGH certainty that the plan will succeed). For the remaining trials, use the numbers 1 to 7 on the keyboard to enter your answers. Do you have any questions before we begin? Things to Remember…

1) Please rate how certain you are that the aircraft will arrive at the runway threshold within the cleared window of time using the following scale.

Very Low Certainty Very High Certainty 1 2 3 4 5 6 7

2) Try to use the entire range of the seven-point scale. 3) Runway time windows refer to the arrival time at the runway, NOT the time to cross the runway. 4) The aircraft should arrive at the runway within the crossing window and neither late, nor

early. - Arriving early will require the aircraft to stop – this doubles the time required for an aircraft to cross a runway as compared. - Arriving late will mean the aircraft has missed the crossing window – this will inevitably cause a delay.

5) Assume that it is a clear day at Dallas Forth Worth and there are no traffic conflicts. 6) Assume that each aircraft has navigation and time-management technology that will enable them to arrive at the runway threshold within a 30 second time window about 50% of the time. 7) Assume that an average taxi speed is 20 knots and that the aircraft is a B-757. 8) Remember that every pilot and aircraft is different and any number of factors may influence their ability to comply with the precise time window. For example:

- A pilot may be slow to start taxi or spool up. - A pilot may slow down more for turns than another. - A pilot may be unfamiliar with the airport. - A pilot may have to stop because a passenger is standing up. - There may be equipment problems that need investigating during taxi.

Page 122: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

115

APPENDIX A2: EXPERIMENT DEBRIEF INTERVIEW QUESTIONS

1. Please rank the importance of the following six factors in determining your level of certainty (or confidence) that a time-based taxi clearance will be carried out successfully. _____ Route Length _____ Route Complexity _____ Aircraft Position Along Route _____ Duration of Runway Arrival Window 2. What other factors contribute to your certainty that a pilot will arrive at the cleared runway within a specified time window? 3. Please rank these factors relative to the factors in question 1. 4) What strategy did you use for assigning certainty ratings? - At what rating would you cancel a clearance or re-route due to lack of confidence? 5) How did you use the runway threshold window.

- i.e., consider if they would make the middle of the window, the beginning, or the end? 6). What percentage, if any, of your clearances today have some sort of time element in them (i.e. expedite etc.). 7) How do you manage uncertainty in today!s surface environment 8. If you are uncertain that an aircraft will comply with its clearance, how does this impact your task of controlling multiple aircraft at the same time. 9. If you have two aircraft, one that you are certain will comply with its clearance, and one that you are less certain, do you treat monitor them differently? If so how? 10. How will time-based clearances change the way you monitor airport traffic? 11. Do you have any suggestions for managing uncertainty in future time-based environments? - changes to communication, information available, technology?

Page 123: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

116

APPENDIX B1: EXPERIMENT 2 PROTOCOL

Welcome and thank you for agreeing to participate in this study. The purpose of this study, which will also form a portion of a PhD research project at the University of Toronto, is to understand how people estimate Time-of-Arrival Uncertainty. Please take a few minutes and read this and the attached informed consent form. It describes the study you are about to participate in, and explains your rights as a subject. If you have any questions at any time, please ask the experimenter. Overview

In this study, I am interested in how you estimate time-of-arrival uncertainty. To start with, I would like you to consider the following scenarios: Scenario A: You travel 10 miles on the freeway to get to work every day for 100 days. Since you work the night shift you don!t have much traffic to worry about, and you always try to stick to the speed limit. Sometimes you go slightly slower or faster when you encounter other traffic along the way, but usually only by 1 or 2 miles per hour. On average, across the 100 days, it takes you 10 minutes to get to work, but of course it is not always exactly 10 minutes because your speed varies slightly, depending on traffic. Scenario B: You travel 10 miles on city streets to get to work every day for 100 days. Since you are always running late, you like to travel as fast as you can, but of course always stopping for red lights. On average, across the 100 days, it takes you 20 minutes to get to work, but sometimes you can make it much faster if you get green lights all the way, and sometimes you take much longer if you get a lot of red lights. Now, imagine your boss is trying to create a very tightly coordinated schedule, and he asks you to predict what time you would be at work. For which scenario would it be easier to estimate a time-of-arrival for your boss? Scenario A might be pretty easy because your time of arrival would be close to the same time every day. Similarly, Scenario B might be more difficult, because your time of arrival will depend on how many red lights you encounter. Instead then, your boss asks you to provide a window of time within which you are very sure (i.e. 95% sure) that you can make it. Of course you!d want to make it as small a window as possible so as to be useful for your boss – but large enough so that you are never considered late. Consider the two scenarios. Would you assign a larger window for Scenario A or Scenario B? Probably Scenario B, because you!ll never know exactly how many traffic lights you might encounter along the way. The Experimental Task

Throughout the study, I am going to show you a screen with 20 "aircraft!, which all move with the same average velocity, and their velocities fluctuate about the same amount across the route. While the aircraft are traversing their routes, you will be able to see a digital read-out of the velocity of each aircraft. As the trial progresses, you will note that, even though the aircraft are all travelling at about the same speed, they each arrive at different times. Sometimes the arrival times will be close together, sometimes they will be more spread out. An elapsed time counter is available at the bottom of the screen.

Page 124: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

117

Once all 20 aircraft have completed their routes, a query screen will appear. It will show another aircraft that is about to begin traversing a route. The screen will present the following question: This aircraft!s speed profile is similar in nature to the aircraft you just observed and it

is about to traverse the same length route.

a) Please estimate this aircraft!s time of arrival.

b) Please provide the smallest TOA window that you are 95% sure will contain the

aircraft!s actual Time of Arrival

When you answer these questions, please consider the aircraft that you just observed and remember that you can assume that this aircraft is representative of the aircraft you observed. Also, you should know that there are no 'right' or 'wrong' answers – we are merely interested in your estimate. The first question “Please estimate this aircraft!s time of arrival” is asking you to make your best guess as to when the aircraft will arrive at its destination. For example, if you were planning to meet the aircraft at the destination point, what time would you expect it to arrive? Please provide your answers in seconds, with accuracy to one decimal point. For example type a number like 60.5 in the box on the screen. Press Enter when you are satisfied with your answer. You will not be able to change it once you have pressed enter. For the second question “Please provide the smallest TOA window that you are 95%

certain will contain the aircraft!s actual TOA,” recall that, even though all aircraft were trying to arrive at the destination at the same time, there was some variability around the arrival time. The window you select should represent the smallest possible window of time that you are 95% sure will contain this aircraft!s actual time of arrival. Use the mouse to move the left slider to represent the lower bound of the time window, and the right slider to represent the upper bound of the time window. Again, please provide answers, in seconds, with precision to one decimal point. The actual numbers are shown above the slider scale. Today!s agenda: We will repeat the sequence described above several times throughout the experiment today. The experiment is divided into four blocks, and you can take a short rest break between each block. If you need to stop for any reason before the break, please let me know, but I will not be able to answer any questions once the experiment has started. Whom to contact for more information:

This study is being conducted jointly by NASA Ames Research Center and the University of Toronto. You may seek further information from the following points of contact:

• Becky Hooey, Principal Investigator (650-604-2399) NASA Ames Research Center, Human System Integration Division University of Toronto, Dept. of Mechanical and Industrial Engineering

• David C. Foyle, Co-Investigator (650-604-3053) NASA Ames Research Center, Human System Integration Division,

• Paul Milgram, Co-Investigator (416-978-3662) University of Toronto, Dept. of Mechanical and Industrial Engineering,

If you have read and understood the above information, please sign the attached informed consent form.

Page 125: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

118

APPENDIX B2: EXPERIMENT 2 DEMOGRAPHIC QUESTIONNAIRE

1. Name: ________________________________________________________

2. Age: _________________________

3. Gender: Male _________ Female ______________

4. Do you normally wear glasses or contact lenses for any of the following activities? (please

check all that apply)

5. Reading _____ Watching TV _____ Working on computer _____ Driving _____

Other tasks, please specify _____________________________________________

6. Are you wearing either glasses or contact lenses today? No _____ Yes _____

7. Are you currently employed? No _____ Yes _____

If yes, are you employed part-time _____, full-time _____, contract/casual ______

8. What is your job title: __________________________________________________

9. Are you currently a student? No _____ Yes _____

If yes, please state program, institution, and your START date (i.e. A.A / De Anza /

2005)

10. What is the highest level of education that you have completed?

_____ High School / GED

_____Community College (2 year degree) Please specify degree: ___AA or ___AS

_____University (4 year degree) Please specify Degree / University (i.e. B.A. Psych / SJSU)

_____ Graduate School. Please specify Degree / University (i.e., Ph.D. Chemistry /

Stanford)

______Other Education / Diploma / Certificate program (please specify)

11. Do you have (or have you ever had) a Pilot!s license? No _____ Yes _____

If yes, Please specify type (e.g., private / commercial; fixed-wing, rotorcraft)

12. Have you ever taken a course in statistics? No _____ Yes _____

If yes, how many statistics courses have you taken? ____________________

13. Please list the Department / Institution (i.e. Psychology / SJSU) for each course.

Page 126: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

119

APPENDIX B3: EXPERIMENT 2 DEBRIEF QUESTIONS

1) Did you use the scratch paper. If so – what did you write down?

2) What strategy did you use to assess the mean velocity?

3) What strategy did you use to assess the velocity window?

4) Were your velocity windows symmetrical or asymmetrical? How did you decide?

5) What strategy did you use to assess TOA

6) What strategy did you use to assess TOA window?

7) Were your TOA windows symmetrical or asymmetrical? How did you decide?

8) Did your strategy change over the course of the experiment?

9) Do you think your windows got smaller with time?

10) Did you find estimating velocity or TOA easier?

Page 127: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

120

APPENDIX B4: METHOD FOR DETERMINING ACTUAL SPEED AND TOA SCORES

Actual Speed Scores

To determine the Actual Mean Speed and the Actual Speed Interval, the following process was used to develop a population distribution comprised of 25,000 aircraft, and calculate the mean and 95% CI. The following steps were conducted 9 times, to create actual values for each of 9 combinations of Speed and Variability.

• Randomly drew a series of 10 !segment speeds" for 25,000 aircraft (one speed for each of 10 segments along the route). The segment speeds were drawn from a normal distribution, with a given mean and standard deviation that mapped to the experimental IV conditions.

• Calculated the mean !route speed" for each of the 25,000 aircraft (by averaging each aircraft"s 10 segment speeds)

• Sorted the list of 25,000 !route speeds" in ascending order

Actual Speed – Calculated the mean !route speed" by averaging the mean route speed of the 25,000 aircraft. Actual Speed Interval - Selected the speed from the sorted list of route speeds that represented the 2.5 and the 97.5 percentile. Note, this is equivalent to calculating the 95% CIs using the following formula: CI = M +/- 1.96 * sd / sqrt n Where M = the mean of the normal distribution from which the 250,000 velocities were selected, sd = standard deviation of that normal distribution, and n = 10. Actual TOA Scores

For TOA, the CIs could not be calculated using the CI formula above given that the TOA distribution was non-normal as discussed previously. The following steps were conducted:

• A segment elapsed time was calculated for each of the 10 segment speeds for each of the 25,000 aircraft (as determined above).

• The 10 segment elapsed times of each aircraft were summed to determine the elapsed time for each of the 25,000 aircraft.

• The 25,000 elapsed times were sorted in ascending order

Actual TOA - Calculated the mean of the 25,000 elapsed times

Actual TOA Interval - Selected the elapsed time from the sorted list of times that represented the 2.5 and the 97.5 percentile. Since the TOA distribution is not normally distributed (it is positively skewed), the equation for 95 % CI reported above for Actual Speed Interval was deemed to be inappropriate. Instead, the actual intervals were determined by selecting the 2.5th percentile and the 97.5th percentile from the distribution of route times.

Page 128: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

121

APPENDIX C1: EXPERIMENT 3 PROTOCOL

Welcome and thank you for agreeing to participate in this study. The purpose of this study, which will also form a portion of a PhD research project at the University of Toronto, is to understand how people estimate Time-of-Arrival Uncertainty. Please take a few minutes and read this and the attached informed consent form. It describes the study you are about to participate in, and explains your rights as a subject. If you have any questions at any time, please ask the experimenter. Overview

In this study, I am interested in how you estimate time-of-arrival uncertainty. To start with, I!d like you to consider the following scenarios: Scenario A: You just got a new job working 6:00 p.m. to midnight at the community hospital which is 30 miles away on a six-lane freeway that is pretty straight, but hilly. Since you work the night shift you travel against the commuting traffic, so you don!t have much traffic to worry about, and you always try to stick to the speed limit. You set your cruise control for 65 mph and, although you maintain a pretty constant speed, you notice that you go a few mph slower up the hills and a few mph faster down the hills and occasionally you have to turn off cruise control to pass a car on the road. On average, across 100 days, it takes you about 30 minutes to get to work, but of course it is not always exactly 30 minutes since your speed varies slightly because of the hills and traffic. Scenario B: Instead of travelling the freeway, you decide to take the rural road that has a slower speed limit of 30 mph, but is only 20 miles. Since there are a lot of curves you don!t use cruise control. On this road, you encounter a lot of traffic, particularly at 6:00 when most people are returning home from their day jobs, and this means you have to slow down and speed up frequently. On average, across 100 days, it takes you 25 minutes to get to work, but sometimes you can make it much faster if traffic is light, and sometimes it takes you much longer if traffic is heavy. Now, imagine your boss is trying to create a very tightly coordinated schedule, and he asks you to predict what time you would be at work. For which scenario would it be easier to estimate a time-of-arrival for your boss? Scenario A might be pretty easy because your time of arrival would be close to the same time every day. Similarly, Scenario B might be more difficult, because your time of arrival will depend on how much traffic you encounter. Instead then, your boss asks you to provide a window of time within which you are very sure (i.e. 95% sure) that you can make it. That is, if you drove the route 100 times, you would arrive within the time window 95 times. Of course you!d want to make it as small a window as possible so as to be useful for your boss. Consider the two scenarios. Would you assign a larger window for Scenario A or Scenario B? Probably Scenario B, because you!ll never know exactly how much traffic you might encounter along the way. Today!s agenda:

The experiment is divided into three parts, with a short rest break between each block. PART 1: LEARNING SPEED PROFILES

First, you will be shown 20 "vehicles!, each traversing a route that is .5 miles long, which all move with about the same average velocity, and their velocities fluctuate about the same

Page 129: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

122

amount across the route. Each vehicle!s speed will change 10 times as it traverses the route (e.g., it will slow down and speed up as if it were encountering varying headwinds). The 20 vehicles in each trial will share the same speed profile – that is, their average velocity is about the same, and their velocities fluctuate about the same amount across the route. There will be four different profiles and they will be identified by colour (Red, Yellow, Blue and Green). Every vehicle with the same colour icon will always share the same speed profile. Your task is to "learn! each of those 4 speed profiles and provide estimates of the average speed and the amount of variability among the speeds. As the trial progresses, you will note that, even though the vehicles are all travelling at about the same speed, each arrives at the destination at a different time. Sometimes the arrival times will be close together, sometimes they will be more spread out. Pay attention to this distribution of arrival times, as it will be important later in the experiment.

PART 2: TIME OF ARRIVAL (TOA) ESTIMATION

In the second part of the experiment, you will be shown only one vehicle at a time. Its speed profile will always be one of the four you saw in Part 1, and will be identified with the same colour code. You will watch the vehicle as it traverses the route to remind you what the speed profile looked like. Part way through the route, the vehicle!s movement will be paused. A query screen will appear that provides the estimated time of arrival (in elapsed seconds) for this vehicle to complete the route, and asks you to estimate the smallest TOA window that you are 95% certain will contain the vehicle!s actual TOA. When answering, you should consider that the vehicle shares the same profile as all of the vehicles with the same colour icon from the first part of the study, and use your knowledge of that speed profile in generating estimates. PART 3: SPEED PROFILE IDENTIFICATION

In the third part of the experiment, you will be shown one vehicle as it traverses the route. Its speed profile colour code will NOT be identified. Rather you will be asked to identify the speed profile from those you observed in PART 1 and 2 of the study, and then estimate the smallest TOA window that you are 95% certain will contain the vehicle!s actual TOA. This study is being conducted jointly by NASA Ames Research Center and the University of Toronto. You may seek further information from the following points of contact for each institution:

• Becky Hooey, Principal Investigator NASA Ames Research Center, Human System Integration Division University of Toronto, Dept. of Mechanical and Industrial Engineering 650-604-2399

• David C. Foyle, Co-Investigator NASA Ames Research Center, Human System Integration Division, 650-604-3053

• Paul Milgram, Co-Investigator University of Toronto, Dept. of Mechanical and Industrial Engineering, 416-978-3662

Detailed instructions will be provided before beginning each part of the experiment. If you have read and understood the above information, please sign the attached informed consent form and tell the experimenter when you are ready to continue.

Page 130: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

123

PART I INSTRUCTIONS

You will be shown a screen with 20 !vehicles", each traversing a route that is .5 miles long

(as shown in the picture below), which all share the same speed profile – that is, they move

with about the same average velocity, and their velocities fluctuate about the same amount

across the route. Each vehicle"s speed will change 10 times as it traverses the route (e.g., it

will slow down and speed up as if it were encountering varying headwinds). As the trial

progresses, you will note that, even though the vehicles share the same speed profile, they

each arrive at the destination at different times. Sometimes the arrival times will be close

together, sometimes they will be more spread out. Pay attention to this distribution of arrival

times, as it will be important later in the experiment.

Once all 20 vehicles have completed their routes, a query screen will appear (as shown

below). You will be asked to enter the identification number of the vehicle that finished the

route first, and the vehicle that finished last. You must enter a vehicle number - even if you

aren"t sure, please provide your best guess. Next, you will be shown a new vehicle that is

about to begin traversing a route and will ask you to predict this vehicle"s average velocity

and then to provide the smallest range that you are 95% sure will contain this vehicle"s

Page 131: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

124

actual average velocity. When you answer these questions, you can assume that this new

vehicle shares the same speed profile of those vehicles you just observed.

The first question “Please estimate this vehicle!s average velocity” is asking you to make

your best guess as to the average speed that this new vehicle will travel based on your

knowledge of the speed profile that it belongs to. Please provide your answers by typing a

number in the box on the screen – you can enter a number with precision up to one decimal

place (i.e. 52.6). Press “Enter” on the keyboard when you are satisfied with your answer.

For the second question, “Please provide the smallest window that you are 95% certain will

contain the vehicle!s average velocity”, use the mouse to move the left slider to represent

the lower bound of the window, and the right slider to represent the upper bound of the

window. The actual numbers of the window that you have selected are shown below the

slider scale. The window you select should represent the smallest possible range of

velocities that you are 95% sure will contain this vehicle!s average velocity. When you are

satisfied with the window you have selected, click “Continue” with the mouse. Once you

press “Continue” you will receive feedback as to whether each of your answers was correct,

too high, or too low. Answers within 1.0 of the actual answer will be considered correct.

Please note this feedback, and consider it in your subsequent answers. You should strive to

provide the most accurate answers that you can.

Page 132: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

125

There will be four different speed profiles – each with a different colour code: Blue, Green,

Yellow, and Red. You will complete five trials of each speed profile. After the fifth trial, you

will be asked to pause and answer a short questionnaire to characterize the profile you just

observed. You will be asked the following three questions for each profile:

1. Based on the average speed, are the vehicles with this profile more likely travelling on a six-lane freeway (i.e. FAST) or a two-lane rural road? (i.e., SLOW) (please circle one) 1 2 3 4 5 6 7

Definitely Definitely two-lane unsure six-lane rural road freeway (i.e., Slow) (i.e., Fast)

2. Based on the speed fluctuations, are the vehicles with this profile more likely traveling with cruise control ON or OFF? (please circle one) 1 2 3 4 5 6 7

Definitely Definitely cruise control unsure cruise control ON OFF 3. Please characterize the time of arrival of vehicles that share this speed profile. Use this to make any notes about the time of arrival that will help you remember later on. Ask the experimenter if you have any questions. We will begin with 4 practice trials, so that you can see one of each of the four speed profiles. Then the experiment will begin.

Page 133: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

126

PART 2 INSTRUCTIONS: TIME OF ARRIVAL ESTIMATION

In the second part of the experiment, you will be shown only one vehicle at a time. Its speed

profile will always be one of the four you saw in Part 1, and will be identified with the same

color code (Blue, Green, Red, or Yellow). You will watch the vehicle as it traverses a route

to remind you what the speed profile looked like. Then a new route will be displayed along

with an estimated time of arrival (in elapsed seconds) specific for the distance of route

shown. This estimated time of arrival is based on the average speed, but as you saw in part

1, the actual TOAs will vary. You will be asked to estimate the smallest window that you are

95% certain will contain the vehicle!s actual TOA”. When answering this question, you

should consider that the vehicle shares the same profile as all of the vehicles with

the same color icon from the first part of the study and use your knowledge of the

speed profile to answer the questions. Recall from Part 1, that even though all vehicles

were trying to arrive at the destination at the same time, there was some variability around

the arrival time. The window you select should represent the smallest possible window of

time that you are 95% sure will contain this vehicle!s actual time of arrival -- that is, that

would capture 95 of 100 vehicles. Use the mouse to move the left slider to represent the

lower bound of the time window, and the right slider to represent the upper bound of the time

window. Again, please provide answers, in seconds, with precision to one decimal point.

The actual numbers of the window you select will be shown below the slider scale. You will

not receive any feedback in this portion of the study.

We!ll start with two trials to let you practice, and then we!ll continue with the experiment.

Please ask the experimenter if you have any questions before commencing.

Page 134: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

127

PART 3 INSTRUCTIONS: SPEED PROFILE IDENTIFICATION

This task will be identical to Part 2, except, this time the vehicle!s speed profile will not be

identified for you – that is, the vehicle!s color code will not be displayed (all vehicles will be

black). Some vehicles will match one of the speed profiles you experienced in the earlier

portions of the study, but some will not. Your first task will be to determine if it matches one

of the color-coded speed profiles, and if so, which one. Then you will continue to provide the

smallest possible window of time that you are 95% sure will contain this vehicle!s actual time

of arrival. Again, you should use your knowledge of the speed profiles that you learned in

Phase 1 to answer this question.

For each trial, you will be asked to answer the following questions:

1. To which profile is this vehicle most similar?

_____ Red _____ Yellow _____ Green _____ Blue

2. Please rate the degree of similarity between the vehicle you just observed and the profile

you selected above:

Very Low Low Moderate High Very High

1 2 3 4 5

Please ask the experimenter if you have any questions before commencing.

Page 135: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

128

APPENDIX C2: EXPERIMENT 3 DEMOGRAPHIC QUESTIONNAIRE

1. Name: ________________________________________________________ 2. Age: _________________________ 3. Gender: Male _________ Female ______________

4. Do you normally wear glasses or contact lenses for any of the following activities?

(please check all that apply) Reading _____ Watching TV _____ Working on the computer _____ Driving _____ Other tasks, please specify _____________________________________________

5. Are you wearing either glasses or contact lenses today? No _____ Yes _____ 6. Are you currently employed? No _____ Yes _____ If yes, are you employed part-time _____, full-time _____, contract or casual hours

______ What is your job title:

________________________________________________________ 7. Are you currently a student? No _____ Yes _____ If yes, please state program, institution, and your START date (e.g. A.A/De Anza/Sept,

2005) ____________________________________________________________________________

8. What is the highest level of education that you have completed? _____ High School / GED _____Community College (2 year degree) Please specify degree: _______AA or

_______AS _____University (4 year degree). Please specify Degree / University (e.g. B.A. Sociology /

SJSU) ________________________________________________________________ ______Graduate School. Please specify Degree / University (i.e Ph.D. Chemistry / Stanford) ________________________________________________________________ ______Other Education / Diploma / Certificate program (please Specify) ________________________________________________________________ 9. Do you have (or have you ever had) a Pilot!s license? No _____ Yes _____

If yes, Please specify type (e.g., private / commercial; fixed-wing, rotorcraft) ___________________________________________________________________

10. Have you ever taken a course in statistics? No _____ Yes _____ If yes, how many statistics courses have you taken? ____________________ Please list the Department / Institution (e.g. Psychology / SJSU) for each course. ____________________________________________________________

Page 136: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

129

APPENDIX C3: EXPERIMENT 3 DEBRIEF QUESTIONS

Consider a new speed profile with an average speed of 45 mph. If the vehicles traveled a

route that was 4 miles long at exactly 45 mph, they would arrive in 320 seconds (5 minutes

and 20 seconds). Since their speeds fluctuate due to traffic, 95% of the vehicles arrive

within an 80 second window around the expected TOA.

a) If the vehicles were then to travel a route that was only 2 miles long, the expected TOA

would be 160 seconds (2 minutes and 40 seconds). How big would the TOA window

have to be to include 95% of the vehicles! arrival times?

b) If the vehicles travel an 8 mile route, the expected TOA would be 640 seconds (10

minutes and 40 seconds). How big would the TOA window have to be to include 95% of

the vehicles! arrival times?

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230

Expected TOA Expected TOA = 160

Seconds TOA Window =

Expected TOA Expected TOA = 320

Seconds TOA Window =

210 220 230 240 250 260 270 280 290 300 310 320 330 340 350 360 370 380 390

530 540 550 560 570 580 590 600 610 620 630 640 650 660 670 680 690 700 710

Expected TOA Expected TOA = 640 Seconds TOA Window = _____ seconds

Page 137: The Development and Evaluation of a Model of Time-of ... · scientific discussions about time of arrival uncertainty, experimental methods, and data analyses; facilitating the development

130

1. With respect to the second block of trials – What percent of vehicle TOAs do you

think you captured within your TOA windows?

2. What strategy did you use to assess the TOA windows?

3. Were your TOA windows symmetrical or asymmetrical? How did you decide?

4. Did your strategy change over the course of the experiment? / Do you think your

TOA windows got smaller with time?

5. How did you factor in the distance of the route?

6. Did you factor in knowledge of the speed profile, i.e. other aircraft with the same

color – or just the one aircraft shown in each trial?

7. With respect to the last block of trials – when a vehicle did not match one of the

previously learned speed profiles – how did you determine the TOA windows?