Intention-based Decision Making via Intention Recognition and its Applications

Introduction Evolution Prospection Intention Recognition Applications Conclusions

Intention-based Decision Making viaIntention Recognition and its Applications

Han The Anh ([email protected])Luıs Moniz Pereira ([email protected])

/UNL

Waren SymposiumUS, September 2012


Outline


Introduction

• Intentions play an important role in our decision making(secure cooperation, coordination; foresee hostile behaviors).

• There has been no real attempt to model and implement therole of intentions in decision making, within a rational choiceframework.

• Intentions of other relevant agents are always assumed to begiven as the input of a decision making process.


Introduction - 2

• We present a Logic-based framework for decision making thattakes into account intentions of other agents (via intentionrecognition).

• It extends our previous work on Evolution Prospection fordecision making, but now empowered to predict intentions ofothers and take them into account when making decision.


Some intended application domains

• Ambient intelligence and elder care: proactive support,dealing with emergency and security issues.

• Moral reasoning: intentionality plays key roles in moral andlegal reasoning.

• Game theory.


A simple scenario ...

Example (Tea-Coffee with Intention Recognition)

• Being thirsty, I consider making tea or coffee.

• I realize that my roommate, John, also wants to have a drink.

• To be friendly, I want to take into account his intention whenmaking my choice.

Example

1. coffee / tea← has intention(john, coffee).

2. tea / coffee ← has intention(john, tea).


Evolution Prospection (EP)

• Enable an agent to look ahead prospectively into itshypothetical futures, to determine the best one to follow.

• We implement several preference constructs.


Constructs of EP

• Active goals

• Abducibles

• Local preferences

• Evolution-level preferences


Active Goal

Definition

At each cycle, the agent has a set of active goals to be satisfied

on observe(AG )← Body

”on observing Body trigger goal AG”

Example

1. on observed(drink)← thirsty .


A Priori Preference

Definition (A priori preferences)

Preferences over abducibles

a / b ← Body

”Prefer abducible a to abducible b”

Example

1. coffee / tea← has intention(john, coffee).

2. tea / coffee ← has intention(john, tea).


A Posteriori Preference

Definition (A posteriori preferences)

Preferences over abductive solutions

Ai � Aj ← holds given(Li , Ai), holds given(Lj , Aj)

”Ai is preferred to Aj if Li and Lj are true consequences of Ai andAj , respectively”

Example

1. unhappy ← make coffee, has intention(john, tea)

2. happy ← make tea, has intention(john, tea)

3. Ai � Aj ← holds given(happy , Ai ),holds given(unhappy , Aj)


Intention Recognition

• Infer intention of other agent based on observed actions.

• Probabilistic approach via Bayesian Network.


Bayesian Network for Intention Recognition

Causes/Reasons

C-2

C-N

I-1

I-M

A-1

C-1

A-P

.

.

.

.

.

.

.

.

.

.

.

Intentions

Actions

Subject to Changes

P(A1|I1,IM)

CPD table for each node X P(X|parents(X))

IR: Compute P(I-i|obs) i = 1,...,M

P(C1)


Fox-Crow Fable

There is a crow,holding a cheese.A fox, being hungry,approaches thecrow and praisesher, hoping thatthe crow will singand the cheese willfall down near him.Unfortunately forthe fox, the crowis very intelligent,having theability of intentionrecognition.


Fox’s Intentions BN for intention recognition


Likelihood of intentions

Probabilities of Fox having intention Food, Territory, Please giventhe observation that Fox praised Crow can be found in P-log1 withqueries

• pr(i(food, t) ′|′ obs(praised(t)), V1). V1 = 0.9317.

• pr(i(territory, t) ′|′obs(praised(t)), V2). V2 = 0.8836.

• pr(i(please, t) ′|′ obs(praised(t)), V3). V3 = 0.0900.

1Details about P-log and how this BN is represented using P-log is inappendix.


Integration

• Technically, knowledge about intentions of others can figure inany EP constructs

• Active goals• Preference rules• Integrity constraints


Intention Triggering Active Goals

• Recall that an active goal has the form

on observe(AG )← L1, ..., Lt (t ≥ 0)

• Help friend to achieve his/her intention

Example

on observe(help achieve goal(make coffee))←

friend(John), has intention(John, drink coffee).


Intention Triggering Active Goals - 2

• Prevent opponents from achieving their goal

on_observe(prevent_achieve_goal(G)) <-opponent(P), has_intention(P,G).

Example

on observe(prevent achieve goal(attack city))←

opponent(David), has intention(David , attack city).


Intention Triggering Preferences

Having recognized an intention of another agent, the agent mayfavor or disfavor an action, a solution, or one evolution to another

• The one providing more support to achieve the intention isfavored in a friendly setting

• The one which provides less support or prevents to achievethe intention is favored in a hostile setting.


Elder Care

• Proactive support: proactively provide suggestions forachieving the recognized intentions.

• Security and emergency: improve Burglary alarms, dealingwith the case where the assisting person have dangerousintentions, etc.


Providing proactive support – Example

• An elder stays alone in his apartment.

• The intention recognition system observes that he is lookingfor something in the living room.

• In order to assist him, the system needs to figure out what heintends to find. The possible things are:

• something to read (Book);• something to drink (Drink);• TV remote control (Rem);• the light switch (Switch).

• The states of the light and TV are observed.



Likelihood of intentions

Probabilities of the Elder having intention book, drink, remotecontrol, switch given the observation that he is looking forsomething and the states of light and TV (on or off), can be foundin P-log with queries 2

?− pr(i(b, t) | (obs(tv(S1)) & obs(lt(S2)) & obs(look(t))), V1).?− pr(i(dr, t) | (obs(tv(S1)) & obs(lt(S2)) & obs(look(t))), V2).?− pr(i(rem, t) | (obs(tv(S1)) & obs(lt(S2)) & obs(look(t))), V3).?− pr(i(sw, t) | (obs(tv(S1))&obs(lt(S2)) & obs(look(t))), V4).

where S1, S2 are boolean values (t or f ) instantiated duringexecution, depending on the states of the light and TV.

2Details of the P-log representation is in Appendix


• If the light is off (S2 = f ), then V1 = V2 = V3 = 0, V4 = 1.0,regardless of the state of the TV.

• If the light is on and TV is off (S1 = t, S2 = f ), thenV1 = 0.7521, V2 = 0.5465, V3 = 0.5036, V4 = 0.0101.

• If both light and TV are on (S1 = t, S2 = t), thenV1 = 0, V2 = 0.6263, V3 = 0.9279, V4 = 0.0102.

Thus, if one observes the light is off, definitely the elder is lookingfor the light switch. Otherwise, if one observes the light is on, nomatter TV is on or off, the first three intentions book, drink,remote control still under consideration in next phase.The intention of looking for the light switch is very unlikelycompared with the others, thus being ruled out.When there is light, one goes directly to the light switch if theintention is to turn it off, without having to look for it.


Providing suggestions using Evolution Prospection system

• Having recognized the intention of an agent, EP can be usedto provide the best courses of evolution (action suggestion)for that agent to achieve its own intention.

• In Elder Care domain, assisting systems should be able toprovide contextually appropriate suggestions for the elders,based on their recognized intentions.

• The assisting system is supposed to be better aware of theenvironment, the elders’ physical states, mental states as wellas their scheduled events, so as to provide good and safesuggestions, or simply warnings.


Suggestions for a Drink – Example (cont.)

Suppose that the final confirmed intention is that of looking for adrink.

The possibilities are natural pure water, tea, coffee and juice.


Representation in EP

• Declaration of program abducibles

abds([water/0, coffee/0, tea/0, juice/0]).

• All hypotheses for drinks are always expectable.

expect(coffee). expect(tea).expect(water). expect(juice).

• Active goal finding a drink is triggered if the ”drink” intentionis confirmed

on_observe(drink) <- has_intention(elder,drink).

drink <- tea. drink <- coffee.drink <- water. drink <- juice.


Representation in EP

• Counter-expectation rules: it is not expected to drink coffee ifhigh blood pressure, ...

expect_not(coffee) <- prolog(blood_high_pressure).expect_not(coffee) <- prolog(sleep_difficulty).expect_not(coffee) <- prolog(late).expect_not(juice) <- prolog(late).

• Some integrity constraints: it is not allowed to have tea andcoffee, tea and juice, coffee and juice, tea and water together

<- tea, coffee. <- coffee, juice.<- tea, juice. <- tea, water.


Representation in EP – A priori preferences

• In the morning, coffee is most preferred

coffee <| tea <- prolog(morning_time).coffee <| water <- prolog(morning_time).coffee <| juice <- prolog(morning_time).

• If weather hot: juice is preferred to coffee, ....

juice <| coffee <- prolog(hot).juice <| tea <- prolog(hot).juice <| water <- prolog(hot).water <| coffee <- prolog(hot).water <| tea <- prolog(hot).

• If it is cold: tea is most preferred

tea <| coffee <- prolog(cold).tea <| juice <- prolog(cold).tea <| water <- prolog(cold).


A Posteriori Preferences

After the elder’s physical states are considered in the a prioripreferences and expectation rules, guaranteeing that only choicescontextually safe for the elder are generated, other aspects such aselder’s pleasure w.r.t. to each kind of drink, are taken intoaccount, by the a posteriori preferences

• Abductive solution providing higher pleasure is preferred

Ai << Aj <- holds_given(pleasure_level(V1),Ai),holds_given(pleasure_level(V2),Aj),V1 > V2.

• Pleasure levels of each abducible

sugar_level(1) <- coffee. sugar_level(1) <- tea.

sugar_level(5) <- juice. sugar_level(0) <- water.


Prolog code

• Prolog code can be embedded in EP programs

beginProlog.late :- time(T), (T > 23; T < 5).morning_time :- time(T), T > 7, T < 10.

hot :- temperature(TM), TM > 32.cold :- temperature(TM), TM < 10.

blood_high_pressure :-physical_state(blood_high_pressure).

sleep_difficulty :- physical_state(sleep_difficulty).endProlog.


Suggestions provided by the EP program

1. time(24) (late); temperature(16) (not hot, not cold); no highblood pressure; no sleep difficulty:

• two abductive solutions: [tea], [water ].• final solutions: [tea].

2. time(8) (morning time); temperature(16) (not hot, not cold);no high blood pressure; no sleep difficulty:

• two abductive solutions: [coffee], [coffee, water ].• final: [coffee], [coffee, water ].

3. time(18) (not late, not morning time); temperature(16) (notcold, not hot); no high blood pressure; no sleep difficulty:

• six abductive solutions: [coffee], [coffee,water], [juice],[juice,water], [tea], and [water].

• final: [coffee], [coffee,water].


4. time(18) (not late, not morning time); temperature(16) (notcold, not hot); no high blood pressure; no sleep difficulty:

• six abductive solutions: [coffee], [coffee,water], [juice],[juice,water], [tea], and [water].

• final: [coffee], [coffee,water].

5. time(18) (not late, not morning time); temperature(16) (notcold, not hot); high blood pressure; no sleep difficulty:

• four abductive solutions: [juice], [juice,water], [tea], [water].• final: [tea].

6. time(18) (not late, not morning time); temperature(16) (notcold, not hot); no high blood pressure; sleep difficulty:

• four abductive solutions: [juice], [juice,water], [tea], [water].• final: [tea].


7. time(18) (not late, not morning time); temperature(8) (cold);no high blood pressure; no sleep difficulty:

• one abductive solution: [tea] (thus, final)

8. time(18) (not late, not morning time); temperature(35)(hot); no high blood pressure; no sleep difficulty:

• two abductive solutions: [juice], [juice,water].• final: [juice], [juice,water].


Security: Burglary alarm example

Example (Elder Intentions)

• An elder stays alone in his apartment.

• One day, the Burglary Alarm is ringing.

• IR system observes that he is looking for something.

• To assist him, it needs to figure out what he intends to find.

• Possible things are:• Alarm button (AlarmB);• Contact Device (ContDev);• Defensible Weapons (Weapon);• Light switch (Switch).


Bayesian Network for Intention Recognition


Intention-based decision making in Moral Reasoning

• A key factor in moral and legal judgements is intention: intentdifferentiates, for instance, murder from manslaughter.

• Intentionality plays central parts in moral rules.

Double effect principle

• Harming another individual is permissible if it is the foreseenconsequence of an act that will lead to a greater good;

• in contrast, it is impermissible to harm someone else as anintended means to a greater good.


Examples: Modelling trolley problem

Trolley problem

• There is a trolley and its conductor has fainted.

• The trolley is headed toward five people walking on the track.

• The banks of the track are so steep that they will not be ableto get off the track in time.

Given this initial circumstance, there exist several cases of moraldilemmas.


Bystander Case


Footbridge Case


Bayesian Network for intentional killing recognition

Intended Means (IM)

Intentional Kill (IK)

observed act (O)

Personal Reason (PR)

• P(O = t|IK ) = 1 for all IK ∈ {t, f }.• P(IK = t|IM = t, PR) = 1; P(IK = t|IM = f , PR = t) =

0.6; P(IK = t|IM = f , PR = f ) = 0.


Bayesian Network for intentional killing recognition

• In original forms, no personal reason is considered:P(PR = t) = 0.

• P(IK = t|O = t) = 0 for the Bystander case and 1 for othertwo cases.


Modelling Bystander Case

• Mere watching or throwing switch

1. abds([watching/0, throwing_switch/0]).2. on_observe(decide) <- train_comming.

decide <- watching.decide <- throwing_switch.<- throwing_switch, watching.expect(watching). expect(throwing_switch).

• Consequences of each choice:

3. train_straight <- watching.end(die(5)) <- train_straight.

4. redirect_train <- throwing_switch.end(die(1)) <- human(X), side_track(X), redirect_train.



• A posteriori preference:

Ai << Aj <- holds_given(end(die(N)), Ai),holds_given(end(die(K)), Aj), N < K.

• Intentional killing is prohibited:• An act is judged as intentional killing if being predicted by the

IR module with a probability greater some given threshold.• This threshold depends on how certain the judgement needs to

be provided, e.g. 0.95 if it is ‘guilty beyond reasonable doubt’.

intentional_kill <- throwing_switch,has_intention(ned, kill, Pr), Prolog(Pr > 0.95).

<- intentional_killing.


Bystander Case: results

• When no personal reason is considered, P(IK = t|O = t) = 0,there are two prior abductive solutions: [watching ],[throwing switch]. Hence, final result: [throwing switch].

• If personal reason is considered, e.g. there is a good chancethat Hank wants to kill the person on the sidetrack:P(PR = t) = 0.85. Hence, P(IK = t|O = t) = 0.51.

• It is not enough to judge that Hank’s action is one ofintentional killing, beyond reasonable doubt, but theprobability is high enough to require further investigation toclarify the case.


Modelling Footbridge Case• Mere watching or shoving the heavy object1. abds([watching/0, shove/1]).

on_observe(decide) <- train_comming.decide <- watching.decide <- shove(X).<- watching, shove(X).expect(watching). expect(shove(X)) <- stand_near(X).

• Consequences of each choice:2. train_straight <- watching.

end(die(5)) <- train_straight.3. on_track(X) <- shove(X).

stop_train(X) <- on_track(X), heavy(X).kill(1) <- human(X), on_track(X).kill(0) <- inanimate_object(X), on_track(X).end(die(N)) <- kill(N).



• A posteriori preference:

Ai << Aj <- holds_given(end(die(N)), Ai),holds_given(end(die(K)), Aj), N < K.

• Intentional killing is prohibited:

intentional_kill <- human(X), shove(X),has_intention(ian, kill, Pr), Prolog(Pr > 0.95).


Footbridge Case: results

• In this case, showing a person is forbidden since it meansusing him/her as intended means, i.e. P(IM = t) = 1.

• Hence P(IK = t|O = t) = 1: there is only one abductivesolutions: [watching ]. Final result: [throwing switch].


Moral Reasoning about Uncertain Actions: Motivation

• Passing moral judgment without actually or fully observingthe situation: No full, certain, information about actions.

• Be able to reason about actions that might have occurred,but uncertainly.

• Pass judgment and verdicts adhering to moral rules, withinpredefined uncertainty limits


Footbridge Case in Court

• The action of shoving was not observed.

• Jurors are presented with facts:• man died on the sidetrack• defendant was seen on the bridge at the occasion

• Defendant is guilty, beyond reasonable doubt?


Modelling Uncertainty

How sure is the verdict?

• Guilty beyond reasonable doubt: if probability of intentionalshoving > 0.95.

• Not guilty: if probability of intentional shoving < 0.6.

highly_probable(guilty_beyond_reasonable_doubt):- pr_iShv(PrG), PrG > 0.9.

highly_probable(not_guilty):- pr_iShv(PrG), PrG < 0.6.


Intention-based decision making in games

• In many strategic situations typically modeled using gamesthe achievement of a player’s goal depends also on otherplayers’ move.

• The knowledge about others’ intentions enables to plan inadvance, either to secure cooperation or to deal with potentialhostile behavior.

• Experimental and theoretical evidence show the importance oftaking into account others’ intentions into the decisionmaking process.


Example - 1

Prefer to cooperate if the co-player intends to cooperate, andprefer to defect otherwise

1. abds([move/1]).2. on observed(decide)← new interaction.3. decide ← move(c).

decide ← move(d).← move(c), move(d).

4. expect(move(X )).5. move(c) / move(d)← has intention(co player , c).move(d) / move(c)← has intention(co player , d).


Example - 2

Based on recognizing a co-player strategies

5. move(d) / move(c)← has intention(co player , allc).move(d) / move(c)← has intention(co player , alld).move(c) / move(d)← has intention(co player , tft)move(c) / move(d)← has intention(co player , wsls),

game state(s), (s = ‘R’; s = ‘P’).move(c) / move(d)← has intention(co player , wsls),

game state(s), (s = ‘T’; s = ‘S’).


Conclusions

• An intention-based decision making system on top ofEvolution Prospection and Intention Recognition systems.

• Application domains: Ambient Intelligence, Elder Care; MoralReasoning; Game Theory.


Future works

• Other application domains: story understanding,human-computer and interface-agents systems, trafficmonitoring, military settings.

• Extend to take into account collective intentions (of group orteam of agents) via collective intention recognition.


Thank you!

QUESTIONS


P-log

The computation in BNs can be automated by using P-log

• P-log – a declarative language based on a logic formalism forprobabilistic reasoning

• It uses Answer Set Programming as its logical and CausalBayesian Nets as its probabilistic foundations.


Fox’s Intentions CBN in P-log

• Two sorts, bool and fox intentions, represent boolean valuesand set of Fox’s intentions

bool = {t,f}.fox_intentions = {food,please,ter}.

• Attributes hungry fox, friendly fox, praised and i state thatfirst 3 have no domain parameter and are boolean, and lastmaps each Fox intention to a boolean

hungry_fox : bool. friendly_fox : bool.i : fox_intentions --> bool. praised : bool.

• These attributes are randomly distributed in their full ranges

random(rh,hungry_fox,full). random(rf,friendly_fox,full).random(ri,i(I),full). random(rp,praised,full).


Representing CPD in P-log: pa rule• Probability distribution of top nodes hungry fox, friendly fox

pa(rh,hungry_fox(t),d(1,2)).pa(rf,friendly_fox(t),d(1,100)).

• Probability distribution of Fox having intention food,conditional on its parents (hungry fox, friendly fox)pa(ri(food),i(food,t),d(8,10))

:- friendly_fox(t),hungry_fox(t).

pa(ri(food),i(food,t),d(9,10)):- friendly_fox(f),hungry_fox(t).

pa(ri(food),i(food,t),d(1,100)):- friendly_fox(t),hungry_fox(f).

pa(ri(food),i(food,t),d(2,10)):-friendly_fox(f),hungry_fox(f).

• Similarly for other intentions.


• Probability distribution of observation Praised, conditional onits parents (i(food), i(please), i(ter))

pa(rp, praised(t),d(95,100)):- i(food, t), i(please, t).

pa(rp, praised(t),d(6,10)):- i(food, t), i(please, f).

pa(rp, praised(t),d(8,10)):- i(food, f), i(please, t).

pa(rp, praised(t),d(1,100)):- i(food, f),i(please,f),i(ter,t).

pa(rp, praised(t),d(1,1000)):- i(food,f),i(please,f),i(ter,f).


Elder’s Intentions CBN in P-log

• Two sorts, bool and elder intentions, represent booleanvalues and set of the Elder’s intentions

bool = {t,f}.elder_intentions = {book,drink,rem,switch}.

• Attributes thsty (thirsty), lr (like reading), lw (like watching),tv (tv on), lt (light on) have no domain parameter and getboolean values

thsty:bool. lr:bool. lw:bool. tv:bool. lt:bool.

• Attribute i maps each Elder’s intention to a boolean value

i:elder_intentions --> bool.

• These attributes are randomly distributed in their full ranges

random(rth,thsty,full). random(rlr,lr,full).random(rlw,lw,full). random(rtv,tv,full).random(rl, lt, full). random(ri, i(I), full).


Representing CPD in P-log: pa rule

• Probability distribution of top nodes

pa(rth,thsty(t),d(1,2)). pa(rlr,lr(t),d(8,10)).pa(rlw,lw(t),d(7,10)). pa(rtv,tv(t),d(1,2)).pa(rl,lt(t),d(1,2)).

• Probability distribution of the Elder having intention book,conditional on its parents (all top nodes)

pa(ri(b),i(b,t),d(0,1)) :-lt(f).pa(ri(b),i(b,t),d(0,1)) :-lt(t),tv(t).pa(ri(b),i(b,t),d(3,5)) :-lt(t),tv(f),lr(t),lw(t),thsty(t).pa(ri(b),i(b,t),d(13,20)):-lt(t),tv(f),lr(t),lw(t),thsty(f).pa(ri(b),i(b,t),d(7,10)) :-lt(t),tv(f),lr(t),lw(f),thsty(t).pa(ri(b),i(b,t),d(4,5)) :-lt(t),tv(f),lr(t),lw(f),thsty(f).pa(ri(b),i(b,t),d(1,10)) :-lt(t),tv(f),lr(f),lw(t).pa(ri(b),i(b,t),d(4,10)) :-lt(t),tv(f),lr(f),lw(f).

• Similarly for other intentions.


• Probability distribution of observation look, conditional on itsparents (i(b), i(dr), i(rem), i(sw))pa(rla,look(t),d(99,100)):-i(b,t),i(dr,t),i(rem,t).

pa(rla,look(t),d(7,10)) :-i(b,t) i(dr,t),i(rem,f).

pa(rla,look(t),d(9,10)) :-i(b,t),i(dr,f),i(rem,t).

pa(rla,look(t),d(6,10)) :-i(b,t),i(dr,f),i(rem,f).

pa(rla,look(t),d(6,10)) :-i(b,f),i(dr,t),i(rem,t).

pa(rla,look(t),d(3,10)) :-i(b,f),i(dr,t), i(rem,f).

pa(rla,look(t),d(4,10)) :-i(b,f),i(dr,f),i(rem,t).

pa(rla,look(t),d(1,10)) :-i(b,f),i(dr,f),i(rem,f),i(sw,t).

pa(rla,look(t),d(1,100)) :-i(b,f),i(dr,f),i(rem,f),i(sw,f).

Documents

Intention-based Decision Making via Intention Recognition and its Applications