Humans in Loops –

Neocybernetic View at Complex Processes

Presentation at Automation Days 2009

Heikki HyötyniemiHelsinki University of Technology

Automation and Systems Engineering

Process levels – EXTENDED

Physical level (formation): Input Qin, output Qout; based on the actual process structure and its energy/matter flows

Information level (metaformation): Inputs Qin and Qout, output h; based on what can be measured and what can be affected

Knowledge level (metainformation): Inputs are now the tensions of the system, outputs are its resulting properties

FC

Qin

Qout

LI

h

FCQin

Qout

h

“How to enhance the measurement?”

Emphasis now

From process flows to information flows ...

Controls

Controllerdynamics

-+

Actuatordynamics

Measurementdynamics

Processdynamics

(zoom)

... From information flows to “knowhowflows”

Large amounts of appropriately chosen external information and understanding

Controllerdynamics

-+

Actuatordynamics

Measurementdynamics

Processdynamics

Referen

ce

Error

Contro

l

Inpu

tOutp

ut

Feedba

ck

Distur

bance

“Controls”

ux+ B

A

(zoom) “Eksoinformation”

“Endoinformation” (found within the system)

Role of neocybernetics

Evolution that is based on local agents: special structure remaining visible in the closed loop on the global level

Interpretation: degrees of freedom in the directions of interest, balance among tensions (physical + economical)

System becomes better and better balanced and robust against environmental disturbances

This applies also to “artificial evolution” implemented by humans when developing industrial plants

Neocybernetics offers tools for implementing “higher-level control” in the spirit of “second-order cybernetics”

Now, there are humans in the system structure...

First, study a simple case: Linearity assumed Process linearized around a single operating point Mental model reduced to a set of quality measures

Then, degrees of freedom reduce to principal component subspace axes

Higher-level view: abstract the signals away, concentrate on qualifiers vs. qualities

Utilize the low-dimensional mapping between the qualifiers and qualities in parameter optimization, etc. ...

Experiments with the IRT method

Veitsiluoto case:Pulp digester

0 10 20 300

1

2

3

4

5Cost function J

0 10 20 303.5

4

4.5

5

5.5

6Quality measure 1

0 10 20 300.012

0.014

0.016

0.018

0.02

0.022

0.024

0.026

0.028Quality measure 3

0 10 20 300.23

0.24

0.25

0.26

0.27

0.28

0.29

0.3

0.31Quality measure 2

0 10 20 300.1

0.12

0.14

0.16

0.18

0.2

0.22Quality measure 4

0 10 20 300.0115

0.012

0.0125

0.013

0.0135

0.014

0.0145Quality measure 5

0 10 20 30200

300

400

500

600

700

800

900

1000

1100Quality measure 6

Optimization in reduced-dimension parameter space

Extending the operator view

PID

A R S

PID

PID

PID

PID

PID

Explicit optimization is often not possible

The slopes of criteria can still be found

The tuning knobs can be made to reflect relevance

Lower-level controllers tuned in

a coordinated manner according to a model where

control parameters affect qualities

“Neocybernetic expertise”

More general case: Process consisting of many submodels, expertise with various categories = prototypes

Observation/measurement data correlations are detected and modeled in terms of sparse coded features

Categories/concepts (subsymbolic) can be implemented using sparse coding in a cognitively plausible way

Reasoning/inference becomes a static pattern matching process in a high-dimensional space

Process/mental structures share the same sparse coded model structures and can thus be combined!

“Conceptual froth types” at Pyhäsalmi

How operators see the froth

“Stiff”

“Wet”

“Dry”

Sparse coding captures categories!

Scores of “hand-coded” features

Scores of sparsely coded features

Delphi method

Exploitation of expertise “The Delphi method is a systematic, interactive forecasting method

which relies on a panel of independent experts. The experts answer questionnaires in two or more rounds. After each round, a facilitator provides an anonymous summary of the experts’ forecasts from the previous round as well as the reasons they provided for their judgments. Thus, participants are encouraged to revise their earlier answers in light of the replies of other members of the group. It is believed that during this process the range of the answers will decrease and the group will converge towards the "correct" answer. …”

Automation of the method (cf. Case-Based Reasoning): In industry, operator expertise is directly available (measurements +

actions) This information only has to be compressed using the cybernetic

models Claim: the neocybernetic model of tensions is just right for this purpose

How about different kind of expertise?

H. Hyötyniemi: Expertise: an optimized data model? (STeP 2000)

Data courtesy of Neles

Idea of “Second-Order Cybernetics” exploited

Symmetry: The complex systems “observe” each other, constructing neocybernetic pattern models of each other

Recognition ofprocess patterns

Recognition ofmental patterns

Process domain

Mental domain

“Second-Order Control” in general

Neocybernetic intuitions combined

Process

StateControl

Processpatterns

Actionpatterns

Subspace

Featureextraction

Featureextraction

Human

Learn

ing

Learn

ing