Vous êtes sur la page 1sur 1213

s ^ WiCi/^

Ip M^r \

Proceeding s

Fourteent h

Annua l

o f

th e Conferenc e

 

o f

th e

Cognitiv e

Scienc e

Societ y

< y .

'f\f,

July 2 9 to Augus t 1, 199 2 Cognitive Science Progra m Indiana University, Bloomingto n

Proceeding s

o f

th e

Fourteent h Annua l

Conferenc e

o f

th e

Cognitiv e Scienc e

Societ y

July 2 9 to Augus t 1,199 2 Cognitive Science Progra m Indiana University, Bloomingto n

Copyright © 1992 by the Cognitive Science Society.

All rights reserved. No part of this book may be reproduced in any form, by photostat, microform, retrieval system, or by any other means, without the prior written permission of the publisher.

Distributed by

Lawrence Erlbaum Associates, Inc. 365 Broadway Hillsdale, New Jersey 07642

ISBN 0-8058-1291-1

ISSN 1047-1316

Printed in the United States of America

TABLE O F CONTENT S The Fourteenth Annual Conference of the Cognitive Science Society Indiana University Bloomington, Indiana 47405 July 29 - August 1, 1992

Conference Organizers

xi

The effects of pattern presentation on interference

Preface

xiii

in backpropagation networks

Reviewers

xiv

  • J. Murre

54

TALKS

DAY ONE, SESSION ONE

Planning

Why intelligent systems should get depressed occasionally and appropriately

Development

Developmental changes in infants' perceptual processing of biomechanical motions

  • J. Pinto, J. Shrager, B. Bertenthal

60

The role of measurement in the construction of

  • C. Webster

1

conservation knowledge

 
  • T. Simon, D. Klahr, A. Newell

66

Reasoning about performance intentions

  • M. Freed, B. Krulwich, L. Birnbaum,

An investigation of balance scale success

  • G. Collins

7

  • W. Schmidt, T. Shultz

72

Preconditions and appropriateness conditions

  • T. Converse, K. Hammond

13

Multi-agent interactions: A vocabulary of engagement

  • P. Goldweic, K. Hammond

18

Psycholinguistics

Chunking processes and context effects in letter perception

E.Servan-Schreiber

78

A vocabulary for indexing plan interactions and Abstractness and transparency in the mental

repairs K, Hammond, C. Seifert

Connectionist Models

24

lexicon W . Marslen-Wilson, L. Tyier, R. Waksler, L.OIder

84

Polysemy and lexical representation: The case of three English prepositions

Dynamic construction of mental models in connectionist networks

S.Rice

89

  • V. Ajjanagadde

30

Grammaticality judgment in Chinese-English

bilinguals: A gating experiment

 

Learning relations in an interactive architecture

  • H. Liu, E. Bates, P. Li

95

  • R. Stark

36

Developing microfeatures by analogy

Understanding English past-tense formation: The shared meaning hypothesis

  • E. Melz

42

  • C. Harris

100

Direct, incremental learning of fuzzy propositions

G. Oden

48

lU

Imagery and Visual Reasoning

Modelling inductive and deductive discovery strategies in Galilean Kinematics

A recognition model of geometry theorem-proving

P.Cheng

171

T. McDougal, K. Hammond

106

Complexity management in a discovery task

Simulating theories of mental imagery J. Glasgow, D. Conklin

112

  • C. Schunn, K. Klahr

177

Fractal (reconstructive analogue) memory

Scientific induction: Individual versus group processes and multiple hypotheses

  • D. Stucki, J. Pollack

118

E.Freedman

183

When can visual images be re-interpreted? Non- chronometric tests of pictorialism

P.Slezak

124

Team cognition in the cockpit: Linguistic control of shared problem solving

J. Orasanu, U. Fischer

189

"Ill-structured representations' for ill-structured

problems

V.Goel

130

Discourse

DAY ONE, SESSION TW O

A unified process model of syntactic and semantic error recovery in sentence understanding

  • J. Holbrook, K. Eiselt, IC Mahesh

195

Memory

Sensory discrimination in a short-term trace memory

  • J. McAuley, S. Anderson, R. Port

136

A speech based connectionist model of human short term memory

  • D. Bairaktaris, K. Stenning

141

What you infer might hurt you - a guiding principle for a discourse planner

I.Zukerman

201

Theme construction from belief conflict and resolution

J.Reeves

207

Communicating abstract advice: The role of stories

Does memory activation grow with list strength and/or length?

  • E. Jones

213

  • D. Huber, H. Ziemer, R. Shiffrin,

K-Marinelli

147

Learning and Reactivity

Misinformed and biased: Genuine memory distortions or artifactual phenomena?

Primacy effects and selective attention in incremental clustering

  • R. Pohl

153

D.Thau

219

Neurally motivated constraints on the working memory capacity of a production system for parallel processing: Implications of a connectionist model based on temporal synchrony

L.Shastri

Scientific Discovery

159

Psychological responses to anomalous data

  • C. Chinn, W . Brewer

165

Some epistemic benefits of action: Tetris, a case study

  • D. Kirsh,P. Maglio

224

Reference features as guides to reasoning about opportunities

  • L. Pryor, G. Collins

230

The evolutionary induction of subroutines

  • P. Angeline, J. Pollack

236

Learning several lessons from one experience

  • B. Krulwich, L. Birnbaum, G. Collins

242

IV

Connectionist Models of Language I

Generic teleological mechanisms and their use in case adaptation

Energy minimization and directionality in phonological theories

D.Touretzky, X. Wang

248

  • E. Stroulia, A. Goel

319

Representing cases as knowledge sources that apply local similarity metrics

Integrating category acquisition with inflectional

D.Skalak

325

marking: A model of the German nominal system

  • P. Gupta, B. MacWhinney

253

Rules or connections? The past tense revisited

  • K. Daugherty, M. Seidenberg

259

Multicases: A case-based representation for procedural knowledge

  • R. Zito-Wolf, R. Alterman

331

A connectionist account of English inflectional Similarity and Categories morphology: Evidence from language change

  • M. Hare, J. Elman

265

Learning language in the service of a task

M. St. John

DAY TWO, SESSION ONE Vision

271

Locally-to-globally consistent processing in similarity

R. Goldstone

337

Goal-directed processes in similarity judgement

  • H. Suzuki, H. Ohnishi, K. Shigemasu

343

Correlated properties in artifact and natural kind concepts

ICMcRae

349

On the unitization of novel, complex visual stimuli

  • N. Lightfoot, R. Shiffrin

277

Extending the domain of a feature-based model of property induction

Discovering and using perceptual grouping principles in visual information processing

  • S. Sloman, E. Wisniewski

355

  • M. Mozer, R. Zemel, M. Behrmann

283

An instantiation model of category typicality and instability

The role of genericity in the perception of illusory E. Heit, L. Barsalou contours

360

  • M. Albert

289

 

Cognitive Neuroscience

Perceiving the size of trees via their form

G.Bingham

295

Inhibition and brain computation

 
  • S. Small, G. Fromm

366

Towards the origins of dyslexia

  • R. Nicolson, A. Fawcett

301

Relearning after damage in connectionist networks:

Case-Based Reasoning

A memory architecture for case-based argumentation

  • E. Shafto, R. Bareiss,

L. Birnbaum

307

Implications for patient rehabilitation

D. Plant

372

Modelling paraphasias in normal and aphasic speech

  • T. Harley, S. MacAndrew

378

Linguistic permeability of unilateral neglect:

Constructive similarity assessment: Using stored Evidence from American sign language

cases to define new situations D.Leake

313

  • D. Corina, M. Kritchevsky, U. Bellugi

384

Hippocampal-system function in stimulus

The nature of expertise in anagram solution

representation

and

generalization:

A

L

Novick, N. Cot6

450

computational theory

 
  • M. Gluck,C. Myers

 

390

Allocation of efl^ort to risky decisions

 
  • K. Smith, P. Johnson

456

Connectionist Models of Language II

A constraint satisfaction model of cognitive dissonance phenomena

Learning distributed representations for syllables

  • T. Shultz, M. Lepper

462

M.Gasser

396

A rational theory of cognitive strategy selection

Finding linguistic structure with recurrent neural and change

networks

  • N. Chater, P. Conkey

Q.Wu, J. Anderson

402

468

A phonologically motivated input representation Sentence Comprehension

for the modelling of auditory word perception in continuous speech

  • R. Shillcock, G. Lindsey, J. Levy,

Simultaneous question comprehension and answer

retrieval

  • N. Chater

408

  • S. Robertson, J. Ullman, A. Mehta

474

A PDP approach to processing center-embedded sentences

Implicit argument inferences in on-line comprehension

  • J. Weckerly, J. Elman

414

  • G. Manner, M. Tanenhaus, G. Carlson

480

Forced simple recurrent neural networks and grammatical inference

  • A. Maskara, A. Noetzel

DAY TWO, SESSION TWO

Event Perception

420

Decomposition of temporal sequences

  • J. Avrahami, Y. Kareev

426

Another context effect in sentence processing:

Implications for the principle of referential support

M.Spivey-Knowlton

486

Consulting temporal context during sentence

comprehension: Evidence ft-om the monitoring of eye movements in reading

  • J. Trueswell, M. Tanenhaus

492

Plausibility and syntactic ambiguity resolution

N. Pearlmutter, M. MacDonald

498

The role of correlational structure in learning event categories

  • A. Kersten, D. Billman

432

Development of schemata during event parsing:

Analogy and Metaphor

The time course of metaphor comprehension

  • P. Wolff, D. Centner

504

Neissei^s perceptual connectionist network

cycle

as

a

recurrent

Is the future always ahead? Evidence for system-

  • C. Hanson, S. Hanson

 

438

mappings in understanding space-time metaphors

  • D. Centner, M. Imai

Expertise, Choice, and Problem Solving

Indirect Analogical Mapping

Skill as the fit between performer resources and task demands: A perspective from software use and learning

J.

Hummel, K. Holyoak

Visual Analogical Mapping

P.

Thagard, D. Gochfeld, S. Hardy.

CAshworth

444

510

516

522

VI

Probing the emergent behavior of tabletop, an

architecture uniting high-level perception with

analogy-making

Explanation

Abductive explanation of emotions

  • D. Hofetadter, R. French

528

  • P. O'Rorke, A. Ortony

594

 

Assessing explanatory coherence: A new method

Associative Learning

for integrating verbal data with models of on-line

belief revision

Concept learning and flexible weighting

  • P. Schank, M. Ranney

599

  • D. Aha, R. Goldstone

534

Adaptation of cue-specific learning rates in

network models of human category learning

M. Gluck, P. Glauthier, R. Sutton

540

Educating migraine patients through on-line

generation of medical explanations

  • J. Moore, S. Ohlsson

605

Abstractional and associative processes in concept Tutoring and Modeling of Students

learning: A simulation of pigeon data

  • H. Matute, E. Alberdi

546

Validating COGNITIO by simulating a student

learning to program in Smalltalk

Multivariable function learning: Applications of

Y. Chee, T. Chan

611

the adaptive regression model to intuitive physics

  • P. Price, D. Meyer, K. Koh

552

Using theory revision to model students and

acquire stereotypical errors

Memory for multiplication facts

  • P. Baffes, R. Mooney

617

  • R. Dallaway

558

 

Knowledge tracing in the ACT programming tutor

  • A. Corbett, J. Anderson

623

DAY THREE, SESSION ONE

 

Integrating case presentation with simulation-based

Feature Salience and Reminding

learning-by-doing

  • R. Burke, A Kass

629

Calculating salience of knowledge

L.Rau

564

Diagnosis can help in intelligent tutoring

 

R.Nicolson

635

The interaction of memory and explicit concepts in

 

learning

S.Epstein

570

Foundations

REMIND: Integrating language understanding and The proper treatment of cognition*

episodic memory retrieval in a connectionist

T.vanGelder

641

network

  • T. Lange, C. Wharton

576

Are computational explanations vacuous?

V.Goel

647

A model of the role of expertise in analog retrieval

L. Branting

582

Taking connectionism seriously: The vague

promise of subsymbolism and an alternative

The story with reminding: Memory retrieval is

influenced by analogical similarity

  • C. Wharton, K. Holyoak, P. Downing,

  • T. Lange, T. Wickens

588

P.Verschure

653

Compositionality and systematicity in connectionist

language learning

R.Hadley

659

Vll

The

(non)necessity

of recursion

in natural

Using cognitive biases to guide feature set

language processing

 

selection

 

M.Christiansen

665

  • C. Cardie

743

 

The interaction of principles and examples in

POSTERS

instructions

  • R. Catrambone, R. Wachman

749

Encxxling and retrieval processes: Separate issues

in problem solving

Strategies for contributing to collaborative

L.

Adams

672

arguments

 
  • V. Cavalli-Sforza, A Lesgold,

A grounded mental model of physical systems: A

AWeiner

755

modular connectionist architecture

 
 

AAlmo r

678

The zoo keeper's paradox

 

ICChan

761

Self-organization of auditory motion detectors

 

S.

Anderson

684

A connectionist architecture for sequential decision

 

learning

Simple + Robust = Pragmatic: A natural

 

Y.Chauvin

767

language query processing model for card-type

 

databases

Early warnings of plan failure, false positives and

S.

Arita, H. Shimazu, Y. Takashima

690

envelopes: Experiments and a model

 
  • P. Cohen, R. St. Amant, D. Hart

773

Integrating reactivity, goals, and emotion in a

 

broad agent

Syllable priming and lexical representations:

J.

Bates, A Loyall, W . Reilly.

696

Evidence from experiments and simulations

 

D.Corina

779

Dedal: Using domain concepts to index

 

engineering design information

 

An empirically based computationally tractable

C.

Baudin, J. Gevins, V. Baya,

dialogue model

AMabogunje

702

  • N. DahlbSck, A JOnsson

785

No logic? No problem! Using a covariation

Learning context-free grammars: Capabilities and

analysis on a deductive task

limitations of a recurrent neural network with an

J.

Best

708

external stack memory

 
  • S. Das, C. Giles, G. Sun

791

Projected meaning, grounded meaning and intrinsic

 

meaning

The role of expertise in the development of

C.

Boyle

714

display-based problem solving strategies

 

S.Davies

797

Seeing is believing: Why vision needs semantics

 

M.

Brand, L. Birnbaum, P. Cooper

720

Taxonomies and part-whole hierarchies in the

 

acquisition of word meaning - a connectionist

A case-based approach to problem formulation

model

LBranting

726

G.Dorffner

803

Orthographic and semantic similarity in auditory Point of view: Modeling the emotions of others

rhyme decisions

 
  • C. Elliott, A Ortony

809

C.

Burgess, M. Tanenhaus, N. Marks

732

 

Using stories to enhance and simplify computer

Analogy and representation: Support for the

simulations for teaching

Copycat model

  • R. Feifer, T. Hinrichs

815

B.

Burns, M. Schreiner

737

vui

Bootstrapping syntactic categories

S. Finch, N. Chater

Inference evaluation in deductive, inductive and

820

analogical reasoning

 

B.

Kokinov

903

Frequency effiects on categorization and

 

recognition

Identifying language from speech: An example of

 

J.Florian

826

high-level, statistically-based feature extraction

 

S.

Kwasny, B. Kalman, W . Wu ,

 

Declarative learning: Cognition without primitives

A.

Engebretson

 

909

E.

Furse, R. Nicolson

832

 

Towards a knowledge representation for simple

Hebbian learning of artificial grammars

 

narratives

 

G.

Ganis, H. Schendan

838

R.

Lang, R. Goldman

 

915

Comparison of well-structured & ill-structured task

Question asking during learning with a point and

environments and problem spaces

 

query interface

 
 

V.Goel

844

M.

Langston, A. Graesser

 

921

Prediction performance as a function of the

Towards a distributed network learning framework:

representation language in concept formation

Theory and technology to support educational

systems

electronic learning environments

 

M.Hadzikadic

850

J.

Levin, M. Jacobson

 

927

Transitions between modes of inquiry in a rule A theory of dynamic selective vigilance and

discovery task

preference reversal, based on the example of new

N.

Halberstadt, Y. Kareev

855

coke

 

D.

Levine, S. Leven

 

933

Are rules a thing of the past? The acquisition of

 

verbal morphology by an attractor network

 

Why are situations hard?

 

J.Hoeffner

861

M.

Lewis

939

Memory and discredited information: Can you

Front-end serial processing of complex and

forget I ever said that?

 

compound words: The APPL E model

 

H.

Johnson, C. Seifert

867

G.Libben

945

A fine-grained model of skill acquisition: Fitting

The search image hypothesis in animal behavior:

cascade to individual subjects

 

Its relevance to analyzing vision at the complexity

R.

Jones, K. VanLehn

873

level

 

D.

Lomas, J. Tsotsos

 

951

A re-examination of graded membership in animal

 

and artifact categories

 

Learning by problem solving versus by examples:

 

CKalish

879

The

benefits

of

generating

and

receiving

 

information

 

Progressions of conceptual models of

 

M.

Lovett

956

cardiovascular physiology and their relationship to

 

expertise

The phase tracker of attention

 

D.

Kaufman, V. Patel, S. Magder

885

E.Lumer

962

Augmenting qualitative simulation with global

An extension of rhetorical structure theory for the

filtering

treatment of retrieval dialogues

 
 

H.Kim

891

E.

Maier, S. Sitter

 

968

A production system model of cognitive

A connectionist solution to the multiple

impairments following frontal lobe damage

instantiation problem using temporal synchrony

D.

Kimberg, M. Farah

897

D.

Mani, L. Shastri

 

974

IX

Acquiring rules for need-based actions aided by perception and language

Learning and problem solving under a memoiy load

G. Mani, L. Uhr

980

  • P. Reber, K. Kotovsky

1068

Genetically generated neural networks I:

Categorization and stimulus structure

Representational effects L-Martf

986

S.Ritter

1074

Genetically generated neural networks II:

"Adaptation" to displacement prisms is sensorimotor learning

Searching for an optimal representation L.Martf

991

  • J. Romack, R, Buss, G, Bingham

1080

Using analogies in natural language generation

  • V. Mittal, C Paris

9%

Perceiving size in events via kinematic form

M. Muchisky.G. Bingham

1002

Parallelism in pronoun comprehension A. Nelson, R. Stevenson, K. Stenning

1008

Constraints on models of recognition and recall imposed by data on the time course of retrieval

  • P. Nobel, R. Shiffrin

1014

A model of knowledge-based skill acquisition

  • S. Ohlsson, E. Rees

1020

Problem-solving stereotypes for an intelligent assistant

An analysis of how students take the initiative in keyboard-to-keyboard tutorial dialogues in a fixed

domain

  • G. Sanders, M. Evens, G. Hume, A Rovick,

J.Michael

1086

Visual attention and manipulator control P.Sandon

1092

Additive modular learning in preemptrons

  • G. Saunders, J. Kolen, P. Angeline,

J. Pollack

1098

MusicSoar: Soar as an architecture for music cognition

  • D. Scarborough, P. Manolios,

J.Jones

1104

Attention, memory, and concepts in autism

C Owens

1026

  • H. Schendan

1110

Communicating properties using salience-induced

Collaborative mediation of the setting of activity

comparisons

  • P. Sibun, J. Shrager

1116

  • T. Pattabhiraman, N. Cercone

1032

 

Incremental reminding: The case-based

E)ynamic gating in vision

elaboration and interpretation of complex problem

  • E. Postma, H. van den Herik,

situations

P.Hudson

1038

  • B. Slator,R. Bareiss

1122

Understanding detective stories

  • I. Pratt, L. Xu, I. Leudar

1044

Imagery as process representation in problem solving

Y. Qin, H. Simon

1050

What does a system need to know to understand a user's plans?

  • B. Raskutti, I. Zukerman

Calculating breadth of knowledge L.Rau

1056

1062

Integrating causal learning rules with

backpropagation in PDS networks

R. Sumida

1128

Fuzzy evidential logic: A model of causality for

commonsense reasoning

R.Sun

1134

Exemplar competition: A variation on category

learning in the competition model

  • R. Taraban, J. Palacios

1140

A view of diagnostic reasoning as a memory-

directed task

R.Tumer

1146

Defining the aaion selection problem

T.TVrrell

1152

Analogical versus rule-based classification

  • W. Wattenmaker, H. McQuaid,

S. Schwertz.

1158

A simple recurrent network model of serial

conditioning: Implications for temporal event

representation

M.Young

1164

The figural effect and a graphical algorithm for

syllogistic reasoning

  • P. Yule, K. Stenning

1170

A computational best-examples model

J.Zhang

1176

Index.

1183

The Fourteenth Annual Conference of the Cognitive Science Society

Conference Chair

John K. Kruschke Department of Psychology and Cognitive Science Program Indiana University, Bloomington, IN 47405

Steering Committee Indiana University, Cognitive Science Program David Chalmers, Center for Research on Concepts and Cognition

J. Michael Dunn, Philosophy Michael Gasser, Computer Science and Linguistics Douglas Hofstadter, Center for Research on Concepts and Cognition

David

Leake, Computer Science

David Pisoni, Psychology Robert Port, Computer Science and Linguistics

Richard Shiffrin, Psychology Timothy van Gelder, Philosophy

Local Arrangements: Candace Shertzer, Cognitive Science Program

Officers of the Cognitive Science Society James L. McClelland, President

1988 - 1996

Geoffrey Hinton 1986 - 1992 David Rumelhart Dedre Centner James Greeno Walter Kintsch Steve Kosslyn George Lakoff Philip Johnson-Laird Wendy Lehnert Janet Kolodner Kurt VanLehn

1986 - 1993 1987 - 1993 1987 - 1993 1988 - 1994 1989 - 1995 1989 - 1995 1990 - 1996 1990 - 1996 1991 - 1997 1991 - 1997

Ex Officio Board Members Martin Ringle, Executive Editor, Cognitive Science

1986 -

Alan Lesgold, Secretary/Treasurer (2nd Term)

1988 - 1994

This conference is supported in part by funds from the Indiana University College of Arts and Sciences, the Indiana University Research and University Graduate School, and the Indiana University Cognitive Science Program.

xui

Preface

The 1992 Conference of the Cognitive Science Society was held at Indiana University in Bloomington,

sponsored in part by the Indiana University Cognitive Science Program, the Indiana University College of Arts

and Sciences, and the Indiana University Research and University Graduate School.

Of the 284 papers submitted to this year's Conference, 113 were accepted for oral presentation and 87 for

poster presentation. Acceptances were based on reviewers' ratings and on the capacity of the 3-day program.

The efforts of the reviewen are gratefully acknowledged; their names appear on the next page.

A paper by Yuri Mashintsev, Moscow State University, entitled "Eliciting implicit theory of personality through

processing questionnaires with Boltzmann machine", was accepted for poster presentation, but because of the

slow mail between the US A and the former USSR, a camera-ready copy of the paper was not received by press

time.

The Conference also featured 7 plenary talks and 9 symposia, the texts of which are not included in these

Proceedings. Plenary speakers and talk titles were as follows: Elizabeth Bates (University of California, San

Diego), Crosslinguistic studies of language breakdown in aphasia; Daniel Dennett (Tufts University), Problems

with some models of consciousness; Martha Farah (Carnegie-Mellon University), Is an object an objea?

Neuropsychological evidence for domain-specificity in visual object recognition; Douglas Hofstadter (Indiana

University), The centrality of analogy-making in human cognition; John Holland (University of Michigan),

Must learning precede cognition?; Richard Shiffrin (Indiana University), Memory representation, storage, and

retrieval; Michael Turvey (University of Connecticut), Ecological foundations of cognition. Symposium topics

and organizers were as follows: Timothy van Gelder (Indiana University) and Beth Preston (University of

Georgia), Representation: Wh o needs it?; Rik Belew (University of California, San Diego), Oomputational

models of evolution as tools for cognitive science; Caroline Palmer (The Ohio State University) and Allen

Winold (Indiana University), Dynamic processes in music cognition; Geo^ey Bingham (Indiana University)

and Bruce Kay (Brown University), Dynamics in the control and coordination of action; David Leake (Indiana

University) and Ashwin Ra m (Georgia Institute of Technology), Goal-driven learning; Mary Jo Rattermann

(Hampshire College), Similarity and representation in early cognitive development; K. Jon Barwise (Indiana

University), Reasoning and visual representations; David Pisoni (Indiana University) and Robert Peterson

(Indiana University), Speech perception and spoken language processing; Douglas HoCstadter (Indiana

University) and Melanie Mitchell (University of Michigan), Analogy, high-level perception, and categorization.

A list of symposium speakers' names was not available at press time. The banquet speaker was Bemhard

Flury, Indiana University Department

of Mathematics.

XV

REVIEWER S

Robert Allen

Jordan Graf man

Michael Pazzani

Richard Alterman John Bamden

Richard Granger Kristian Hammond

Herbert Pick David Plaut

K.

Jon Barwise

Reid Hastie

Kim Plunkett

William Bechtel

Geoffrey Hinton

Robert Port

Thomas Bever

Steven Hirtle

Bruce Porter

Dorrit Billman

Douglas Hofstadter

Mitchell Rabinowitz

Geoffrey Bingham Daniel Bobrow Gary Bradshaw Bruce Britton Mark Burstein Richard Catrambone David Chalmers Yves Chauvin Catherine Clement Axel Cleeremans Gregg Collins

Keith Holyoak William Hoyer Edwin Hutchins Susan Kemper Janet Kolodner Kenneth Kotovsky Bruce Krulwich John Kruschke John Laird David Leake Jill Fain Lehman

Ashwin Ram Michael Ranney William Rapaport Michael Redmond Brian Reiser Scott Robertson Paul Rosenbloom Brian Ross Roger Schank Walter Schneider Alberto Segre

Cynthia Connine Paul Cooper Kenneth Deffenbacher

Wendy Lehnert Daniel Levine Clayton Lewis

Colleen Seifert Martin Sereno Stuart Shapiro

S.

Farnham Diggory

Robert Lindsay

Lokendra Shastri

Stephanie Doane Charles E>olan Carolyn Drake Kevin Dunbar Kurt Eiselt Martha Evens Brian Falkenhainer Martha Farah Paul Feltovitch R. James Firby Ken Forbus Bob French Susan Gamsey Michael Gasser Dedre Gentner Morton Gemsbacher Randy Gobbel Ashok Goel Richard Golden Robert Goldstone Art Graesser

Steven Lytinen Colin MacLeod Pattie Maes James Martin Michael Masson Richard Mayer James McClelland Doug Medin Melanie Mitchell Michael Mozer Paul Munro Gregory Murphy Sheldon Nicholl Donald Norman Gregg Oden Stellan Ohlsson Gary Olson Judith Orasanu Alice O Toole Christopher Owens Caroline Palmer Vimla Patel

Richard Shiffrin Jeff Shrager Peter Slezak Steven Small Linda Smith Elliot Soloway Mark St.John Michael Swain David Swinney Katia Sycara David Touretzky David Townsend John Tsotsos Elise Turner Roy Turner Timothy van Gelder DeLiang Wang Ed Wisniewski Dekai W u Ivan Yaniv Ingrid Zukerman

XVI

TALK S

W h y

Intelligent Systems Should Ge t Depressed Occasionally an d Appropriately

Charles Webster

Intelligent Systems Program

University of Pittsburgh

Pittsburgh. P A 15261

Abstract

Some researchers suggest that depression may be

adaptive. For example, depression may provide an

opportunity to assess our capabilities, learn from

past failures, u-igger personal change, and allocate

activity away from futile goals. There are a variety

of signature phenomena associated with depression,

such as stable, global, and internal styles of failure

explanation, a cognitive loop of failure-related

rumination, lowered self-esteem and self-efficacy, and

increased negative generalization and depressive

realism. DEPlanner is presented, a simulated agent

that adapts to failure in a simulated environment and

exhibits eight targeted signature phenomena of

depression.

Introduction

Some types of depression in response to personal

failure may be adaptive. Taylor (1989, p. 225)

suggests that depression serves as a reality check, an

opportunity to take "realistic stock of what one is and

where one is going" and make "accurate assessment of

his or her capabilities." Williams et al. (1988, p.

183) suggest that depression facilitates coping with

long-term problems by "strategic access to previous

problem-solving attempts." Flach (1974) and Gut

(1989) argue that depression is a normal response to

personal failure, and is necessary for personal change.

Nesse (1991) hypothesizes that decreased mood reduces

activity in siuiations where effort will not be rewarded,

allocating energy away from bad investments. If

depression can serve an adaptive function, then perh^)s

a simulated agent can be designed in such a way as to

adapt to failure while exhibiting signature phenomena

of depression. DEPlanner, based on ideas discussed in

Webster et al. (1988), is such an adaptive agent.

This research was supported by the University of

Pittsburgh Alzheimer Disease Research Center.

Overview of DEPlanne r Simulation

DEPlanner is a nonlinear planner (Chapman, 1987)

that generates sequences of instantiated STRIPS-style

operators in response to initial and goal states. Since

representing the social environment of a human is so

complex, and since something like depression may

be

generally useful for adapting to some kinds of

environmental change, planning and learning occur in

a Blocksworld micro-world. DEPlanner's environment

is a description of a set of blocks stacked on each other

and scattered among locations on a table. Goals are

conjunctions of targeted block configurations. A plan

is a sequence of block movements. Successes have

positive utilities, and failures negative utilities.

Interactions between DEPlanner and its environment

occur through the use of a problem generator that

produces a stream of randomly constructed

Blocksworld problems. The problem generator

periodically changes internal rules that determine

whether one block can be stacked on another block or

location. A plan violating such a rule results in

failure, and a deduction from DEPlanner's score.

DEPlanner's task is to adapt in such a way as to

maximize total utility accrued during interactions with

the problem generator.

/declarative^

vJCnowledge^

Isodic

now ed

/''ProceduraTN

VjOTowledge^

Information Flow in DEPlanner

Fig. 1

Like Soar (Rosenbloom et al. 1991) DEPlanner

relies on declarative, procedural, and episodic

representations. Fig. 1 depicts the flow of

information among them. Declarative knowledge

corresponds to representations of which block can be

stacked on which block or location. (ASSUM E

(STACKONABL E B ?X)) means that block A can be

stacked anywhere. Fig. 2 illustrates a plan operator.

The "SENSE" preconditions match against the problem

description provided by the problem generator, while

the "ASSUME " precondition matchs against

DEPlanner's revisable assumptions in declarative

memory. Thus, changing declarative assumptions can

result in the production of different plans.

#S(OP

J4AMEMOVE

;move

:VARS (?X ?Y ?Z)

:x from y to z

:PRE(X>NDS

((ASSUME (STACKONABLE ?X ?Z))

;inatches revisable assumption

(SENSE (CLEAR ?X));matches problem description

(SENSE (CLEAR ?Z)):

"

(SENSE (ON ?X?Y))); "

:ADD-UST ((SENSE (ON ?X ?Z))

(SENSE (CLEAR ?Y))

:DEL-UST ((SENSE (ON ?X ?Y))

(SENSE (CLEAR ?Z))))

DEPlanner Operator

Fig. 2

Procedural knowledge consists of plans like

((MOVE B LOCI ) (MOV E A B)) (along with

appropriate preconditions and effects). Episodic

knowledge is based on a time stamp, problem initial

state and goals, the plan used, and outcome (success or

failure).

DEPlanner's declarative knowledge changes in

response to patterns of failure and success observed in

episodic knowledge (A in Fig. I). Procedural

knowledge arises from the application of declarative

knowledge (B in Fig. 1). creating a practice effect

(DEPlanner becomes faster with experience). In order

to change declarative knowledge on the basis of

experience, the results of using procedural knowledge

are recorded in the form of episodic knowledge (C in

Fig. 1). When failures in episodic memory trigger

changes in declarative memory, procedural knowledge

must also change in order to remain a compiled

version of declarative knowledge. DEPlanner's

depression is the computationally expensive process of

consulting episodic knowledge (retrieving past

failures), modifying declarative knowledge (explaining

patterns of failure), and compiling new procedural

knowledge (creating new plans). Specialized control

mechanisms suppress depression unless, and until,

future benefits of depression are estimated to outweigh

current costs.

II <M:

I ClwmiStlcUMlHg

G*Mr(«i PrtkltM

Pn-«1illn|P)MiT

Mtri n*n I

 

EXKUtePlM

PUnF«l1iir>7

ExildnFilliirt

r*liM«i4*M CxtliMHtii

Ctlciilato CUD

flW Pwt«v«7

I

ClMtmte»u»n»tl»M I

Rttrltvt CfTKM CHMm It

1

C*Mrik wv Plan 2

^

X

S«*nPUn2

C>4M*« Lan?

Interaction Between DEPlanner (left) and Problem Generator (right)

Fig. 3

The flowchart in Fig. 3 depicts the flow of

information and control between, and within,

DEPlanner and its problem generator. The problem

generator produces a problem (1-4, Fig. 3).

DEPlanner returns a plan (5-7, Fig. 3). The problem

generator determines whether the plan is a success (8-

9, Fig. 3). And DEPlanner, if certain criteria are met,

gets "depressed" (10-18. Fig. 3).

DEPlanner gradually increases its average response

speed by reusing plans (5, Fig. 3). Plans can be

stored (7. Fig. 3) after generation (6, Fig. 3) in

response to problems posed by the problem generator,

or they can be regenerated an stored (16 and 17, Fig. 3)

assumptions, DEPlanner uses Eq. 1 to calculate the

Expected Utility of Depression (EUD).

If EU D is

positive for forgoing current opportunities (PGI times

PAU), but preparing for future opportunities (PSI

times PMU), then DEPlanner gets "depressed." If the

E U D is negative then DEPlanner "shrags off* failure

and moves on.

PMU = average Plan Marginal Utility per unit time

{difference in utility accrued per unit time

between having and not having a plan)

PSI = average Plan Soundness Interval

(length of time over which a plan will prove

under new assumptions in response to memories of

useful)

old problems (15, Fig. 3). The problem generator

P A U = average Plan Achievable Utility per unit time

periodically changes internal rules about which block

{utility accrued per unit time from current

can be stacked on which block or location (2 and 3,

repertoire of plans)

Fig. 3). If DEPlanner finds a set of assumptions that

explains the observed pattern of recent failure (10 and

11, Fig. 3) it calculates a numerical heuristic Expected

Utility of Depression (EUD) (12, Fig. 3). If EU D is

positive (13, Fig. 3) then DEPlanner changes its

assumptions (14, Fig.3) and retrieves previous

problem-solving episodes now predicted to result in

failure (15, Fig. 3). For

each episode (18, Fig. 3)

DEPlanner generates a new plan under new

assumptions (16, Fig. 3) and stores it (17, Fig. 3).

Signature Phenomena of Depression

as Design Constraints on DEPlanne r

I used eight signature phenomena, culled from the

research literature on depression, as a set of design

constraints on the construction of DEPlanner: stable,

global, and internal styles of failure explanation

(Peterson & Seligman, 1984); a cognitive loop of

failure-related rumination (Ingram, 1984); lowered self-

esteem (Musson & Alloy, 1988) and self-efficacy

(Bandura, 1977; Rehm, 1988); and increased negative

generalization (Beck et al. 1979) and depressive

realism (Alloy & Abramson, 1988). M y task was

constructing a simulated agent that adapts to changes

in a simulated environment, and exhibits as many

signature phenomena of depression as possible.

The first three phenomena (stability, globality, and

intemality) concern the kinds of events that are most

likely to trigger a depression. It is not failure itself,

but rather explanation for failure that serves as a

trigger. Failure atuibutions that are stable (continuing

to be true in the future), global (affecting many

important goals), and internal (having been avoidable)

are most likely to precipitate depression (Peterson &

Seligman, 1984). Each dimension can be mapped to a

different aspect of the formula (Eq. I) used to trigger

or suppress DEPlanner's analog of adaptive deiH'ession.

If DEPlanner finds a parsimonious explanation for a

recent pattern of failures, in terms of a change in

PGI = average Plan Generation Interval

{length of time required to generate apian)

E U D = (PMU * PSI) - (PAU * PGI) Eq.l

The parameters in Eq. 1 can be associated with the

three dimensions of failure attribution in such a way

that increasing stability, globality, or intemality

correspond to increasing EU D and likelihood of

depression. The more stable an environment, the

longer a period of time over which the benefits of a

precomputed plan can be amortized. Therefore

stability is associated with the PSI parameter. The

more globally important an environmental change, the

larger the number of important affected goals. Both

P M U and PA U can be associated with globality (PMU

positively, and PA U negatively) because they both

vary with the total utility affected (in PMU's case) or

unaffected (in PAU's case). As for intemality, an

internal failure attribution is one in which we believe

that we could have avoided failure if we had pursued an

alternative course of action. In DEPlanner's case an

alternative course of action is an alternative plan. The

less effort or time required to generate an alternative

plan, the more internal should be the attribution.

Conversely, impossible plans, which have infinitely

long generation times, should result in large PGI's.

Therefore intemality is mapped

to 1.0 / PGI.

When DEPlanner's depression ensues, five more

signature phenomena appear: the cognitive loop of

failure-related rumination, lowered self-esteem and self-

efficacy, and increased negative generalization and

depressive realism.

Depressed people experience a cognitive loop of

failure-related thoughts that distract them from

normally enjoyable activities (Ingram, 1984). During

the cognitive loop people think about past failures,

construct possible explanations, and consider future

implications. When the cognitive loop recedes,

people often make important changes in their lives

(such as changing their expectations or focusing on

different goals). In DEPlanner the cognitive loop

consists of retrieving failures from a memory of past

problem-solving episodes (accounting for William et

al.'s (1988, p. 183) "strategic access to previous

problem-solving attempts."), searching for failure

explanations (accounting for Taylor's (1989, p. 225)

"accurate assessment of his or her capabilities"), and

generating new plans under new assumptions. Thus,

DEPlanner's analog of the cognitive loop is at the core

of DEPlanner's adaptive ability.

William James defined self-esteem to be total

successes divided by total pretensions (James, 1890).

In DEPlanner's case, self-esteem is the total utility of

achievable goals divided by the total utility of all

goals in episodic memory. Self-esteem ranges from

0.0 (no goals are possible) to 1.0 (all goals are

possible).

Self-efficacy, roughly the subjective probability that

a goal can be accomplished, is thought to drop during

depression (Rehm, 1988). In DEPlanner self-efficacy

is defined in a similar fashion to self-esteem, except

that utility is not taken into account: total number of

attainable goals divided by total number of goals.

People who are vulnerable to depression are more

likely to arrive at negative self-deprecating

generalizations in response to small setbacks (Beck et

al. 1979). For example a mother whose child

complains of a cold breakfast may decide she is a bad

mother, rather than she is not a perfect cook.

DEPlanner's assumptions, about which block can be

stacked on which block or location, are indexed in a

hierarchy, with more general assumptions toward the

top, and more specific assumptions toward the

bottom. (Retracting (ASSUM E (STACKONABL E B

?X)) is more general than retracting (ASSUM E

(STACKONABL E B A)) because retracting

(ASSUM E (STACKONABL E B ?X)) effectively

retracts all assumptions beneath it in the hierarchy.)

Given a set of successful and failing past problem-

solving attempts, DEPlanner finds positions in the

generalization hierarchy that predict the pattern of

successes and failures. The more general a failure

attribution, the more "ASSUME " preconditions are

unsatisfied, and the longer DEPlanner spends

precomputing new plans. Thus DEPlanner's behavior

is consistent with the correlation between human

negative generalization and depression. DEPlanner's

measure of negative generalization is the number of

retracted assumptions divided by the total number of

possible assumptions.

The depressive realism phenomenon is particularly

problematic for other theories of depression. In some

ways, mildly depressed people appear to be more

accurate information processors than nondepressed

people (Alloy &. Abramson, 1988). For example,

nondepressed people seem to over-estimate their

chances of success at a variety of tasks, while

depressed people are relatively more accurate. A

simple measure of DEPlanner's depressive realism is

total actual goal utility (according to correct

assumptions available in the problem generator)

divided by total predicted goal utility (according to

DEPlanner's own assumptions and plans). During

depression, DEPlanner's measure of depressive realism

increases because DEPlanner's assumptions become

more accurate.

Comparin g DEPlanner's Behavior to

Signature Phenomen a of Depression

Cognitive modellers aspire to detailed statistical

comparisons between the behavior of their simulations

and human subjects. However many simulations,

especially those that are the first to model a

psychological process, do well to qualitatively match a

set of signature behaviors obtain^ from the research

literature. DEPlanner is in this latter category. In

order to confirm that DEPlanner behaves according to

its intended design, an experiment (Fig. 4) was

devised.

The value of each number, or change in value, is

not important for the purpose of this qualitative

assessment. It is the order and direction of change that

is relevant. Each of the two decision boxes

corresponds to an experimental manipulation. Box 1

corresponds to the presence or absence of a stable,

global, and internal environmental change. Box 2

corresponds to the presence or absence of DEPlanner's

analog of adaptive depression.

DEPlanner's problem generator has "knobs" that

influence the frequency of change (stability), the

frequency and total utility of goals affected (globality),

and length of time required to construct plans

(internality). In order to set these parameters,

DEPlanner's assumptions, plans, and past problem-

solving attempts are accessed, but not changed. The

resulting statistics are used to determine which rules to

change in the iM^oblem generator, and how often.

Beneath the paths leaving box 1 are tables of

numbers representing stability (St), globality (Gl), and

internality (In). They are higher on the upper path

(see "Compare 1", Fig. 4), consistent with higher

stability, globaUty, and internality. Here the Expected

Utility of Dqjression (EUD) is positive and triggers

DEPlanner's adaptive depression (see "Compare 2",

Fig. 4). In the low stability, globality, and internality

condition EU D is negative (but almost positive) and

DEPlanner suppresses its depression.

Box 2 corresponds to enabling (upper path) or

disabling (lower path) DEPlanner's adaptive

depression. Where adaptive depression is enabled, we

see the signature phenomena of decreased self-esteem

(Es) and self-efficacy (Ef), and increased negative

generalization (Ge) and depressive realism (Re) (see

"Compare 3", Fig. 4). Self-esteem and self-efficacy

Interwl, Global and stable

Depression Begins

Depression Ends

 

Explanation for Failure)

Es 0.527

0.927

Compares

  • 0.527 0.927

 
  • 0.667 0.667

  • 1.000 1.000

Total Utilltg

 

-15010.01

 
  • 1.000 1.000

  • 1.000 Y»s

1.000

  • 0.000 EUD-II2023

0.000

  • 1.000 Depression Enabled

0.733

 

Compare 4

1242.0

0.162

1.000

Compare 2

0.157

If 1.000

Internal,

G> 0.000

f

Global,

R» 0.733

Total UtIlltM

Stable

»| 4250.01

Failures?

Es

Es

  • 1.000 1.000

Ef

t.OOO

Ef

1.000

  • 0.000 0.000

Gt

R*

  • 1.000 0.945

 

• EUD'1-16.691

St

201.1

CI

0.083

In

0.028

-Compare I

Effects of Environmental and Architectural Change on

Fig. 4

Signature Phenomena

have identical values because the current version of

DEPlanner's problem generator assigns the same

utilities to all problem-solving attempts (S.O for

successes and -5.0 for failures). These numbers would

diverge if different problems had different utilities.

Self-esteem and self-efficacy drop because the total

utility of unthreatened goals drops. Negative

generalization increases in this particular case because

DEPlanner retracts (ASSUM E (STACKONABL E B

?X)), effectively retracting a large number of

assumptions. Correcting DEPlanner's assumptions

increases DEPlanner's measure of depression realism.

Eventually all plans that can be precomputed, have

been precomputed, and self-esteem and self-efficacy

move almost, but not quite, back to their original

values. They fail to reach 1.0 because some plans are

impossible in light of the change in assumptions.

For instance, a goal set that includes

(ON B A) is

impossible because (ASSUME (STACKONABL E B

A) has been retracted.

In either condition, dejMiession-enabled or depression-

disabled, the DEPlanner simulation continues until the

next change in the environment, whereupon the value

of total utility accrued is calculated (see "Compare 4",

Fig. 4). When DEPlanner is allowed to get

"depressed" the total utility is 5010.0. When

DEPlanner's depression is disabled the total utility is

  • 4250.0. Thus DEPlanner achieves a 18 per cent

higher total utility in the depression-enabled condition.

DEPlanner exhibits analogs for signature phenomena

of depression and shows how they may be generated

by a functionally useful mechanism.

Conclusio n

To become a better model of depression, DEPlanner

requires redesigning. Relevant issues include:

  • 1. A more plausible model of goal-driven problem

solving and learning should replace the nonlinear

planning algorithm. Computer models of knowledge

acquisition and compilation in human students will be

relevant (e.g., VanLehn et al. 1992).

  • 2. The domain should be changed to reflect

concerns of depressed people. Representing changing

career and family circumstances will require a more

expressive knowledge representation.

  • 3. DEPlanner should be better motivated in terms

of concepts like self-esteem and self-schema, as well

as general theories of emotional information

processing.

  • 4. While DEPlanner is not a model of abnormally

triggered, sustained, w intense depression, DEPlanner

might be "broken" or "lesioned" in order to provide

such an account.

  • 5. The effects of external interventions, analogous

to cognitive therapy for depression (Beck et al. 1979),

should be modeled with respect to reducing likelihood,

intensity, or duration of DEPlanner's depression.

  • 6. Other kinds of cognitive dysfunction, such as

Alzheimer Disease, can masquerade as depression, and

visa versa (Caine, 81; Merriam et al. 1988). Perhaps

DEPlanner can be broken in different ways in order to

account for different but related syndromes.

7.

DEPlanner uses a heuristic to modulate its

Ingram, R. (1984). Toward an Information-Processing

cognitive loop. A sophisticated control regime should

be based on an explicitly normative metareasoning

approach (Horvitz et al. 1991).

  • 8. DEPlanner should be situated with respect to

existing systems in the space of cognitive

architectures, such as case-based reasoners and

explanation-based learners.

  • 9. Extensive sensitivity analysis will be required to

understand the complex interactions among

DEPlanner's parameters, and between DEPlanner and

its environment.

Analysis of Depression. Cognitive Therapy and

Research, 8,443-478.

James, W . (1890). The Principles of Psychology.

Vol. 1. Henry Holt & Co.

Merriam, A., Aronson, K., Gaston, P., Wey, S., &

Katz, I. (1988). The Psychiatric Symptoms of

Alzheimer's Disease. Journal of the American

Geriatric Society, 36,7-12.

Musson, R. & Alloy, L. (1988). Depression and Self-

Directed Attention. In L. Alloy (Ed.), Cognitive

Processes in Depression. New York: Guilford

Nevertheless, the current version of DEPlanner is

Press.

consistent with a relatively large set of signature

Nesse, R. (1991). What is Mood for? Psycoloquy,

phenomena associated with depression, and the

Vol. 2, Issue 9.2.

hypothesis that some forms of depression may be

adaptive. Rudimentary computational mechanisms

can be assembled into a model of depression that

explains a large set of previously unrelated signature

phenomena.

Acknowledgements

Peterson, C. & Seligman, M. (1984). Causal

Explanations as a Risk Factor for Depression:

Theory and Evidence. Psychological Review, 91,

347-374.

Rehm, L. (1988). Self-Management and Cognitive

Processes in Depression. In L. Alloy (Ed.),

Cognitive Processes in Depression. New Yoric:

Guilford Press.

Rosenbloom, P., Newell, A. & Laird, J. (1991).

I thank Kurt VanLehn, Greg Cooper, Gordon Banks,

Toward the Knowledge Level In Soar: The Role

Johanna Moore, Richmond Thomason, Martha Pollack, and two anonymous reviewers, for helpful

of the Architecture in the Use of Knowledge. In

K. VanLehn (Ed.) Architectures for Intelligence,

comments and useful criticisms. Thanks also to Steve

Hillsdale, New Jersey: Lawrence Erlbaum.

Small for access to computing facilities at the

Cognitive Modelling Labcvatory.

Taylor, S. (1989). Positive Illusions: Creative Self-

Deception and the Healthy Mind. Basic Books.

VanLehn, K., Jones, R., & Chi, M. (1992). A Model

References

of the Self-Explanation Effect. In M. Posner

(Ed.), Journal of the Learning Sciences, 2, 1,1-

59.

Alloy, L. & Abramson, L. (1988). Depressive

Realism: Four Theoretical Perspectives. In L.

Alloy (Ed.), Cognitive Processes in Depression.

New Yoric: Guilford Press.

Bandura, A. (1977), Self-Efficacy: Toward a Unifying

Theory of Behavioral Change. Psychological

Review, 84, 191-215.

Beck, A., Rush, A., Shaw, B. & Emery, G. (1979).

Cognitive Therapy for Depression: A Treatment

Manual. New Yoifc Guilford Press.

Caine, E. (1981). Pseudodementia: Current Concepts

and Future Directions. Archives of General

Psychiatry, 38,1359-1364.

Chapman, D. (1987). Planning for Conjunctive

Goals, Artificial Intelligence, 32, 333-377.

Flach, F. (1974). The Secret Strength of Depression.

New York: Lippincott.

Gut, E, (1989). Productive and Unproductive

Depression: Success or Failure of a Vital Process.

New York: Basic Books.

Horvitz. E., Cooper. G. & Heckerman, D. (1989).

Reflection and Action Under Scarce Resources:

Webster. C . Glass, R. «& Banks, G. (1988). A

Computational Model of Reactive Depression.

Proc. lOth Annual Conf. Cognitive Science

Society, Montreal, Canada: Lawrence Erlbaum.

Williams, J., Watts.

F..

MacLeod. C. & Mathews, A.

(1988). Cognitive Psychology and Emotional

Disorders. New York: John Wiley & Sons.

Theoretical Principles and Empirical Study. Uth

Int. Journal Conference on Artificial Intelligence.

Reasonin g abou t Performanc e

Intention s

Michael Freed, Bruce Krulwich, Lawrence Birnbaum, and Gregg Collins Northwestern University, Tiie Institute for the Learning Sciences

  • 1890 Maple Avenue; Evanston, Illinois 60201

Electronic mail: {freed,krulwich,birnbaum,collins}(Dils.nwu.edu

Introduction

For an agent to find and repair the faults that underly

a planning failure, it must be able to reason about the

intended behavior of its planning and decision-making

mechanisms. Representations of intended decision-

making behaviors, which we refer to as intentions,

provide a basis for generating testable hypotheses about

the source of a failure in the absence of complete

information about its cause. Moreover, intentions

provide measures by which beneficial modifications to

cognitive machinery can be differentiated from harmful

or useless ones. This paper presents several examples of

these intentions and discusses how they may be used to

extend the range of circumstances in which agents can

learn.

Since our claim concerns the representations needed to

learn from failure, we describe in section 2 a situation in

which an agent should be able to learn, and discuss some

of the obstacles to doing so. In section 3, we present

the idea of intentions and show, using the example

from the previous section, how they may be used to

extend the range of circumstances in which machines

can learn. Section 4 presents an implementation of our

theory of failure-driven learning and discusses the role

of intentions in several stages of the learning process.

Finally, in section 5, we relate our claim to the work of

other researchers and summarize the argument that the

intended behaviors of decision-making mechanisms must

be represented explicitly in order to learn from failure.

quickly, a driver may notice many collision threats and

choose not to enter. Alternately, s/he may employ some

plan that avoids or neutralizes all of these threats, and

proceed to enter traffic.

For the task of entering the flow of traffic, the category

collision-threat is useful for constraining the set of plans

which may be safely employed. However, the novice

in our example may have been employing an overbroad

definition of collision-threat. As a result, the set of

seemingly safe plans will be overconstrained, thereby

increasing the average amount of time before some plan

seems safe. It is reasonable to blame the category

definition for overconstraining the set of plans which can

be used to enter traffic safely, thereby preventing timely

action. Learning from the failure thus means narrowing

the faulty definition of collision-threat.

There may of course be many ways in which the

definition of collision-threat could be usefully narrowed.

For instance, a novice may learn that other drivers

tend to stay in

their own lanes between exits; thus,

the possibility that a vehicle will make a sudden and

inexplicable lane change should not be considered a

collision-threat.

Tracing the delay to its underlying cause

In our example, the novice's failure to expeditiously

enter the flow of traffic stems from a failure to generate

a safe plan for entry.

This failure stems, in turn,

An every-day example

Consider the situation of a novice driver attempting to

traverse a crowded and confusing road feature such as

the rotary shown in figure 1. Lacking experience, the

novice may end up waiting longer than more experienced

drivers before entering the flow of traffic. It is not

unreasonable to suppose that the novice notices the

undesirably long wait, perhaps with the assistance

of impatient drivers waiting behind. The situation

implicates some shortcoming in the novice's driving skill,

and so warrants an attempt to learn some improvement.

What should be learned?

Drivers can identify collision-threats such as road

obstacles and moving vehicles, and must take them into

account in deciding how and when to enter traffic. For

instance, if at some moment traffic is heavy and moving

Figure 1: Flow of traffic on a rotary

from having noticed more threats than the novice

could plan to counter or avoid. The novice can get

better at entering traffic by learning to more accurately

differentiate between threats and non-threats.

One way that this diagnosis could proceed is as

follows: Suppose the novice receives feedback indicating

that a particular perceived threat was not in fact a

threat over some time interval. The circumstances

during this interval could then be reviewed to show

that a safe plan to enter the rotary would have been

discovered had the misperceived threat (a "near miss")

been ignored. Finally, the misperceived

threat could be

used to narrow the collision-threat category definition.

The preceding process is unrealistic for several

reasons. First, it assumes that the agent can receive

specific feedback regwding a decision that was made

incorrectly, as opposed to an action that was incorrectly

taken. Moreover, since the novice had no way to know

a priori which of its decisions should be monitored,

it would require that the novice receive

continuous

feedback on each of its decisions.

Finally, it is highly

implausible that the novice will be able to receive

detailed feedback about particular perceived threats,

since novices typically lack the knowledge necessary to

evaluate such situations post-hoc [Fitts, 1964; Starkes

and Deakin, 1985).

A

more realistic diagnostic process begins as the driver

of zmother car honks his horn and thereby leads the

novice to question why no safe traffic entry plans have

been generated. Hypotheses are developed as to why

this might be the case, including for example that either

the planning mechanism is inadequate, or its set of

collision-threats is faulty. These hypotheses enable the

novice to seek specific feedback through experimentation

or advice.

Reasoning about agent intentions

As previously argued in [Collins et al., 1991], it is

is useful to divide a planner into components, each

responsible for a task-independent function such as

detecting threats or selecting among competing plans.

The effectiveness of a component depends on whether

it performs its function quickly, how reliably it attends

to relevant input items, whether it avoids pathological

interactions with other components, and so on. W e call

these measures of component effectiveness intentions.

Whe n an agent implemented as a component

architecture learns, one or more modifications are made

to its constituent components. Component intentions

(see figure 2 for examples) are a measure of the

value of a modification. For instance, a modification

is valuable if it helps avoid pathological interactions

between components and harmful to the extent that

it aggravates such interactions or introduces new ones.

Because the choice of what to learn depends on factors

measured by intentions, intentions should be explicitly

represented so that learning processes can reason about

them. T o understand this point, consider again the

Within components

Inter-component

No f&lse positives

Don't flood other components

No false negatives

Don't monopolize resources

Efficient computation

Don't reproduce computation

Output values within

Don't focus on areas of

acceptable ranges

ultimate irrelevance

Don't be a bottleneck

Figure 2: Sample planner component intentions

rotary example of section .

Recall that the principal difficulty in speeding traffic-

entry performance lay in locating a sample misperceived

threat (a near miss) on which to base a refinement of

the collision threat classification. The most effective

solution was apparently to hypothesize that the category

is overbroad and to seek out specific feedback on future

threat-detection performance through experimentation

or advice. In the absence of an instance of a

misclassification, this solution requires some other basis

for formulating the fault hypothesis. Intentions provide

this alternate basis.

Consider the component intention: avoid flooding

another component with output. The negation of this

intention represents a situation in which a component is

producing too much

some output-definer

output; this in turn indicates that

(category)

ma y be too broad and

that fixing the problem requires locating a false positive

with which to narrow the category.

A second role of intentions is to evaluate candidate

component modification for preventing failure recur-

rence. In our example, drastically narrowing the

collision threat category (so that no collision threats

are generated) would solve the problem of flooding the

traffic-entry planner with output, but would lead to

catastrophically faulty plans. The basis for rejecting this

candidate modification is the intention that the threat-

detection component's output be free of false-positives

Implementation and second example

In this section, we describe a system, castle', which

implements important aspects of our theory. The

system, which operates in the domain of chess, detects

situations that are contrary to its expectations, and

responds to these expectation failures by repairing the

faulty planner components which were responsible for

the failure. W e view this learning as a knowledge-based

process, in which the system uses knowledge of its own

planning components to learn from events which led to

expectation failures. More specifically, the system must

reason along different dimensions of intentionality to

determine what repairs should be made to its planning

'Castle stands for Concocting i4bstracl Strategies

Tlirough Learning from i^xpectation-failures.

Events

Situation

 

^

"

^

Focusing

Detection

New hreats

Figure 3: Incremental threat detection

(d«f-brule focus-new-source

(focus focus-Boved-piece ?player

(•ove Tplayer ?BOve-type ?piece ?locl ?loc2)

(world-at-ti«e Ttime))

<•

(Bov«-to-make (Bove ?player ?prev-Bove-type

?piece ?old-loc ?locl)

?player ?goal (1- ?tima)) )

Figure 4: Focusing on new moves by a moved piece

rules.

The knowledge necessary for this repair process is

expressed in the form of a planner self-model, which

is used to diagnose and repair expectation failures

[Davis, 1984; deKleer and Williams, 1987; Simmons,

1988]. More specifically, when the expectation fails, the

system first examines an explicit justification structure

which encodes the the reasoning which let to its belief

in the incorrect expectation [deKleer el o/., 1977;

Doyle, 1979]. This justification is used to isolate the

faulty components of its architecture, each of which

implements a particular sub-task in the decision-making

process [Collins et al, 1991; Krulwich, 1991]. It then

uses a specification of the faulty components to guide the

learning of new rules in response to the failure [Krulwich,

1992]. Each of these information sources must explicitly

reference the planner's intentions.

Detection focusing

A central cognitive task in which CASTLE engages is that

of noticing threats and opportunities as they become

available [Collins et al., 1991]. Rather than recomputing

these at each turn, CASTLE maintains a set of active

threats and opportunities that is updated over time. To

accomplish this incremental threat detection, the system

uses a detection focusing component, which consists

of focus rules that specify the areas in which new

threats may have been enabled. Then, a separate threat

detection component, consisting of rules for noticing

specific types of threats, detects the threats that have

in fact been enabled. This relationship between the two

components is shown pictorially in figure 3. A sample

focus rule is shown in figure 4. This rule embodies

the system's knowledge that the most recently moved

piece, in its new location, may be a source of new

threats. Another focus rule, not shown, specifies that

the more recently moved piece can also be a target of

newly enabled attacks. Using focus rules such as these,

the actual threat detector rules will only be invoked on

areas of the board which can potentially contain new

threats.

Focusing intentions

What intentions does the system have regarding its

detection focusing component? The primary intention

that the system has is that there not be any newly

enabled threats that are not within the scope of

the bindings generated by the focusing component.

This condition is clearly necessary for the incremental

detection scheme to work. A more subtle intention is

that the focusing component not generate too many

bindings in which threats do not exist. If this intention

is not met, the detection component will be invoked

more than is necessary, and

in the extreme case the

entire point of the detection focusing is lost. Clearly

the savings gained by only applying the threat detection

rules in constrained ways (and not over the entire board)

must be greater than the cost of applying the focusing

rules. This will not be the case if the constraints given

by the focusing component are too weak. It will also not

be the the case if the computational cost of applying the

focusing component is too high.

Another planner intention regarding the focusing

component is that the division of the tasks shown in

figure 3 be enforced. This means that the system siiould

not incorporate information about different types of

threats into the focusing rules.

Discovered attacks

To see how CASTLE uses representations of planner

intentions in learning, let's first see an example of

CASTLE enforcing its simplest intention, that there be

no false negatives of its focusing component. Consider,

in particular, the example of discovered attacks in chess,

in which the movement of one piece opens a line of attack

for another piece. Novices often fall prey to such attacks,

not because they fail to understand the mechanism of

the threat (i.e., the way in which the piece can move to

Intention

Application to focusing

No false negatives

Don't let a threat be enabled

without detectors being invoked

No false positives

Detectors not over-applied

Efficiency

Incremental scheme shouldn't

be less efficient than

brute-force

No redundancy

Don't encode information about

specific threats in focus rules

Figure 5: Planner intentions in detection focusing

Vj

(c)

(b)

Figure 6: Example: Opponent (white) to move

(daf-brul* leamed-focus-method25

(focus leam«d-focu>-Dethod25 ?player

(move ?play«r (capture ?taken-piece)

?taking-piece (loc ?rowl ?coll)

(loc ?row2 ?col2)) (world-at-time ?tiine2))

<-

(and (move-to-make

(move ?other-player move ?interm-piece

(loc ?r-interm ?c-interm)

(loc ?r-other ?c-other))

?player ?goal ?timel)

(loc-on-line ?r-lnterm ?c-interm

?roBl ?coll ?row2 ?col2)

(at-loc ?player ?taking-piece

(loc ?roBl ?coll)

(-

gen-tiBe2.24 2)) ))

Figure 7: Learned focus rule for discovered attacks

make the capture), but rather because they simply fail

to consider new threats arising from pieces other than

the one just moved. The same is true of castle if it is

equipped only with the two focus rules described above.

The example in figure 6 shows the system falling prey

to a discovered attack due to its lack of a necessary

detection focusing rule. In the situation shown in

figure 6(a), the opponent advances its pawn and thereby

enables an attack by its bishop on the computer's rook.

When the system updates its set of active threats and

opportunities, its threat focusing rules will enable it to

detect its own ability to attack the opponent's pawn,

but it will not detect the threat to its rook. Because

of this, when faced with the situation in figure 6(b),

the computer will capture the opponent's pawn instead

of rescuing its own rook, and it will expect that the

opponent's response will be to execute the attack which

it believes to the only one available, namely to capture

the computer's pawn. Then, in the situation shown in

figure 6(c), when the opponent captures the computer's

rook, the system has the task of diagnosing and learning

from its failure to detect the threat which the opponent

executed.

Learning from the failure

To diagnose the failure, CASTLE examines an explicit

justification structure [deKleer ei ai, 1977; Doyle, 1979],

which record how the planner's expectation was inferred

from the rules that constitute its decision-making

mechanisms, in conjunction with the policies and under-

lying assumptions which it has adopted. Diagnosing

the failure then involves "backing up" through the

justification structure, recursively explaining the failure

in terms of faulty rule antecedents [Smith et ai, 1985;

Simmons, 1988; Birnbaum et ai, 1990; Collins et ai,

1991]. This diagnosis process will "bottom out" by

faulting either an incorrect planner rule or an incorrect

assumption that underlies the planning mechanism. In

our example, the fault lies in an assumption that the

planner could enforce its first intention regarding its

focusing component, that it would generate bindings

for all enabled threats. Castle concludes from this

that its set of focusing rules is incomplete and must be

augmented.

To construct the new rule, castle retrieves a com-

ponent performance specification for each component.

These performance specifications, a form of planner

self-knowledge, describe the correct behavior of each

component. The specification of the detection focusing

component says roughly that the focusing componeni

will generate bindings that include any capture that is

enabled by a given move. This specification enables

CASTLE to focus on the details of the example that are

relevant to the component being repaired, by serving as

an explanation-based learning target concept [Krulwich,

1991; Krulwich, 1992]. After retrieving the specification,

CASTLE invokes its deductive inference engine to

construct an explanation of why the possible capture

of the rook should have been in the set of constraints

generated by the focusing component. This explanation

says roughly that the opponent's move should have

been generated by the focusing component, because the

opponent's previous move enabled the attack, because it

was on a square between the bishop

and the rook, and

there were no other pieces along the line of attack, and

emptying the line of attack is an enabling condition for

the capture to be made. Castle then uses explanation-

based learning techniques [Mitchell et ai, 1986; DeJong

and Mooney, 1986] to generalize this explanation and

to construct a new detection focusing rule shown in

figure 7.

Back to intentionality

In the

example of learning discovered attacks, tiie

system is able to correct the failure of the detection

10

focusing component to generate bindings that included

the new attack. The learning process that we

described involves the construction of a new rule to

enforce the intention that the focus component not

generate any false negatives. Suppose, however, that

instead of adding the rule shown in figure 7, the

system's learning component added a focusing rule

that returned completely unconstrained bindings. This

would cause the detection rules to be applied to all

board positions in computing the newly enabled threats.

This would, of course, enforce the system's intention

to have no false negatives, and would also satisfy

the component specification, because all threats that

could possibly be enabled are by definition within the

unconstrained bindings. However, the whole purpose

of the incremental threat detection scheme (figure 3)

would be undermined, because not only will the system

apply the threat detection rules over the entire board,

but it will then proceed to integrate the threats that it

finds into the set of previously available threats, which

is clearly a waste of time. In short, this is a violation of

the system's no-false-posUives intention.

Unfortunately, while the violation of the no-false-

negatives intention could be easily noticed by observing

an enabled threat that was not in the focus bindings, it is

much more difficult to notice the failure of the no-false-

positives intention. In our example, after the opponent

moves his pawn in figure 6(b), three classes of bindings

constraints should be generated: threats by the moved

piece at its new location, which are generated by the

rule

in figure 4, threats against the moved piece at its

neiu location, and threats through the square vacated by

the moved piece, generated by the learned discovered

attacks rule in figure 7. Tw o of these in fact reflect

new threats that have been enabled, but one of them,

threats by the moved piece, in fact do not reflect any

new threats.

W e can see that this must be the case, because the

division of labor between the focusing and detection

components requires that no information about the

types of threats themselves be present in the focus rules.

Since the focusing component is only determining where

to look for new threats, it is clear that there will be

times when a correct place to look for new threats will

not in fact contain any.

This complicates the problem of detecting false

positives, since there is no direct test to determine

whether the focus component was generating false

positives. One approach would be for this intention to

only come into play when the system is learning new

focus rules. The no-false-positives intention could be

used to force the learning mechanism to generate the

most specific possible rule. Of course, this approach

does not allow the system to reason explicitly about this

intention.

Another approach would be to have the system

generate expectations about the performance of its

component that do not relate directly to false positives,

but that are good indicators of the system's false

positive rate. As we discussed above, the intention

not to have any false positives is in service of system

efficiency, because the design of the incremental threat

detection scheme is based on the focusing component's

sufficiently narrowing down the scope of the detection

rule application. It follows from this that the system

could monitor the computational effort spent on the

detection focusing and compare it with the savings in

threat detector application. If this tradeoff turned out

not to be worthwhile, the system

false positive rate in more

detail.

could examine its

This is similar to

the example in which the driver was unable to enter

the intersection, causing him to examine his detection

mechanism for sources of false threats. This requires

that the system be able to make utility judgements

about different tradeoffs between false positives, false

negatives, and efficiency.

A similar learning process could be invoked if the

system noticed that too much time was being spent

considering pointless opportunities. This could arise

if the system found that it was spending too much

time considering pawn captures that were always being

discarded by the plan selection component.

If this

were the case, the system could infer that its focusing

component should be further constrained not to generate

bindings for captures of pawns.

Discussion

W e have shown that reasoning about the faults under-

lying a planning failure requires that an agent explicitly

represent performance intentions which describe the

desired behavior of its components. When one of its

components is faulty, the agent must reason explicitly

about its intentions to diagnose the failure and make a

repair which is to its overall benefit.

This paper presents several examples of single-

component intentions, such as completeness, soundness,

and efficiency, as well as intercomponent intentions

such as avoiding flooding and competition for global

resources. W e have discussed aspects of the learning

process which require explicit reasoning about these

intentions, thereby extending the range of concepts an

agent can learn, and allowing it to learn in circumstances

in which it could not otherwise learn. This work thus

builds on previous research in failure-driven acquisition

of new planning knowledge [Hammond, 1989; Birnbaum

et a/., 1990; Collins et ai, 1991].

Previous research has dealt with several of the issues

we have discussed. Minton [1988] discussed the need

for learned planner rules to be sensitive to the global

efficiency of the system. Our work builds on this idea

by explicitly modeling and a variety of such intentions.

Hunter's system [1989] recisoned about shortcomings

in its diagnostic knowledge and explicitly modeled the

intentions involved in that task to guide learning.

Similarly, Cox and Ra m [l99l] have modeled several

intentions of the case retrieval process for use in the task

11

of understanding. Our research extends these ideas to

model intentions for a more genereil problem solver, as

well {IS modeling the intercomponent intentions. Others

have used representations of the system's intentions for

plauining [Jones, 1991] and understanding (Ram, 1989].

Our previous research has involved extending our

model of planning and decision-making to include a

variety of tasks and components, all in the domun

of competitive geimes. To date we have developed

models of threat detection, counterplanning, schema

application, goal regression, lookahead search, and

execution scheduling. Future research will elucidate

the breadth of planner intentions, and will demonstrate

the benefits of explicitly representing them for use in

learning.

Acknowledgements: Thanks to Matt Brand, Bill

Ferguson, Eric Jones, and Louise Pryoi for many discussions

on the research presented here. This work was supported

in part by the Air Force Office of Scientific Research under

grant number AFOSR-91-0341-DEF, and by the Defense

Advanced Research Projects Agency, monitored by the Office

of Naval Research under contract N-00014-91-J-4092. The

Institute for the Learning Sciences was established in 1989

with the support of Andersen Consulting, part of The Arthur

Andersen Worldwide Organization. The Institute receives

additional support from Ameritech, an Institute Partner, and

from IBM.

References

(Birnbaum

et al., 1990] L. Birnbaum, G. Collins, M. Freed,

and B. Krulwich. Model-based diagnosis of planning

failures. In Proceedings of the Eighth National Conference

on Artificial Intelligence, pages 318-323, Boston, MA,

1990.

[Collins et al., 199l] G. Collins, L. Birnbaum, B. Krulwich,

and M. Freed. Plan debugging in an intentional

system. In Proceedings of the Twelfth International

Joint Conference on Artificial Intelligence, pages 353-358,

Sydney, Australia, 1991.

 

[Davis, 1984] R.

Davis.

Diagnostic reasoning based on

structure and

function: Paths of interaction and the

locality principle. Artificial Intelligence, 24(l-3):347-410,

1984.

[DeJong and Mooney, 1986] G. DeJong and R. Mooney.

Explanation-based learning: An alternative view.

Machine Learning, 1:145-176, January 1986.

[deKleer and Williams, 1987] J. deKleer and B.C. Williams.

Diagnosing multiple faults.

Artificial Intelligence,

32(1):97-129, April 1987.

[deKleer et al., 1977] J. deKleer, J. Doyle, G.L. Steele, and

G.J. Sussman. Explicit control of reasoning. SIGPLAN

Notices, 12(8), 1977.

 

[Doyle, 1979] J. Doyle.

A truth maintenance system.

Artificial Intelligence, 12(3):231-272, 1979.

[Fitts, 1964] P.M. Fitts. Perceptual-motor skill learning. In

Categories of human learning. Academic Press, New York,

1964.

Thirteenth Annual Conference of The Cognitive Science

Society, pages 132-136, Chicago, IL, 1991.

[Hammond, 1989] K. Hammond.

Case-based planning:

Viewing planning as a memory task. Academic Press, San

Diego, CA, 1989.

[Hunter, 1989] L. Hunter. Knowledge-acquisition planning:

Gaining expertise through experience. PhD thesis, Yale

University, 1989.

[Jones, 1991] E. Jones. The flexible use of abstract knowledge

in planning. PhD thesis, Yale University, 1991.

[Krulwich, 199l] B. Krulwich. Determining what to learn in

a multi-component planning system. In Proceedings of the

Thirteenth Annual Conference of The Cognitive Science

Society, pages 102-107, Chicago, IL, 1991.

[Krulwich, 1992] B. Krulwich. Learning New Methods for

Multiple Cognitive Tasks. PhD thesis, Northwestern

University, Institute for the Learning Sciences, 1992. (in

preparation).

[Minton, 1988] S. Minton. Learning Effective Search Control

Knowledge: An Explanation-Based Approach. PhD thesis,

Carnegie Mellon University, 1988.

[Mitchell et al., 1986] T.M. Mitchell, R.M. Keller, and S.T.

Kedar-Cabelli.

Explanation-based generalization: A

unifying view. Machine Learning, 1(1), January 1986.

[Ram and Cox, 199l] A. Ram and M. Cox. Using introspec-

tive reasoning to select learning strategies. In Proceedings

of the First International Workshop on Multistrategy

Learning, 1991.

[Ram, 1989] A. Ram. Question-driven Understanding: An

integrated theory of story understanding, memory, and

learning. PhD thesis, Yale University, 1989.

[Simmons, 1988] R.G. Simmons. A theory of debugging

plans and interpretations. In Proceedings of the Seventh

National Conference on Artificial Intelligence, St. Paul,

MN., 1988.

[Smith et al., 1985] Reid

G.

Smith,

HA.

Winston,

T. Mitchell, and B.G. Buchanan. Representation and use

of explicit justifications for knowledge base refinement. In

Proc. IJCAI-85, Los Angeles, August 1985.

[Starkes and Deakin, 1985] J.L. Starkes and J.M. Deakin.

Perception in sport: A cognitive approach to skilled

performance. In Cognitive Sport Psychology. Sports

Science Associates, Lansing, N.Y., 1985.

[Freed, 1991] M. Freed. Learning strategic concepts from

experience: A seven-stage process. In Proceedings of the

12

Precondition s an d appropriatenes s conditions ^

Timothy M. Converse and Kristian J. Hammond Artificial Intelligence Laboratory Computer Science Department University of Chicago 1100 E. 58th St. Chicago, IL 60637 (312) 702-8584

con ver se@cs. uchi cago. edu

Abstract

Classical plan preconditions implicitly play a

dual role, both documenting the facts neces-

sary for a plan to be sound and listing the

conditions under which it should be used. As

the closed-world assumption is relaxed these

two roles begin to diverge, particularly when

attempts are made to use plans in situations

other than those for which they were origi-

nally constructed. Rosenschein and Kaelbling

exploit one aspect of the divergence by sug-

gesting that some logiceil preconditions can be

considered in the design phase of building an

agent, but "compiled away" so that the agent

need not explicitly consider them [Rosenschein

and Kaelbling, 1986]. W e suggest an alterna-

tive view whereby an agent can explicitly rea-

son and learn about which conditions are the

best cues for employing standard plans, and

discuss the idea in the context of the Runner

project.

Introduction

Plan execution and the role of

preconditions

The closed-world assumption of classical planning

allowed the assumption that plans could have a

small and explicit set of preconditions which, if

true, would ensure that the plan worked. The most

straightforward application of this idea to the exe-

cution of such plans is that an executor should know

or verify the truth of the preconditions before start-

ing the plan. This is particularly unproblematic if,

•This work was supported in part by AFOSR grant

number AFOSR-91-0112, DARPA contract number

F30602-91-C-0028 monitored by Rome Laboratories,

DARPA contract number N00014-91-J-4092 monitored

by the Office of Naval Research, Office of Naval Re-

search grant number N00014-91-J-1185

in addition to a closed world assumption, the plan

is being generated for exactly the circumstances in

which it is to be used; in that case, everything that

is known about the state of the world can be taken

into account during the synthesis of the plan, and

the process of construction can (given the closed-

world assumption) itself guarantee the soundness

of the plan.

In recent years, greater awareness of the in-

tractability of generative planning [Chapman,

1985], coupled with greater concern about time

pressure in activity, has led to attempts to amor-

tize the cost of planning over repeated instances

of activity, either by "pre-compiling" action de-

cisions [Rosenschein and Kaelbling, 1986, Drum-

mond, 1989], or by re-using the fruits of previous

planning attempts [Hammond, 1989].

At the same time it has been widely acknowl-

edged that the set of logical preconditions for plans

in many real-world situations is effectively inflnite.

This been called the "qualification problem" (de-

fined variously in [McCarthy, 1977, Shoam, 1986,

Ginsberg and Smith, 1987]). That is, given any

attempt at enumeration of logical statements that

need to be true for a given plan to be guaranteed

to work, it is usually possible to come up with an

additional potential fact that would render the plan

unworkable.

Early planning research tried to confront this di-

rectly using large numbers of frame axioms. Most

generative planning systems since [Pikes and Nils-

son, 1971] have used the more optimistic and

tractable STRIPS assumption that primitive ac-

tions can have associated lists of the facts that are

changed by applying them.

Precondition sets for classical plans implicitly

play a dual role. They

1. describe the initial conditions under which the

plan as described can be expected to be sound

13

(under certain assumptions, and because the

facts were used in the plan's construction), and

2. describe the facts that an executor should know

to be true before beginning execution of the plan.

Let us keep the existing term of precondition for

the first sort of fact above, and use the term appro-

priateness condition for the second sort.

Appropriateness conditions

Under assumptions of perfect knowledge and a

closed world, there is little divergence between pre-

conditions and appropriateness conditions. When

these assumptions are relaxed, however, there are

several different ways in which the divergence can

become important in plan execution and reuse:

• A precondition can effectively be "always true".

This means that the plan may depend upon it for

correctness, but an executor will never run into

trouble by not worrying about its truth value.

This sort of fact should not be an "appropriate-

ness condition", since consideration of it cannot

help in the decision whether to use the plan.

• A precondition may

be almost always true, and it

may be difficult to know or check in advance. If

the consequences of an abortive attempt at per-

forming the plan are not too severe, then this sort

of fact should not be an appropriateness condi-

tion, since the utility of knowing its truth is out-

weighed by the cost of acquiring the knowledge.

• A precondition may be intermittently true, but

may be easily ''subgoaled on" in execution,

and achieved if false. (This of course depends

strongly on representation of plans, and how flex-

ible the execution is.) To the extent this can be

handled in "execution", the condition should not

be an appropriateness condition, since whether

or not the condition holds the plan is likely to

succeed.

• A particular condition may not be a precondition

per se, but may be evidence that the plan will

be particularly easy to perform, or will produce

results that are preferable to the usual default

plan for the relevant goals. This should be an

appropriateness condition, even though it is not

a precondition.

Action nets and appropriateness

conditions

Rosenschein and Kaelbling noted the first possibil-

ity in the above list, that some preconditions might

be "always true", and realized that, while such facts

may need to be explicitly considered in the design

of an agent for some domain and task, there is

no

rejison why the agent itself need consider them.

[Rosenschein and Kaelbling, 1986]. Such facts can

essentially be "compiled away" in the design of an

agent that will behave appropriately.

To summarize the argument so far:

• The set of facts that an agent should consider

before embarking on a given plan is interest-

ingly different from both the (possibly infinite)

set of facts that need to be true for the plan

to work, and the set of facts explicitly used in

the plan's construction. This is true particularly

when de novo plan construction is impossible or

too costly, and plans must be reused.

• One possible approach that recognizes this is to

explicitly design an agent so that it only con-

siders the conditions that are actually relevant

for action, either by hand-crafting its decision

procedure, or by a mixture of hand-crafting and

clever compilation of declarative specifications,

as in Rosenschein and Kaelbling's work.

And the point we want to make (which will oc-

cupy us

for the rest of the paper):

There is a large potential middle ground between

an approach that requires explicit reasoning about

all preconditions on the one hand, and approaches

that compile in any needed reasoning of that sort

in advance. In particular, even if an agent is as-

sumed to have a largely immutable library of plans

and behaviors that will determine its competence,

there is still room for learning the particular appro-

priateness conditions that govern when to invoke

particular plans.

An example

To make these distinctions clearer, let's look at a

common sense example: the task of making but-

tered toast, in a well-equipped kitchen, with an

electric toaster.

If we start to enumerate the preconditions that

we can think of that are associated with this task,

the most available ones have to do with the re-

sources we would normally worry about it conjunc-

tion with it: possessing bread, possessing butter.

Others that come to mind might have to do with

available time, instruments (the toaster, a knife), or

knowledge about these things (do we know where to

find a knife?). As we strain to think of things that

are not part of the concerns eissociated with the

plan, we might think of possible "external" prob-

lems like an interruption in electric service. Finally,

imaginable "preconditions" start to be explicitly

counterfactual; what if gravity no

longer operated,

or heat conduction worked in a different way?

In practice, when deciding whether to make

toast, one is probably aware of only the first few

14