Cognitive Models for Problem Gambling Marvin Schiller and Fernand - - PowerPoint PPT Presentation

Cognitive Models for Problem Gambling

Marvin Schiller and Fernand Gobet

Centre for the Study of Expertise Brunel University, UK

2011 London Workshop on Problem Gambling – Theory and (Best) Practice

Overview

• Towards Cognitive Models for Problem Gambling
• Modelling using CHREST

– Iowa Gambling Task – Near Wins

• Discussion

Problem Gambling

Various fields provide theories/hypotheses/data on PG

• Psychiatric & Biological Theories: Interactions between neural,

genetic and social factors; comorbidity (anxiety, depression, alcoholism)

• Psychological Theories: Conditioning, personality, cognitive biases,

e.g. gambler’s fallacy, reinforcement history (near wins, early wins), emotion as a modulator

• Integrative Theories: pathways models (e.g. Blaszczynski and Nower,

2002, Sharpe, 2002)

Motivation

• Cognitive Modelling

– Uses precise formal techniques (e.g. equation systems, computer simulations) to model/explain cognitive processes and behaviour (qualitatively & quantitatively) – Fosters theory development and coherence – Generates testable predictions

• Proposed Approach

– Models three levels (neural, cognitive, integrative) – Relates PG to established models of perception, learning and decision making

CHREST

• A cognitive architecture with a particular focus on visual

processing and memory

• Computer implementation allows one to develop, run and test

models for cognitive processes

• Based on chunking theory and template theory
• Models of human learning and expertise in various domains,

including:

– Board games: chess and awale – Language acquisition in children – Physics: creation of diagrams for electric circuits

Components of PG Model

STMs

BAR

Perceptual Input Mechanism

Anticipation = perception + LTM retrieval

Simulation of Environment Attention Memory Prediction Action Selection

Decision Making Component

CHREST LTM

• Discrimination

network

• Emotion tags +

association learning

Current Modeling

• Ensures fundamental results are adequately modeled:

– Iowa Gambling Task – Near wins prolong slot machine gambling (e.g. Cote et al., 2003)

Iowa Gambling Task

• Models for reward and decision making:

– Each deck evaluated, evaluations updated with each selection (via association/reinforcement learning) – Exploration vs. evaluation determined e.g. by Boltzmann exploration A B C D

+100 +100 +50 +50 +100 +100 +50 +50 +100 /-150 +100 +50/

• 50

+50

Expected value/trial: -25 +25 (adapted from Bechara et al., 2000)

Current Modelling

+100/-150 +100 +100 +100/

• 150

100 150 A B C D

…

100

A B C D

STM LTM

Perception

Current Modelling

+100/-150 +100 +100 +100/

• 150

100 150 A B C D

…

100 100 100 50 50 150 20 150 25

A +100

• 150

A B C D

STM A

+100/

• 150

∆V=α*(λ-V)

LTM

Perception

Choices in the Iowa Gambling Task

Healthy Patients

Selection of 100 cards

Slot Machine Gambling

• Addictive (cf. e.g. Griffiths et al., 1999)
• Persistently popular and highly

available

• Relatively easy to simulate
• Important revenue-generator

(cf. Ghezzi et al., 2000)

Slot Machine Modelling

+100

• 1

STM +100/-1 LTM

+100/-1 0/-1 0/-1 +100/-1 100 1 1 20 0.2 0.2

association learning

• Cote et al (2003): during a losing streak, a higher proportion of

near wins leads to more persistence

• Dependent variable: persistence in part 2

Near Wins Prolong Gambling

Bar

7 7

sequence including

• 9 wins
• 12 near wins
• 27 losses

sequence consisting of

• 25% near wins
• 75% plain losses

sequence consisting of

• 100% plain

losses Condition 1 (n=29) Condition 2 (n=30) Part 1 Part 2 Games played in part 2

Data from Cote et al (2003)

Near Wins Prolong Gambling (II)

• Tentative explanation: anticipation when recognising two

“nearly winning” symbols

Bar

7 7 0.1 0.2 0.1 0.4 0.1

Perspectives

• Modelling of further aspects of PG and their interactions

– Modulating effect of emotions on processing (and possibly, bias) – Investigating effect of early wins, further structural characteristics, and their interplay – Question: can systematic biases be learned – or sustained – via specific combinations of parameters?

• Connect the model to online (slot-machine) games to make

qualitative and quantitative predictions

Discussion

• Development of PG is a complex phenomenon on several

dimensions

• Cognitive models for PG are still lacking, despite benefits
• This work allows one to investigate the development of PG as a

phenomenon of learning, in particular implicit learning