SIMULATING HUMAN BEHAVIOR: Agent Based Models for ... · AGENT-BASED MODELS • Allow heterogeneity...

SIMULATING HUMAN BEHAVIOR: Agent Based Models for Simulating Human Behavior

IFORS 2017 CONFERENCEQUEBEC CITY, QC, 17 APRIL 2017

Duncan A. Robertson including work with L. Alberto FrancoLoughborough University, UK

INTRODUCTION

1

Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference

4

MODELLING APPROACHES

EXAMPLE APPLICATION:GROUP DECISION MAKING

2


NEED FOR CLOSURE

• A desire for unambiguous information, as opposed to uncertainty or unambiguity. (Webster & Kruglanski 1996)

• Goal oriented• High NClos individuals typically

are inclined to (Kruglanski & Fishman 2009):– attain closure as quickly as

possible, and maintain it for as long as possible;

– they achieve this by relying on past knowledge and avoiding new information.


AGENT-BASED MODELS

• Allow heterogeneity of agents (e.g. participants in a group process), moving on from representative individuals or aggregate measures: the state of each agent can be inspected at any time

• Allow study of the interactions between agents• Allow modelling of the temporal: study of the dynamics of a

system rather than just an equilibrium• ‘bottom up’ rather than ‘top down’ Modelling• Previous work:

• Larson (2007) N dimensional hill climbing• Rousseau and van der Veen (2006) cellular automata• Deffuant (2000)


PREVIOUS SIMULATION STUDIES OF GROUP PROCESSES

Larson (2007) N dimensional problem space (hill climbing algorithm)- Results: heterogeneous groups

(different searching / ‘flipset’ heuristics) produce better solutions

- Stylized heuristics – limited to ways that a relatively simple solution space is searched

- Little interaction between group members

8Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference


Rousseau and van der Veen (2005)- Model of the emergence of a shared identity- Limited repertoire of possible

outcomes- Actually a cellular automata

model – agents are confined to grid locations



Deffuant et al (2000) Continuous beliefs (rather than binary). Adjustment based on random meeting of agents who readjust when opinion difference is lower than a threshold d. Either random mixing or (percolation) on a grid to reflect a fixed social network (cf Schelling segregation model, Bak et al forest fire model)

10

OUR TWO APPROACHES:EXTERNAL & INTERNAL

1A


TWO APPROACHES TO THE PROBLEM

Two approaches to the problem of behavioural modelling:1 Individuals Moving Over

Congitive LandscapeDecision based on externalcomparison with next best neighbor

2 Individuals Updating their InformationDecision based on internal comparison with my next best alternative

12

APPROACH 1:INDIVIDUALS MOVING OVER COGNITIVE LANDSCAPE

DECISION BASED ON EXTERNAL COMPARISON WITHNEXT BEST NEIGHBOR

2A


OUR MODEL

Combines– Thresholds of Deffunant et al (2000)– Hill Climbing of Larson (2007)– Social network of both Rousseau and van der Veen (2006) and

Deffuant et al (2000) (2nd model) but:• Makes the social network endogenously constructed rather than

cellular automata

14

CONSTRUCTION OF THE FITNESS LANDSCAPE

The fitness landscape is created by adding M Gaussians at random positions

This is a fitness landscape for M = 1 – one central peak, and is the landscape we will use for our experiments


PARTICIPATOR AGENTS

N participants – agents in our agent-based model –are randomly positioned on a fitness landscape.

Fitness landscapes (Wright 1932) widely used in the evolutionary biology community. ‘Fitness’ reduces the output into the ‘height’ of the landscape.

Widely used in strategic management (Levinthal 1997 and many more recent publications)


CONSTRUCTION OF THE FITNESS LANDSCAPE

We add N agents to the landscape

Participant agents can move around a fitness landscape.


MODEL

The Landscape isConstructed

Optimal Solution Point18


MODEL

Participants

N participants areplaced randomly onthe fitness landscape


MODEL

Each participant will compare its height with that of its nearest neighbor. If it is within the NFClo threshold, then the participant reaches closure.If not within this threshold, the participant will search the landscape with a hill climbing random walk (they will take a random walk but will move only if their height increases)


MODEL: A DYNAMIC SOCIAL NETWORK

In effect we are using an agent-based model to construct a dynamic social network model (arrows show reference participant and whether participants have reached closure).


Difference inheight betweenAgent 1 and Agent 2

24

HEIGHT THRESHOLDS: ± THRESHOLDS

Agent 1’sheight threshold

Agent 1Agent 2

Agent 1’s height threshold > Difference in height between Agent 1 and its nearest neighbor⇒ Agent 1 reaches closure and stops searching

Agent 2 is Agent 1’s nearest neighbor

25

CONSTRUCTION OF THEDYNAMIC SOCIAL NETWORK

0

10

20

30

40

50

60

70

80

90

100

0 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 25 30 35 40 45 50 60 70 80 90

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

31

GROUPS HEIGHT AS A FUNCTION OFNFCLO THRESHOLD

Sigma^-1 = 2; random walk hill climbing; monomodal; plus thresholdData: Groups Model v35 160315 IFORS 2017 Experiment 3A-table

participants

threshold (note scale)

mea

n he

ight

of p

artic

ipan

ts

Medium performance at Low NFClo– High Variance with Group Size

High Performance at Medium NFClo – Low Variance

Low Performance

at High NFClo

APPROACH 2:INFORMATION FLOWS IN GROUPS

DECISION BASED ON INTERNAL COMPARISON WITHMY NEXT BEST ALTERNATIVE

2B


INDIVIDUAL INFORMATION

• Based on Stasser and Titus (1985) model of Hidden Profiles but adding a Need for Closure parameter for each participant

• Agents (participants) are trying to choose a candidate for a job. They have private information (known only to them) and public information (known to all participants)

• Over the course of discussion, this private information is discussed and becomes public.

36

PARTICIPANTS AND MEMES

Participants Information Memes

37

}Private

Public

Private

Closure when differencebetween my first choice

and second choice exceeds a threshold

PARTICIPANTS AND MEMES

Participants Information Memes

38

} Private

Public

Private

Closure when differencebetween my first choice

and second choice exceeds a threshold

39

ALTERNATIVE APPROACH:INDIVIDUAL AGENT DECISIONS

40

RESULTS: TIME TO COMPLETE TASK

0

100

200

300

400

0-9

10-1

920

-29

30-3

940

-49

50-5

960

-69

70-7

980

-89

90-9

910

0-10

911

0-11

912

0-12

913

0-13

914

0-14

915

0-15

916

0-16

917

0-17

918

0-18

919

0-19

920

0-20

921

0-21

922

0-22

923

0-23

924

0-24

925

0-25

926

0-26

927

0-27

928

0-28

929

0-29

9

freq

uenc

y

Time to Completion of Task where Hidden Profile Found

1

2

3

4

5

NFC Threshold

Altering Participants’ Need For Closure: Trade Off Between FasterCompletion of the Problem and Reduced Number of Hidden Profiles Found

QUESTION

• At the moment, discussions end when the next best alternative is more than a certain threshold from our currentdecision.

• But when should the discussion end in our model? At the moment, all information could potentially become public.

• How to model when conversations stop?

FRAMEWORK FOR ABM & BOR

Introducing a framework for ABM + BOR

43

‘Agent-Based Modeling and Behavioral Operational Research’,In: Kunc, Malpass, White Behavioral Operational Research: Theory, Methodology, and Practice, Palgrave Macmillanwww.duncanrobertson.com

SIMULATING HUMAN BEHAVIOR: Agent Based Models for Simulating Human Behavior

IFORS 2017 CONFERENCEQUEBEC CITY, QC, 17 APRIL 2017

Duncan A. Robertson including work with L. Alberto FrancoLoughborough University, UK

SIMULATING HUMAN BEHAVIOR: Agent Based Models for ... · AGENT-BASED MODELS • Allow heterogeneity...

Documents

Transcript of SIMULATING HUMAN BEHAVIOR: Agent Based Models for ... · AGENT-BASED MODELS • Allow heterogeneity...