SIMULATING HUMAN BEHAVIOR: Agent Based Models for ... · AGENT-BASED MODELS • Allow heterogeneity...
Transcript of 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
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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.
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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)
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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
PREVIOUS SIMULATION STUDIES OF GROUP PROCESSES
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
9Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
PREVIOUS SIMULATION STUDIES OF GROUP PROCESSES
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
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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
15Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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)
16Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
CONSTRUCTION OF THE FITNESS LANDSCAPE
We add N agents to the landscape
Participant agents can move around a fitness landscape.
17Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
MODEL
The Landscape isConstructed
Optimal Solution Point18
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
MODEL
Participants
N participants areplaced randomly onthe fitness landscape
19Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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)
20Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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).
21Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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
Duncan A. Robertson and L. Alberto Franco | IFORS 2017 Conference
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