Transcript ETHOS - centria - Universidade Nova de Lisboa

Neuro-Psychological
Social Theorizing and Simulation
with the Computational
Multi-Agent System ETHOS
Jorge P. F. Simão
[email protected]
Luís Moniz Pereira
[email protected]
Centro de Inteligência Artificial – CENTRIA
Universidade Nova de Lisboa
Congresso em Neurociências Cognitivas
Évora, 10-12 Novembro 2003
Talk Outline
 ETHOS Simulation Framework Design Goals
 Current Agent Based Model Simulation Frameworks
 ETHOS Simulation Framework Overview
 Human Mate Choice: case study I
 The Cultural Evolution of Preferences: case study II
 Conclusions and Future Work
ETHOS Simulation Framework:
design goals (1)
ETHOS is an Object-Oriented Simulation Framework


Implemented in Java
Download from: http://centria.di.fct.unl.pt/~jsimao/ethos
Gives Computational Support for Social Theory Building to:

Reify in software useful theoretical constructs
(shared and/or plausible)

Experiment with variations of theoretical constructs

Re-use theoretical constructs

Easily (re-)implement and extend a large array of models

Easily explore the model and theory spaces of possibilities
ETHOS Simulation Framework:
design goals (2)
General Computational Requirements of Frameworks:
 Expressiveness and Flexibility
 Extensibility and Modifiability
 Transparency
 Performance
 Scalability
 Portability
 Ease of Use
Current ABM Simulation Frameworks
• Swarm, RePast, Ascape
+ Good Support for General Computational Service
- Lack Specific Support for Social Theory Building
• PS-I
+ Support for Social Theory
- Targeted only to a Specific Set of Mid-RangeTheories:
constructivist identity theories
• Evo
+ Support for Evolutionary Discovery of Behaviour Strategies
- Limited Plausible set of Mechanisms (Evolutionary Programming)
• Starlog, AgentSheets
+ Easy to Use
- Mostly Limited to “Toy” Models
• Sugarscape, Consumat, . . . (and other highly parameterized models)
+ Interesting Case Studies
- Not a Generic Simulation Framework
ETHOS Simulation Framework Overview (1)
Physical Environment Structure:
– Space is the unit of spatial layout; provides
topological arrangement of Site
– Site have any number of Body
– Body represents a physical entity:
(Human) Agent, Resource, Organization
– World as aggregation of Space
ETHOS Simulation Framework Overview (2)
ETHOS Simulation Framework Overview (3)
(Human) Agent Structure:





Agent = Genome + Visible Attributes +
Social Networks + Control
Genome is a set of inherited traits
Attr is a visible agent attribute (e.g. sex, quality)
Tie is a connection between agents in a
SocialNet
Selector objects used as reusable selection criteria
mechanism: SocialNet, . . .
Control is the behaviour control mechanism, on the
basis of the Task Env
ETHOS Simulation Framework Overview (4)
Ethos’s
Class
Hierarchy:
ETHOS Simulation Framework Overview (5)
Event Scheduling and Population Structure:
 Population are aggregations of Body; coordinates their activities
 Population can contain other Population; composite structure
 Population also place-holder for operations at aggregate level
 Each Space contains a top level Population to add other Population
 Population set associated with a Space
 Selectable Scheduling Policy:
• single or multi-phase
• syncronous or asyncronous
• fixed or variable time, per agent
ETHOS Simulation Framework Overview (6)
ETHOS’s GUI look-and-feel
Human Mate Choice: case study I
Emergent population-level patterns in human
mating systems:
 Assortative Mating
• Couples highly correlated in attractiveness (0.4 - 0.6)
• (But) Individuals prefer more attractive partners
• Matching hypothesis?
 Distribution of age at mating time
• Right-skewed bell-curve (robust cross-culturally)
• Explanation ?
Previous Models of Mate Choice
• S. Kalick and T. Hamilton
”The matching hypothesis re-examined”,
in Journal of Personality and Social Psychology, 4:(51), 1986.
• P. Todd and G. Miller
”From pride and prejudice to persuasion: satisficing in mate search”,
in Simple Heuristics that Make Us Smart, Oxford UP, 1999.
• Rufus Johnstone
”The tactics of mutual mate choice and competitive search”,
in Behavioral Ecology and Sociobiology, 1:(40), 1997.
Courtship Based Model: social ecology (1)
Parameter
P
L
µ, 2
Y
K
Description
population size/2
reproductive lifetime
quality distribution
meeting rate
courtship time
Value(s)
50
200
10, 4
0.1 – 1.0
5 - 50
Courtship Based Model: social ecology (2)
• Fixed population size (2 x P) and sex ratio (50%)
• (Quasi) normal distribution of qualities:
mean µ and variance 2 (0 < Qmin ≤ q ≤ Qmax).
• Meeting rate Y (0.1 – 1.0). Discrete time steps.
• List of alternatives: one has ”special status” -- the date.
• (Age depended) Courtship time K before mating; current time ct .
• Limited reproductive life time L (> K) = 200.
Individual mate choice strategies
Fitness function:
F(qm, t) = qm · (L - t)/L
Decision rules:
• Partner switching (risk insensitive):
F(qa, t + Ki) > F(qd, t + Ki - ct)
• Partner acceptance/aspiration level setting:
q*i new = q*i old · (1 - ) +  · qj · 
• Aspiration level dropping with time:
tmax =  · (L – t)/L · (1 – qb / q*)
• Age dependent minimum courtship time:
Ki = K · (1 – ti / L)
Simulation results (1)
Robust Empirically Validated Results:
• Mean correlation of qualities in mated pairs: 0.6 - 0.8
• Mean number of alternatives seen before settling with the last date: 2 - 10
• Percentage of individuals in the population that are able to mate: ≥ 90%
Simulation results (2)
Distribution of age at mating (marriage) time
-- right-skewed bell-curve --
Conclusions from Model
 More realistic results than previous models
 Model assumptions more psychologically
plausible and more relevant to humans
 Future work:
• Other mating systems: Serial Monogamy, and Divorce
• More complex preferences: structure and dynamics
The Cultural Evolution of Preferences:
case study II
What do miniskirts, afro haircuts, and body tattoos
have in common?
• They are all forms of body accessories that have had a characteristic
fashion-like career.
• They emerge out of obscurity and spread through a population very fast.
• Shortly after they have reached their maximum popularity:
• vanish again from the cultural landscape
• sometimes surge again long after
Current explanations:
– Simmel Effect
– Externalities
– Information cascades
– Decay of value
Our proposal: Individual Conditioning drives collective behaviour
An agent-based model of fashion:
emergence (1)
Agent attributes: ai = < qi , ti , v0i , v1i >.
Model pseudo-code:
repeat (T) {
for all agent {
update trait values ;
switch to most preferred trait ;
}
}
An agent-based model of fashion:
emergence (2)
Trait value update rules:
v1i (t) = v1i (t-1) ·  + 1/N ·
v0i (t) = v0i (t-1) ·  + 1/N ·

qj · (1- )

qj · (1- )
aj: ajMi ∧tj=1
aj: ajMi ∧tj=0
Parameter settings:
Parameter
P
N
E


D
Description
population size
number of role models
assortment
1 - learning rate
standard deviation
delay
Value(s)
50
5
4
0.2
2
4
Note
small sample
small
r  0.75
fast learning
cognitive or material
Simulation Results (1)
Bit map of trait usage
across time (D = 4):
Frequency of
trait usage across
time (D = 4):
Simulation Results (2): deterministic model
Bit map of trait usage
across time (D = 4)
with deterministic
selection of model:
Notes:
• Small deterministic neighborhood changes behaviour of model
• Propagation of trait usage / avoidance is more regular
• General caveat: spatial analogies of social strata can bias results
Simulation Results (3): sensitivity analyses
Bit map of trait usage
across time (D = 10):
Bit map of trait usage
across time (D = 0):
Conclusions from Model
 Fashion like collective behaviour can emerge
from individual conditioning
 Model is very sensitive to delay parameter D
 Complex networks of traits may have more
complex dynamics
 Models with multi-valued trait may also have
more complex dynamics
Conclusions and Future Work
 ABM Software support for Social Theory Building
• Is Feasible: Identifies Key Foundational Abstractions
• Is Useful: Simplifies Theory Building, Comparison, and Testing
• Is Desirable: Contributes to the Unification of the Social Sciences
 Further Developments in ETHOS
• (Re-) Implement Additional Models
• Refine and Add Abstractions (if and as needed)
• Make Software Publicly Available