Professional Documents
Culture Documents
Guven Demirel
Guven Demirel
MODELING
Güven Demirel
1
OUTLINE
2
OUTLINE (ctd.)
3
ABM Definitions
“The main concern and interest of Agent-Based Modeling is the
experimental study and modeling of agents in a common world of
compound and unpredictable effects of a population of agents in a
common world.”
(Castelfranchi, C., Simulating with Cognitive Agents: The Importance of Cognitive Agents, MABS, 1998)
4
Complexity Science & ABM
The modeling domain of ABM: complex systems.
Complexity Science: the science of modeling and
analysis of complex systems.
“Science of Emergence”
Examples of emergence in sand pile, el Farol bar
and firefly models
Unexpectedness and unpredictability
Focus on nonlinear interactions
Wholistic nature
(Ref.: Castelfranchi, C., Simulating with Cognitive Agents: The Importance of
Cognitive Emergence, MABS, 1998)
5
Complexity Science & ABM
Application Areas:
– Agent-Based Computational Economics
– Agent-Based Computational Demography
– Individual-Based Modeling in Ecology and Biology
– Political Science
– Anthropology
– Sociology
– Physics, etc.
(Ref.: Billari, C., F., et. al. Agent-Based Computational Modeling, 2006)
6
Complexity Science & ABM
Research Institutes:
– Santa Fe
– New England
– Brookings
7
Methodological Status of ABM
(Ref.: Axtell, R., Epstein, J. Growing Artificial Societies: Social Science From Bottom
Up,1996)
9
Methodological Status of ABM
Deductive or Inductive
– Deductive in the sense that behavioral rules are given
– Inductive in the sense that these rules can be adaptive and if so should be
analyzed
– Inductive in the sense that behavior is analyzed
(Ref.: Gilbert, N., Terna, P. How to build and use agent-based models in social science,1999
Gilbert N. Quality, Quantity and Third Way, 2001)
10
Building Blocks of ABM
Multiple Agents
– Individual Agent Layer
– Interactions Among Agents Layer
Environment
11
What is an Agent?
Perceive-Reason-Act Approach:
– perceive the environment, reason about its perceptions, and
act based on the reasoning (traditional AI).
13
More on Agents
Beliefs
Goals
Plans
Agent Function& Agent Program
Emergence of ABM from Distributed AI.
Agents as computational processes implemented on
a computer that have:
– Autonomy
– Social ability
– Reactivity
– Proactivity
14
Agent Characteristics
15
Types of Agents
16
Modeling of Simple Reflex Agents
17
Model-Based Reflex Agents
18
Goal-Based Agents
19
Utility-Based Agents
20
Learning Agents
21
Different Approaches to
Agents in ABM
KISS Principle
Complex Agents Perspective
Agent type depends on model purpose and domain.
Dominance of usage of collection of condition-action
rules in ABM.
More use of AI algorithms, techniques. For the
design of adaptive (learning) agents:
– Learning by Induction
– Reinforcement Learning
– Learning through Artificial Neural Networks
– Learning by Genetic Algorithms
(Ref.: Billari, C., F., et. al. Agent-Based Computational Modeling, 2006)
22
Genetic Algorithm
Evolutionary algorithm
Biological analogy: Evolution by natural selection
A population of individuals, each with some degree of “fitness”,
a metric explicitly defined by the modeler.
The fittest individuals are “reproduced” by breeding them with
other fit individuals to produce offspring that share the features
from parents.
Average fitness increases as population adapts to environment.
Individuals of GA may be:
– Agents: Evolution of the agent population as a whole
– Agent Algorithms: Evolution of better algorithms, i.e. increases
utility.
(Ref.: Gilbert, N., Terna, P. How to build and use agent-based models in social
science,1999)
23
Multiple Agents & Interactions
Among Agents
Interaction Types:
– Coordination
– Cooperation
– Negotiation
– Competition
24
Environment Types
25
The Process of ABM Building
Determination of whether the model to be constructed will be a
specific or generic model
Determination of actual actors in the real system and software
agents are representations of a subset of these actors.
Determination of model abstraction/aggregation level.
Determination of what cognitively oriented computations the
agents can perform
Selection of agent architecture
Selection of software platform and model implementation
26
An Example Model
SUGARSCAPE
27
ABM vs. SD
28
ABM vs. SD
SD
Stock-flow structure
Differential/difference equations, thus also called
equation-based modeling.
Feedback theory, circular causality
ABM
Agent Program
Behavioral Rules & Interactions
Adaptation and Learning
29
ABM vs. SD
Aggregation:
– SD is based on aggregation philosophy. SD makes an
abstraction from single events and individuals; and forms a
macro level modeling approach. Focus on system-level
variables, observables.
– ABM is a micro level modeling approach. ABM focuses on
the individual agents’ actions. ABM defines behavior at
individual level.
Policy Analysis:
– It is more natural to model multiagent systems by ABM for
policy analysis regarding individual agents.
30
ABM vs. SD
Adaptation:
– SD model structures are typically static.
– Agent decisions, reasoning, goals, etc. change by learning:
adaptation. As a result system structure may show
adaptation to changing conditions.
Heterogeneity:
– SD does differentiate among individuals.
– Each agent has its own design, need not expose its internal
structure to others in the system.
31
ABM vs. SD
Implementation:
– ABM may have more direct consequences for implementation
compared to SD, reference to direct inferences about individual
agent behaviors.
Validation:
– Structural Validation
– Behavioral Validation
(Ref.: Scholl, H., J., Agent-Based and System Dynamics Modeling: A Call for Cross study and Joint
Research, 2001
Scholl, H., J., Looking Across the Fence: Comparing Findings From SD Modeling Efforts with
those of Other Modeling Techniques
Parunak, H., V., D. et. al., Agent-Based Modeling vs. Equation-Based Modeling: A Case Study and
User’s Guide, MABS, 1998
Rahmandad, H., Heterogeneity and Network Structure in the Dynamics of Contagion: Comparing
Agent-Based and Differential Equation Models, ISDC, 2004
32 Demirel, G., Aggregated and Disaggregated Approaches to Multiple Agent Dynamics, ISDC,
2006 )
THANKS…
33