COMPUTATIONAL SOCIAL SCIENCE

CSS SEMINAR - OCTOBER 16 - AXTELL

Friday, October 16, 3:00 p.m.
Center for Social Complexity Suite
Research Hall, Third Floor

Agentization: Relaxing Unrealistic Assumptions and Simplistic Specifications in Social Science Models with Agent Computing

Robert Axtell, Prorfessor
Computational Social Science Program
Department of Computational and Data Sciences
George Mason University

ABSTRACT: Consider a conventional model in the formal (e.g., mathematical) social sciences, producing explanations E of some social phenomenon P. Such models commonly make ‘heroic’ assumptions (e.g., all agents can interact with one another) and employ highly idealized behavioral specifications (e.g., rationality) in order to make analysis tractable and derivation of results possible. Because such models are typically solved for agent-level equilibrium (e.g., Nash, Walras, refinements) aspects of agent behavior away from equilibrium may be unspecified. In ‘agentizing’ a conventional model the highly expressive power of software agents can be used to relax unrealistic assumptions and fill-in missing behavioral specifications. Because of these alternative specifications, execution of the agent model may or may not yield the conventional explanations, E. I exhaustively characterize the similarities and differences that can arise as a result of agentization, and give examples of each from the extant literature. While it may be the case that an agent model simply reproduces E, it is common for alternative explanations, E’, to result, based on more realistic representation of P. Agentization can thus be used to test the generality and robustness of standard results. I will give several examples, based on well-known economic models, in which relaxation of mathematically convenient but unrealistic model specifications clearly demonstrate that conventional results are sometimes robust, sometimes brittle, and sometimes simply not credible. I will argue that agentization of social science models is a general approach for (1) assessing conventional models and (2) for building models that are more realistic than analytical ones