Friday, October 10, 3:00 p.m.
Center for Social Complexity Suite
Research Hall, Third Floor

Simulating Conflict Decision-Making and Event Data

David Masad, PhD Student
Computational Social Science
George Mason University

ABSTRACT: International conflict (and political conflict in general) involves competition and cooperation between multiple actors, who differ from one another not only in their interests and capabilities, but often in their decision-making process as well. It is, in a word, complex. Much of the research on conflict has sought to reduce this complexity. Statistical modeling has been used to find regularities to data on political conflict, and has proven successful both at testing some qualitative theories and making predictions. Game theory can explicitly model at least stylized interaction, but are confined by a need for analytic tractability and assumptions of rationality. Nevertheless, some game-theoretic models have been successfully fit to data, though occasionally-extravagant claims have been made as to their predictive power. Computational models and simulations (particularly agent-based ones) are often relatively abstract as well, intended like the formal models to provide qualitative insight. Both in game-theoretic and agent-based models, all actors are generally assumed to be utilizing the same decision-making process.

I will present my proposed dissertation research, which involves using political event data as a bridge between these methodologies, and a way of validating and comparing different models. I propose to develop a framework for modeling international political contests with two key features: it will separate agents' external behaviors from their internal decision-making process, and it will produce event data as an output. This approach will allow for explicit testing of different decision-making models, identifying which is the best fit for real-world event data. I then propose to implement several models of decision-making, including game-theoretic optimization, rule-based decision-making, reinforcement learning, and BDI-type planning. By testing these rules against one another, as well as against historic event data, I will identify the explanatory and predictive power of each, and as well as emergent properties of each or of the system within which they are interacting.