Department of Computational Social Science Thesis Proposals/Defenses

Thursday, July 21 - 11:00 am
Location: Center for Social Complexity Suite, 3rd Floor Research Hall (formerly Research 1),


On Bounded Rationality In Multi-Agent Environments

Maciej Latek

Department of Computational Social Science
George Mason University

Decision-makers who face intelligent adversaries or competitors have used statistical, game-theoretic and war-gaming approaches to gain insight into the most likely and effective opponent strategies and to determine which of their own strategies can best be used to respond. Much of the criticism levied against those approaches can be traced in their limited capacity to deal with strategic and temporal uncertainty:

  1. Strategic uncertainty decision makers suffer from strategic uncertainty, of which the classical equilibrium selection problem is a special case, when their adversary enjoys multiple equally desirable strategies. For example, incomplete knowledge of evolving organizational and cultural heuristics that shape adversary objectives and payoffs (a) makes any extrapolation from history dangerous and (b) causes response surfaces to be too noisy and flat to hope for a single equilibrium to exist.
  2. Temporal uncertainty Inherently continuous and asynchronous, adversarial behaviors transcend “my move, your move” abstractions of game-theoretic models. For example, forces of nature or adversary disruption may compel adversaries to replan their strategies mid-course. Moreover, the need to replan may be anticipated and planned for by opponents.

To curtail strategic and temporal uncertainty decision makers need flexible and extensible models of their operating environments, with explicit treatment of adversarial decision making and an apparatus to translate those into robust and farsighted courses of action. This dissertation summarizes results of an ongoing research program centered around questions of using agent-based models for decision support in strategic interactions. The research program is composed of four research axes: (a) language and concepts, (b) theory and methodology, (c) software and technology and (d) computational and experimental demonstrations.

This thesis demonstrates the value of anticipation in strategic environments and reviews a body of existing approaches. The key result is an efficient computational implementation of a farsighted, N-th order rationality in a recursive simulation framework where simulated decision-makers each use simulation themselves to devise robust courses of action. The architectural and algorithmic issues related to this approach are discussed and tested using a set of small agent-based models.