In a repeated game, a memory-bounded agent selects its next action by basing its policy on a fixed window of past L plays. Traditionally, approaches that attempt to model memory-bounded agents, do so by modeling them based on the past L joint actions. Since the number of possible L sized joint actions grows exponentially with L, these approaches are restricted to modeling agents with a small L. This paper explores an alternative, more efficient mechanism for modeling memory-bounded agents based on high-level features derived from the past L plays. Called Targeted Opponent Modeler against Memory-Bounded Agents, or TOMMBA, our approach successfully models memory-bounded agents, in a sample efficient manner, given a priori knowledge of a feature set that includes the correct features. TOMMBA is fully implemented, with successful empirical results in a couple of challenging surveillance based tasks.
In Proceedings of the Adaptive Learning Agents Workshop (ALA), May 2013.

Doran Chakraborty Ph.D. Alumni chakrado [at] cs utexas edu
Peter Stone Faculty pstone [at] cs utexas edu