In this talk we will discuss two problems related to automated agents, one from each side of the behavior spectrum: Coordination and Competition.
Dr. Inon Zuckerman
Bio: Inon Zuckerman is a Ph.D. candidate at the department of Computer Science of Bar-Ilan University (expected spring 2009), where he was granted the President's scholarship to outstanding PhD students. Its dissertation includes work in the areas of human-agent coordination, BDI models and adversarial search methods. In 2005 he received his Master of AI degree (Magna Cum Laude, ranked 3rd out of 127) in K.U.Leuven, Belgium, where he was granted the Flemish community scholarship. In 2003 he received his B.A. in Computer Science (Cum Laude) from the Interdisciplinary Center in Herzliya, Israel.
Abstract: We will first consider an automated agent that needs to coordinate with a human partner when communication between them is not possible or is undesirable (tactic coordination games). Specifically, we examine situations where an agent and a human attempt to coordinate their choices among several alternatives with equivalent utilities. We use machine learning algorithms to help the agent predict human choices in these tactic coordination domains. Learning to classify general human choices tends to be very difficult. Nevertheless, experiments have shown that often humans are able to coordinate with one another in communication-free games, by using focal points, as ``prominent'' solutions to coordination problems. We will present Focal Points Learning by way of transforming raw domain data into a new hypothesis space based on focal points properties. This learning method will help classifiers achieve higher classification percentages in three experimental domains.
In the competition based problem we will explore competitive zero-sum interactions using classical adversarial search techniques. There are two basic approaches to generalize the propagation mechanism of the two-player Mini-Max search algorithm to multi-player (3 or more) games: the MaxN algorithm and the Paranoid algorithm. The main shortcoming of these approaches is that their strategy is fixed. In this paper we suggest a continuum approach (called MP-Mix) that dynamically changes the propagation strategy based on the players' relative strengths between MaxN, Paranoid and a newly presented offensive strategy. In addition, we introduce the Opponent Impact factor for multi-player games, which measures the player's ability to impact their opponent's score, and show its relation to the relative performance of our new MP-Mix strategy. Experimental results show that our new MP-Mix algorithm outperforms all other approaches under most circumstances.