Clustering Player Profiles to Improve Opponent Modelling in Simulation Based Poker Agents
by: Jack Pimbert
Abstract
Games have been a focus for research in Artificial Intelligence for several decades as many real life situations can be modelled through them. More recently, Texas Hold’em Poker has been identified as an ideal testing ground as it addresses some of the most challenging aspects of the research; it is a non-cooperative, partially observable, imperfect information game with stochastic outcomes. The latest work in multiplayer Poker relies on using opponent models generated from historical data to build counter strategies. This project focuses on using cluster- ing as a method to improve the performance of such simulation based systems in partially observable, stochastic, imperfect information domains. Historical data has been collected from online gameplay and is here used to develop an exploitative no-limit multiplayer Texas Hold’em Poker agent. The agent uses opponent models for action prediction and range estimation that are integrated into a selective sampling simulation in order to determine the optimal action for the agent to take. In comparison with agents that do not cluster opponent data, analysis of the agent’s performance over 20,000 hands shows a clear increase in winnings and thus also shows the advantage of clustering data in these domains. This project shows that clustering is an effective means to improve simulation based agents in multiplayer Poker.
Paper (2014)Statistics: Posted by ibot — Tue Sep 09, 2014 8:35 pm
]]>