CADIA-Player : a general game playing agent

The aim of General Game Playing (GGP) is to create intelligent agents that can automatically learn how to play many different games well without any human intervention, given only a description of the game rules. This forces the agents to be able to learn a strategy without having any domain-specific knowledge provided by their developers. The most successful GGP agents have so far been based on the traditional approach of using game-tree search augmented with an automatically learned evaluation function for encapsulating the domain-specific knowledge. In this thesis we describe CADIAPlayer, a GGP agent that instead uses a simulation-based approach to reason about its actions. More specifically, it uses Monte Carlo rollouts with upper confidence bounds for trees (UCT) as its main search procedure. CADIAPlayer has already proven the effectiveness of this simulation-based approach in the context of GGP by winning the Third Annual GGP Competition. We describe its implementation as well as several algorithmic improvements for making the simulations more effective. Empirical data is presented showing that CADIA-Player outperforms naïve Monte Carlo by close to 90% winning ratio on average on a wide range of games, including Checkers and Othello. We further investigate the relative importance of UCT’s actionselection rule, its memory model, and the various enhancements in achieving this result.

Author dc:creator

• Hilmar Finnsson 1974-

Contributors dc:contributor

• Háskólinn í Reykjavík

• Gervigreind
• Tölvuleikir
• Vitvísindi
• Tölvunarfræði
• Meistaraprófsritgerðir
• Computer science

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/1946/7478

OAI identifier oai:identifier

oai:skemman.is:1946/7478