Monte Carlo Tree Search in Poker Games

Andrej Babič (2013) Monte Carlo Tree Search in Poker Games. EngD thesis.

Preview

PDF Download (4Mb)

Abstract

As a method of artificial intelligence, Monte Carlo tree search has been widely adopted in the last few years. The main reasons for this are its relatively simple implementation and a very low domain specific knowledge needed for it to work properly. In my thesis, I will present its implementation and analyse the key parameters for the Texas Hold'em poker game. The main challenge was to develop an agent which would be able to play poker on an acceptable level without consuming too much time for each decision. One of the main problems of the Monte Carlo tree search, compared with other methods that rely more heavily on domain specific knowledge, is its relative slowness. There are a lot of poker frameworks available online. One of the most popular is Meerkat API. Its online community is large, thus there are a lot of poker playing algorithms using it. Since my agent has been developed with Meerkat API as well, I have had no problems finding suitable opponents for the simulations. Within this thesis, a generic Java framework for Monte Carlo Tree Search algorithms was developed. The framework is open source and freely available online. It enables us to customize each portion of the Monte Carlo Tree Search algorithm, as well as gives us the possibility to monitor our simulations with several data visualizers. The simulation results have shown that a Monte Carlo Tree Search simulation can be a very challenging opponent in a game of Texas Hold'em. Unfortunately, its decision-making speed cannot be compared to other algorithms with more domain specific knowledge. Its speed is also further impaired by the fact that it is written in Java.

Actions (login required)