Signal Processing and Speech Communication Laboratory
hometheses & projects › Evolutionary Neural Network Poker Agents and the Effects of Opponent Modeling

Evolutionary Neural Network Poker Agents and the Effects of Opponent Modeling

Status
Finished
Type
Bachelor Project
Announcement date
01 Oct 2018
Student
Fabian Moik
Mentors
Research Areas

Abstract

Poker is a card game with imperfect information where players have to deal with randomness, hidden information, opponent modeling, risk manage- ment and deception. These properties turn the game into a very interesting test-bed for artificial intelligence research. Evolutionary algorithms have been used in many fields of machine learning to find solutions to problems in a large decision space, while neural networks nowadays are widely used to find solutions in non-linear decision spaces. A combined version of these two models is used in this thesis to create No- Limit Texas Hold’em poker agents capable of developing a profitable playing style and learning the fundamental principles of a successful poker strategy. To counter some problems inherent in evolutionary algorithms, such as Evolutionary Forgetting a concept called Hall of Fame is used to improve the performance of the evolved agents. Opponent modeling is an essential part of the decision-making process of a poker player and is imperative to achieve a high skill in the game. The results of the conducted experiments show that a hall of fame greatly increases the overall performance of agents, while opponent modeling allowed to develop a versatile strategy that works especially well against static opponent types. The resulting program successfully understood the most important principles for playing profitable poker but there remains further research to be done to achieve the skill of professional tournament poker players.