University of Alberta expert poker agent: A survey

Games have always been a natural topic for Artificial Intelligence researchers to study and poker has proven to be a game that is both interesting and challenging. Part of the challenge of poker comes from the fact that it is a game of imperfect knowledge where multiple competing agents must deal with risk management, agent modeling, unreliable information and deception, much like decision-making applications in the real world. In order to produce an expert level poker intelligent agent, researchers from the University of Alberta identified five characteristics of an expert poker player that must be modeled in the agent. Of these five characteristics, the characteristic that has been shown to improve quality of the play the most is opponent modeling. Initially, the researchers built their opponent models using a weight table and opponent action frequencies. This proved to be an improvement over previous versions with no opponent modeling, but required domain specific knowledge and proved to be difficult to maintain. In an effort to fix the problems with their previous opponent models, researchers began using neural networks and a training technique called backpropagation to accurately predict an opponent’s actions. After running the neural network training system on game played previously, the use of neural networks proved to be more accurate than previous attempts at opponent modeling.