Games generally have well-defined rules and goals and therefore represent an ideal environment for examining complex topics in machine intelligence. Previous research has mainly focused on deterministic zero-sum games providing perfect information such as checkers, Othello and, above all, chess. These types of games have one crucial thing in common: every player has complete knowledge of the entire game state at all times.

The game of poker represents a challenging domain, since it contains imperfect information, deception, and non-deterministic elements. An extensive degree of uncertainty is caused by the random shuffling of the card deck. Private cards held by opponents and concealed cards remaining in the deck represent the imperfect information elements of the game. Opponents are provoked to play deceptively for the purpose of hiding their pocket cards. These characteristics prevent traditional game-tree search methods from working properly.

We present a new approach for dynamically modelling the opponent’s behaviour in Texas Hold’em poker. The model highlights a way to partially solve the problem of imperfect information elements. This objective is primarily achieved by applying artificial neural networks.