Modern computer games place different and more diverse demands on the behavior of non-player characters in comparison to computers playing classical board games like chess. Especially the necessity for a long-term strategy conflicts often with game situations that are unsteady, i.e. many non-deterministic factors might change the possible actions. As a consequence, a computer player is needed who might take into account the danger or the chance of his actions. This work examines whether it is possible to train such a player by evolutionary algorithms. For the sake of controllable game situations, the board game Kalah is turned into an unsteady version and used to examine the problem.
展开▼