Predicting Success in an Imperfect-Information Game

One of the most challenging tasks when creating an adaptation mechanism is to transform domain knowledge into an evaluation function that adequately measures the quality of the generated solutions. The high complexity of modern video games makes the task to generate a suitable evaluation function for adaptive game AI even more difficult. Still, our aim is to fully automatically generate an evaluation function for adaptive game AI. This paper describes our approach, and discusses the experiments performed in the RTS game Spring. TD-learning is applied for establishing a unit-based evaluation term. In addition, we define a term that evaluates tactical positions. From our results we may conclude that an evaluation function based on the defined terms is able to predict the outcome of a Spring game reasonably well. That is, for a unit-based evaluation the evaluation function is correct in about 76% of all games played, and when evaluating tactical positions it is correct in about 97% of all games played. A straightforward combination of the two terms did not produce improved results.