Learning Position Evaluation Functions for Tactical Behaviors of Game Characters
Kyeonah Yu · Jeongbo gwahaghoe nonmunji. so'peuteuweeo mich eung'yong · 2011
The position evaluation function which is originally introduced to evaluate the board configuration of computer chess can be used as a key element to determine the tactical behavior of AI characters in computer games. The position evaluation function is defined as a weighted sum of features, and it is important to assign weights properly according to the importance of features. In this paper the weight factor of a position evaluation function is learned by using a supervised learning technique. The weight learning is performed in a way similar to the single-layer Perceptron learning which minimizes the difference between the output of the current weights and the target, and the convergence of this weight update method is proved. The simulation result shows that the weights can be modified automatically according to the designer's intention and the learned position evaluation function can be applied to path-finding considering tactical features.