In this seminar, we present the evaluation of chess positions. We discuss the mathematical background of evaluation and the methods used to address this problem in practice. The evaluation methods can be clustered in two groups of static evaluation and search. In static evaluation, we initially focus on a simple example of assessment based on the value of the pieces, and by the end, we introduce efficiently updatable neural networks. In the search part, we limit ourselves to algorithms based on the minimax principle. In both parts, there are parameters whose values can be optimized using various machine learning techniques. We cover logistic regression, gradient descent and simultaneous perturbation stochastic approximation.
|
