Assessing the positional values of chess pieces by tuning neural networks' weights with an evolutionary algorithm

Finding a method that can automatically set the weights of the evaluation function of a chess engine is an important research topic, since the use of manual settings requires a significant amount of time and expertise, which are not always available. The specialized literature reports several works in which the weights of the positional values of the chess pieces are evolved based on values stored in tables. Here, however, we propose to use a neural network architecture to obtain the positional values of the chess pieces based on specific features of each position. The neural networks that we adopt for this sake are relatively small and we argue that they constitute a robust way of obtaining the positional values of the chess pieces. The adjustment of weights of such neural networks was done through the use of an evolutionary algorithm, producing an increase of 433 points of the ranking of our chess engine (from 1745 to 2178 points, reaching a value close to that of a chessmaster).