Umělá inteligence v šachu – přístup stroje versus přístup člověka

This bachelor’s thesis explores the possibilities of using artificial intelligence principles to capture and simulate the human style of play in chess, with the aim of comparing machine decisionmaking to human one. It analyzes selected algorithms and approaches used in the development of chess programs – from traditional search methods such as Minimax and Alpha-Beta pruning to modern deep learning techniques employing convolutional neural networks (CNNs) and Transformer architectures. The theoretical part compares these methods with findings about human cognition and thought processes based on the research of Adriaan de Groot, the chunking and template theories of Simon, Chase, and Gobet, Herbert Simon’s concept of bounded rationality, the CHREST model, and the views of Tikhomirov and Saariluoma on apperception and holistic perception – concepts that originally emerged from the study of chess. The analysis shows that there are significant parallels as well as differences between contemporary AI models and human thinking. Convolutional networks exhibit similarities to the human approach described by Simon, Chase, and Gobet in their chunking theory. The key difference between the human thought model according to Simon, Chase, and Gobet, and the CNN approach lies mainly in the adaptive nature of human memory, which chess engines do not possess. Saariluoma’s theory of apperception, describing the abstraction and subsequent analysis of a problem, diverges from the computationally intensive procedures of classical chess AI models. Although Transformer architectures partially resemble the process of apperception through their attention mechanism, they are not yet capable of fully modeling the dynamics of human thought. The practical part of the thesis focuses on the design, implementation, and testing of a convolutional neural network intended to imitate human chess behavior. The model was trained on a large dataset of games from Lichess.org using a tensor representation of chess positions (8×8×20). The architecture includes an input convolutional layer, six residual blocks with Squeeze-andExcitation mechanisms, and two output branches — one for move prediction and one for position evaluation. Experiments demonstrated that the model achieves a Top-1 accuracy of 50.79% and a Top-10 accuracy of 96.61% in predicting human moves. Over 99% of its Top-1 predictions in openings were legal, and its position evaluations show high reliability. Although the model occasionally errs on less common tactical motifs, it successfully imitates the human playing style, including typical strategic principles and mistakes, reaching a performance roughly equivalent to 1500 ELO points on the Lichess.org platform.