Aplikace pro predikci šachových tahů pomocí neuronových sítí

Abstract

This thesis deals with the design, implementation, and experimental evaluation of a system for predicting chess moves using deep neural networks. The HouskaNet model was proposed, a convo- lutional residual network inspired by the AlphaZero architecture, which simultaneously predicts the probability distribution of moves and estimates the position value. A Monte Carlo Tree Search algo- rithm was implemented for move selection. Training took place in three phases: initial reinforcement learning with a curriculum-based approach, fine-tuning on a dataset of real games from the Lichess platform, and subsequent continuation of self-play training. Experiments showed that reinforcement learning from scratch alone was insufficient given limited computational resources, while combining it with fine-tuning on human games led to a significant improvement in the model’s performance. The resulting system was integrated into a web application that allows users to visualize games and play against the implemented AI player.