Application of Reinforcement Learning to Develop an AI Model for Game “Togyz Qumalaq”

Logical games often require players to solve various puzzles and strategic challenges. Thanks to the active implementation of Artificial Intelligence, it has possible to use Deep Learning models in such games, which has led to a significant breakthrough in solving other related tasks in the field. This paper presents a study on the development and training of two Reinforcement Learning algorithms, Q-learning and Deep Q-Network (DQN) for playing the game Togyz qumalaq. Both models were trained and evaluated in a game against MiniMax, which was also implemented by the authors of this research. The experiments were conducted with MiniMax recursion depths of 2, 3, and 4, respectively. The article presents the training parameters of models based on Q-learning and DQN, which achieved the best results. For each of the models, reward and training episode graphs are provided. The paper also includes the architecture of DQN, which demonstrated promising results. The results show that DQN has achieved significant success in solving this task, demonstrating notable performance. A comparative analysis also revealed that, unlike DQN, Q-learning requires more significant computational and memory resources for training.