| Abstract [eng] |
In recent decades, artificial intelligence agents have achieved excellent results in specific games, but most of them are highly specialized and difficult to adapt to different rules or environments. The field of General Game Playing (GGP) aims to address this shortcoming. This paper investigates how GGP agents based on the Ludii language can be adapted to games with imperfect information and how their performance can be improved using Monte Carlo Tree Search (MCTS) and neural networks. The paper analyzes game description languages and agent development methods, primarily focusing on Ludii, MCTS, and neural network methods. Six games were prepared for the experiments: "The Monty Hall Problem", "Unfair Rock, Paper, Scissors", "Meier", "2048", "Battleship", and "Durak". A mechanism was developed that limits the information visible to the agent and generates possible perfect-information contexts consistent with the player’s observations. Several agents were also implemented to compare different approaches to imperfect-information games. Experimental results showed that the MCTS-based agent was able to play all selected games and served as a suitable baseline solution. Neural network outputs improved performance when used to guide the MCTS selection phase, but replacing or guiding the simulations with these outputs often reduced performance. The results confirm that Ludii can be used for GGP research for imperfect information games when additional information hiding and context generation mechanisms are applied. |
