Work

Monte Carlo -puuhaku hyödyntäen neuroverkkoon perustuvaa vastustajanmallinnusta piilotetun informaation Moska-korttipelissä

Year: 2025
Type: dissertation
Abstract: When playing a game, humans often think ahead about the potential effect of their actions on the current game state, often doing so subconsciously. In our minds, we are creating a tree-like structure of allowed moves and the possible responses of our opponents. This look-ahead reasoning is straightforward in games, where all the information is visible, such as chess. But what happens when some of the information is hidden, like in many card games? This study researches the question in the context of a hidden information multiplayer card game Moska. We create an application of the Monte Carlo Tree Search (MCTS) algorithm within this environment to develop an agent capable of beating a human player. The hidden information was handled with a determinization approach, using random and neural network-based methods. We trained the neural network with 50,000 games of data from self-play to model opponents hand cards, which was then used in the determinization process. Our experiments showed that while a chance based random approach performed well in this environment, our neural network was able to model opponents’ hands with around 89% accuracy and significantly increase our agent’s playing strength. These results indicate that while a tree search algorithm like MCTS is applicable to hidden information environments, the quality of the information given is key to better performance.
Author: Törö, Arno
Language: English
Open Access status: green
OpenAlex ID: W7077237749
Cites: 14
Cited by:
Related to: 10