We explore different approaches for solving problems with incomplete information. As an example a card game Tressette is chosen where the incomplete information is presented as cards still in the deck. This means that players cannot make deterministic strategies on possible outcomes or predict the moves of an opponent, because such moves are not guaranteed, but are possible with certain probability. The main emphasis is on the Monte Carlo tree search method (MCTS), which uses random sampling and simulates only a part of the search space. MCTS has proven to be a successful method in this domain. A prototype of an intelligent agent was developed for playing the game. The agent was gradually improved by tuning MCTS method parameters and by introducing new heuristics into the search. We used knowledge extracted from the database of human-played games in the agent to improve its efficiency. The agent was tested by different approaches playing against each other and against human players.
|
