| Abstract: |  Most game playing algorithms are specifically designed to play one single game. Therefore, transferring such programs into another context is not possible. On the other hand, general game playing (GGP) is a concept of playing vast number of games within a one concrete game environment. This emerging field of general artificial intelligence proficiently serves as a challenging testbed for research in domain-independent algorithms of computational game theory. If one can design a general game playing system capable to play more than one game successfully, such algorithm can be used in other areas which require a real-time domain-independent deciding, such as in providing intelligence for search and rescue missions or navigating remote autonomous robotic units in unknown territory.
Uppermost intention of this work is to focus on selecting the reasoner capable to properly and rigorously interpret the game; and employing the algorithm which meets the theoretical guarantees to work proficiently in the imperfect-information games. The most prominent outcome of this thesis is Shodan, one of the few GGP agents for playing general games with imperfect information. Shodan is based on Monte Carlo tree search with EXP3 selection criterion, working with the games represented as propositional networks.
|
|---|