A model free approach to general video game playing

Precise decision making in games has always been one of the complex and exciting problems in artificial intelligence. General video game playing is a new branch of artificial intelligence and its purpose is to design agents who are able to carry out intelligent behavior in every unknown environment. Two of the biggest problems of recently introduced algorithms is the assumption of availability of an exact model of the world (for conducting search) and performing search in an online way (not using previous experiences). In this paper, we tried to eliminate these two problems with introducing an offline method for learning model of the world. Actor attempts to learn the model of the environment of a specific game, by repeating game and gaining experience about the environment of that game. In each play, the actor uses the updated model and after each play the model will be updated. Final model is general and is able to be used in every model based search algorithm. Finally, we analyzed the performance of the proposed method.