Exploring a Learning Architecture for General Game Playing

摘要

General Game Playing (GGP) is a platform for developing general Artificial Intelligence algorithms to play a large variety of games that are unknown to players in advance. This paper describes and analyses GGPZero, a learning architecture for GGP, inspired by the success of AlphaGo and AlphaZero. GGPZero takes as input a previously unknown game description and constructs a deep neural network to be trained using self-play together with Monte-Carlo Tree Search. The general architecture of GGPZero is similar to that of Goldwaser and Thielscher (2020) with the main differences in the choice of the GGP reasoner and the neural network construction; furthermore, we explore additional experimental evaluation strategies. Our main contributions are: confirming the feasibility of deep reinforcement for GGP, analysing the impact of the type and depth of the underlying neural network, and investigating simulation vs. time limitations on training.