Graph embeddings for deep learning in general game playing

Abstract

General Game Playing agents can play many different games. They take as an input a description of a game written in a Game Description Language (GDL) and then infer a proper strategy for playing that game. Typically this includes learning a value function for evaluating the merit of game states, for example, by using a neural network. In this work we investigate whether more effective neural networks-based state evaluations can be built by feeding the networks not only with positions describing the game states but also a graph-based embedding of the game rules (derived from GDL). The game rules are encoded as either Rule Graphs or Propositional Networks, and then we experiment with several different graph-based embeddings for encoding the graphs. The result shows that such an approach has promise, but care must be taken in choosing an appropriate embedding.