An Improved GNN-Reasoner for Game Description Language

Abstract

The intersection of Artificial Intelligence (AI) with computer game playing has yielded significant breakthroughs, exemplified by Deep Blue’s victory over a world chess champion and AlphaGo’s triumph in Go through self-learning. The focus on General Game Playing (GGP) [17] seeks to develop AI systems capable of playing multiple games based solely on their rules, aiming for a broader intelligence application. A recent development in GGP introduced AlphaGo-style neural network learning approach in game playing [7, 10]. In these works, reasoning of game rules in the Game Description Language (GDL) is handeled by a logical reasoner based on Prolog or Propnet. The introduction of the GNN-Reasoner [9] shows a promising approach of using neural networks by employing a graph-based approach for game rules and state representation, though it faced challenges such as dataset feature design and inference task completion. Addressing these limitations, this thesis introduces an enhanced model, GNN-Reasoner- V2. Our contributions include a more efficient graph structure, a method for flattening negation rules, an improved neural network architecture that reduces the demand of hardware resource, a faster and modular dataset generation process, and a suite of analytical tools for comprehensive model evaluation.