War Chess as Hierarchical Learning Environment

Abstract

This paper introduces GWCLE (General War Chess Learning Environment), a general machine learning environment based on hexagonal wargaming. Hexagonal war chess, when utilized as machine learning challenge, is naturally a multi-agent problem with the intelligent interaction of human or machine. The GWCLE supports hybrid engine, allowing credible simulation for kinds of war chess, which provides hierarchical training framework for massive agents control problem. The agent can be trained with designated level of war chess data and transferred bottom-up or top-down. For training on the whole deduction, we build the database to store refined replay data. Our framework is able to support agents to be trained in tactical and strategic level simultaneously. GWCLE offers a hierarchical perspective of the war chess simulation, allowing researchers controlling the granularity of action and time step.