The Application of AlphaZero to Wargaming

In this paper, we explore the process of automatically learning to play wargames using AlphaZero deep reinforcement learning. We consider a simple wargame, Coral Sea, which is a turn-based game played on a hexagonal grid between two players. We explore the differences between Coral Sea and traditional board games, where the successful use of AlphaZero has been demonstrated. Key differences include: problem representation, wargame asymmetry, limited strategic depth, and the requirement for significant hardware resources. We demonstrate how bootstrapping AlphaZero with supervised learning can overcome these challenges. In the context of Coral Sea, this enables AlphaZero to learn optimal play and outperform the supervised examples on which it was trained.