Monotonic Model Improvement Self-play Algorithm for Adversarial Games

The problem of solving strategy games has intrigued the scientific community for centuries. In this paper, we consider two-player adversarial zero-sum symmetric games with zero information loss. Here, both players are continuously attempting to make decisions that will change the current game state to his/her advantage and hence the gains of one player are always equal to the losses of the other player. In this paper, we propose a model improvement self-play algorithm, where the agent iteratively switches roles to subdue the current adversary strategy. This monotonic improvement sequence leads to the ultimate development of a monolithic, competent absolute no-loss policy for the game environment. This tactic is the first of its kind in the setting of two-player adversarial games. Our approach could perform competitively and sometimes expertly in games such as 4x4 tic-tac-toe, 5x5 domineering, cram, and dots & boxes with a minimum number of moves.