DAGS: Data-Augmented Game Starts Accelerate Self-Play in Imperfect Information Games

A novel strategy for initiating multi-agent sampling, known as Data-Augmented Game Starts (DAGS), seeks to enhance online exploration within regularized policy-gradient methods for two-player zero-sum games. This technique leverages offline demonstrations from proficient human players to kickstart reinforcement learning data gathering at key intermediate states, thereby promoting the exploration of strategically important subgames. Experiments were carried out using synthetic datasets and analytically manageable, long-horizon control variants. This research tackles the computational challenges faced in large-scale imperfect-information competitive games such as StarCraft, Dota, and CounterStrike, which are hindered by sparse rewards and difficult long-horizon exploration. The paper can be accessed on arXiv under ID 2605.14379.