DAGS algorithm slashes AI training time in complex games like StarCraft

Imperfect-information games like StarCraft and Dota remain notoriously hard for AI due to sparse rewards and long horizons that make exploration in self-play computationally infeasible. A new paper from JB Lanier, Nathan Monette, Pierre Baldi, and Roy Fox introduces Data-Augmented Game Starts (DAGS), which leverages offline human demonstrations to provide high-level strategic coverage. Instead of starting each self-play episode from scratch, DAGS samples intermediate states from the offline data, allowing the agent to focus on strategically relevant subgames. The method is tested on analytically tractable variants of Kuhn Poker and Goofspiel, as well as a custom counterexample game that penalizes biased beliefs. Under fixed compute budgets, DAGS significantly reduces exploitability—a key measure of equilibrium approximation—compared to standard self-play.

To address the risk of biased equilibria from non-uniform starting distributions, the authors introduce multi-task observation flags that let the agent distinguish augmented starts from natural ones, preserving convergence guarantees. They also release new OpenSpiel benchmark environments with drastically increased exploration challenges and state counts while keeping exploitability analytically tractable. While the paper focuses on two-player zero-sum games, the approach could extend to multi-agent settings. DAGS offers a practical, data-efficient way to train game AI that previously required massive computational resources, potentially accelerating progress in complex strategic domains.