New algorithm for zero-sum games needs no learning rate tuning
First traditional online method to achieve rate-free convergence and stability in zero-sum games
Researchers Gaoqi Xue and James P. Bailey introduce implicit midpoint gradient descent, a new algorithm for unconstrained bilinear zero-sum games—a foundational model in online learning, adversarial optimization, and multi-agent decision-making. The method is derived from continuous-time follow-the-regularized leader dynamics using symplectic integration, bridging numerical analysis and game theory.
The algorithm proves three key properties: bounded orbits, fast ergodic convergence to Nash equilibria, and learning-rate-independent stability—all achieved without manual hyperparameter tuning. Computational experiments show it significantly outperforms standard methods like optimistic gradient descent and alternating gradient descent. This marks the first time a traditional online optimization approach has simultaneously delivered these guarantees in zero-sum games, opening avenues for more robust adversarial training and multi-agent systems.
- Derived from continuous-time follow-the-regularized leader via symplectic integration methods
- First traditional online optimization to achieve bounded orbits, fast ergodic convergence, and rate-independent stability in unconstrained bilinear zero-sum games
- Outperforms optimistic and alternating gradient descent in computational experiments
Why It Matters
Enables stable, tuning-free training for adversarial systems and multi-agent AI without manual learning rate schedules.