A Unified Approach to Reinforcement Learning, Quantal Response Equilibria, and Two-Player Zero-Sum Games
About
This work studies an algorithm, which we call magnetic mirror descent, that is inspired by mirror descent and the non-Euclidean proximal gradient algorithm. Our contribution is demonstrating the virtues of magnetic mirror descent as both an equilibrium solver and as an approach to reinforcement learning in two-player zero-sum games. These virtues include: 1) Being the first quantal response equilibria solver to achieve linear convergence for extensive-form games with first order feedback; 2) Being the first standard reinforcement learning algorithm to achieve empirically competitive results with CFR in tabular settings; 3) Achieving favorable performance in 3x3 Dark Hex and Phantom Tic-Tac-Toe as a self-play deep reinforcement learning algorithm.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Matrix Game Strategy Learning | Rock-Paper-Scissors (RPS) | Exploitability0.00e+0 | 6 | |
| Matrix Game Strategy Learning | Matching pennies | Last Iterate Exploitability0.00e+0 | 6 | |
| Matrix Game Strategy Learning | Random 10 x 10 Matrix Game | Exploitability (Last Iterate)0.1207 | 6 | |
| Matrix Game Strategy Learning | Random 12 x 6 Matrix Game | Exploitability (Last Iterate)0.14 | 6 | |
| Board-game self-play | Animal Shogi | Best-Response Win Rate87 | 5 | |
| Board-game self-play | Hex 10k | Best-Response Win Rate96 | 5 | |
| Board-game self-play | Othello | Best-Response Win Rate92 | 5 | |
| Board-game self-play | Connect Four | Best-Response Win Rate76 | 5 | |
| Equilibrium finding | Leduc Hold’em | Exploitability0.552 | 5 | |
| Exploitability | Neural Kuhn poker | Exploitability53.8 | 5 |