Reevaluating Policy Gradient Methods for Imperfect-Information Games

TL;DR

This paper compares various deep reinforcement learning algorithms in imperfect-information games and finds that simpler policy gradient methods like PPO outperform more complex approaches based on fictitious play, double oracle, and CFR.

Contribution

It provides the first large-scale exploitability comparison of DRL algorithms in imperfect-information games, demonstrating the competitiveness of generic policy gradient methods.

Findings

PPO and similar methods outperform FP, DO, and CFR-based approaches.

Conducted over 7000 training runs for robust comparison.

Provided the first accessible exploitability computation framework for large games.

Abstract

In the past decade, motivated by the putative failure of naive self-play deep reinforcement learning (DRL) in adversarial imperfect-information games, researchers have developed numerous DRL algorithms based on fictitious play (FP), double oracle (DO), and counterfactual regret minimization (CFR). In light of recent results of the magnetic mirror descent algorithm, we hypothesize that simpler generic policy gradient methods like PPO are competitive with or superior to these FP-, DO-, and CFR-based DRL approaches. To facilitate the resolution of this hypothesis, we implement and release the first broadly accessible exact exploitability computations for five large games. Using these games, we conduct the largest-ever exploitability comparison of DRL algorithms for imperfect-information games. Over 7000 training runs, we find that FP-, DO-, and CFR-based approaches fail to outperform generic policy gradient methods. Code is available at https://github.com/nathanlct/IIG-RL-Benchmark and https://github.com/gabrfarina/exp-a-spiel .