Multi-Agent Reinforcement Learning on Google Research Football (Scenario Win Rates)
97Win Rate: Pass and ShootExpert
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ExpertTask Specificity=Single-task expert2026.04 | 97 | 40 | 89 | 99 | 100 | 94 | |
| MARL-GPTTraining Scope=Multi-Env, Memory Architecture=Memory-based, Number of Parameters=7M, History Window Size=62026.04 | 96 | 43 | 89 | 98 | 98 | 68 | |
| BC-LSTMTraining Scope=Single-Env, Memory Architecture=Memory-based2026.04 | 90 | 22 | 88 | 43 | 30 | 24 | |
| DTTraining Scope=Single-Env, Memory Architecture=Memory-based2026.04 | 80 | 60 | 88 | 0 | 0 | 0 | |
| RATETraining Scope=Single-Env, Memory Architecture=Memory-based2026.04 | 78 | 58 | 85 | 4 | 1 | 1 | |
| CQLTraining Scope=Single-Env, Memory Architecture=Memory-free2026.04 | 60 | 30 | 87 | 38 | 35 | 34 | |
| BCTraining Scope=Single-Env, Memory Architecture=Memory-free2026.04 | 44 | 37 | 86 | 40 | 41 | 40 |