Multi-agent reinforcement learning on Google Research Football Academy scenarios
100Pass & Shoot Success RateRL-D2
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| RL-D2divergence=Reverse KL (RKL)2025.09 | 100 | 97.9 | 99.2 | 98.5 | 98.2 | 97.1 | 99.1 | 99.4 | 0 | |
| MAPPO2025.09 | 98.8 | 74.5 | 92.6 | 71.4 | 64.2 | 54.7 | 94.8 | 53.9 | 4.8 | |
| RL-D2divergence=Forward KL (FKL)2025.09 | 98.4 | 99.2 | 96.8 | 97.4 | 96.5 | 94.1 | 97.9 | 98.1 | 68.5 | |
| MAT2025.09 | 96.8 | 97.6 | 91.8 | 88.5 | 87.1 | 94.6 | 92.9 | 91.3 | 10.2 | |
| Autoregressive2025.09 | 79.4 | 68.9 | 74.1 | 70.8 | 35.2 | 0 | 72.5 | 33.1 | 0 |