Multi-Agent Reinforcement Learning on SMAC 3s5z v1 (test)
92.7Win RateMAGI
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MAGIUnit Composition=2-type, Training Budget=1.2M2026.05 | 92.7 | — | |
| AIB-onlyUnit Composition=2-type, Training Budget=1.2M2026.05 | 92.6 | — | |
| HIBCGUnit Composition=2-type, Training Budget=1.2M2026.05 | 92.6 | — | |
| BVMEBandwidth Ratio (r)=0.05, Message Reduction vs 0.30=83.3%2025.12 | 91.56 | 0.5517 | |
| CommFormerUnit Composition=2-type, Training Budget=1.2M2026.05 | 91.2 | — | |
| HIB-flatUnit Composition=2-type, Training Budget=1.2M2026.05 | 90.8 | — | |
| QMIXUnit Composition=2-type, Training Budget=1.2M2026.05 | 89.7 | — | |
| GACGUnit Composition=2-type, Training Budget=1.2M2026.05 | 89.7 | — | |
| BVMEUnit Composition=2-type, Training Budget=1.2M2026.05 | 89.6 | — | |
| BVMEBandwidth Ratio (r)=0.075, Message Reduction vs 0.30=75.0%2025.12 | 87.29 | 0.5427 | |
| GACGBandwidth Ratio (r)=0.32025.12 | 86.77 | 0.5201 | |
| BVMEBandwidth Ratio (r)=0.1, Message Reduction vs 0.30=66.7%2025.12 | 86.77 | 0.557 | |
| ExpoCommUnit Composition=2-type, Training Budget=1.2M2026.05 | 85.5 | — |