Zero-shot Coordination on Overcooked-AI Standard Five Layouts Evaluation
111.5CRoom ScoreUPD w/o bias
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| UPD w/o biasAblation=uniformly random policy πr for partner mixing2025.08 | 111.5 | 159.9 | 66.2 | 57.1 | 44.4 | 87.9 | 10.9 | |
| E3THyperparameters=ϵ=0.5 for CRoom/AA/CR/CC, ϵ=0.0 for FC2025.08 | 111.3 | 127.9 | 58.4 | 55 | 41.3 | 78.8 | — | |
| UPD w/o ℓAblation=removes learnability scoring2025.08 | 110.8 | 164 | 68.9 | 64.5 | 45.7 | 90.8 | 14.1 | |
| FCPPopulation size=482025.08 | 109.9 | 117.4 | 64.7 | 30.8 | 27.2 | 70 | -11.8 | |
| MEPPopulation size=482025.08 | 109.5 | 142.8 | 64 | 13.7 | 46.2 | 75.3 | -4.5 | |
| UPD (Ours)Partner buffer size |B|=512, Evaluation rollouts N=102025.08 | 107.5 | 181.4 | 69.2 | 64.5 | 48.7 | 94.4 | 18 | |
| SPTraining algorithm=IPPO2025.08 | 68.9 | 64.4 | 25.3 | 14.5 | 33.9 | 41.4 | -62.2 |