Multi-task Reinforcement Learning on DMLab-30 (test)
72.8Mean Capped Human ScorePopArt-IMPALA
Evaluation Results
| Method | Links | |
|---|---|---|
| PopArt-IMPALA2018.09 | 72.8 | |
| IMPALA2018.09 | 58.4 | |
| IMPALAarchitecture=deep, PBT=true, learners=12018.02 | 49.4 | |
| IMPALAarchitecture=deep, PBT=true, learners=82018.02 | 49.1 | |
| IMPALAarchitecture=deep, actors=150, learners=12018.02 | 46.5 | |
| IMPALA-Expertsarchitecture=deep, training_regime=individual (separate agent per task)2018.02 | 44.5 | |
| IMPALAarchitecture=shallow, actors=210, learners=12018.02 | 37.1 | |
| A3Carchitecture=deep, workers=2102018.02 | 23.8 |