Offline Reinforcement Learning on D4RL Gym (medium-replay, medium-expert)
45.7HalfCheetah (medium-replay)CQL
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| CQLLearning Source=Task rewards, Policy optimization algorithm=CQL2023.01 | 45.7 | 84.1 | 80 | 88.5 | 103.7 | 108.4 | 85.1 | |
| IQLLearning Source=Task rewards, Policy optimization algorithm=IQL2023.01 | 44.3 | 100.5 | 74.8 | 85.2 | 84.1 | 107.5 | 82.7 | |
| PT+IQLLearning Source=Preference only, Reward modeling=PreferenceTransformer, Policy optimization algorithm=IQL2023.01 | 42.3 | 59.7 | 43.3 | 83.6 | 67.8 | 109.8 | 67.8 | |
| DPPOLearning Source=Preference only, Policy optimization algorithm=DPPO2023.01 | 40.8 | 73.2 | 50.9 | 92.6 | 107.2 | 108.6 | 78.8 | |
| PT+CQLLearning Source=Preference only, Reward modeling=PreferenceTransformer, Policy optimization algorithm=CQL2023.01 | 27.1 | 49.1 | 52.8 | 77.1 | 89.2 | 77.7 | 62.2 |