Reinforcement Learning on Atari 49 games (train)
32Score (Subset 100)POP3D
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| POP3Dseeds=3, training steps=40M frames, network architecture=CNN with one action head and one value head, optimizer=Adam, penalty coefficient beta=5, initialization=PPO strategy, frame stacking=Enabled, no-ops=302018.07 | 32 | 20 | |
| PPOseeds=3, training steps=40M frames, network architecture=CNN with one action head and one value head, optimizer=Adam, implementation=OpenAI Baselines, frame stacking=Enabled, no-ops=302018.07 | 11 | 18 | |
| BASELINEseeds=3, training steps=40M frames, network architecture=CNN with one action head and one value head, optimizer=Adam, penalty coefficient beta=10, strategy=fixed KLD, frame stacking=Enabled, no-ops=302018.07 | 5 | 6 | |
| TRPOseeds=3, training steps=40M frames, network architecture=CNN with one action head and one value head, optimizer=Adam, implementation=OpenAI Baselines, frame stacking=Enabled, no-ops=302018.07 | 1 | 5 |