Reinforcement Learning on Atari 57 full suite 2600
57Games Above Human CountAgent57
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Agent572020.03 | 57 | 100 | 4,766.25 | 1,933.49 | 1,091.07 | 614.65 | 324.78 | 184.35 | 116.67 | |
| R2D2variant=bandit2020.03 | 54 | 96.93 | 5,461.66 | 2,357.92 | 1,298.8 | 648.17 | 303.61 | 116.82 | 93.25 | |
| R2D2variant=Retrace2020.03 | 52 | 94.2 | 3,518.36 | 1,457.63 | 817.77 | 420.67 | 267.25 | 116.03 | 48.32 | |
| R2D22020.03 | 52 | 94.33 | 4,622.09 | 1,935.86 | 1,176.05 | 529.23 | 215.31 | 115.33 | 50.27 | |
| NGU2020.03 | 51 | 95.07 | 3,421.8 | 1,359.78 | 610.44 | 267.1 | 226.43 | 107.78 | 64.1 | |
| MuZero2020.03 | 51 | 89.92 | 5,661.84 | 2,381.51 | 1,172.9 | 503.05 | 171.39 | 75.74 | 0.03 | |
| APTRL interaction steps=100K, Contrastive Learning (CL)=true2021.03 | 12 | — | 47.73 | 33.41 | — | — | — | — | — | |
| VISRRL interaction steps=100K2021.03 | 11 | — | 102.31 | 6.81 | — | — | — | — | — | |
| APTRL interaction steps=100K, Contrastive Learning (CL)=false2021.03 | 9 | — | 41.12 | 28.65 | — | — | — | — | — | |
| DIAYNRL interaction steps=100K, Contrastive Learning (CL)=false2021.03 | 6 | — | 23.9 | 2.95 | — | — | — | — | — | |
| DIAYNRL interaction steps=100K, Contrastive Learning (CL)=true2021.03 | 6 | — | 25.14 | 3.28 | — | — | — | — | — |