Reinforcement Learning on Atari-57 (test)
2,041.12Median Human Norm ReturnMuZero
Evaluation results
| Method | Links | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MuZeroScale=20B2023.05 | 2,041.12 | 4,994.97 | — | — | — | 741,812.63 | 28,634.39 | 143,972.03 | 998,425 | 678,558.64 | 1,674,767.2 | 1,278.98 | 848,623 | 454,993.53 | 85,932.6 | 260.13 | 100 | 864 | 1,159,049.27 | 991,039.7 | 458,315.4 | 839,642.95 | 143,964.26 | 23.94 | 2,382.44 | 91.16 | 33.03 | 631,378.53 | 130,345.58 | 6,682.7 | 49,244.11 | 67.04 | 41,063.25 | 16,763.6 | 269,358.27 | 204,824 | 0 | 243,401.1 | 157,177.85 | 955,137.84 | 0 | 21 | 15,299.98 | 72,276 | 323,417.18 | 613,411.8 | 131.13 | 999,976.52 | -29,968.36 | 56.62 | 74,335.3 | 549,271.7 | 9.99 | 0 | 476,763.9 | 491.48 | 715,545.61 | 0.4 | 981,791.88 | 197,126 | 553,311.46 | 725,853.9 | |
| LBCScale=1B2023.05 | 1,665.6 | 10,077.52 | — | — | — | 279,703.5 | 12,996.3 | 62,025.7 | 999,999 | 1,106,603.5 | 3,824,506.3 | 1,410 | 857,369 | 457,321 | 35,340 | 233.1 | 100 | 864 | 728,080 | 999,999 | 233,090 | 995,950 | 900,170 | 24 | 14,332.5 | 75 | 34 | 13,792.4 | 488,900 | 6,372.5 | 37,545.6 | 47.53 | 623,300.5 | 14,372.6 | 593,679.5 | 1,666,665 | 2,500 | 31,403 | 81,473 | 999,999 | -1 | 21 | 15,100 | 151,730 | 27,964.3 | 999,999 | 144 | 1,000,000 | -5,903.34 | 10,732.5 | 159,999.6 | 999,999 | 2.726 | 24 | 531,614 | 436.2 | 999,999 | 2,200 | 999,999 | 118,900 | 998,970 | 241,570.6 | |
| IMPALAArchitecture=deep, Training Strategy=experts, Evaluation Protocol=Up to 30 no-ops2018.02 | 191.8 | 957.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ReactorTraining Strategy=experts, Evaluation Protocol=Up to 30 no-ops2018.02 | 187 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoisyNet-DuelingNetwork architecture=NoisyNet2017.06 | 172 | 633 | 30 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DuelingNetwork architecture=Baseline2017.06 | 132 | 524 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoisyNet-DQNNetwork architecture=NoisyNet2017.06 | 123 | 379 | 48 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| A3CArchitecture=deep, Training Strategy=experts, Evaluation Protocol=Up to 30 no-ops2018.02 | 117.9 | 503.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EfficientZeroScale=100K2023.05 | 109 | 194.3 | — | — | — | 808.5 | 148.6 | 1,263.1 | 25,557.8 | — | — | 351 | 13,871.2 | — | — | — | 52.7 | 414.1 | — | 1,117.3 | 83,940.2 | — | 13,003.9 | — | — | — | 21.8 | 296.3 | 3,260.3 | — | 3,915.9 | — | 517 | 724.1 | 5,663.3 | 30,944.8 | — | 1,281.2 | — | — | — | 20.1 | 96.7 | 14,448.5 | — | 17,751.3 | — | 1,100.2 | — | — | — | — | — | — | — | — | 17,264.2 | — | — | — | — | — | |
| NoisyNet-A3CNetwork architecture=NoisyNet2017.06 | 94 | 347 | 18 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IMPALAArchitecture=shallow, Training Strategy=experts, Evaluation Protocol=Up to 30 no-ops2018.02 | 93.2 | 466.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DQNNetwork architecture=Baseline2017.06 | 83 | 319 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| A3CNetwork architecture=Baseline2017.06 | 80 | 293 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IMPALAArchitecture=deep, Training Strategy=multi-task, Evaluation Protocol=Up to 30 no-ops2018.02 | 59.7 | 176.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| A3CArchitecture=shallow, Training Strategy=experts, Evaluation Protocol=Up to 30 no-ops2018.02 | 54.9 | 285.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |