Online and Offline Reinforcement Learning by Planning with a Learned Model

Learning efficiently from small amounts of data has long been the focus of model-based reinforcement learning, both for the online case when interacting with the environment and the offline case when learning from a fixed dataset. However, to date no single unified algorithm could demonstrate state-of-the-art results in both settings. In this work, we describe the Reanalyse algorithm which uses model-based policy and value improvement operators to compute new improved training targets on existing data points, allowing efficient learning for data budgets varying by several orders of magnitude. We further show that Reanalyse can also be used to learn entirely from demonstrations without any environment interactions, as in the case of offline Reinforcement Learning (offline RL). Combining Reanalyse with the MuZero algorithm, we introduce MuZero Unplugged, a single unified algorithm for any data budget, including offline RL. In contrast to previous work, our algorithm does not require any special adaptations for the off-policy or offline RL settings. MuZero Unplugged sets new state-of-the-art results in the RL Unplugged offline RL benchmark as well as in the online RL benchmark of Atari in the standard 200 million frame setting.

PDF Abstract NeurIPS 2021 PDF NeurIPS 2021 Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Atari Games Atari 2600 Alien MuZero (Res2 Adam) Score 70192.35 # 5
Atari Games Atari 2600 Amidar MuZero (Res2 Adam) Score 1197.38 # 22
Atari Games Atari 2600 Assault MuZero (Res2 Adam) Score 33292.22 # 6
Atari Games Atari 2600 Asterix MuZero (Res2 Adam) Score 862406.65 # 5
Atari Games Atari 2600 Asteroids MuZero (Res2 Adam) Score 476412 # 4
Atari Games Atari 2600 Atlantis MuZero (Res2 Adam) Score 1137475.12 # 10
Atari Games Atari 2600 Bank Heist MuZero (Res2 Adam) Score 27219.8 # 1
Atari Games Atari 2600 Battle Zone MuZero (Res2 Adam) Score 178716.9 # 6
Atari Games Atari 2600 Beam Rider MuZero (Res2 Adam) Score 333077.44 # 3
Atari Games Atari 2600 Berzerk MuZero (Res2 Adam) Score 2705.82 # 13
Atari Games Atari 2600 Bowling MuZero (Res2 Adam) Score 131.65 # 10
Atari Games Atari 2600 Boxing MuZero (Res2 Adam) Score 100 # 1
Atari Games Atari 2600 Breakout MuZero (Res2 Adam) Score 758.04 # 12
Atari Games Atari 2600 Centipede MuZero (Res2 Adam) Score 874301.64 # 4
Atari Games Atari 2600 Chopper Command MuZero (Res2 Adam) Score 5989.55 # 28
Atari Games Atari 2600 Crazy Climber MuZero (Res2 Adam) Score 158541.58 # 17
Atari Games Atari 2600 Defender MuZero (Res2 Adam) Score 557200.75 # 8
Atari Games Atari 2600 Demon Attack MuZero (Res2 Adam) Score 143838.04 # 6
Atari Games Atari 2600 Double Dunk MuZero (Res2 Adam) Score 23.91 # 7
Atari Games Atari 2600 Enduro MuZero (Res2 Adam) Score 2365.81 # 8
Atari Games Atari 2600 Fishing Derby MuZero (Res2 Adam) Score 73.94 # 4
Atari Games Atari 2600 Freeway MuZero (Res2 Adam) Score 33.87 # 13
Atari Games Atari 2600 Frostbite MuZero (Res2 Adam) Score 374769.76 # 3
Atari Games Atari 2600 Gopher MuZero (Res2 Adam) Score 122882.5 # 5
Atari Games Atari 2600 Gravitar MuZero (Res2 Adam) Score 8006.93 # 3
Atari Games Atari 2600 HERO MuZero (Res2 Adam) Score 37234.31 # 8
Atari Games Atari 2600 Ice Hockey MuZero (Res2 Adam) Score 41.66 # 7
Atari Games Atari 2600 James Bond MuZero (Res2 Adam) Score 28626.23 # 9
Atari Games Atari 2600 Kangaroo MuZero (Res2 Adam) Score 13838 # 15
Atari Games Atari 2600 Krull MuZero (Res2 Adam) Score 72570.5 # 7
Atari Games Atari 2600 Kung-Fu Master MuZero (Res2 Adam) Score 116726.96 # 7
Atari Games Atari 2600 Montezuma's Revenge MuZero (Res2 Adam) Score 2500 # 15
Atari Games Atari 2600 Ms. Pacman MuZero (Res2 Adam) Score 70659.76 # 2
Atari Games Atari 2600 Name This Game MuZero (Res2 Adam) Score 101197.71 # 2
Atari Games Atari 2600 Phoenix MuZero (Res2 Adam) Score 815728.7 # 7
Atari Games Atari 2600 Pitfall! MuZero (Res2 Adam) Score 0 # 4
Atari Games Atari 2600 Pong MuZero (Res2 Adam) Score 20.95 # 16
Atari Games Atari 2600 Private Eye MuZero (Res2 Adam) Score 100 # 43
Atari Games Atari 2600 Q*Bert MuZero (Res2 Adam) Score 94906.25 # 7
Atari Games Atari 2600 River Raid MuZero (Res2 Adam) Score 171673.78 # 2
Atari Games Atari 2600 Road Runner MuZero (Res2 Adam) Score 531097 # 6
Atari Games Atari 2600 Robotank MuZero (Res2 Adam) Score 100.59 # 6
Atari Games Atari 2600 Seaquest MuZero (Res2 Adam) Score 999659.18 # 6
Atari Games Atari 2600 Skiing MuZero (Res2 Adam) Score -30000 # 3
Atari Games Atari 2600 Solaris MuZero (Res2 Adam) Score 5132.95 # 13
Atari Games Atari 2600 Space Invaders MuZero (Res2 Adam) Score 3645.63 # 23
Atari Games Atari 2600 Star Gunner MuZero (Res2 Adam) Score 154548.26 # 10
Atari Games Atari 2600 Surround MuZero (Res2 Adam) Score 9.9 # 3
Atari Games Atari 2600 Tennis MuZero (Res2 Adam) Score 0 # 20
Atari Games Atari 2600 Time Pilot MuZero (Res2 Adam) Score 424011.16 # 4
Atari Games Atari 2600 Tutankham MuZero (Res2 Adam) Score 347.99 # 7
Atari Games Atari 2600 Up and Down MuZero (Res2 Adam) Score 634898.18 # 6
Atari Games Atari 2600 Venture MuZero (Res2 Adam) Score 1731.47 # 12
Atari Games Atari 2600 Video Pinball MuZero (Res2 Adam) Score 865543.44 # 8
Atari Games Atari 2600 Wizard of Wor MuZero (Res2 Adam) Score 100096.6 # 5
Atari Games Atari 2600 Yars Revenge MuZero (Res2 Adam) Score 219838.09 # 7
Atari Games Atari 2600 Zaxxon MuZero (Res2 Adam) Score 154131.86 # 5
Continuous Control cartpole.swingup MuZero Unplugged Return 594.3 # 2
Continuous Control cheetah.run MuZero Unplugged Return 869.9 # 2
Continuous Control finger.turn_hard MuZero Unplugged Return 759 # 2
Continuous Control fish.swim MuZero Unplugged Return 681.6 # 1
Continuous Control humanoid.run MuZero Unplugged Return 643.1 # 1
Continuous Control manipulator.insert_ball MuZero Unplugged Return 659.2 # 1
Continuous Control manipulator.insert_peg MuZero Unplugged Return 556 # 1
Continuous Control walker.stand MuZero Unplugged Return 887.2 # 2
Continuous Control walker.walk MuZero Unplugged Return 949.5 # 2

Methods