Uncertainty-Based Offline Reinforcement Learning with Diversified Q-Ensemble

Offline reinforcement learning (offline RL), which aims to find an optimal policy from a previously collected static dataset, bears algorithmic difficulties due to function approximation errors from out-of-distribution (OOD) data points. To this end, offline RL algorithms adopt either a constraint or a penalty term that explicitly guides the policy to stay close to the given dataset. However, prior methods typically require accurate estimation of the behavior policy or sampling from OOD data points, which themselves can be a non-trivial problem. Moreover, these methods under-utilize the generalization ability of deep neural networks and often fall into suboptimal solutions too close to the given dataset. In this work, we propose an uncertainty-based offline RL method that takes into account the confidence of the Q-value prediction and does not require any estimation or sampling of the data distribution. We show that the clipped Q-learning, a technique widely used in online RL, can be leveraged to successfully penalize OOD data points with high prediction uncertainties. Surprisingly, we find that it is possible to substantially outperform existing offline RL methods on various tasks by simply increasing the number of Q-networks along with the clipped Q-learning. Based on this observation, we propose an ensemble-diversified actor-critic algorithm that reduces the number of required ensemble networks down to a tenth compared to the naive ensemble while achieving state-of-the-art performance on most of the D4RL benchmarks considered.

PDF Abstract NeurIPS 2021 PDF NeurIPS 2021 Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Gym halfcheetah-random D4RL EDAC Normalized Average Return 28.4 # 1
Adroid door-cloned D4RL SAC-N Normalized Average Return -0.3 # 2
Gym halfcheetah-random D4RL SAC-N Normalized Average Return 28 # 2
Gym halfcheetah-medium D4RL EDAC Normalized Average Return 65.9 # 2
Gym halfcheetah-expert D4RL SAC-N Normalized Average Return 105.2 # 2
Gym halfcheetah-medium-expert D4RL EDAC Normalized Average Return 106.3 # 2
Gym halfcheetah-medium-replay D4RL SAC-N Normalized Average Return 63.9 # 1
Gym halfcheetah-medium-replay D4RL EDAC Normalized Average Return 61.3 # 2
Gym halfcheetah-full-replay D4RL EDAC Normalized Average Return 84.6 # 1
Adroid pen-human D4RL EDAC Normalized Average Return 52.1 # 1
Adroid door-human D4RL EDAC Normalized Average Return 10.7 # 1
Adroid pen-cloned D4RL SAC-N Normalized Average Return 64.1 # 2
Adroid hammer-cloned D4RL SAC-N Normalized Average Return 0.2 # 2
Adroid relocate-human D4RL EDAC Normalized Average Return 0.1 # 1
Adroid hammer-human D4RL EDAC Normalized Average Return 0.8 # 1
Adroid hammer-human D4RL SAC-N Normalized Average Return 0.3 # 2
Adroid relocate-cloned D4RL EDAC Normalized Average Return 0 # 1
Adroid relocate-cloned D4RL SAC-N Normalized Average Return 0 # 1
Adroid door-cloned D4RL EDAC Normalized Average Return 9.6 # 1
Adroid hammer-cloned D4RL EDAC Normalized Average Return 0.3 # 1
Adroid pen-cloned D4RL EDAC Normalized Average Return 68.2 # 1
Adroid relocate-human D4RL SAC-N Normalized Average Return -0.1 # 2
Adroid door-human D4RL SAC-N Normalized Average Return -0.3 # 2
Adroid pen-human D4RL SAC-N Average Return 9.5 # 1
Gym halfcheetah-full-replay D4RL SAC-N Normalized Average Return 84.5 # 2
Gym halfcheetah-medium-expert D4RL SAC-N Normalized Average Return 107.1 # 1
Gym halfcheetah-expert D4RL EDAC Normalized Average Return 106.8 # 1
Gym halfcheetah-medium D4RL SAC-N Normalized Average Return 67.5 # 1

Methods