On Pathologies in KL-Regularized Reinforcement Learning from Expert Demonstrations

Tim G. J. Rudner, Cong Lu, Michael A. Osborne, Yarin Gal, Yee Whye Teh

in 34th Advances of Neural Information Processing Systems (NeurIPS), 2021.

Abstract

KL-regularized reinforcement learning from expert demonstrations has proved successful in improving the sample efficiency of deep reinforcement learning algorithms, allowing them to be applied to challenging physical real-world tasks. However, we show that KL-regularized reinforcement learning with behavioral policies derived from expert demonstrations suffers from hitherto unrecognized pathological behavior that can lead to slow, unstable, and suboptimal online training. We show empirically that the pathology occurs for commonly chosen behavioral policy classes and demonstrate its impact on sample efficiency and online policy performance. Finally, we show that the pathology can be remedied by specifying non-parametric behavioral policies and that doing so allows KL-regularized RL to significantly outperform state-of-the-art approaches on a variety of challenging locomotion and dexterous hand manipulation tasks.

Link to Paper

Link to Code

Behaviors Learned with N-PPAC

door-binary-v0

We show that the predictive variance of the non-parametric behavioral prior is crucial to the success of KL-regularized RL by performing an ablation study where the behavioral prior's predictive mean is fixed to a GP mean and the predictive variance is set to that of a parametric behavioral policy (a neural network trained via MLE) and that of a non-parametric behavioral policy (a GP posterior predictive), respectively.

PArametric (NN)

Non-Parametric (GP)

pen-binary-v0

N-PPAC @ 0 TIMESTEPS

Pre-training from the offline prior gives 70% success rate which is then fine-tuned by N-PPAC.

N-PPAC @ 100k TIMESTEPS

Online fine-tuning improves the success rate rapidly after 100k timesteps.

For the following three MuJoCo environments, we show the online performance of KL-regularized RL with a non-parametric behavioral policy after a small number of time steps, showing rapid adaptation from and improvement to the expert demonstrations.