Sample Efficient Imitation Learning for Continuous Control

The goal of imitation learning (IL) is to enable a learner to imitate expert behavior given expert demonstrations. Recently, generative adversarial imitation learning (GAIL) has shown significant progress on IL for complex continuous tasks. However, GAIL and its extensions require a large number of environment interactions during training. In real-world environments, the more an IL method requires the learner to interact with the environment for better imitation, the more training time it requires, and the more damage it causes to the environments and the learner itself. We believe that IL algorithms could be more applicable to real-world problems if the number of interactions could be reduced. In this paper, we propose a model-free IL algorithm for continuous control. Our algorithm is made up mainly three changes to the existing adversarial imitation learning (AIL) methods – (a) adopting off-policy actor-critic (Off-PAC) algorithm to optimize the learner policy, (b) estimating the state-action value using off-policy samples without learning reward functions, and (c) representing the stochastic policy function so that its outputs are bounded. Experimental results show that our algorithm achieves competitive results with GAIL while significantly reducing the environment interactions.