Acgail: Imitation Learning About Multiple Intentions With Auxiliary Classifier Gans

As an important solution to decision-making problems, imitation learning learns expert behavior from example demonstrations provided by experts, without the necessity of a predefined reward function as in reinforcement learning. Traditionally, imitation learning assumes that demonstrations are generated from single latent expert intention. One promising method in this line is generative adversarial imitation learning (GAIL), designed to work in large environments. It can be thought as a model-free imitation learning built on top of generative adversarial networks (GANs). However, GAIL fails to learn well when handling expert demonstrations under multiple intentions, which can be labeled by latent intentions. In this paper, we propose to add an auxiliary classifier model to GAIL, from which we derive a novel variant of GAIL, named ACGAIL, allowing label conditioning in imitation learning about multiple intentions. Experimental results on several MuJoCo tasks indicate that ACGAIL can achieve significant performance improvements over existing methods, e.g., GAIL and InfoGAIL, when dealing with label-conditional imitation learning about multiple intentions.