Variational Inverse Control with Events: A General Framework for Data-Driven Reward Definition
This addresses the practical problem of reward specification for RL practitioners, particularly in domains with high-dimensional observations, though it builds incrementally on inverse reinforcement learning approaches.
The paper tackles the challenge of reward function design in reinforcement learning by proposing VICE, a framework that requires only goal state samples instead of full expert demonstrations. The method achieved effective performance on continuous control tasks with high-dimensional image observations where reward specification is difficult.
The design of a reward function often poses a major practical challenge to real-world applications of reinforcement learning. Approaches such as inverse reinforcement learning attempt to overcome this challenge, but require expert demonstrations, which can be difficult or expensive to obtain in practice. We propose variational inverse control with events (VICE), which generalizes inverse reinforcement learning methods to cases where full demonstrations are not needed, such as when only samples of desired goal states are available. Our method is grounded in an alternative perspective on control and reinforcement learning, where an agent's goal is to maximize the probability that one or more events will happen at some point in the future, rather than maximizing cumulative rewards. We demonstrate the effectiveness of our methods on continuous control tasks, with a focus on high-dimensional observations like images where rewards are hard or even impossible to specify.