Variational Inverse Control with Events: A General Framework for Data-Driven Reward Definition

TL;DR

This paper introduces VICE, a framework that learns reward functions from goal state samples rather than full demonstrations, enabling reinforcement learning in complex, high-dimensional environments.

Contribution

The paper presents VICE, a novel variational inverse control method that relaxes the need for full demonstrations, focusing on event-based goal achievement in reinforcement learning.

Findings

Effective in continuous control tasks with high-dimensional observations

Requires only goal state samples, not full demonstrations

Outperforms traditional inverse reinforcement learning methods

Abstract

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.