What question did this study set out to answer?

This research aims to develop a new algorithm for imitation learning that operates without expert actions.

February 4, 2022Open Access

Versatile Offline Imitation from Observations and Examples via Regularized State-Occupancy Matching

Key Points

This research aims to develop a new algorithm for imitation learning that operates without expert actions.
Developed the SMODICE algorithm based on state-occupancy matching techniques.
Applied SMODICE to three distinct offline imitation learning settings.
Evaluated performance across gridworld environments and high-dimensional benchmarks.
SMODICE significantly outperforms previous state-of-the-art methods in all evaluated settings.
Demonstrated successful implementation of imitation from observations, even with mismatched experts.
Provided effective solutions for both state-occupancy matching problems and example-based reinforcement learning.

Abstract

We propose State Matching Offline DIstribution Correction Estimation (SMODICE), a novel and versatile regression-based offline imitation learning (IL) algorithm derived via state-occupancy matching. We show that the SMODICE objective admits a simple optimization procedure through an application of Fenchel duality and an analytic solution in tabular MDPs. Without requiring access to expert actions, SMODICE can be effectively applied to three offline IL settings: (i) imitation from observations (IfO), (ii) IfO with dynamics or morphologically mismatched expert, and (iii) example-based reinforcement learning, which we show can be formulated as a state-occupancy matching problem. We extensively evaluate SMODICE on both gridworld environments as well as on high-dimensional offline benchmarks. Our results demonstrate that SMODICE is effective for all three problem settings and significantly outperforms prior state-of-art.

Bookmark

View Full Paper