Data-Efficient Offline Reinforcement Learning with Approximate Symmetries

Giorgio Angelotti, Nicolas Drougard, Caroline P. C. Chanel

Abstract: The performance of Offline Reinforcement Learning (ORL) models in Markov Decision Processes (MDPs) is heavily contingent upon the quality and diversity of the training data. This research furthers the exploration of expert-guided symmetry detection and data augmentation techniques by considering approximate symmetries in discrete MDPs, providing a fresh perspective on data efficiency in the domain of ORL. We scrutinize the adaptability and resilience of these established methodologies in varied stochastic environments, featuring alterations in transition probabilities with respect to the already tested stochastic environments. Key findings from these investigations elucidate the potential of approximate symmetries for the data augmentation process and confirm the robustness of the existing methods under altered stochastic conditions. Our analysis reinforces the applicability of the established symmetry detection techniques in diverse scenarios while opening new horizons for enhancing the efficiency of ORL models.

External IDs:dblp:conf/icaart/AngelottiDC23a