Enhancing Multi-Objective Offline RL with Adaptive Preference Integration
Keywords: Reinforcement Learning, Multi-objective Optimization, Offline Reinforcement Learning, Decision Transformer
TL;DR: We introduce a new transformer based architecture which provides suitable preference integration with decision transformer for multi-objective reinforcement learning tasks.
Abstract: Multi-objective reinforcement learning (MORL) is crucial for real-world applications where multiple conflicting goals must be optimized, such as in healthcare or autonomous systems. Offline MORL extends these benefits by using pre-collected datasets, allowing for effective learning without continuous interaction with the environment. However, existing offline MORL algorithms often struggle with scaling across large preference spaces and handling unknown preferences during evaluation. To address these challenges, we propose the Preference-Attended Multi-Objective Decision Transformer (PA-MODT), a novel architecture that integrates a preference-attention block with a modular transformer structure. This design enables effective generalization over different preferences and trajectories, providing a more robust approach to generating optimal Pareto fronts. We tested PA-MODT on five D4MORL datasets with millions of trajectories representing various objectives and found that it consistently outperforms existing models, achieving Pareto fronts that align closely with behavioral policy. This demonstrates PA-MODT's potential to effectively manage complex multi-objective reinforcement learning tasks.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2847
Loading