Summary of Policy-regularized Offline Multi-objective Reinforcement Learning, by Qian Lin et al.
Policy-regularized Offline Multi-objective Reinforcement Learning
by Qian Lin, Chao Yu, Zongkai Liu, Zifan Wu
First submitted to arxiv on: 4 Jan 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper explores the use of offline trajectory data to train a policy for multi-objective reinforcement learning (RL) tasks. Building upon existing methods for single-objective offline RL, the authors extend these approaches to handle multi-objective scenarios. A key challenge in this setting is the preference-inconsistent demonstration problem, which the authors address through two proposed solutions: filtering out inconsistent demonstrations and adopting regularization techniques with high policy expressiveness. Additionally, the paper introduces a preference-conditioned scalarized update method that enables simultaneous learning of multiple policies using a single policy network, reducing computational costs. Finally, the authors present Regularization Weight Adaptation to dynamically determine suitable regularization weights for target preferences during deployment. Empirical results demonstrate the effectiveness of this approach on various multi-objective datasets. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps machines learn to make good decisions when given incomplete information. It’s like trying to figure out how a robot should behave in different situations without actually seeing it move. The researchers developed new ways to teach robots using data collected beforehand, which is really useful because it saves time and energy. They also came up with clever solutions to overcome some tricky problems that can happen when teaching robots multiple tasks at once. Overall, the paper shows that machines can learn to make good decisions based on past experiences, which has many potential applications. |
Keywords
* Artificial intelligence * Regularization * Reinforcement learning
