Summary of Offline Safe Reinforcement Learning Using Trajectory Classification, by Ze Gong et al.
Offline Safe Reinforcement Learning Using Trajectory Classification
by Ze Gong, Akshat Kumar, Pradeep Varakantham
First submitted to arxiv on: 19 Dec 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 proposes an approach to offline safe reinforcement learning (RL) that learns a policy to generate desirable trajectories while avoiding undesirable ones. The existing methods in offline safe RL rely on cost constraints at each time step, which can result in overly conservative policies or the violation of safety constraints. To address this issue, the authors partition the pre-collected dataset of state-action trajectories into desirable and undesirable subsets based on their rewards and safety. Then, they learn a policy that generates desirable trajectories and avoids undesirable ones using a classifier learned from the dataset. This approach bypasses the computational complexity and stability issues associated with min-max objectives used in existing methods. The authors also show theoretical connections to learning paradigms involving human feedback. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making sure machines learn good behaviors without taking risks online. Right now, most ways of doing this rely on setting a limit at each step based on a global limit. This can make the machine too careful or not safe enough. The authors suggest a new way to teach the machine what good and bad behaviors look like by partitioning the data into “good” and “bad” sets. Then, they use a special tool (classifier) learned from this data to guide the machine’s decisions. This approach avoids some big problems with current methods. They also show how their method is related to other ways of teaching machines. |
Keywords
» Artificial intelligence » Reinforcement learning
