Reinforcement learning typically lets an agent query any state–action pair to learn a good policy. But in many real-world settings, the agent only observes a single continuing trajectory of experience. A new arXiv preprint, Average-Reward Policy Identification from a Single Continuing Trajectory, examines this harder problem.
The abstract notes that unlike methods that can freely sample, an agent learning from a single trajectory must spend its limited interactions carefully. The paper focuses on identifying an optimal policy under the average-reward objective, which is relevant for continuing tasks without a fixed horizon.
Because the source is a brief abstract, the specific method and results are not yet clear. The significance lies in formalizing a realistic constraint and asking whether optimal policies can still be identified from one ongoing stream of experience.