Title: High-Confidence Performance Bounds for Multi-Task Reinforcement Learning

Abstract: Multi-task reinforcement learning aims to develop generalist policies capable of handling a variety of tasks. Despite substantial advancements in this field over recent years, current methodologies seldom offer formal performance assurances. Such guarantees are essential for deploying these policies in environments where safety is paramount. This work introduces a method for calculating high-confidence performance bounds for multi-task policies applied to tasks that were not part of the training set. Specifically, we propose a novel generalization bound that integrates two components: (i) per-task lower confidence limits derived from a finite number of rollouts, and (ii) task-level generalization estimates based on a finite set of sampled tasks. This combination results in a robust, high-confidence guarantee for new tasks sampled from an arbitrary, unknown distribution. Our evaluation across leading multi-task RL algorithms demonstrates that these guarantees are both theoretically rigorous and practically informative, even at realistic sample sizes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC