Multi-task Reinforcement Learning in Reproducing Kernel Hilbert Spaces via Cross-learning

TL;DR

This paper introduces a multi-task reinforcement learning approach using cross-learning in reproducing kernel Hilbert spaces, enabling quick adaptation to related tasks and environments, with theoretical convergence guarantees and practical navigation experiments.

Contribution

It proposes a novel cross-learning framework for multi-task RL in RKHS, providing convergence guarantees and a method for rapid policy adaptation across related tasks.

Findings

The approach produces a central policy useful for quick adaptation.

The method converges to a near-optimal solution with high probability.

Agents successfully navigate new environments with unseen obstacles.

Abstract

Reinforcement learning (RL) is a framework to optimize a control policy using rewards that are revealed by the system as a response to a control action. In its standard form, RL involves a single agent that uses its policy to accomplish a specific task. These methods require large amounts of reward samples to achieve good performance, and may not generalize well when the task is modified, even if the new task is related. In this paper we are interested in a collaborative scheme in which multiple agents with different tasks optimize their policies jointly. To this end, we introduce cross-learning, in which agents tackling related tasks have their policies constrained to be close to one another. Two properties make our new approach attractive: (i) it produces a multi-task central policy that can be used as a starting point to adapt quickly to one of the tasks trained for, in a situation when the agent does not know which task is currently facing, and (ii) as in meta-learning, it adapts to environments related but different to those seen during training. We focus on continuous policies belonging to reproducing kernel Hilbert spaces for which we bound the distance between the task-specific policies and the cross-learned policy. To solve the resulting optimization problem, we resort to a projected policy gradient algorithm and prove that it converges to a near-optimal solution with high probability. We evaluate our methodology with a navigation example in which agents can move through environments with obstacles of multiple shapes and avoid obstacles not trained for.