[2403.01636] Pattern Environment friendly Myopic Exploration By way of Multitask Reinforcement Studying with Numerous Duties

0
18


Obtain a PDF of the paper titled Pattern Environment friendly Myopic Exploration By way of Multitask Reinforcement Studying with Numerous Duties, by Ziping Xu and 4 different authors

Obtain PDF
HTML (experimental)

Summary:Multitask Reinforcement Studying (MTRL) approaches have gained growing consideration for its large purposes in lots of necessary Reinforcement Studying (RL) duties. Nevertheless, whereas latest developments in MTRL idea have targeted on the improved statistical effectivity by assuming a shared construction throughout duties, exploration–a essential facet of RL–has been largely missed. This paper addresses this hole by displaying that when an agent is skilled on a sufficiently various set of duties, a generic policy-sharing algorithm with myopic exploration design like $epsilon$-greedy which might be inefficient basically will be sample-efficient for MTRL. To the most effective of our data, that is the primary theoretical demonstration of the “exploration advantages” of MTRL. It could additionally make clear the enigmatic success of the large purposes of myopic exploration in follow. To validate the function of range, we conduct experiments on artificial robotic management environments, the place the varied activity set aligns with the duty choice by automated curriculum studying, which is empirically proven to enhance sample-efficiency.

Submission historical past

From: Ziping Xu [view email]
[v1]
Solar, 3 Mar 2024 22:57:44 UTC (1,668 KB)
[v2]
Wed, 6 Mar 2024 04:34:01 UTC (1,668 KB)



Supply hyperlink

LEAVE A REPLY

Please enter your comment!
Please enter your name here