English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Reinforcement learning algorithms are typically designed to learn aperformant policy that can repeatedly and autonomously complete a task, usuallystarting from scratch. However, in many real-world situations, the goal mightnot be to learn a policy that can do the task repeatedly, but simply to performa new task successfully once in a single trial. For example, imagine a disasterrelief robot tasked with retrieving an item from a fallen building, where itcannot get direct supervision from humans. It must retrieve this object withinone test-time trial, and must do so while tackling unknown obstacles, though itmay leverage knowledge it has of the building before the disaster. We formalizethis problem setting, which we call single-life reinforcement learning (SLRL),where an agent must complete a task within a single episode withoutinterventions, utilizing its prior experience while contending with some formof novelty. SLRL provides a natural setting to study the challenge ofautonomously adapting to unfamiliar situations, and we find that algorithmsdesigned for standard episodic reinforcement learning often struggle to recoverfrom out-of-distribution states in this setting. Motivated by this observation,we propose an algorithm, $Q$-weighted adversarial learning (QWALE), whichemploys a distribution matching strategy that leverages the agent's priorexperience as guidance in novel situations. Our experiments on severalsingle-life continuous control problems indicate that methods based on ourdistribution matching formulation are 20-60% more successful because they canmore quickly recover from novel states.

You Only Live Once: Single-Life Reinforcement Learning