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

A key challenge in training generally-capable agents is the design oftraining tasks that facilitate broad generalization and robustness toenvironment variations. This challenge motivates the problem setting ofUnsupervised Environment Design (UED), whereby a student agent trains on anadaptive distribution of tasks proposed by a teacher agent. A pioneeringapproach for UED is PAIRED, which uses reinforcement learning (RL) to train ateacher policy to design tasks from scratch, making it possible to directlygenerate tasks that are adapted to the agent's current capabilities. Despiteits strong theoretical backing, PAIRED suffers from a variety of challengesthat hinder its practical performance. Thus, state-of-the-art methods currentlyrely on curation and mutation rather than generation of new tasks. In thiswork, we investigate several key shortcomings of PAIRED and propose solutionsfor each shortcoming. As a result, we make it possible for PAIRED to match orexceed state-of-the-art methods, producing robust agents in several establishedchallenging procedurally-generated environments, including a partially-observedmaze navigation task and a continuous-control car racing environment. Webelieve this work motivates a renewed emphasis on UED methods based on learnedmodels that directly generate challenging environments, potentially unlockingmore open-ended RL training and, as a result, more general agents.

Stabilizing Unsupervised Environment Design with a Learned Adversary