AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Goal-Aware Cross-Entropy for Multi-Target Reinforcement Learning

Kibeom Kim, Min Whoo Lee, Yoonsung Kim, JeHwan Ryu, Minsu Lee, Byoung-Tak Zhang

NeurIPS 2021 · Label-efficient and robust learning › Recognition and long-tail · approach: Loss design and classifier theory · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add GACE goal-state classification lossA3C · V3success ratio: 33.45 ± 4.72 -> 52.58 ± 4.09Improved
Add GDAN attention using the goal-discriminator query (ablation)A3C+GACE · V3success ratio: 52.6% -> 78.2%Improved
Add GACE goal-state classification lossSAC · R2 Unseensuccess ratio: 53.4 ± 6.9 -> 63.0 ± 9.0Improved
Add GDAN attention using the goal-discriminator query (ablation)SAC+GACE · R2 Unseensuccess ratio: 63.0 ± 9.0 -> 73.3 ± 5.8Improved
Add GACE goal-state classification lossSAC · R1 sample efficiency at 63.1% successSEI (%): 485.41Improved
Add GDAN attention to GACE (ablation)SAC+GACE · R1 sample efficiency at 63.1% successSEI (%): 485.41 -> 398.57Hurt
Freeze GACE discriminator weights after 10K or 20K updates instead of continuing GACE training (ablation)A3C+GACE · V1learning speedHurt
Replace GDAN attention with simple concatenation of the state encoding and discriminator query (ablation)GACE ActorCritic · R1success ratio: 81.5 ± 10.0Hurt
Train separate single-target policies rather than one multi-target policy (ablation)SAC · R1success ratio: 47±11% vs 63.1±6.9%Hurt
Use SAC+CURL instead of SAC on the unseen checkered-background taskSAC · R2 Unseensuccess ratio: 53.4 ± 6.9 -> 36.6 ± 3.4Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy