AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Improving Generalization in Meta-RL with Imaginary Tasks from Latent Dynamics Mixture

Suyoung Lee, Sae-Young Chung

NeurIPS 2021 · Reinforcement learning › Visual and representation RL · approach: Task structure and curricula for meta-RL · 3 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 latent dynamics mixture imaginary tasks to RL^2/variBAD-style agentLDM · Gridworld M_testtest goals reached: 19 of 27Improved
apply input dropout to reward decoder (p_drop 0.0 to higher rates) (ablation)LDM · Gridworldtest returnImproved
increase decoder dropout rate to 0.9 (ablation)LDM · Gridworldtest returnHurt
no decoder input dropout (p_drop=0.0) (ablation)LDM · Gridworldtest return vs RL^2No clear effect
apply dropout to policy encoder state input (ablation)RL^2 · Gridworldtraining stabilityHurt
share one encoder for policy and VAE with decoder dropout (ablation)LDM · Gridworldpolicy training stabilityHurt
increase extrapolation level beta from 1.0 to larger values (ablation)LDM · Gridworld-extrapolationreturn on M_test2 vs M_test1Mixed
use beta=2 instead of 1 (ablation)LDM · Ant-goaltest returnImproved
vary number of mixture workers (2,3,4,5 of 16) (ablation)LDM · Half-cheetah-velocitytest returnMixed
mix observations and rewards in observation space (Mixreg) instead of latent spaceRL^2 · Gridworld M_testtest returnMixed
Mixreg on MuJoCo tasksRL^2 · MuJoCo M_evaltest returnHurt

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