AI Paper Atlas

Research map / Reinforcement learning / Imitation learning

Error Bounds of Imitating Policies and Environments

Tian Xu, Ziniu Li, Yang Yu

NeurIPS 2020 · Reinforcement learning › Imitation learning · approach: Offline imitation from imperfect data · 17 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
Match state-action occupancy distributions with GAIL instead of cloning actions with BCPolicy imitation · Infinite-horizon MDPspolicy value gap bound: BC: 1/(1-γ)^2; GAIL: 1/(1-γ)Improved
Learn the transition model with a GAIL-style distribution-matching objective instead of one-step BCEnvironment imitation · Policy evaluation in learned environmentsmodel-bias term in policy evaluation error bound: BC: 1/(1-γ)^2; GAIL: 1/(1-γ)Improved
Use GAIL rather than BC to learn the transition modelEnvironment imitation · Three MuJoCo environmentspolicy evaluation errorImproved
Use gradient penalty with a moderate coefficient rather than no penalty or a very large coefficient (ablation)GAIL · MuJoCo policy imitationperformanceImproved
Reduce the number of discriminator network layers (ablation)GAIL · MuJoCo policy imitationperformanceNo clear effect
Provide BC with many expert demonstrations rather than scant demonstrationsBC · MuJoCo policy imitationperformance relative to expertImproved
Use BC rather than GAIL as the transition-model learnerEnvironment imitation · HalfCheetah-v2policy evaluation error during learningHurt

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