AI Paper Atlas

Research map / Reinforcement learning / Imitation learning

Identifiability and generalizability from multiple experts in Inverse Reinforcement Learning

Paul Rolland, Luca Viano, Norman Schuerhoff, Boris Nikolov, Volkan Cevher

NeurIPS 2022 · Reinforcement learning › Imitation learning · approach: Theory of inverse RL · 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
Use two experts in the randomly generated environmentsIRL on Random-Matrices · Random-Matricesreward identifiability up to a constant: rank 35 = 2|S| -1 = 35Improved
Use experts with different transition noise levelsIRL on Gridworld · Gridworldreward identifiability up to a constant: α^1=0.4 and α^2=0.2Improved
Vary discount factors while keeping the transition dynamics fixedIRL on Gridworld · Gridworldreward identifiability up to a constantImproved
Observe additional experts in environments with different wind distributionsIRL on WindyGridworld · WindyGridworldgeneralizability: 4 expertsImproved
Use two experts in a setting with an exogenous productivity shockIRL on Strebulaev-Whited · Strebulaev-Whitedreward identifiability up to a constant: 552 < 2|S| -1 = 799Hurt
Use the recovered reward to train in a third environment with σ_ε=0.6IRL on Strebulaev-Whited · Strebulaev-Whitedgeneralizability: σ_ε=0.6Improved
Restrict the reward to linear combinations of three known featuresIRL on Strebulaev-Whited · Strebulaev-Whitedexact reward identifiability: 803 = 2|S| + dImproved

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