AI Paper Atlas

Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Pragmatically Learning from Pedagogical Demonstrations in Multi-Goal Environments

Hugo Caselles-Dupré, Olivier Sigaud, Mohamed CHETOUANI

NeurIPS 2022 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Boundedly rational agent modeling · 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 both a pedagogical teacher and pragmatic learner instead of a naive teacher and literal learnerGANGSTR teacher-learner training · Fetch Block Stacking, 100 demos per goalGIA x GRA: 0.15 -> 0.62Improved
Use both a pedagogical teacher and pragmatic learner instead of a naive teacher and literal learnerGANGSTR teacher-learner training · Fetch Block Stacking, 500 demos per goalGIA x GRA: 0.62 -> 0.80Improved
Use both a pedagogical teacher and pragmatic learner instead of a naive teacher and literal learnerGANGSTR teacher-learner training · Fetch Block Stacking, 1000 demos per goalGIA x GRA: 0.70 -> 0.82Improved
Train the teacher to infer its own goals and reward correctly inferred demonstrationsNaive teacher · Fetch Block Stacking teacher demonstrationsOwn Goal Inference Accuracy: 83.4% ± 0.2 -> 95.6% ± 0.1Improved
Train the teacher to produce demonstrations it can disambiguateNaive teacher · Fetch Block Stacking, 500 ambiguous situationsAmbiguity Score: 64% ± 2 -> 9% ± 1Improved
Train the learner to infer its own goals using BGI and reward correct inferencesLiteral learner · Fetch Block StackingOwn Goal Inference Accuracy: 12.4% relative differenceImproved
Replace BGI with a separately trained LSTM goal prediction network (ablation)BGI goal inference · Fetch Block Stacking, pedagogical demos, 100 demos per goalGIA: 69.2% -> 46.2%Hurt
Replace BGI with a separately trained LSTM goal prediction network (ablation)BGI goal inference · Fetch Block Stacking, pedagogical demos, 500 demos per goalGIA: 82.2% -> 70.0%Hurt
Replace BGI with a separately trained LSTM goal prediction network (ablation)BGI goal inference · Fetch Block Stacking, naive demos, 100 demos per goalGIA: 61.3% -> 38.4%Hurt
Use demonstrations only and discard learner-collected experience (ablation)Modified SQIL learner · Fetch Block Stacking, pedagogical demos, 100 demos per goalGIA: 69.2±4.2 -> 5.3 ± 1.5Hurt
Add behavioral-cloning updates on demonstrations (ablation)Modified SQIL learner · Fetch Block Stacking, pedagogical demos, 100 demos per goalGIA: 69.2±4.2 -> 2.3 ± 1.8Hurt
Use original SQIL instead of modified SQIL (ablation)Modified SQIL learner · Fetch Block Stacking, pedagogical demos, 100 demos per goalGIA: 69.2±4.2 -> 59.1 ± 2.4Hurt

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