Conditional Imitation Learning for Multi-Agent Games

on YouTube ↗10m

" summary

Proposes a conditional imitation framework that learns a latent low-rank subspace over ego and partner strategies and uses tensor decomposition to infer a new partner’s strategy from offline paired trajectories. Builds policies that interpolate in that subspace, tests on bandits, particle, Hanabi and an Overcooked user study with 50 humans, and reports better adaptation than meta, multitask and latent baselines. Finds partner behaviors fit a small basis (a rank 4 tensor explains the bandit data) and the adapted policy improves post-adaptation.

AI-written summary · about

" chapters (6)

1. 0:00Start
2. 0:40The Goal of Our Work Is To Study the Framework of Coordinating with Different Partners
3. 3:28Multi-Agent Imitation Learning Setup
4. 7:10Collaborative Contextual Bandit Environment
5. 8:32The Overcooked Environment
6. 9:31Summary

" related talks

• 28mPreference Learning from Minimal Human Feedback for InteractiveArdam Vik · ODSC West 2024
• 30mLearning Multi-Agent Collaborations With DecompositionYuandong Tian · Reinforcement Learning from Batch Data and Simulation
• 56mOf Moments and Matching: Trade-offs and Treatments in Imitation LearningSteven Wu · Adversarial Approaches in Machine Learning 2022
• 59mComplementarity and Delegation in Multi-Agent ModelsJon Kleinberg · Agency in Collaborative Learning 2026
• 15mSequential Decision Making Problems in Multi-Agent EnvironmentsSumitra Ganesh · Eighth Annual Industry Day
• 48mPractical Model-based Algorithms for Reinforcement Learning and Imitation Learning, with...Tengyu Ma · Frontiers of Deep Learning 2019
• 29mCombining Imitation + Reinforcement Learning to win the Bot Bowl competitionNicola Pezzotti · PyData Eindhoven 2021
• 37mMultiagent Reinforcement Learning: Rollout and Policy IterationDimitri Bertsekas · Reinforcement Learning from Batch Data and Simulation