Complementarity and Delegation in Multi-Agent Models
summary
Builds Maia by retraining Leela’s head on billions of Lichess games so Maia mimics human moves (Maia 1100 becomes the site’s most-played bot and its argmax behaves like a ~1400 player); fine-tunes Leela into a PartnerBot that raises team score to ~0.61265 vs Leela-Maia teams and avoids rook-sacrifice traps; proves a two-by-two delegation model where optimal delegation equals row/column variance minimization (problem is NP‑hard); and models a weather-app feature choice tradeoff (sort by a_i^2 − (a_i−h_i)^2), showing a fixed-sequence learning policy and at most n choose 2 + k switches.
AI-written summary · about
chapters (7)
- 0:00Human-AI delegation models
- 3:35Superhuman AI in chess
- 18:04Training Maia and PartnerBot
- 25:30Theoretical delegation model
- 37:15Weather app informativeness
- 48:29Future directions in delegation
- 59:07Q&A
related talks
30mLearning Multi-Agent Collaborations With Decomposition
10mConditional Imitation Learning for Multi-Agent Games
31mFrom Single AI to Multi AI Systems
31mLayla El Asri, Research Scientist, Maluuba, NYC 2017
32mThe Road to General AI
20mWhat can AI do for me?: evaluating machine learning interpretations in cooperative play
42mReasonable AI
29mCombining Imitation + Reinforcement Learning to win the Bot Bowl competition