Learning Multi-Agent Collaborations With Decomposition
summary
Three papers present concrete algorithms, proofs and results: CollabQ decomposes joint Q into a local Q-alone and a co-collaboration term learned from local reward assignments, runs on StarCraft II multi-agent challenge and raises win rates and robustness to swapped or missing units (ad hoc team play). A policy change density theorem enables a depth-first joint policy search for imperfect-information games and improves bridge bidding, reportedly beating WR5 in double-dummy evaluation. B-BOLT (vs RND) uses episodic delta visitation intrinsic rewards, solves 12 hardest MiniGrid tasks and shows strong NetHack gains.
AI-written summary · about
chapters (27)
- 0:00Intro
- 1:22Research Directions
- 2:31Multi-Agent Reinforcement Learning
- 2:58Research Target
- 4:54The Assigned Reward for each agenti
- 5:36Reward Assignment Problems
- 6:40Approximate decentralized perceived rewa
- 8:21Using end-to-end Training instead of getting
- 10:20Starcraft II Multi-Agent Challenge
- 11:26performs well in ad hoc team play
- 12:04Ablation Studies
- 13:17An illustrative Example
- 15:31Another example
- 15:51Communication Game
- 16:32Dependency between policies
- 17:47Policy-change Density
- 19:02Value Changes w.rt Localized Policy Change
- 19:37JPS Joint Policy Search
- 20:27Performance
- 21:02Contract Bridge Bidding
- 22:43Environment with Sparse Reward
- 23:51Random Network Distillations (RND)
- 25:07Issues in RND
- 25:30Multi-Corridor Problems
- 27:38MiniGrid
- 28:34Pure Exploration
- 29:046 Tasks in NetHack
related talks
15mMulti-Agent Reinforcement Learning In Stochastic Games: From Alphago To Robust Control
10mConditional Imitation Learning for Multi-Agent Games
15mSequential Decision Making Problems in Multi-Agent Environments
24mDeepCoMP: Multi-agent Reinforcement Learning for Multi-cell Selection in 5G and Beyond
59mComplementarity and Delegation in Multi-Agent Models
31mFrom Single AI to Multi AI Systems
21mScalable agent-based models for optimized policy design: applications to the ec...
49mML: Teaching machine to learn from its own mistakes (...)