Too Many Cooks: Coordinating Multi-Agent Collaboration Through Inverse Planning

Abstract

AAMAS 2020, May 9–13, Auckland, New Zealand Too Many Cooks: Coordinating Multi-agent Collaboration Through Inverse Planning∗ Extended Abstract Rose E. Wang∗ Sarah A. Wu∗ James A. Evans MIT MIT U. Chicago rewang@mit.edu sarahawu@mit.edu jevans@uchicago.edu Joshua B. Tenenbaum David C. Parkes Max Kleiman-Weiner MIT Harvard Harvard, MIT, & Diffeo jbt@mit.edu parkes@eecs.harvard.edu maxkleimanweiner@fas.harvard.edu ABSTRACT work in parallel when sub-tasks can be carried out individually, Humans collaborate in dynamic and flexible ways. Collaboration (B) Cooperation: agents work together on the same sub-task when requires agents to coordinate their behavior on the fly, sometimes required or most efficient, (C) Spatio-temporal movement: agents jointly solving a single task together and other times dividing it up avoid getting in each others way while working separately or to- into sub-tasks to work on in parallel. We develop Bayesian Delega- gether. tion, a learning mechanism for decentralized multi-agent coordi- We study decentralized Markov decision processes (Dec-MDPs) nation that enables agents to rapidly infer the sub-tasks that other with a partial order of sub-tasks over object-object interactions agents are working on by inverse planning. These inferences enable [3, 4]. We use this formalism to develop a test suite based on simple agents to determine, in the absence of communication, whether kitchen cooking tasks where sub-tasks are different parts of a non- to plan jointly with others or work on complementary sub-tasks. linear recipe. The left column of Figure 1 shows the different kitchen We test this model in a suite of decentralized multi-agent envi- layouts and the top row shows the different recipes which can be ronments inspired by cooking problems. To succeed, agents must composed together. Agents move simultaneously in any cardinal coordinate both their high-level plans (sub-task) and their low-level direction or can stay still. The kitchens h