Tutorial 05 - Orchestrating Autonomous Agents: Reinforcement Learning for Hierarchical Planning with COACH
Abstract
This tutorial will provide an introduction to COACH - a suite of Python tools for recasting Gymnasium and PettingZoo-compatible Multi-Agent Reinforcement Learning (MARL) problems as orchestration-style planning problems. Traditional Reinforcement Learning (RL) focuses on training low level agents to interact with an environment in a high frequency feedback loop. Once policies have been trained, human direction becomes an orchestration problem, especially with large numbers of agents. COACH provides tools for researchers to bridge this gap: given any simulation compatible with Gymnasium or PettingZoo, trained agents can be set up to interface with a director agent who tackles high level scheduling, policy selection, or coordination for generating autonomously executed plans. COACH was created in association with the US Air Force Research Laboratory’s Autonomy Capability Team (ACT3).
Resources
About the authors
Nate Bade is an award-winning educator and former teaching professor and program director of the MS in Applied Mathematics (MSAM) program at Northeastern University. He specialized in project based education and designed the MSAM’s graduate machine learning program. He is currently a Senior Data Scientist at Mobius Logic and works in coordination with ACT3 on hierarchical methods in multi-agent reinforcement learning and automated planning.