Deep Coordination Graphs

TL;DR

Deep Coordination Graphs (DCG) provide a flexible, neural network-based approach for multi-agent reinforcement learning, enabling efficient training and effective solutions to complex tasks like predator-prey and StarCraft II micromanagement.

Contribution

The paper introduces Deep Coordination Graphs, a novel method that combines factorization of joint value functions with deep neural networks for scalable multi-agent reinforcement learning.

Findings

DCG effectively solves predator-prey tasks with overgeneralization issues.

DCG achieves strong performance on StarCraft II micromanagement tasks.

Parameter sharing and low-rank approximations improve sample efficiency.

Abstract

This paper introduces the deep coordination graph (DCG) for collaborative multi-agent reinforcement learning. DCG strikes a flexible trade-off between representational capacity and generalization by factoring the joint value function of all agents according to a coordination graph into payoffs between pairs of agents. The value can be maximized by local message passing along the graph, which allows training of the value function end-to-end with Q-learning. Payoff functions are approximated with deep neural networks that employ parameter sharing and low-rank approximations to significantly improve sample efficiency. We show that DCG can solve predator-prey tasks that highlight the relative overgeneralization pathology, as well as challenging StarCraft II micromanagement tasks.