Skip to content

LUKELIEM/team_marl

Repository files navigation

Team-based Multi-agent Reinforcement Learning

Research Report

Abstract

In multi-agent reinforcement learning (MARL), differentiating between agent intelligence and organization intelligence may hold the key to major breakthroughs.

This project separates the encoding of agent intelligence from organization intelligence. Agents are programmed with a simple naive algorithm, but they are organized under teams and provided with US versus THEM context. The organization intelligence is separately encoded in the team’s culture, which determines how team rewards are doled out to its agents on top of the environmental reward they gather during training.

With the separation of agent and organization intelligences, the methodology becomes mathematically and computationally simple. It can scale easily with the number of agents and teams and it enables teams of agents to achieve a wide range of desired results and behaviors with only slight changes to the team culture and no change to the agents’ policy algorithm.

The new approach enables teams of agents to easily exceed the performance of agents trained under “state-of-the art” MARL algorithms. In addition, the use of team reward in culture can lead to agent specialization, which enables a team of specialized agents to build a dominating strategy to a game which is previously intransitive to multiple individual agents.

Installation

pip install -r requirements.txt