MARLlib

A multi-agent reinforcement learning benchmark for research and industry

Despite the fast development of multi-agent reinforcement learning (MARL) algorithms, there is a lack of commonly-acknowledged baseline implementation and evaluation benchmarks.

An urgent need for MARL researchers is to develop a unified benchmark suite, similar to the role of RLlib in single-agent RL, that can support both high-performance MARL implementations and replicable evaluations in various testing environments.

We introduce a counterpart of RLlib in MARL studies: Multi-Agent RLlib

Enter MARLlib

URL: https://github.com/Replicable-MARL/MARLlib

Four benefits of MARLlib

Algorithms

MARLlib includes by far the most comprehensive list of MARL algorithms, covering different categories on either the game type (cooperative, competitive, mixed), the action space (discrete, continues, multi-discrete), or the decision mode (turn-based, simultaneous).

Environments

MARLlib introduces a unified interface for MARL implementations.

10 diverse environments are available, with new ones easy to be incorporated.

Easy-to-read

MARLlib decouples algorithms, neural architectures, and environments, thereby offering great flexibility for systematic benchmarking with enhanced debugging tools and easy-to-read codes.

Replicable and reloadable

MARLlib provides replicable and reloadable performance results across tens of environments, each with detailed hyper-parameter settings and training logs

Benchmark Overview

MARLlib unify both the multi-agent environment and MARL algorithms in one framework

Algorithms Dataflow

In MARLlib, we mainly implement and unify three types of algorithms that belong to independent learning (IL), centralized critic (CC), and value decomposition (VD) considering their broad topic coverage.

Environment Interface

In MARLlib environments, agents are not required to act simultaneously. No transition data is shared among agents except terminal signal done. Multiple environments/tasks are supported. See the figure below for details.

Page updated

Google Sites

Report abuse