Skip to content

instadeepai/amld-africa-2021

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 

Repository files navigation

AMLD Africa 2021 Workshop

Hands-on Multi-Agent Reinforcement Learning using Mava

This workshop aims to provide an overview of multi-agent reinforcement learning theory and its applications.

MARL systems have gained more interest in the last years as it extends the decision-making capabilities of single-agent reinforcement learning to larger and more complex systems.

In real world applications, most of the problems are designed as a multi-agent decision systems: This is the case for managing a fleet of autonomous vehicles, or augmenting the capacity of train management systems, etc. - the list is long.

Having said that, this workshop will guide you from the theory and foundations of RL & MARL to hands-on experience with introductory examples using the mava framework.

RL foundations

We start by giving an introduction to the RL field, explaining the value based and policy methods, and providing the ideas behind some algorithms like DQN, DDPG, D4PG.

Deep RL in practice

In this section, we talk about Multi agent RL via the Deutsche Bahn use case. We provide intuitions on the different MARL training architectures as well as details on their machinery.

Mava: Open-Source Framework for Multi-Agent Reinforcement Learning

We finish the presentation part by explaining the mava framework, how it works, and the panoply of features it provides.


In the second part of the workshop, we will walk through two notebooks and showcase the flexibility of mava in implementing and training multi agent systems.

Part I: Flatland notebook

In this notebook, we use MADQN from mava to learn the route of trains using the open sourced flatland environnement.

For more information regarding flatland, we recommend visiting their official documentation.

Part II: Mava overview notebook

In this notebook, we showcase the mava features and capabilities. We demonstrate the simplicity of using different networks, architectures, agents, and an easy-to-use integrated loggers like tensorboard to benchmark results.

We also provide insights on the distributed training, which is one of the essential building blocks of mava that enhance scaling properties.