Author(s): Tan Pengshi Alvin

Originally published on Towards AI.

Multi-agent reinforcement learning with 3 MADQN frameworks on the ma-gym’s “Switch4” environment
Image by Nik Korba on Unsplash

With the introduction of Function Approximation methods in Part 4, we are ready to solve Reinforcement Learning problems in more complicated environments with dynamic and continuous states. In this article, we will further extend this knowledge to solve Multi-Agent Reinforcement Learning problems in these environments. To refresh our understanding of Function Approximation and Deep Q-Networks for single-agent settings, check out the previous article (Part 4) below:

Reinforcement Learning with continuous state spaces and gradient descent techniques

pub.towardsai.net

So far, we have dealt with a single decision-making agent that aims to act optimally under uncertainty in an environment to produce maximum long-term reward relative to a task. In Multi-Agent Reinforcement Learning, however, there is an added layer of complexity in which more than one agent is present, and these agents could either be cooperative or adversarial, or a mix of both. In these settings, each agent’s state includes observation not only about itself but also about other agents’ positions and their activities.

In training adversarial multi-agent models, the goal is for all competing agents to discover the optimal strategies against opposing parties by reaching a game state called Nash Equilibrium. As such, adversarial Multi-Agent Reinforcement Learning can be adapted and… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI