Deep Multi-Agent RL for Complex Swarm Behavior

Published in Extended Abstract Submitted for EEML Application. The preliminary results outlined are very much work-in-progress and originated during a lab rotation from January to March of 2019 at Henning Sprekeler’s Lab at the TU Berlin., 2019

Recommended citation: Lange, Robert Tjarko. (2019). "Multi-Agent RL for Complex Swarm Dynamics." Do not cite. Preliminary work.

Download paper here.

Download poster here.

This working paper is the result of my ECN Berlin lab rotation at Henning Sprekeler’s lab at TU Berlin. We use tools from Multi-Agent Reinforcement Learning to learn collective behavior in a gradient-driven fashion. More specifically, we introduce a multi-agent environment with a reward function design based on phenomenological observations of fish schools. These include survival, alignment, attraction and repulsion. Checkout the Swarm MARL environment in OpenAI Gym style.