Aim

Develop novel algorithms for multi-agent reinforcement learning (MARL), enabling multiple autonomous assets to learn how to interact and collaborate towards a specified task.

Objectives

1. Explore the current state-of-the-art in MARL research, identify key limitations.
2. Develop novel MARL algorithms to address identified limitations.
3. Implement and evaluate the algorithms in challenging environments and learning tasks.

Description

Multi-agent reinforcement learning (MARL) algorithms are designed to train the decision policies of multiple autonomous agents acting in a shared environment to accomplish specified tasks. MARL can be applied to a broad range of applications, including many tasks related to defence and security scenarios, including: multiple drones or mobile assets patrolling in a specified area, distributed sensing and intelligence gathering in areas of interest, and human-agent teamwork scenarios. This project will identify key limitations in current MARL capabilities, such as relating to scalability, learning stability, and communication capabilities; and will develop novel algorithms to address such limitations.

References

1. Stefano V. Albrecht, Filippos Christianos, and Lukas Schäfer. Multi-Agent Reinforcement Learning: Foundations and Modern Approaches. MIT Press, 2024. https://www.marl-book.com