SHF: Small: High-Performance Multi-Agent Reinforcement Learning

George Washington University, Washington DC

Investigators

Abstract

Artificial Intelligence (AI) has rapidly become a critical domain with applications in autonomous driving, robotics, aerospace, healthcare, and others. For a machine (AI agent) to closely mimic human behavior and operate effectively, it should possess the capabilities of robust decision making and learning simultaneously as it operates in the environment. Multi-Agent Reinforcement Learning (MARL) is a promising research area that can model and control multiple distributed decision-making AI agents. However, recent studies have shown that the MARL algorithms suffer from inefficiencies that can severely limit their adoption in real-world systems. These problems occur due to complexities in decision-making processes arising from having to observe and act upon a large number of events present in the environment, along with the growth in the number of AI agents needed to interact with each other. To ameliorate the learning efficiency and scalability issues of MARL algorithms, the project investigators adopt a novel interdisciplinary solution approach, harnessing computer architecture, machine-learning theory and optimization. Specifically, the project will seek techniques to improve neural-network throughput, to efficiently manage the state-action space in a dynamic fashion and to scalably encode states and observations of a large and varying number of agents. A hardware-software co-design approach is adopted to accelerate the concurrent optimization of software and hardware layers. The research outcomes of this project will significantly enhance the adoption of MARL frameworks in real-world applications and positively impact university curricular development and the computing industry. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.