Applied Mathematics Colloquium: Dr Thu Nguyen

Abstract: Multi-agent reinforcement learning (MARL) is concerned with the sequential decision-making problems where multiple agents operate in a shared stochastic environment and aim to maximise some long-term rewards through interacting with the environment and other agents. These problems cover a wide range of possible applications with the potential to impact many domains such as robotics, health care, smart grids, smart transportation systems, finance, self-driving cars, and many more. Compared with single agent RL, the multi-agent aspect of MARL creates many additional challenges for developing and analyzing efficient learning algorithms for those systems. In this talk, we will give a selective overview of MARL with focus on networked agents MARL. In this context, agents are considered heterogeneous rather than homogeneous; they may have different reward functions. Furthermore, the agents can only share/exchange information with their neighbours via a time-varying and spare communication network. This setting is believed to be more realistic for many engineering systems and real-world applications.