This May London Reinforcement Learning will be bringing you two brilliant talks with topics such as how can reinforcement learning agents learn to cooperate and solve problems and more! Register now!
How can reinforcement learning agents learn to cooperate and solve problems? In this talk, I will introduce the field of multi-agent reinforcement learning and various approaches that have been taken to address this challenge. I will then introduce our own framework, called Feudal Multi-Agent Hierarchies in which a 'manager' agent learns to communicate sub-goals to multiple, simultaneously-operating 'worker' agents. Finally, I will discuss the future outlook, both of our method and multi-agent reinforcement learning more generally.
Sanjeevan Ahilan is a PhD student at the Gatsby Computational Neuroscience Unit working with Peter Dayan. His primary interests are in multi-agent and hierarchical reinforcement learning, and its connections to diverse fields such as neuroscience, psychology and economics. He previously completed a Masters in Neuroscience at UCL (2014) and a undergraduate degree in Natural Sciences at Cambridge University (2013), specialising in Experimental and Theoretical Physics.
How can reinforcement learning algorithms benefit from knowledge learned from previous tasks? In this talk, we will dive into recent works in transfer learning, multi-task learning and meta-learning. We will look at several transfer approaches and compare their differences. After the talk, you will understand the current state-of-the-art of knowledge transfer in reinforcement learning and their applications.
Jin Cong Ho is a final year computer science student at the University of Nottingham, specializing in machine learning. He recently completed his undergraduate dissertation on multi-task reinforcement learning. His primary interest is to deploy the latest research advances in the larger production environment to create value.