Abstract: 

Reinforcement Learning (RL) is the study of how agents (ought to) behave under regimes of rewards and punishments. It forms the core of Prescriptive Analytics, i.e. methods that try to identify optimal sequences of actions. It is also closely aligned to Game Theory, Operations Research, Planning and Machine Learning, with most advances coming from cross-pollination across different aligned fields (e.g. Deep Reinforcement Learning, Neuro-evolution). Advances in Reinforcement Learning are often seen as advances in Artificial Intelligence. In recent years we have seen an influx of algorithms that tackle Reinforcement Learning and Game Theoretic problems. These algorithms operate on certain sets of assumptions about the environmental structure and agent capabilities. This talk will provide a unifying framework one could use to identify which algorithms are better suited for what environments, what are some research ""black spots"" that have not been explored enough and what new algorithms could be tried by mixing and matching. More specifically, we will explore algorithmic bias, method suitability and practical results along the following axes:

a) The agent's sensory abilities (e.g. partial or full observability) b) The type of action (e.g. discrete, continuous, textual) c) The number of agents that populate an environment d) The availability (and exploitation) of forward and inverse models e) The use of global and local function approximators f) Ways to explore and exploit available information. Some of the themes explored in this talk have been discussed in Vodopivec, Tom, Spyridon Samothrakis, and Branko Ster. "On Monte Carlo tree search and reinforcement learning." Journal of Artificial Intelligence Research 60 (2017): 881-936.

Bio: 

Dr Spyros Samothrakis is Assistant Director for the Institute for Analytics and Data Science at the University of Essex. Spyros’ research interests include reinforcement learning, neural networks and causality.

He obtained his PhD from the University of Essex (2014) and has published numerous papers on the topics above. He is currently working closely with industrial partners, helping bridge the gap between data science concepts and business applications.