Reinforcement Learning for complex sequential descision problems. 

The field of reinforcement learning began to be explored more than 35 years ago as an approach for the control of agents interacting with an external environment, modeled as a Markov Decision Process. Since then, impressive progress has been made, particularly as "deep" reinforcement learning methods have been implemented to solve increasingly complex problems. In this talk, I will describe some recent work carried out in the SCIDyn team (Inrae-MIAT) around reinforcement learning, in "complex" problems (beyond the control of "classic" MDPs). In particular, I will describe how we used deep RL for the control of "non-Markovian" processes, in the context of a medical application. I will also describe uses of deep RL methods for the control of systems in a multi-agent framework, in the areas of public policy design or the fight against illegal poaching.