On the 9th of November 2009, Matthieu Geist received the PhD degree es Mathematics from the University Paul Verlaine of Metz.
Reinforcement learning is the response of machine learning to the problem of optimal control. In this paradigm, an agent learns to control an environment by interacting with it. It receives evenly a numeric reward (or reinforcement signal), which is a local information about the quality of the control. The agent objective is to maximize a cumulative function of these rewards, generally modelled as a so-called value function. A policy specifies the action to be chosen in a particular configuration of the environment to be controlled, and thus the value function quantifies the quality of this policy. This paragon is very general, and it allows taking into account many applications. In this manuscript, we apply it to a gas flow management problem in the iron and steel industry. However, its application can be quite difficult. Notably, if the environment description is too large, an exact representation of the value function (or of the policy) is not possible. This problem is known as generalization (or value function approximation) : on the one hand, one has to design algorithms with low computational complexity, and on the other hand, one has to infer the behaviour the agent should have in an unknown configuration of the environment when close configurations have been experimented. This is the main problem we address in this manuscript, by introducing a family of algorithms inspired from Kalman filtering.