Deep reinforcement learning control of white-light continuum generation

Deep Reinforcement Learning (deep RL) is a branch of Machine Learning dealing with the solution of optimization problems, usually formalized as Markov Decision Processes [1]. At discrete temporal steps, an agent takes an action on the system receiving back a reward that depends on the state reached by the system. The goal of the agent is to determine an optimal policy, i.e. a map between states and actions, to maximise the future rewards, by directly experiencing and sampling the environment without any a-priori knowledge of the system. Among the various optimization algorithms, deep RL is particularly versatile thanks to the underlying neural networks [2] , a powerful universal approximator.