Lorenzo Croissant

Topic; Continuous time reinforcement learning with stochastic control. In control of discrete time problems in continuous state/action spaces, it is generally expensive to compute the value function or an optimal policy. This is a problem for RL tasks where repeated computation of policies is needed. For problems with correct scaling, it is possible to approximate via a limiting process, in particular a diffusion, which yields much simpler objects. This opens the way for new tools and methods using control theory. Approximation results are derived in terms of value function, the value of an optimal policy in the diffusion limit, and numerical resolutions thereof. Finite horizon and ergodic control settings studied. Results are leveraged in Reinforcement Learning theory in the form of algorithms and their regret bounds.