École Nationale Supérieure de
Techniques Avancées

Abstract

In this bibliographic study, we aim to solve finite-horizon optimal control problems with or without constraint using the principles of reinforcement learning for infinite-horizon problems. Optimal control problem is about finding the control to apply to the system to optimize a criterion function (for example, the search for the trajectory of a rocket which will minimize the fuel cost). We compared the performance of three reinforcement learning algorithms OP(Optimistic Planning), SOP(Simultaneous Optimistic Planning) and SOPMS(Simultaneous Optimistic Planning Multiple Step) applied to finite-horizon optimal control problems through numerous numerical simulations (applied to differential equation and partial differential equation). We concluded that the algorithm OP is less efficient (result and CPU time) than the SOP and SOPMS algorithms. keywords : optimal control, optimistic planning, reinforcement learning, tree structure