This thesis proposes a reinforcement learning–based procedure for tuning the gains of the Super-Twisting sliding mode controller. The gain selection problem is formulated as a Markov Decision Process, and a deterministic actor–critic architecture is employed to learn gain values directly from the system state, without modifying the structure of the underlying control law. The methodology is developed within the scope and limitations of reinforcement learning methods and its effectiveness is demonstrated on a nonlinear benchmark system. The evaluation includes single initial-condition experiments together with an analysis over multiple initial conditions, allowing the behavior of the learned gains to be examined both at specific operating points and across a range of scenarios.