In this work, a torque controller for a variable rotational speed wind turbine has been modelled using Reinforcement Learning and considering the Optimal Torque - Maximum Power Point Tracking problem as one of optimization. The reward optimization function is designed as a non-linear function depending mainly on the rotor power variation. Based on this, an optimal action (electromagnetic torque variation) regulates the turbine rotational speed. A simulated 1.5 MW three bladed wind turbine operation is managed by the torque controller. It keeps the turbine working at optimal operational conditions after a successful training process, which is carried out using the Proximal Policy Optimization algorithm. For the controller training, the turbine confronts constant and then randomly staggered wind speed behaviour. Time series of rotor angular speed, torque and power are presented. Our result...