Abstract: A method for training a computerized mechanical device, comprising: receiving data documenting actions of an actuator performing a task in a plurality of iterations; calculating using the data a neural network dataset and used for performing the task; gathering in a plurality of reward iterations a plurality of scores given by an instructor to a plurality of states, each comprising at least one sensor value, while a robotic actuator performs the task according to the neural network; calculating using the plurality of scores a reward dataset used for computing a reward function; updating at least some of the neural network's plurality of parameters by receiving in each of a plurality of policy iterations a reward value computed by applying the reward function to another state comprising at least one sensor value, while the robotic actuator performs the task according to the neural network; and outputting the updated neural network.