Design of robot controller based on evolutionary algorithms and neural networks

This paper addresses application of evolutionary algorithms for training of ANN based controller for robotic arm. The first part presents the existing obstacles when trying to improve neural network controller trained based on the existing controller. In the second part, evolutionary algorithms are used for that purpose. Weights and biases of ANN controller are coded into a vector-variable, and genetic algorithm (GA) and evolutionary strategy (ES) are applied to improve controller's performance, with results of measurements on the robotic arm presented. In order to improve performance of GA, a modification of mutation probability in was also introduced, using a couple of functions by means of which mutation probability for a chromosome was determined.