COMPARATIVE STUDY OF THE MICROPROCESSOR IMPLEMENTED NEURAL CONTROLLERS

摘要

Neural controllers are used to give solution to control problems that cannot be described with complex mathematical models. This paper compares the performances of two classes of neural controllers, cascade and multi layer perceptron networks, implemented in the HC711 micro controller in emulating a given control surface. Both layered and fully interconnected neural networks were Investigated. The Elliot activation function was used for microprocessor implementation, instead of tangent hyperbolic activation function. This leads to neural networks implementation with shorter code and faster control operation. All structures were implemented and optimized using Stuttgart Neural Network Simulator software for Windows. This software has the Elliot function included in algorithm. The error buck propagation algorithm was used to train the networks. This algorithm uses a single pattern vector with one set of inputs and outputs for each cycle of training. The purpose was to obtain the lowest control error while keeping the neural network architecture as simple as possible. The results showed better performance for fully connected network in terms of computational power. The cascade network produced better results than the multi layer perceptron network using fewer neurons, that meaning its overall size is smaller with less weights and connections to calculate. The paper presents as conclusions the obtained control surfaces of neural controllers, code lengths, computation time and squared error sum comparison.