Finding optimal architecture and weights using evolutionary least square based learning

摘要: In this paper, we present a novel idea of implementing growing neural network architecture using an evolutionary least square based algorithm. This paper focuses mainly on the following aspects, such as heuristics updating weights algorithm, finding number hidden neurons for two layer feed forward multilayered perceptron (MLP), stopping criteria algorithm and finally comparisons results with other traditional methods searching optimal or near solution in multidimensional complex search space comprising weight variables. We applied our proposed XOR data set, 10 bit odd parity problem many real bench mark set like handwriting dataset from CEDAR breast cancer, heart disease UCI ML repository. The comparison results, classification accuracy time complexity are discussed. also discuss issues probabilistic starting point method address problems involving fitness breaking.