A variable structure neural network model and its applications

The paper presents a neural network model called variable structure neural network model (VSNNM), also named improved multilayer perceptron (IMLP). In view of the back propagation algorithm (BPA), it is a time-consuming algorithm and its learning time is about O(n/sup 3/). In contrast to BPA, the speed of the learning algorithm proposed is much faster. Taking XOR for example, the speed of the learning algorithm is about 30 times faster than BPA. Moreover, hard limiters as the activation functions of neurons and only integer connection weights are used in VSNNM. Both the number of hidden layers and the number of hidden neurons in each hidden layer are variable, along with the demands of problems, but they are always kept minimum. Thus, this will greatly facilitate actual hardware implementation of training VSNNM. Hence, considering its speed and the number of neurons, the VSNNM is a successful attempt.<<ETX>>