Structure Identification And Modified Algorithm Of Multilayer Feedforward Neural Network

Research on neural network has been developing rapidly in recent years. Neural network's features display its great capacity in solving highly nonlinear and indefinite control system. It has important significance in all kinds of science areas. In neural network models, multilayer feedforward neural network is a kind of neural network, which is in use widely and has been used in many science areas such as function approximation, image processing, pattern recognition, self-adaptive control.Although multilayer feedforward neural networks are widely used, a few problems are still remained unsolved, which are in dire need of solution. In these issues, the selection of network topology is one of the most troublesome problems. This problem has been an obstacle, which interferes with the application of neural network. It is well known that the precondition of using neural network is to determine the neural network topology. However, for a long time, the structure of neural networks is determined by trial-and-error method in most cases due to the lack of theoretical guidance. This results in intensive computation and low efficiency. This paper briefly introduces the basic concept of neural network, systematically looks back the development of the structure identification methods of multilayer feedforward neural network, and discusses the merit and limitation of the structure identification methods of multilayer feedforward neural network. By the systematic analysis of mechanism of the hidden neurons, this paper shows that vectors, which consist of the output of every neuron in same hidden layer corresponding to different samples, should be nonlinearly correlated with the optimized structure. A method to determine the tolerance and a new algorithm to determine the MFN's topology are presented in this thesis. The simulation results show that the topology of neural network has been simplied. Not only the error goal is met with, but also the generalization capability is improved.The BP algorithm is a kind of algorithm which is in use widely in multilayer neural network, because the BP algorithm is a kind of gradient descended searching algorithm in essence, it has weaknesses such as slow convergent speed, liable to fell into error function's local extreme value point, insurable to find global extreme valuepoint for multi-modal and non-differential function in larger searching zone, which restrict neural network's application in all fields. So the modification to BP algorithm is necessary. A modified training algorithm based on reverse transformation and singular value decomposition is proposed, it can improve the convergent speed and precision. The simulation results confirm its efficiency.