The Research Of Input Space Self-Optimal Neural Network And Its Application On Identifying The Vehicle Model

The neural network is bionic model that bases on physiology. In recent years, it has beenapplied in more and more areas, among which pattern recognition is the most successful.But the majority of the training processes of neural network only optimize the parameter ofthe network. Under the condition that the topology structure is fixed. Also there is a littlenetwork that has the thought of structure optimize network, that is the cascade-correlationlearning architecture and network pruning, but to optimize input space is paid little attentionin literature.This dissertation puts forward the new conception of structure self-optimal neuralnetwork, and mainly introduces the definition of input space self-optimal as well as the basicthought and basic algorithm of input space self-optimal neural network. The dissertationcompile program for the input space self-optimal neural network with Matlab6.5, and appliesit to identifying the vehicle model on highway, which can increase classification correctnessand decreases the calculation of neural network, thus solves the actual problem.Compared with the traditional neural network, the main contributions of the dissertationconsists are:(1) Put forward the concept of input space self-optimal neural network. A.E.Fahlman andC.Lebiere put forward "The cascade-correlation learning architecture" in 1990, Y. Le Cun etc.put forward "Optimal brain damage" in 1990, and B.Hassibi etc.put forward "Optimal brainsurgeon" in 1992. In fact, these kinds of networks supply the thought of neural network'sstructure self-optimal, this dissertation defines the network, which has this function asstructure self-optimal neural network that is shortened as self-optimal neural network.(2)The dissertation puts forward a new type of self-optimal neural network that is inputspace self-optimal neural network. It optimizes not only the parameter of the neural networkbut also the input space With continuously adjusting the input space of neural network, thestructure of neural network is adjusted continuously, and causes the result of neural network'sclassification optimize.(3)This research combines input space self-optimal neural network with actualapplication, and proves the thought of input space self-optimal neural network is practicableand accurate.