Research And Application Of Interval Random Weights Neural Network

In the real world,many datasets are often inaccurate which we got in the production process,such as scientific experiments or industry.Therefore it's an important problem to study how to describe inaccurate information and model.Grain is a new concept that is used primarily to describe imprecise information.It derives a number of methods,of which set theory and interval analysis have been proved with an effective result and have been widely studied.On the other hand,interval neural network is also an effective method to solve the problem of inaccurate data modeling which apply interval data presenting inaccurate data and use neural network to complete the task of modeling.The random weight neural network adopts the manner whose input weight randomly given,and the output layer weight is trained by the least squares method.This method effectively overcomes the problem that the BP neural network is slow to converge,easily into the local optimum and sensitive to the learning rate.Combining the interval theory with stochastic weight neural network to solve the problem of inaccurate information modeling has high research value both in theoretical and practical aspects.In this thesis,a random weight neural network is selected and combined with the theory of interval analysis.After a large number of references and books are reviewed,the structure of random weights neural networks with intervals are designed based on the random vector function-link network and the learning algorithm is also studied.The main research work of this paper includes the following contents.In this thesis,the interval random weight neural networks are divided into two categories,and the forward weights and inverse learning processes of these two kinds of intervals are deduced.The model make full use of the weight learning method of RVFL network,and the weights of the hidden layer to the output layer are adjusted by the principle of least squares method.Through the simulation experiment,it is proved that the interval RVFL network structure is reasonable,and the convergence performance is good.In order to achieve a better effect of the interval random weight neural network proposed in this paper,we propose a structural improvement for the proposed interval RVFL network.First,using the structure of parallel layer perceptrons and the learning method of interval RVFL network propose the PLP-IRVFLN(parallel layer perceptrons--interval RVFL network weight),which by increasing the training parameters and joining the input impact to improve the network performance;the second is using the integrated neural network structure,and adopting the negative correlation learning algorithm to obtain better performance.Finally,the simulation experiment verifies the superiority of the network.Combined with the random weight network proposed in this paper,we put forward two aspects of the application.First aspect is modeling for UBBE problems with unknown errors,bounded problem is handled by interval and modeling process is achieved by interval RVFL network.Making improvements in the weight of the network for learning methods,we effectively solve two cases that the error bounded is known or unknown.Second,the random weight neural network is applied to the classification problem,and the interval RVFL network proposed in this paper is applied to the modeling of the greenhouse environmental quality evaluation system,and the good classification effect is obtained.