Mixed Uncertainty Modular Neural Networks And Universities Benefit Forecast

Raise the marginal benefit is the ultimate goal of universities asset and resources allocation optimization process, the annual budget of the Ministry of Education invested into various colleges and universities, which is to ensure that schools engaged in teaching and researchs most directly, and is the foundation of talents training, teachers team building and the social service infrastructure. With the rapid development of higher education, how to evaluate the configuration of university resources, is the primary problem of using funds rationaly and maximizing outputs. Factors that influence benefit often have complex nonlinear relationship with benefits.Factors affecting the effectiveness of the evaluation, not just only can be identified by the fuzziness and randomness, variables often contain vague uncertainty and stochastic uncertainty, we called hybrid uncertainty variables. This chapter discusses variables mixing properties of random uncertainty and fuzzy uncertainty, we are discussing to build "hybrid uncertainty neural networks" system of fuzzy random characteristics, useing Mamdani type fuzzy logic system model of neural networks system modeling, and established an integrated system block neural network output modular approach, the main work is as follows:1, The single structure of the neural network for samples has a bad generalization ability, slow convergence and easy to fall into the minimum value problem. Faced with a complex mixture of events, we use fuzzy K-means clustering algorithm to construct modular neural networks, its fault tolerance, robustness, computing power has greatly been improved.2, Variables contains random vague uncertainty, are proposed taking into account variables fuzziness and randomness. For example, in the model indicators such as "investment", "national identity", "media attention" and other variables not as clear-type variable,we need to dig its fuzziness and randomness and to describe its probability density functions of the fuzzy domain mixed uncertainties.3, Combining fuzzy stochastic systems and neural networks. To get fuzzy variables parameters of fuzzy membership functions based on fuzzy K-means clustering center, fuzzy rules extracted from the samples belonging to the largest membership category, use Madamni fuzzy logic systems, to integrate inference logic rule sets which mixed different uncertainties.With the establishment of a single layer of nonlinear BP neural network, I use a gradient descent method, by correcting network weight parameters, to get the output of each sub-network, finally obtain the total neural network integrated module output.4, Using modular hybrid uncertainty neural network to forecast resources allocation of Universities. According to the nature of the sample characteristics, the samples are divided into several classes, to set up several sub-module neural network systems, to separate the complex problem into several sub-issues relatively parts, which improve the system self-learning ability effectively, increasing the accuracy on universities benefit forecast problem. Then it proves mixed uncertainty Modular neural network availability in a continuous problem.