神经网络复合推理技术的研究

Research on Neural Network Composite Inference

提出了神经网络复合推理的思想，阐述了神经网络数值推理与传统符号逻辑推理的差异，提出了神经网络前置处理器和神经网络后置处理器的设计方法，建立了复合推理的系统结构，结合大型工程课题进行了验证．实践检验证明，该方法对解决非结构化、半结构化。

In traditional IDSS, the procedures of solving and inferring in decision problems consist mainly of sign matching, searching and backtracking in solving space. The efficiency of inferring will be affected to some extent, and it may even have some porblems such as the “combination explosion” of inferring space. On the contrary, the numerical calculation is mainly carried out in the neural net work. Therefore the “combination explosion” in solving space can be avoided. Furthermore, as there exists the potential paralled calculation in neural net work, the effciency of calculation is high, so the net work has high degree of solving ability. In intellectual decision support system, how to make combination inference in neural net is the leading problem studied now in the world, and it is also the emphasis of this paper. This paper discusses the idea of the neural net composite inference. It also illustrates the difference between the neural net work numerical inference and traditional sign logical inference, and the design method of the prefix and suffix processor is brought out. It also offers a structure of system for composite inference, which is proved by the engineering project with large scale. It is very effective to solve the unstructured, half structured, and structured problems.