Neural network approach to R&D projects termination decision and its application

Summary form only given. In this paper, a multi-layer neural network is presented to make termination decision for ongoing R&D projects. The strengths of neural networks accrue from the fact that they do not make assumptions of models in advance and from their capability to infer complex underlying relationships. Further some improvements are also presented as follows. A trial algorithm used to optimize the number of nodes in the hidden layer and a genetic algorithm employed to optimize the initial values of the parameters of the neural network are put forward. In addition, an accelerated gradient descent algorithm (the improved back-propagation algorithm) is employed to adjust the values of the parameters of the neural network. At last, in the FANN (feedforward artificial neural networks) determined in this paper, there are three types of processing neurons: fourteen input neurons, five hidden neurons, and one output neuron. In addition, to identify the importance of each variable that influences the implementation of the R&D projects, we introduce the conception of the weight contribution rate