A pseudoinverse learning algorithm for feedforward neural networks with stacked generalization applications to software reliability growth data

摘要

A supervised learning algorithm, Pseudoinverse Learning Algorithm (PIL), for feedforward neural networks is developed. The algorithm is based on generalized linear algebraic methods, and it adopts matrix inner products and pseudoinverse operations. Incorporating with network architecture of which the number of hidden layer neuron is equal to the number of examples to be learned, the algorithm eliminates learning errors by adding hidden layers and will give an exact solution (perfect learning). Unlike the existing gradient descent algorithm, the PIL is a feedforward only, fully automated algorithm, including no critical user-dependent parameters such as learning rate or momentum constant. The algorithm is tested on case studies with stacked generalization applications to software reliability growth data. The results indicate that the proposed algorithm is very efficient for the investigation on the computation-intensive generalization techniques.