Exploration of Characteristics of the Ensemble net Approach in learning and generalization

摘要

In this article the characteristics of new method ofr learing nad generalization with a general one-hidden layer feedforward neural network are explored for task in material science. This scheme encomplases the use of a linear combination of nodes of randomly selected type and having randomly presciribed parameter values. The learning of types and parameters is realized through adaptive stochastic optimization using a generalization data set and an ensemble of nodes. The learning of the linear coefficients in the liner combination of nodes is achieved with a linear regression metod usin data from the training set. One node is learnedat a time. the method allows for choosing the proper number and types of net nodes, and its computationally efficient. The method was tested on data from a real material science research task.