Nonlinear system identification using deep learning and randomized algorithms

Abstract

Randomized algorithms have good performances for regression and classification problems by using random hidden weights and pseudoinverse computing for the output weights. They have one single hidden layer structure. On the other hand, deep learning techniques have been successfully used for pattern recognition due to their deep structure and effective unsupervised learning. In this paper, the randomized algorithm is modified by the deep learning method. There are multiple hidden layers, and the hidden weights are decided by the input data and modified restricted Boltzmann machines. The output weights are trained by normal randomized algorithms. The proposed deep learning with the randomized algorithms are validated with three benchmark datasets.