Stochastic configuration broad learning system and its approximation capability analysis

In this paper, a kind of stochastic configuration broad learning system (SCBLS) is proposed for data modeling. The proposed SCBLS is established in the form of a flat network and its architecture is determined by a constructive learning approach. The input parameters of feature nodes and enhancement nodes of SCBLS are randomly assigned in the light of a supervisory mechanism. Inequality constraints are used to randomly assign the hidden parameters and adaptively select the scopes of random parameters. The output parameters of SCBLS are determined either by a constructive manner or by solving a global least squares problem. It is proved that the proposed SCBLS possesses universal approximation properties. The performances of the proposed SCBLS are evaluated by function approximation, benchmark datasets and time series prediction. Numerical examples show that SCBLS can achieve satisfactory approximation accuracy.