Broad learning system based on driving amount and optimization solution

Broad learning system (BLS) was proposed by C. L. Philip Chen to overcome the time-consuming problem of traditional deep learning. However, the prediction precision of BLS is mainly dependent on its regularized parameter λ. Usually, λ is calculated by the trial and error method, which often suffers from the problem of too much calculation. To alleviate this issue, we propose an improved BLS with the driving amount and optimization solution (i.e., DA-BLS) in the study. The contributions of this study include: First, we use the iterative least square method to replace the ridge regression calculation of BLS, which avoids the selection of λ. Second, we provide the formulas of the driving amount and optimization solution under specific conditions. Third, the universal approximation property of DA-BLS is given. Last but not the least, extensive experimental results on the 1-D nonlinear function, UCI data-sets, and fault diagnosis of TEP show that DA-BLS outperforms the relevant methods such as BLS and the stochastic configuration network.