부산대학교 도서관

Broad Learning Systems (BLS) have recently been shown to be effective and efficient. This paper reviews the basic structure of BLS and proposes a BLS based on improved PSO. The proposed model optimizes the parameter $\lambda$ in the ridge regression method used when the BLS directly computes the pseudo-inverse to obtain the network weights. In many problems, $\lambda$ is usually set as close to zero as possible, but in practice, different $\lambda$ values are used, and the accuracy of the model will be different. If a method can be found, and it is different from the tedious and complex grid analysis, it can find the value of $\lambda$ that makes the model as accurate as possible while taking into account the accuracy and efficiency, then the overall performance of the network can be improved. Finally, experimental results on multiple fraud and cancer datasets demonstrate, collected from the UCI repository, that this method is better than selected state-of-the-art methods, such as stacked autoencoders, deep belief networks, etc., in terms of accuracy and training speed. And the results of this paper also show the conclusion that the method of parameter selection can greatly decrease the training time of the traditional BLS on the classification scene on the UCI dataset.