Research On A Broad Learning System Based On Reservoir Computing

In recent years,with the development of algorithm theory and computers,the application fields of neural networks have become more and more extensive.More and more different neural network models have been proposed.Deep learning,which has developed rapidly in the past decade,is a large-scale,complex and large-scale neural network.The performance difference of different neural networks depends on how the neurons are connected.According to the connection mode,it can be classified into forward network and recursive network.According to the network architecture,it can be divided into deep network and shallow network.Nowadays,the research on deep network applications is very popular,but in theory it has encountered a bottleneck.Unlike deep networks,wide networks are shallow networks.With the deepening of deep network research,its drawbacks gradually appeared.Recent studies have shown that wide networks can also achieve the performance of deep networks and wide networks have advantages that deep networks do not.The width learning system BLS is an efficient incremental learning system based on RVFLN,which has the characteristics of fast and high precision.The traditional BLS is a kind of forward network,which can achieve the performance equivalent to the deep network in image classification,and has become a research hotspot.Due to the lack of dynamic characteristics,BLS cannot be compared with recursive networks in time series problems.In order to realize the accurate prediction of the time series by BLS,combined with the reserve pool structure of the echo state network(ESN),a width learning system RCBLS based on pool calculation is proposed.The system introduces a simple ring-shaped reserve pool connection in the reinforcement layer and replaces the forward connection in the original system with a parallel reserve pool,so that RCBLS has certain echo state characteristics and is easy to design.At the same time,the application of incremental learning ensures the real-time performance of the system.Based on the MSO time series prediction problem,the performance of different reservoir structures RCBLS was studied for data samples of different sizes.The results show that the RCBLS with multiple reservoir greatly improves the generalization ability and stability of the model.