Application of Hybrid Neural Networks based on bagging and Group Method of Data Handling for forecasting

In this paper hybrid Deep learning network based on bagging and Group Method of Data Handling is considered. This network differs from earlier developed hybrid DL networks by application of bagging method. The system consists of a sequence of stacks, wherein odd stacks are essentially ensembles formed by parallel connected different models that solve the same problem, while even ones are metamodels that implement the bagging procedure and calculate the levels of fuzzy membership of each member of the ensemble to the optimal result. The process of increasing the number of stacks is based on GMDH principles until the desired accuracy of the final results is achieved. The proposed system does not require large volumes of training samples, provides online work and automatically determines the number of its layers - stacks in the learning process. The algorithm of training hybrid DL network based on bagging in on-line mode is developed. The experimental investigations of suggested Hybrid DL network based on bagging were carried out at the problem of forecasting Dow Jones Ind. Aver and comparison by accuracy with models ARIMA, GMDH, GMDH-neo-fuzzy network was performed.