Aanlyzing Time Series Using Multi-objective Diversified Echo State Network

The analysis and prediction of time series movement are considered as challeng-ing tasks.The difficulty of predicting the trends lies in the dynamic temporality and noise in the time series data.The Echo State Network(ESN)is a popular time series prediction model that considers the temporality of the stock time series,but ESN often falls into the dilemma of over-fitting due to the existence of many unnecessary neurons in the hidden layer.Some parameters in echo state network are randomly generated and it is difficult to ensure that we can get an appropriate ESN following the process of trial and error.In order to obtain the appropriate ESN,the traditional approach is randomly generating ESN constantly until we obtain an appropriate ESN.Traditional methods can not guarantee that the new generated ESN is better than the previous ESN.In this paper,we propose Multi-objective Diversified Echo State Network(MODESN).MODESN defines the ESN diversity in this model.By optimizing the ESN diversity,we can optimize the structure of ESN,so that it can avoid over-fitting.And then,the multi-objective genetic algorithm is used to optimize the diversity and accuracy,it can control the direction of ESN generation.The main contributions of this paper are:(1)This paper define the diversity of ESN.We can improve the system's general-ization ability and decrease the occurrence of over-fitting by optimizing the diversity of ESN.(2)This paper proposed a multi-objective diversified ESN(MODESN)model that is based on an improved multi-objective genetic algorithms.By optimizing the accuracy and the diversity of ESN at the same time,MODESN converges in each iteration.(3)This paper reduce the features by combining the MODESN and the Japanese Candlestick to improve the prediction accuracy.