Improvement Of Radial Basis Function Neural Network And Its Application To Cash Flow Prediction

Radial basis function neural network is a common feed-forward neural network based on the radial basis functions. It is mainly used in pattern recognition, function approximation, forecasting and other fields. RBF neural network is a local approximation network, and has been favored to become the artificial neural network, a research focus because of its fast learning speed, generalization ability and strong characteristics.RBF neural network construction process is through the study to determine the parameters of the network, mainly including the center of hidden layer nodes, width, and output layer weights of each node. The existing RBF neural network learning algorithms are divided into two phases: The first stage is learning of the hidden layer centers and width; The second stage is the output layer weights training. The continuous improvement of RBF neural network is mainly concentrated in the learning algorithm for the selection of the node centers.This paper presents two improved RBF neural network learning algorithm, the first one is a hybrid learning algorithm based on PSO, K-means clustering and subtractive clustering algorithm, hereinafter referred to as DPSO algorithm. The other one is a dynamic clustering method based on PSO, hereinafter referred to DCPSO algorithm.DPSO algorithm uses K-means clustering algorithm to determine the number of clusters named K, then uses K-means clustering algorithm to generate a more reasonable multiple of the initial particle swarm PSO algorithm, detailed steps are as follows: Step 1, use subtractive clustering algorithm to determine the number of clusters, namely, number of the center of RBF networks K.Step 2, determine the initial particle PSO algorithm by K-means algorithm, if the initial population size of PSO algorithm is P (artificial), we need to do P times K-means clustering, in order to generate P initial particles.Step 3, use the PSO algorithm for training the parameters of RBF neural network. First, we combined the P groups of basis function center generated by K-means algorithm, as well as P groups of randomly generated weights of RBF neural network wi, and Gaussian radius ri together as(Cp1, Cp2,……, Cpk, wp1, wp2,……, wpK, rp1, rp2,……, rpK). When p = (1,2, ... ..., P), here are totally P particles like this as the initial particle swarm of the PSO algorithm. Then initialize each particle's velocity Vi randomly, comparing the position of each particle with its own best and its group's best. Adjust the position and velocity of each particle until meeting the iteration termination conditions, and all the parameters of the RBF network were determined.The number of initial cluster centers for the DPSO method does not need to select artificially, which can reduce the disturbance to the accuracy of the algorithm generated by EXP. Using the K-means clustering algorithm several times to form the initial particle swarm PSO algorithm, which reduces the randomness of K-means clustering algorithm itself, and also makes the PSO algorithm to obtain a reasonable initial particle swarm.The improvements of the DPSO algorithm focused on the role of pre-PSO algorithm, the corresponding, and the improvements of DCPSO algorithm mainly concentrated on the role of the PSO algorithm after. DCPSO algorithm specific steps are as follows:Step 1, determine the number of the center of K-means clustering algorithm. The PSO algorithm can update the number of clusters after the calculation of PSO, so we choose the square root of the number of training samples as the initial number of classes.Step 2, determine the initial particle PSO algorithm through the K-means algorithm, the same as step 2 of DPSO algorithm.Step 3, use PSO algorithm training the parameters of the RBF network, the same as step 2 of DPSO algorithm.Step 4, Optimizing cluster center and its corresponding parameters through merging and splitting classification center. The merging is the consolidation of the two nearest centers among all classes; splitting is to split the largest class into two new classes; The premise is both the merger and split must improve the fitness. After the termination of iteration, all the parameters of RBF networks have been determined.DPSO algorithm reduces the randomness of PSO algorithm and achieves "dynamic" characteristics of the dynamic clustering algorithm by merging and splitting the classification center.In order to examine the classification results, the two improved algorithms were applied to the classification data set chosen from the UCI dataset and compared with RBF neural network based on PSO. Experimental results show that two kinds of improved learning algorithm are obviously superior to the existing PSO algorithm. The two improved algorithms are not only a high accuracy rate, but good stability. Between the two improved algorithms, the stability of the DCPSO algorithm is relatively higher; the network structure training by DCPSO was relatively simple.This paper has also established a cash flow-based early warning indicator of financial analysis system, and used RBF network based on DPSO and DCPSO to this problem. Select the Shanghai and Shenzhen A-share listed companies which are special treatment because of financial conditions unusual during 2001 to 2008 and their matching companies have a total of 210 samples, Chosen a current capacity analysis indicators, analysis of indicators of solvency and financial flexibility analysis of indicators of three types six indicators. The experimental results not only illustrate that the cash flow-based financial targets has a good recognition in the financial early-warning analysis, RBF neural network method can be applied to solve this type of problems. Also described the improved algorithms in terms of training and prediction accuracy rate or on the stability of the algorithms are better than the PSO algorithm as a learning algorithm of RBF network.