The experiments showed that the structure of artificial neural networks and parameters strongly influenced the performance of the network [1]. If the network structure is too simple, it will lower the classification capacity of the network, thus affecting the final recognition result. If the network structure is too complex it will slower the learning speed and eventually the final error may increase. Also the momentum, learning rate parameters will also affect the speed of network learning and the final accuracy. If the most suitable network structure and network parameters are used it will get a best result [2]. In this paper, we will analysis the number of hidden units, learning rate on the impact of network learning speed and final accuracy, based on BP network in the character image recognition as an example. Keywords-BP network ; learning rate; the number of hidden units I. INTRODUCTIONCharacter recognition is one of the most important pattern recognition applications. Main methods are: simple template matching; outer contour matching; [3] projection sequence feature matching; external contour projection matching; Hausdorff distance based template matching, Character stroke feature matching[4], etc. Chai Zhi had proposed a multi-reader identification fusion [5], LIU XiaoJing had proposed a projection -transform coefficients feature extraction method [6]. For the standard font, the above identification methods can achieve the desired recognition results. But it is difficult to achieve the desired recognition results when there are interference and noise.Artificial neural networks (ANN) has been widely used in pattern recognition, signal, information processing, and many other fields because of its self-organize ability, selfadapt ability, robustness and fault tolerance, and which are coming from its parallel processing structure and distribute information storage. But at the same time, they also bring some problems such as the difficulty of getting optimized topology structure and the great training time-consuming, which restrict its reality use. Scollich and Krogh [7] put forward an integrated neural network method for the

Natural rubber is mainly dependent on import in China, its domestic market price is influenced by the Natural Rubber Customs Declaration Price (NRCDP). Considering the fluctuating properties of the NRCDP, a method of the NRCDP based on Wavelet and the optimized Back Propagation (BP) neural network Group using a Genetic Algorithm (W-GA-BPG) is proposed. First, an NRCDP dataset is established based on the original Customs Declaration Price (CDP) dataset collected by Qingdao Customs, in which the commodity types are selected consistently according to the sampling intervals, and the features are deleted if they are less affected by the fluctuation of NRCDP. Secondly, the selected features in NRCDP are decomposed using wavelet transform to obtain a group of feature sequences with different scales. Then, a Group of BP neural networks (BPG) optimized by Genetic Algorithm (GA) is used to predict multiple decomposition sub-sequences, respectively. Finally, the predicted values are obtained through wavelet reconstruction. Combined with the NRCDP dataset, the W-GA-BPG model is established by comparing and analyzing experiments by evaluating the Mean Square Error (MSE) and determination coefficient of the prediction results. The MSE and determination coefficient predicted using the proposed model are 0.0043 and 0.9302, respectively, which is the best prediction effect.

Abstract. BP neural network is wildly used because of its strong nonlinear processing ability, self-learning capability, fault tolerance capability. Therefore, the structure and parameters of artificial neural network determine the performance of neural networks. The performance of a BP neural network is not only affected by the network structure, but also affected by its parameters. In this article we will discuss the learning rate and momentum parameters matching relationship and its impact on network performance. The experimental results show that for the curve fitting problem there will be an optimal structure and parameters for the BP neural network. Artificial Neural Network and Its Structural CharacteristicsBP neural network is wildly used because of its strong nonlinear processing ability, self-learning capability, fault tolerance capability. Curve fitting especially for higher-order function is a complex process. The processing of BP neural network curve fitting, in fact is adjusting the connection weights of all artificial neurons in the BP neural networks. It is a process of matching optimal state of the network and curve function. The training convergence process of the BP neural network is the matching process of parameters adjustment. Generally with the size and complexity of the problem is growing, the number of neurons in the neural network will be more and more, the network structure will be more complex. If a neural network contains very little neurons, there are not enough weights to be adjusted, so it cannot remember the appropriate mode, it cannot be expected to complete the mappings. But too complex network structure will lead to neural network training convergence slows down, sometimes even lead to the training process cannot converge, making it impossible to complete the mapping of expected. Therefore, the structure and parameters of artificial neural network determine the performance of neural networks. In the curve fitting problems, fitting precisely and quickly is an important performance indicator of BP neural network. This includes samples that have been learned and not been learned, that is called the generalization ability. In addition, in practical applications it should also consider the cost of time and space.In theory, the number of the hidden layers of BP neural network may be 0, 1, 2…n. It has been proved that only one input layer and output layer BP neural network cannot solve nonlinear problems [1]. It also has been proved that a BP neural network with a single hidden layer structure can approximate any derivable function and its derivatives at any precision [2,3]. Therefore, the BP neural networks were selected for the single hidden layer structure in most practical applications. The number of the neurons of the input layer and the output layer depends on the dimensions of the problem to be solved. In general, these two parameters are determined according to the actual problems. Therefore, to determine the BP neural network topological structure can be attributed to the se...

No abstract