017 nd International Conference on est, Measurement and Computational Method (MCM 017) ISBN: 978-1-60595-465-3 Research Enterprise Performance Evaluation Based on BP-LM Neural Network Meng-meng YUAN 1, Feng-shan XIONG and Xue-quan YANG 1,* 1 College of Information Science and echnology, Agriculture University of Hebei, Baoding 071000, Hebei, China College of Business, Agriculture University of Hebei, Baoding 071000, Hebei, China *Corresponding author Keywords: Neural network, Enterprise performance evaluation, LM algorithm, Evaluating indicator. Abstract. [Objective]For the better development of the enterprise, enterprise performance evaluation is the symbol of the overall enterprise success.[methods]in this study, we use Levenberg Marquardt (LM) algorithm to improve the BP neural network to evaluate the performance of enterprises.[result]hrough the simulation test of the relevant data of a car company in 010-016, it is proved that the improved BP neural network can improve the performance of the enterprise more quickly and accurately.[conclusion]lm - BP neural network of enterprise performance evaluation can make enterprises better understand themselves. Introduction At this stage, more and more people choose to venture, leading to the increasingly fierce competition in enterprises. Enterprise performance evaluation is a symbol of the overall success of the enterprise, and it is a major theoretical and practical problem in the business community. It has always been a hot issue that the Chinese and western theories and business circles are concerned. herefore, the enterprise needs to carry out comprehensive and concrete evaluation. he results of the evaluation adjust the direction and strategy of enterprise development. he enterprise performance evaluation system has experienced three periods: cost performance evaluation, financial performance evaluation and strategic performance evaluation. he evaluation of the traditional corporate performance is only to consider the short-term gains, not consider the long-term development, "eyes are short", this evaluation is usually not a true reflection of the company's performance and operating conditions. At the same time, it is easy to mislead business operators to pursue short-term interests at the expense of long-term interests, Formed the idea of quick success. At present, there are many researches on the performance evaluation in the enterprise at home and abroad, and there are some imperfections in each research. After the analysis of the factors that affect the evaluation of enterprise performance, there is a nonlinear relationship between various factors and there is a hidden layer, so the general evaluation method cannot be more perfect to make the evaluation. It is proved that the BP neural network can take into account the factors of the hidden layer, so BP neural network model is suitable for the evaluation of enterprise performance. It is proved that the artificial neural network is effective in solving the complicated nonlinear system. Evaluation of Enterprise Performance Based on Improved BP Neural Network Standard BP Neural Network Principle BP (Back Propagation) neural network is an algorithm that can effectively adjust the connection weight of hidden layer in artificial neural network. BP neural network has the characteristics of nonlinear, distributed storage, parallel processing, self-learning, self-organizing, adaptive and has good fault tolerance ability of neural network. It is a good way to simulate the human brain and is free from interference. BP neural network is composed of input layer, hidden layer, output layer. Where in 14
the hidden layer may have one or more layers. Each layer is composed of a number of neurons, and the same layer of neurons are not connected, but between the adjacent layers of neurons are fully linked state. BP neural network is the positive propagation of information that is input information through the input layer to reach the hidden layer, after the treatment of the hidden layer to the output layer, if the results of the expected error in the expected are error, the error of the reverse propagation error signal in the process of transmission, adjust the weight between the layers and the threshold between the neurons, and thus the error signal continues to decrease until it reaches the acceptable range. LM Algorithm Improved BP Network he basic idea of the LM optimization algorithm is to allow the error to be searched in the opposite direction so that it no longer iterates along a single negative ladder direction,by using of the Gauss Newton method and the steepest descent method, we can adjust the weights of the BP network, improve the convergence speed and generalization ability of BP network, and make the BP network converge effectively. he LM-BP algorithm can effectively suppress the local minimum of the network and reduce the sensitivity of the local details of the error surface. he LM algorithm is combined with the Newton method and the gradient descent method to correct the BP network model. he formula for the LM algorithm is: xk + 1 = xk [ J J + µ I] 1 J e (1) Where I is the unit matrix, J is the network error for the weight derivative Jacobian matrix, and e is the error vector of the BP network model synthesis output. µ is the constraint coefficient of the negative gradient between the neurons. When µ is large, it becomes a small step gradient method, which is very close to the negative gradient algorithm of the standard BP neural network inverse error iterative method, the purpose of this algorithm is to converge as soon as possible in the error minimum faster and more accurate Newton method, after each successful iteration, the error performance of the BP network model is reduced, then reducing the value of µ, in contrast increase the value of µ. When µ is small, it is corrected by error: w E A = η wsp So that its network model is close to the Gauss - Newton error iterative algorithm. () Experimental Analysis Selection of Evaluation Indicators After the study, there are one-sided, vulnerable to subjective factors, the selection of indicators is not enough science and other issues by most of the evaluation of the evaluation of enterprises in the selection of indicators. And now the selection of evaluation indicators mainly focus on the financial performance of enterprises, that is not a good reflection of the actual level of business [6]. he selection of evaluation indicators plays an important role in the evaluation of the performance of enterprises. he selection of scientific and practical evaluation index is the basis of enterprise performance evaluation. he evaluation indexes should be comprehensive, representative, scientific, operable, and hierarchical. Because the enterprise not only requires that the performance evaluation system to systematically reflect the effect of the previous stage of business activities, but also requires that performance evaluation to more fully reflect the overall situation of enterprises and the future development trend. Financial aspects. he financial position of the enterprise can well reflect the business situation of the enterprise, and it can reflect the use of funds and profitability. herefore, the financial indicators are: sales profit rate = sales profit / total sales * 100%, net assets = Asset profit / total assets * 100%, 143
asset-liability ratio = total liabilities / total assets * 100%, asset flow = total current assets / total current liabilities * 100%, net profit growth rate = income tax / total profit * 100 %. Customer satisfaction aspects. Customer is the subjective factor in determining the survival of the enterprise, the so-called 'customer is God', is a lot of business to follow the purpose, so in the selection of indicators, customer satisfaction is essential. Specific indicators are: product repair rate, punctual delivery rate, customer complaints rate, product qualification rate. Innovation and development. he innovation and development of the enterprise is related to whether the enterprise can survive for a long time, and only innovation can make every time the enterprise move forward. Innovative development indicators are: technology cost rate = technical cost / total assets * 100%, domestic market share, development success rate, product competitiveness, new product market share [7,8]. Structural Design of the Evaluation Model Network layer setting. BP neural network is composed of input layer, hidden layer, output layer, and research shows that a hidden layer can achieve any nonlinear input and output mapping, and the hidden layer will extend the network learning time. It is not good to observe and adjust the training effect. Robert Hertht-Nielson, as early as 1989, has demonstrated that a BP neural network with three layers of layers can achieve approximation of any nonlinear continuous function composed of finite discontinuities by arbitrary precision, which can achieve the mapping in finite multidimensional space. he 3 layer BP neural network can satisfy the input and output model. [9] herefore, this experiment is carried out by using the three layer network with a hidden layer. he number of input neurons. In this paper, the number of input neurons of BP neural network is determined as 15 by selecting the evaluation index. he number of hidden neurons. he number of neurons in the hidden layer is related to the accuracy of the training, but too many neurons complicate the network and increase the training time, so the number of neurons is also significantly critical, according to an empirical formula: p = m + n + a (3) (P is the number of hidden nodes, m is the number of input nodes, n is the number of output nodes, and a Constant for 1-10). he number of neurons in the hidden layer is 1 through empirical formula and training he he number of neurons in the output layer. he output layer is the output of the evaluation results of the enterprise, and the evaluation result of the enterprise can be displayed with one data. herefore, the neuron of the output layer is set to one, and the result of this neuron can be divided into five levels: excellent [0.8 1], good [0.6 0.8), general [0.4 0.6), poor [0. 0.4), poor [0 0.). BP-LM Algorithm Enterprise Evaluation Steps Based on improved LM algorithm of neural network enterprise evaluation steps are as follows: Step 1: Set the range of allowable error of BP neural network ε, and θ> 1, µ0; Step : Initialize the network weight as µ = µ0, k = 0; Step 3: Enter the training sample: ( x, x,... ) X = 1 x n Step 4: Calculate the network error and error index function E (X) : 1 E( X ) = n i= 1 e i ( X ) Step 5: Calculate the Jacobian matrix J (X); Step 6: Calculate the amount of adjustment of the weight vector W, (4) (5) 144
W = 1 [ J ( X ) J( X ) + I] J ( X ) e( X ) µ (6) Where µ is a constant greater than 0, I is a unit matrix; Step7: Weight vector adjustment, the formula is: W k+1 = W k + W Step 8: Judge E( X ) < ε is established, if it is established then end, otherwise, go to the ninth step; k +1 Step 9: Calculate the error indicator function E( X ); k +1 k Step 10: If E ( X ) < E( X ) is established, you need to make k = k + 1, µ = µ / θ, turn to the fourth step, otherwise µ = µ θ, and go to the sixth step. Network Simulation In this paper, the research on enterprise performance evaluation based on LM-BP algorithm is carried out in MALAB. he initial model of network model is designed, including design accuracy and number of times, the initial weight value and threshold value, and the sample data, and the sample data must be between [0-1]. he experimental data are derived from the data of a car industry from 006 to 016, and the experimental data are standardized into learning samples in the [0,1] range. he target vector of its output vector is t = [0.340 0.180 0.30 0.8360 0.59 0.6740 0.8630 0.790 0.630 0.1970 0.540]. he input vector is the data in able 1, each of which is a vector, and the range of input vectors is between [0,1]. he training function is an improved RAINLM algorithm for LM algorithm. he transfer function between the input layer and the hidden layer is the logarithmic SIGMOID, that is LOGSIG. he transfer function between the hidden layer and the output layer is selected as tangent SIGMNID or ANSIG, and the other parameters are the default. (7) Figure 1. Convergence results of BP-LM network training. Figure. Convergence results of the standard BP network training. When the training network is trained, the number of training is 5000; the training accuracy is 0.000001, the learning speed is 0.01, the momentum coefficient is 0.00001, the coefficient MUDEC is 0.1; the increase coefficient MUINC is 10; the other parameters are missing.when the coefficient of performance is increased, the mu value multiplies the coefficient MUINE, and if the mu exceeds the maximum value MUMAX, the training stops. he network has been trained 43 times to reach the error requirement. he training process is shown in Figure 1, and Figure shows the standard BP neural network. he standard BP neural network is trained by 645 times to meet the error requirement. After the neural network training is completed, the results of the training and expected values are shown in able 1 below: 145
able 1. wo comparison of the output results and the target values of 3 kinds of algorithms. 006 007 008 009 010 011 01 013 014 015 01 6 target value t 0.340 0.180 0.30 0.8360 0.590 0.6740 0.8630 0.790 0.630 0.1970 0.540 Standard BP network actual output t1 0.347 0.170 0.315 0.8355 0.591 0.6751 0.8605 0.801 0.633 0.1968 0.5419 Actual output 0.340 0.180 0.319 0.8360 0.589 0.6740 0.8630 0.790 0.630 0.1969 0.540 of BP-LM network t From able 1, we can see that the improved BP algorithm results in closer to the target value than the standard BP algorithm, and it also shows that the improved BP-LM algorithm learns more fast than the standard BP algorithm. Summary he experimental results show that the improved BP network model based on LM algorithm not only improves the convergence speed, but also avoids the occurrence of local minima and achieves higher accuracy. he network model is applied to the evaluation of enterprise performance, its adaptive, self-learning and fault-tolerant ability, not only to avoid the enterprise performance evaluation process in the human subjective factors, but also solve the other calculation methods involved in the complex issues, Formed a more rigorous and accurate evaluation system. In the evaluation of enterprise performance, only need to be trained in the network model to enter the corresponding evaluation index of the data matrix, the network can give a professional evaluation results.we can see that BP-LM algorithm is suitable for the evaluation of corporate performance. References [1] Y.Z. Cai. Strategic Management Based on Balanced Scorecard - aking ENN Group as an Example [D].Nankai University, 008. []L. Zhao, Based on the Balanced Scorecard (BSC) of the Implementation of Enterprise Information Strategy - HX Group as an Example [D]. Anhui University, 014. [3] Robert S. Kaplan. & David P. Norton. Balanced Scorecard ~ ranslating Strategy In to Action [J]. Harvard Business School Press, 000 (5): 9-1. [4] D.F. Zhang, MALAB Neural Network Application Design. Beijing Machinery Industry Press, 009. [5] D.W. Cui. Application of Improved BP Neural Network Model in Comprehensive Evaluation of Well-off Water Conservancy [J]. Journal of Hohai University, 014.7, 306-313. [6] Y.K., Guo. Based on BP Neural Network of a Petrochemical Enterprise Supplier Evaluation Method Research [D]. Beijing Jiaotong University, 015.4. [7] X.Chen. Based on the Knowledge Management of Innovative Enterprise Evaluation Index System Construction and Verification [J]. Library Science Research, 014, 76-89. [8] M.X. Wang, D. Li, Innovation Enterprise Evaluation Method Improvement and Innovation Ability Promotion Path [J]. Science and echnology Progress and Countermeasures, 014.08, 18-13. [9] R.Hong.Based on BP Neural Network Market Value of Enterprise Evaluation [D]. Yunnan University, 015.4. 146