CN102737285A - Back propagation (BP) neural network-based appropriation budgeting method for scientific research project - Google Patents

Abstract

The invention discloses a neural-network-based appropriation budgeting method for a scientific research project, and aims to the problem that scientific research appropriation budgeting accuracy and practicability cannot be combined. The method comprises the following steps of: analyzing components of appropriation of the scientific research project and key factors influencing budgeting, wherein the key factors comprise a research cycle, the number of researchers, a key technical coefficient, a project result coefficient and the innovativeness and complexity of the project; and establishing a quantified expression of the influencing factors, and establishing a nonlinear expression between a quantification result and an appropriation budgeting result of the project through a neural network by taking the quantification result as input and taking the appropriation budgeting result of the project as output. An application result shows that the method is reliable and high in evaluation accuracy.

Description

A kind of scientific research item funds budget method based on the BP neural network

Technical field

The present invention relates to a kind of scientific research item funds budget method, can realize new scientific research item funds budget based on history item funds data based on the BP neural network.

Background technology

At present, scientific research item funds budget method commonly used comprises parameter estimation method, engineering estimation algorithm, empirical estimation method and analogy cost estimation method etc.

The parameter estimation method is according to the historical data of model development, and as variable, the utilization mathematical statistics method is estimated the new model funds through setting up costimating relational expression (CER) the characteristic parameter relevant with development process.The parameter estimation method can be carried out the funds estimation quick and objectively, and under accurate, the complete situation of historical data, precision of prediction is higher, but requires very high at aspects such as Model Selection, model adaptability, basic data accuracys.

The engineering estimation algorithm is according to work breakdown structure (WBS) (WBS); With system decomposition is some subsystems, and each subsystem is decomposed into the plurality of sub system again, and decomposing down always ends until not being further divided into; Press the inscape estimated cost of each operation then; The establishment cost manual, the expense with estimation adds up from bottom to top, finally obtains total research fund of system.Each subsystem branch of this method solves thin more, and the result is also accurate more in estimation, and workload is just big more.

The expert judgments method is a kind of method the most frequently used when formulating project resource plans, normally by the cost management expert according to the experience of similar item in the past with to the judgement of this project, through thinking through, carry out reasonable prediction, formulate the way of project resource plans.This method is based on experience in the past to be estimated, therefore is a too subjective method.

The basic thought of analogy cost estimation method (ABE) is: when new projects are provided, with it and the most similar history item analogy that retrieves, through relatively predicting the cost of new projects.But the research fund of disparity items relatively is the ten minutes complicated problems, often can not reach a conclusion through the simple contrast of indivedual model research funds.

Summary of the invention

Traditional appropriation budget method model fitness is poor in order to overcome, workload greatly, deficiency such as subjectivity too, the present invention proposes scientific research item funds budget method based on the BP neural network.The learning ability that the BP neural network is powerful can merge the data of the multiple scientific research item funds influence factor that needs to consider; Set up the nonlinear relationship between influence factor and the project funds according to historical funds data through training, export a more accurate scientific research item funds predicted value after nonlinear transformation.The present invention utilizes unique premium properties such as BP neural network concurrent distribution process, self-organization, self-adaptation, self study and its fault-tolerance; The deficiency of several kinds of evaluation methods more than having overcome has better solved this multifactor, nonlinear problem of scientific research item funds budget.

For solving the problems of the technologies described above, concrete grammar of the present invention is following:

The 1st step: analyze scientific research item funds and constitute, confirm to influence the key factor of budget;

Said key factor comprises quantitative target and qualitative index two parts; Quantitative target comprises research cycle, joins and grind personnel amount, gordian technique coefficient amount, project output achievement coefficient, and qualitative index comprises project novelty and project complexity;

The 2nd step: the quantification expression formula of setting up influence factor;

The quantification expression formula of quantitative target is: Wherein, X' is for quantizing the back data, and X is data before quantizing, X _MinBe the minimum value of similar influence factor, X _MaxMaximal value for similar influence factor data;

Qualitative index adopts degree of membership to carry out assignment, and the set of degree of membership value is (1,0.75,0.5,0.25), and the performance of the high more expression qualitative index of degree of membership is high more;

The 3rd step: as input, project appropriation budget result makes up the BP neural network as output with each quantized result that influences budget; The input layer of this BP neural network, hidden layer and output layer neuron number get 6,7,1 respectively; Input layer and hidden layer neuron adopt Sigmoid type activation function, and the output layer neuron adopts linear activation function; Neuronic initial weight is got the random number between (1,1); Learning rate gets 0.7, and the network precision gets 0.01%;

The 4th step: 6 key factors obtaining from historical scientific research project that the 1st step confirmed and the actual funds of project, and the quantification expression formula that adopted for the 2nd step defined quantizes the composing training sample respectively; Adopt a plurality of training samples that the BP neural network is trained;

The 5th step: during actual prediction, scientific research project to be predicted is confirmed 6 key factors, the BP neural network that input trains, the output of neural network is the appropriation budget result.

Beneficial effect:

The invention has the beneficial effects as follows, can effectively carry out the scientific research item funds budget, simple to operate, the budget precision is higher, overcome that the operation of traditional budget method is too complicated, prediction accuracy is undesirable and the appraiser is required high defective.Specifically:

1) the present invention analyzes the main composition of project funds; Determining design charges, direct cost expense, overhead cost and wage and service charge is the main subject of decision scientific research item funds; And analyzed these subjects and which factor is closely related, thereby these factors have been confirmed as influence factor.The BP neural network is input with the influence factor, can adopt the parameter of limited quantity to realize comparatively comprehensively item description like this, thereby makes up neural network simple, that be easy to realize.

2) the present invention carries out the degree of membership assignment to the qualitative influence factor in the influence factor, thereby has realized the quantification of qualitative impression factor.

3) the present invention is directed to the project appropriation budget; Neuronal activation function in the neural network, initial weight, learning rate, network precision are carried out numeral to be chosen; Thereby the assurance neural network is approached the I/O relation of scientific research item funds as much as possible; And shorten convergence time as far as possible, improve prediction accuracy.

Description of drawings

Fig. 1 is a basic scientific research project funds major influence factors structural drawing.

Fig. 2 is the appropriation budget illustraton of model.

Fig. 3 is an appropriation budget BP neural network model structure.

Fig. 4 is the training data input interface.

Fig. 5 is an appropriation budget BP neural metwork training conditional curve.

Embodiment

Below in conjunction with accompanying drawing and embodiment the present invention is done further explain.

The 1st step: analyze scientific research item funds and constitute, confirm to influence the key factor of budget;

Scientific research item funds are made up of project cost, project income two parts.The core of scientific research item funds budget is the budget items costs.Project cost is meant according to the provisions of the relevant regulations issued by the State and comprises design charges, special-purpose expense, fee of material, external coordination expense, fuels and energy expense, expense for using fixed assets, wage and service charge, travel expenses, meeting expense, transaction fee, expert consulting expense, overhead cost, unexpected expenses etc. for carrying out the expense that scientific and technological industrial research work takes place.Wherein design charges and wage expense and labor service are to multiply by per capita expense with the person-time who estimates to calculate; Fee of material, external coordination expense, special-purpose expense, test fee, expense for using fixed assets are the estimations that expense takes place in the project development process; Overhead cost is that the certain proportion by above-mentioned expense (do not contain wage expense) calculates, and generally is no more than 15%.

According to the characteristics of scientific research project with to the statistical study of historical scientific research project, scientific research item funds depend primarily on design charges, direct cost expense (comprising fee of material, external coordination expense, special-purpose expense, test fee and expense for using fixed assets), overhead cost and wage and service charge.Wherein, the gordian technique number of design charges and breakthrough, project novelty, project complexity and the ginseng personnel of grinding count positive correlation; Direct cost is taken main relevant with entry property, achievement quantity and achievement form; Overhead cost takes decision by design charges and direct cost; Wage expense and research cycle and the ginseng personnel of grinding count positive correlation.

The analysis-by-synthesis project cost constitutes, and in conjunction with basic scientific research item types character, the scientific research project influence factor is divided into two types the most at last, promptly qualitative influence factor and quantitative effect factor.The quantitative effect factor comprises research cycle, joins and grind personnel's number, gordian technique coefficient and project output achievement coefficient etc., and qualitative influence factor comprises that item types (application and development class, through engineering approaches class, invention class, innovation class, international co-operation class), project novelty, project complexity are project technical sophistication degree etc.The influence factor classification is as shown in Figure 1.

The 2nd step: the quantification expression formula of setting up influence factor;

Each is very big to figureofmerit difference and data level difference in the original sample, for convenience of calculation and prevent that partial nerve unit from reaching hypersaturated state, carry out normalization to it and handle.The disposal route of each index is following:

1. the normalization of quantitative effect factor index is handled

Grind personnel amount (C), gordian technique quantity (K), these 4 quantitative targets of output achievement quantity (O) for research cycle (T), ginseng, carry out normalization by following formula unification and handle:

$X^{\′}=\frac{X-X_{\min }}{X_{\max }-X_{\min }}---\left(1\right)$

Wherein, X' representes the data after the normalization, and X representes T, C, K, the O before the normalization, X _MaxIndex maximal value in the expression historical data, X _MinIndex minimum value in the expression historical data.

Wherein the expression of output achievement quantity O is following:

O＝0.4×N _i+0.3×N _t+0.2×N _m+0.1×N _s (2)

Wherein, N _i, N _t, N _m, N _sRepresent number of devices, principle prototype quantity, material sample quantity and amount of software in the output achievement respectively.

2. qualitative influence factor index is handled

Novelty to scientific research project is estimated, and its objective is the low-level repetition of avoiding scientific research project, guarantees novelty, advance and the applicability of scientific research project.The novelty of scientific research project is determined by factors such as Project Study content, project gordian techniquies.

The scientific research project technical complexity is the key factor that influences the project complexity.Design charges in the project cost, fee of material, wage and service charge etc. all there is direct or indirect influence.The technical sophistication degree mainly shows academic level, technical merit, research difficulty, relates to aspects such as subject.

According to the form of expression of scientific research project novelty, can the project novelty be divided into following four grades: strong innovation (academic novel, technology creative high, do not see pertinent literature both at home and abroad), novelty strong (science is novel, creative higher, the domestic pertinent literature of not seeing of technology), novelty general (academic novelty is general, creative general, the domestic a small amount of pertinent literature that has of technology), novelty poor (academic novelty is poor, creative poor, the domestic relative literature that has of technology).Corresponding membership function value is set at 1,0.75,0.5,0.25.

According to the form of expression of scientific research project technical complexity, can the project novelty be divided into following four grades: technical complexity strong (academic level is high, technical merit is high, the research difficulty is big, relate to subject wide), technical complexity strong (academic level is higher, technical merit is higher, the research difficulty is big, relate to subject wider), technical complexity general (academic level is general, technical merit is general, the research difficulty is general, relate to single subject), technical complexity low (academic level is low, technical merit is low, study difficulty low, relate to single subject).Corresponding membership function value is set at 1,0.75,0.5,0.25.

Wherein, rank is drawn by domain expert's evaluation under project novelty and the technical complexity.Computing formula is:

$P=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{P}_i}{n}---\left(3\right)$

$Q=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{Q}_i}{n}---\left(4\right)$

Wherein, P _i, Q _iRepresent the marking of i position expert to project novelty and technical complexity, n representes expert's number of participating in estimating.

Equally, to through marking and average project novelty and technical complexity index, also to carry out normalization and handle.

In sum; Input comprises six of T', C', K', O', P', Q' based on the funds of BP neural network intelligence budget model, representes the current corresponding index of estimating that Project Study cycle, ginseng are ground personnel's number, gordian technique number, weight quantization output achievement coefficient, project novelty and technical complexity waited respectively.

The 3rd step: as input, project appropriation budget result makes up the BP neural network as output with each quantized result that influences budget.

1. input layer, output layer design

The input layer number of BP neural network equals the influence factor number of application problem, and output layer node number depends on the desired output result, and the present invention is scientific research item funds.Therefore, confirm that BP neural network input layer number is N _i, corresponding input is respectively research cycle, ginseng is ground corresponding index T', C', K', O', P', the Q' statement after treatment of personnel's number, gordian technique number, output achievement coefficient, project novelty and technical complexity.The output layer neuron number is made as N _o, corresponding scientific research item funds W.

2. hidden layer structural design

Under the situation that does not limit implicit node number, the BP network that only contains a hidden layer can be realized any Nonlinear Mapping.So select to comprise the BP neural network of one deck.

The experimental formula that the hidden layer node number calculates is:

$N_y=\sqrt{N_i+N_o}+\mathrm{\α};$

Wherein the span of α is the integer in 1 ~ 10.When α gets 1 ~ 10 successively, then calculate N through after rounding up _yValue be 4 ~ 13.

Hidden layer node crosses that the network precision is low at least, crosses that network convergence speed is slow at most.In conjunction with experimental formula gained result, through attempting obtaining the hidden layer node number: increase gradually from 4 through the hidden layer neuron number, time 7, reach accuracy requirement basically, thereby selected 7 as the hidden layer neuron number.

The appropriation budget network model structure of design is as shown in Figure 3.

Carry out the network parameter design to this appropriation budget network model then:

1.1 the design of neuronal function function

The Sigmoid type function has non-linear amplification coefficient function; It can be the signal of input from the minus infinity to positive infinity; Become the output between-1 to 1, because scientific research item funds have characteristics such as continuous variation, mobility scale are big, so; In this model, adopt S type activation function can approach the I/O relation of scientific research item funds as much as possible to guarantee model at hidden layer.Use linear activation function at output layer, with the budget result that guarantees that output is correct.

1.2 the design of initial weight

Because system is non-linear, initial value for study whether reach local minimum, whether can restrain and the length relation of training time very big.If initial value is too big, make that the input after the weighting drops on the saturation region of activation function, so generally get the random number of initial weight between (1,1).Between input layer and the hidden layer neuron, between hidden layer and the output layer neuron, weights are arranged all, all press this design.

1.3 choosing of learning rate and precision

The scope of learning rate generally is chosen between the 0.01-0.7.In order to guarantee that network can realize the appropriation budget function efficiently, this model chooses 0.7 as learning rate.The precision of neural network has determined that network operations result is appropriation budget result's accuracy.In this model, select 0.01% as the network precision.

The 4th step: 6 key factors obtaining from historical scientific research project that the 1st step confirmed and the actual funds of project, and the quantification expression formula that adopted for the 2nd step defined quantizes the composing training sample respectively; Adopt a plurality of training samples that the BP neural network is trained.

In this step, utilize the propagated forward of BP neural network to calculate and backward algorithm, adjustment neuron weights and threshold value reach precision or cycle index requirement up to error.

The 5th step: during actual prediction, scientific research project to be predicted is confirmed 6 key factors, the BP neural network that input trains, the output of neural network is the appropriation budget result.

Validity for check institute budget speech method; Scientific research project assessment and appropriation budget system have been developed based on JAVA; And to utilize this system be example with budget problem of scientific research item funds of certain research classification, on the basis of 200 given history item funds data, carried out appropriation budget.

Concrete steps are following:

Step 1. is utilized scientific research project assessment and appropriation budget system that 200 history item data are carried out normalization and is handled, and is converted into the input parameter that is fit to Processing with Neural Network.The interface of system's typing training data is as shown in Figure 4.The normalization result of partly importing data is as shown in table 1.

Table 1

Step 2. is utilized the budget model of software building network structure as shown in Figure 3, and to this model training, training process is as shown in Figure 5 according to the normalization input parameter that obtains in the table 1.

The training process of BP neural network is the process according to historical data adjustment network weight, and the weights value after this BP network training is accomplished is following.

Input layer is to the weight matrix of hidden layer:

$W_1={\left[\begin{array}{ccccccc}0.1225& 0.3567& 0.8631& -1.1263& -0.3184& 0.7521& -0.7462\\ 0.8717& 0.7776& -0.1004& -0.8132& 0.2020& 0.6876& 0.1266\\ 0.3409& -0.5247& 0.3364& -1.0795& -0.4709& -2.1109& -0.1485\\ 0.7939& -1.1100& 1.5144& -0.0997& -0.8885& -2.4819& -0.6286\\ 0.1308& -0.2616& 0.2369& -0.0313& -0.1912& 1.1175& -0.5938\\ 1.4758& -1.5664& -1.5789& -1.3703& -0.7493& 1.1331& -1.5221\end{array}\right]}^T$

Hidden layer is to the weight matrix of output layer:

W ₂＝[3.0264 -4.3904 3.7586 -2.4933 -4.5019 -4.5453 -2.8143]

The network that step 3. utilization is accomplished training carries out appropriation budget to new projects.New projects' information comprises: research cycle 5, gordian technique is several 5, and it is several 23 that ginseng is ground personnel, and the output achievement is several 1.4, project novelty 0.75, technical complexity 0.75.The estimation budget of BP neural network output is 1,429 ten thousand.Be more or less the same with expected results 1,500 ten thousand.

Present embodiment is selected a plurality of history item data, and the validity of budget model is tested.Verification msg and checking result are as shown in table 2.

Table 2

Hence one can see that, and test result meets the scientific research item funds actual conditions, and conclusion is directly perceived.

Can find out by the above, the present invention proposes a kind of scientific research item funds budget method, make up appropriation budget model, through instance model validity verified at last based on the BP neural network based on the BP neural network.Simple to operate based on the scientific research item funds of BP neural network intelligence budget method, the budget precision is higher, has overcome that the operation of traditional budget method is too complicated, prediction accuracy is undesirable and the appraiser is required high defective.Can find out that from test result budget model of the present invention has reached the expection requirement, can effectively carry out the scientific research item funds budget.

In sum, more than being merely preferred embodiment of the present invention, is not to be used to limit protection scope of the present invention.All within spirit of the present invention and principle, any modification of being done, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (1)

1. the scientific research item funds budget method based on the BP neural network is characterized in that, comprising:

The 1st step: analyze scientific research item funds and constitute, confirm to influence the key factor of budget;

Said key factor comprises quantitative target and qualitative index two parts; Quantitative target comprises research cycle, joins and grind personnel amount, gordian technique coefficient amount, project output achievement coefficient, and qualitative index comprises project novelty and project complexity;

The 2nd step: the quantification expression formula of setting up influence factor;

The quantification expression formula of quantitative target is: Wherein, X' is for quantizing the back data, and X is data before quantizing, X _MinBe the minimum value of similar influence factor, X _MaxMaximal value for similar influence factor data;

Qualitative index adopts degree of membership to carry out assignment, and the set of degree of membership value is (1,0.75,0.5,0.25), and the performance of the high more expression qualitative index of degree of membership is high more;

The 3rd step: as input, project appropriation budget result makes up the BP neural network as output with each quantized result that influences budget; The input layer of this BP neural network, hidden layer and output layer neuron number get 6,7,1 respectively; Input layer and hidden layer neuron adopt Sigmoid type activation function, and the output layer neuron adopts linear activation function; Neuronic initial weight is got the random number between (1,1); Learning rate gets 0.7, and the network precision gets 0.01%;

The 4th step: 6 key factors obtaining from historical scientific research project that the 1st step confirmed and the actual funds of project, and the quantification expression formula that adopted for the 2nd step defined quantizes the composing training sample respectively; Adopt a plurality of training samples that the BP neural network is trained;

The 5th step: during actual prediction, scientific research project to be predicted is confirmed 6 key factors, the BP neural network that input trains, the output of neural network is the appropriation budget result.

Images