JPH11282823A - Land evaluation system and method, and medium storing land evaluation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically generate a land value model expression which can well represent the actually appraised land value. SOLUTION: A land value model expression is automatically generated by applying the learning function of a neural network to the land value which is appraised by a real estate appraiser and the land value forming factors including the road width, dwelling environment, distance from the nearest transit station, etc. Then the land value is calculated by means of the said land value model expression and a land value comparison table (route value comparison table) is produced to explain the calculation result of the land value. The produced route value comparison table is visually shown in a one-dimensional or two-dimensional graph. In this land value model expression, the land value forming factors are automatically corrected to reduce the error caused between the estimated land value and the land value that is previously appraised.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a land evaluation system and method, and a medium storing a land evaluation program. More specifically, the present invention relates to a method for automatically generating a land price model formula by applying a learning function of a neural network, and The present invention relates to an apparatus, a method, and a program recording medium for creating a road price comparison table for explaining accuracy or correcting land price formation factor data used for creating a land price model formula.

[0002]

[Background] Since the enforcement of the Land Basic Law, taxable valuations, such as property tax and inheritance tax property valuation, have been significantly raised, and land prices have started to fall due to the bursting of the bubble almost in tandem with this period.

[0003] Of these, in particular, with respect to the valuation of property tax, since the disclosure of land valuation has been made completely since 1997, higher valuation than ever has been required. In addition, regarding the evaluation of the published land price, which is the basis for such taxable land evaluation, it is expected that the appraisal report will be released from the flow of information disclosure, and the evaluation criteria used in the appraisal will be clarified. Is desired.

In such a situation, a two-dimensional land evaluation that evaluates the entire municipalities is regarded as a system evaluation business, and has been actively performed in recent years by real estate appraisers and aeronautical surveyors. I have.

[0005] With regard to property tax, a general inspection system, an examination application system, and the like are defined, and it is expected that opportunities for asking the basis for determining the land valuation will increase in the future.
Therefore, development of a land evaluation system with high prediction accuracy is desired.

[0006]

[Prior art and its problems] Multivariate analysis systems are mainly used for evaluating the land price itself in the system evaluation work. Or, it is limited to quantification type I analysis.

[0007] These multivariate analysis methods have a reasonable explanation power from a statistical point of view, but do not have such high accuracy that they can be put to practical use, and are modified with reference to the advice of a practitioner. It is the fact that it is doing.

As an index for evaluating land prices, apart from a land price evaluation system using a multivariate analysis system, for the purpose of properly implementing the National Land Use Planning Act,
The “Land Price Comparison Table” has been prepared under the supervision of the Land Price Survey Division, Land Bureau, National Land Agency.

[0009] This "land price comparison table" uses the published price based on the Land Price Public Notice Law and the standard price based on the National Land Use Planning Law, and allows the prefectural governor to purchase and sell land that is subject to the National Land Use Planning Law. It was created for the purpose of utilizing it for proper price screening.

[0010] The "land price comparison table" is open to the public and is highly reliable, for example, when it is used as a reference for real estate appraisers and land acquisition for public works. However, when looking at this “land price comparison table” from the viewpoint of tax assessment, there are the following problems.

1. Since it has been prepared so that it can be applied nationwide, professional judgment is probably required in order to reflect the local land price formation factors. 2. Since it was created by consolidating the rules of thumb of land price valuation experts, it is not possible to specify the basis of the disparity rate. 3. Insufficient methods to verify the accuracy of land prices estimated from the "land price table". Since it is created from the viewpoint of comprehensively examining the land price formation factors, it is said that mechanically applying the “land price comparison table” causes a problem similar to multicollinearity in multivariate analysis. Here, the problem of multicollinearity is a phenomenon in which the regression coefficient error becomes very large when the explanatory variables have a high correlation with each other, and is listed in the land price ratio table of the National Land Agency. There are several factors that have such high correlation, and if all factors covered in the level table are applied mechanically, price errors may increase. .

[0012]

DISCLOSURE OF THE INVENTION The present invention applies a land price model formula to a land price formation factor such as a road width, a living environment, a distance from a nearest station and a land price evaluated by a real estate appraiser by applying a learning function of a neural network. An object of the present invention is to provide a computer-readable storage medium storing a method, an apparatus, and a program for automatically generating the program.

The present invention also calculates a land price using the land price model formula created as described above, and prepares a land price comparison table (also called a road price comparison table) to explain the calculation result. And a storage medium storing a method, an apparatus and a program for visually displaying as a one-dimensional or two-dimensional graph.

Further, the present invention provides a method and an apparatus for automatically correcting a land price formation factor so that an error between an estimated land price and a land price evaluated in advance is reduced in the land price model formula created above. And a storage medium storing the program.

The method for automatically generating a land price model according to the present invention takes input data of a plurality of records consisting of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data and stores them. The initial value of the weighting factor of the neural network is set in the storage device, the land price formation factor data of a plurality of records stored in the storage device is input to the neural network, and the appraisal land value data is used as a teacher signal, and the neural network is used. The learning is performed until the solution converges, and a set of weighting coefficients of the neural network obtained by the learning is output as a land price model formula.

An apparatus for automatically generating a land price model formula according to the present invention fetches input data of a plurality of records comprising a set of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data. Means for storing in the storage device, means for setting the initial value of the weighting factor of the neural network in the storage device, land price formation factor data of a plurality of records stored in the storage device as input to the neural network, and appraisal land value data as a teacher signal And means for causing the neural network to perform learning until the solution converges, and means for outputting a set of weighting coefficients of the neural network obtained by the learning as a land price model formula.

A program for automatically generating a land price model formula according to the present invention is a program for acquiring input data of a plurality of records comprising a plurality of land price formation factor data and appraisal land price data corresponding to the land price formation factor data. The initial value of the weighting factor of the neural network is set in the storage device, the land price formation factor data of a plurality of records stored in the storage device is input to the neural network, and the appraisal land value data is used as a teacher signal. , The neural network performs learning until the solution converges, and controls the computer so as to output a set of weighting factors of the neural network obtained by learning as a land price model formula.

The recording medium includes all computer-readable recording media such as a semiconductor memory, a magnetic storage device, a magneto-optical storage device, and an optical storage device.

According to the present invention, the land price model formula is generated by using the learning function of the neural network using a large number of record data (land price factor data and appraised land price data). Suitable for many other cases, and reliable land evaluation is possible. It is also possible to generate a regional land price model formula.

In the above-described method, apparatus and program, input data of a plurality of records composed of a plurality of sets of land price formation factor data and corresponding appraisal land price data are converted into variables according to a predetermined function. Learning of the neural network may be performed using the data obtained by the conversion. In this case, the land price model formula is composed of data or formulas for variable transformation, a set of weighting factors obtained by learning using the input data transformed with variables, and data or formulas for inverse transformation of variables. Will be composed of

Since the input data is variable-converted by a predetermined function, it is possible to equalize the influence of the neural network on the data of all the land price forming factors. Examples of variable conversion include variable conversion to a normal distribution with mean 0 and variance 1, variable conversion to a uniform distribution with a maximum value of +1 and a minimum value of -1, a maximum value of +1 and a minimum value of 0. There is a variable change to the like distribution.

In order to obtain estimated land price data using this land price model formula, it is sufficient to input land price formation factor data to a neural network, obtain an output of the neural network, and inversely transform the obtained output. Data representing the estimated land price can be obtained.

The present invention also provides a method for preparing a land price ratio table,
Provides equipment and programs.

According to the method of preparing a land price ratio table according to the present invention, data of one or more of a plurality of land price formation factor items forming a set is changed discretely in one direction and data of other items is changed. Creates land price ratio table creation data with fixed reference data and stores it in a storage device. According to the land price model formula for calculating the estimated land price, the land price formation data included in the land price ratio table creation data Estimated land price data is calculated for all sets of factor data, and land price ratio table data including the land price ratio table creation data and the calculated estimated land price data is output.

The land price comparison table creating apparatus according to the present invention changes the data of one or more of a plurality of land price formation factor items forming a set in one direction and discretely, and changes the data of other items in one direction. Is a means for creating land price ratio table creation data in which reference data is fixed and storing the data in a storage device. According to a land price model formula for calculating an estimated land price, the land price included in the land price ratio table creation data is used. Means for calculating estimated land price data for all sets of formation factor data, and means for outputting land price ratio table data including the land price ratio table creation data and the calculated estimated land price data Things.

The program for preparing a land price comparison table according to the present invention is a program for changing one or more items of one or more of the plurality of land price forming factor items in a set discretely in one direction. For the items of the above, the data for preparing the land price ratio table with the fixed reference data is created and stored in the storage device, and included in the data for preparing the land price ratio table according to the land price model formula for calculating the estimated land price. The computer is controlled to calculate estimated land price data for all sets of land price formation factor data to be output, and to output land price comparison table data composed of the above-mentioned land price comparison table creation data and the calculated estimated land price data. It is for doing.

A land price comparison graph can be displayed based on the created land price comparison table data.

If the land price model specific to the region is used, the land price comparison table created reflects the factors forming the land price specific to the region, has a small error, and can be used as a basis for a rational explanation of the land price.

Using the above-described variable conversion of input data, the land price ratio table creation method according to the present invention can be expressed as follows.

That is, in this method, the data of one or more of the plurality of land price formation factor items forming a set is changed discretely in one direction,
For other items, land price ratio table creation data with fixed reference data is created and stored in a storage device, and the created land price ratio table creation data is variable-converted according to a predetermined function, and the estimated land price is calculated. According to the land price model formula for calculation, the estimated land price data is calculated for all the sets of land price formation factor data included in the converted data for preparing the land price ratio table, and the data for preparing the land price ratio table is calculated. The calculated estimated land price data is inversely converted and output as land price ratio table data.

The land price comparison table creating device and program can be expressed in the same manner.

The method of correcting the land price formation factor according to the present invention is to fetch and store input data of a plurality of records consisting of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data. The device is stored, an item to be corrected among a plurality of land price forming factor items is designated, and an estimated land price for each set of land price forming factor data is calculated for each record according to a land price model formula for calculating an estimated land price. The data of the specified correction target item is corrected so that the difference between the calculated estimated land price and the appraisal land price corresponding to the land price formation factor item is reduced.

Preferably, if the difference is within a predetermined value, no correction is made or the correction is terminated.

An apparatus for correcting a land price formation factor according to the present invention fetches and stores input data of a plurality of records comprising a set of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data. Means to be stored in the device, Means to specify items to be corrected among a plurality of land price formation factor items, Estimated land price for each set of land price formation factor data is calculated for each record according to the land price model formula for calculating the estimated land price Means for correcting the data of the specified correction target item so that the difference between the calculated estimated land price and the appraisal land price corresponding to the land price formation factor item is reduced.

A program for correcting a land price formation factor according to the present invention fetches input data of a plurality of records consisting of a plurality of land price formation factor data and appraisal land price data corresponding to the land price formation factor data. And let the user specify the correction target item among the multiple land price formation factor items, and calculate the estimated land price for each set of land price formation factor data according to the land price model formula for calculating the estimated land price for each record Then, the computer is controlled to correct the data of the specified correction target item so that the difference between the calculated estimated land price and the appraisal land price corresponding to the land price formation factor item is reduced.

If the land price calculated according to the land model formula does not always sufficiently match the appraised land price, on the contrary, the land price formation factors which may be changed are automatically determined according to the above-mentioned method or the like. To make it more reasonable overall.

[0037]

[Embodiment] 1. Generation of land price model formula by neural network In advance, a set of a plurality of land price formation factors and land price data is prepared in a table format. This data is a real estate appraiser,
Created by other experts with reference to actual land prices. A set of these data is called analysis data. The following are the land price formation factors.

・ Width: Road width ・ Road type ・ System ・ Pavement ・ Pavement ・ Sideway ・ Stopping ・ Gradient ・ Station distance: Distance from the nearest station ・ Station sphere: Characteristics of the nearest station (such as whether it is an express stop or not)・ Bus stop distance ・ Primary school distance ・ Water supply ・ City gas ・ Sewerage ・ Land use: Characteristics of the land (residential, commercial, industrial) ・ Topography ・ Commercial density ・ Space ratio ・ Building ratio ・ Floor ratio ・ Terminal approach ・Living environment / railway

There are more than 20 factors that can be considered as land price formation factors. In practice, about 10 to 15 are selected from these to express the relationship between land price formation factors and land prices. Example in which width, station distance, station area, city gas, land use, and floor area ratio were selected as land price formation factors 1, Example 2 in which street width, terminal approach, living environment, station distance, and alongside were selected as land price formation factors These are shown in FIGS. 1 and 2, respectively.

The n land price formation factor data x ₁ , x ₂ ,
..., relationship of x _n and appraisal land values _{z z = F (x 1,} x 2,
..., _xn ) are called land price model formulas. In the following, a procedure for estimating the land price model formula using the function of the neural network will be described.

In order to improve the accuracy of the land price model formula, variable conversion is performed so that the average is zero and the variance is 1 for each of the land price formation factor and the appraisal price. In the following description of this chapter, x, y, t, and s are used as variable names, and f and g are used as function names.

The procedure of the variable conversion is as follows. 1. Sort the data (land price formation factor data or appraisal price data) for one variable and calculate the cumulative distribution. 2. If the variable is regarded as a random variable, the random variable x
Is obtained. FIG. 3 shows an example of the cumulative distribution of the width. This function is set as y = g (x). 3. On the other hand, the probability density function of the normal distribution with mean 0 and variance 1 is known. FIG. 4 shows a cumulative distribution function of a normal distribution having a mean of zero and a variance of one. If this function is set as y = f (x), it is expressed by the following equation.

(Equation 1) This function is also the inverse function. This is x = f ^-1 (y). 4. From the obtained cumulative distribution function, a distribution having a mean of zero and a variance of 1 is obtained by mapping to a probability density function of a normal distribution having a mean of zero and a variance of one. How to map

(Equation 2) It may be performed in the order of.

An example of variable conversion will be described with reference to FIGS. (a) The value of the width is x (= 7.5m). (b) The cumulative distribution value (0.78) is obtained from the cumulative distribution function (FIG. 3).

(Equation 3) (c) Given the probability density (0.78) from the probability density curve (FIG. 4) of distribution 0 with zero mean, the value of the random variable (0.79) is determined.

(Equation 4) (d) Therefore, the variable width of 7.5 m is 0.79 when converted to zero mean and unit variance.

After the n land price formation factor data and the appraisal land prices are variable-converted into a mean of zero and a variance of one, a land price model formula is estimated using a neural network.

As is well known, the neural network has a three-layer structure including an input layer, a hidden layer, and an output layer. The structure of the neural network is shown in FIG. N units in the input layer, m units in the hidden layer
There are one unit in the output layer and one unit in the output layer.
Each unit has characteristics like a multi-input, one-output element.

There are n × m links between each unit of the input layer and each unit of the hidden layer, and m links between each unit of the hidden layer and each unit of the output layer.

The weight between the i-th unit in the input layer and the j-th unit in the hidden layer is represented by v _ji (i = 1, 2,..., N;
j = 1, 2,..., m).

Let G be the output function of each unit.
The input to the j-th unit of the hidden layer is given by [sigma] v _ji × x _i, the use of the output function G, the output of the j-th unit of the hidden layer, can be expressed by G (Σv _ji × x _i) .

Similarly, if the weight between each unit in the hidden layer and one unit in the output layer is w _k (k = 1, 2,..., M), the input to one unit in the output layer is Σw _k ×
expressed as y _k . The output from one output unit can be represented by G (Σw _k × y _k ).

As described above, the input to each unit in the input layer is x _i (i = 1, 2,..., N), and the weight between units is v _ji
(I = 1, 2,..., N; j = 1, 2,..., M), w _k
(K = 1, 2,..., M), the output z from the neural network is given by the following equation.

[0051]

(Equation 5)

The learning of the neural network is to correct the weight between the units so that the output from the neural network becomes a given output result.

Since learning of the neural network is an iterative calculation, an initial value of the weight must be determined. In a normal system, several patterns of initial values determined by random numbers are prepared in advance, and convergence calculation is performed for each pattern. From these multiple patterns, the pattern with the highest accuracy is adopted as the initial value. That is, a pattern in which the error between the output from the neural network and the given output is minimized is adopted, and the weight coefficient between the units obtained by learning starting from this pattern is often set as the optimal solution for convenience. This is because the learning result may be a local solution instead of the optimal solution depending on the initial value.

In this embodiment, as described later, an algorithm for controlling so as not to fall into a local optimum solution is employed. Therefore, it is not necessary to prepare a plurality of initial values. I have.

When there are N sets of land price forming factors and appraisal land prices, one record of the land price forming factors is called a pattern vector, and the appraisal result is called a teacher signal.

The s (s = 1, 2,..., N) -th pattern vector and the teacher signal are expressed as x ^s and d ^s , respectively. In the example of FIG. 1, when only s = 1 and s = 2 are shown, x ^s and d ^s are as follows. (Because the data of the land price formation factor and the appraisal land price that have already been transformed are handled here, It is not appropriate to take the example of FIG. 1, but it will be taken up for clarity).

[0057]

(Equation 6)

The output result of the neural network for each pattern vector is represented by z ^s (v, w) (s =
1, 2,..., N).

(Equation 7) _So that the coefficients v _ji , w _k
Is the ultimate goal.

Assuming that the output of the neural network for the s-th pattern vector is z ^s (v, w) and the output of the j-th hidden layer is y ^s _j ,

(Equation 8) And using this value, z ^s (v, w) becomes

(Equation 9) Becomes

This equation generally has no solution because the number of variables is mn + m and the number of equations is N. Therefore, the weight coefficient between the units is determined so that the square error is minimized.

The square error is calculated by the following equation.

[0062]

(Equation 10)

Assuming that the sum of N errors is an objective function,

[Equation 11] It suffices if the weight coefficient between the units can be determined so that is minimized.

The sigmoid function G _s (x) is often used as the output function G of the unit. By using the sigmoid function, the objective function can be partially differentiated in each of v and w, and the steepest descent algorithm can be applied.

The sigmoid function is

(Equation 12) And is differentiable, which is

(Equation 13) Becomes

When the steepest descent method is used, the total error becomes a local optimal solution by convergence calculation. In the convergence calculation, the weighting factor when the iterative calculation is performed t times
^(t) _ij and w ^(t) _i . The weighting factor in the t + 1-th iterative calculation can be determined by the following equation.

[0067]

[Equation 14]

Here,

(Equation 15) And γ is called a step size and is a value that determines how much the gradient of the function is reflected. If this value is small, a large number of iterations are required to converge, and if it is large, it drifts around the local optimal solution.

The following is the partial derivative of each of the above equations.

(Equation 16) These are formulas for obtaining Δv ^(t) and Δw ^(t) in the steepest descent method. The above is based on a method called a batch correction method, but there is another method called a sequential correction method. The sequential correction method corrects the weight for each pattern vector.

[Equation 17] Calculate with the formula.

The batch correction method has a drawback that the correction of the weights is averaged, so that it is impossible to cope with subtle changes in the pattern vector. On the other hand, the sequential correction method has a disadvantage that when the total number of pattern vectors is large, the weight coefficient changes considerably while learning is performed using all the pattern vectors.

In the analysis data for the land price model formula, the number of pattern vectors is generally about several hundred, so the batch correction method is used.

In the steepest descent method, a local optimal solution can be obtained with a large number of iterations, but there is no guarantee that a global optimal solution can be obtained. Can not).

In order to solve these problems, in this embodiment, the inertia method and the annealing method are used in combination with the learning algorithm.

The inertia method is an algorithm for determining the amount of change by the sum of the weight of the amount of change one step before and the amount of change by the steepest descent method in the iterative calculation for minimizing the error.
That is, Δv ^(t) and Δw ^(t) are

(Equation 18) Is obtained by

Here, ωΔv ^(t-1) and ωΔw ^(t-1) are terms called inertia terms, which can accelerate the convergence calculation.

The annealing method is an algorithm for compensating for a defect that cannot be escaped once it falls into a minimum solution by the steepest descent method, and introduces stochastic fluctuation when determining the amount of change. The fluctuation is set so that the fluctuation is reduced in the process of repeated calculation so that the high-temperature metal is cooled.

[0077]

[Equation 19]

Δ is a value generated by a normal random number at an average of zero so that the deviation is reduced in the process of repeated calculation.

[0079]

(Equation 20)

R is a random number whose average is zero and whose maximum and minimum values are each ± 1. α is a constant representing the strength of the effect of the annealing method, and β is a constant representing the annealing speed.

By using the algorithm shown in equations (39) and (40), the value of the weight coefficient at the time of the t-th repetition calculation can be calculated.

Using equation (18), the sum of errors for N pattern vectors can be calculated.

The total error at the time of the t-th convergence calculation is expressed by E
^(t) . To determine whether or not convergence has occurred,

(Equation 21) Performed by ε is a sufficiently small value. If the convergence has not occurred, the weight coefficient is obtained again.

If it is determined that convergence has occurred, the output result of the neural network for the pattern vector is output using the weight between the learned units. The correlation coefficient is also output as an indication of the accuracy of the neural network. An output example of the learning result is shown in FIG. Here, the opening ratio = {[(appraisal price) / (estimated price)]-1} × 100
It is represented by If the convergence is determined, the land price model formula is also output.

2. Land Price Estimation Function Using Land Price Model Formula Since the land price model formula is a relational expression between n land price formation factors and appraisal prices, as described above, z = F (x ₁ , x ₂ ,
.., X _n ).

In this embodiment, since both the land price formation factor and the appraisal price are variable-converted to a normal distribution with a mean of zero and a variance of 1, the land price formation factors after the variable conversion are x ₁ , x ₂ ,.
_xn , and the appraisal price is z.

The learning result of the neural network is z
= F (x ₁ , x ₂ ,..., X _n ). Therefore,

(Equation 22) According to the procedure described above, the appraisal land price can be obtained from the land price formation factors.

The land price model formula is as follows: 1. Probability distribution curve of land price formation factors x to x (x with “〜” above: hereafter “~ ant”)
(See the example of the cumulative distribution function (width) in FIG. 7) 2. Calculate z (（ant) from the coefficients v _ij , w _k x (〜 ant) of the weight between units (see FIG. 8). (See the example of weight coefficient between units.) 3. Probability distribution curve of appraisal land price z is calculated from z (~ ant) (see example of cumulative distribution function (appraisal price) in FIG. 9) (FIG. 10). reference).

Since the price obtained by the neural network is a value on a normal distribution having a mean of 0 and a variance of 1, the formula (2) can be inversely transformed to the original value. That is, in FIG. 4, the value of the cumulative distribution function on the y-axis is obtained from the value of the random variable on the x-axis on the normal distribution with mean 0 and variance 1. Next, the original value can be obtained by obtaining the value of the x-axis with the obtained value of the cumulative distribution function as the y-axis on the cumulative distribution graph of the appraisal price in FIG.

The method of estimating land prices using the estimated land price model formulas follows the following procedure. 1. Variable conversion of the input land price formation factor data. 2. Calculate the price using the learning result of the neural network. 3. Since the price obtained earlier is obtained as the value of the normal distribution with zero mean and variance of 1, it is inversely converted to the original value.

3. Preparation of Road Price Comparison Table and Comparison Graph As described above, the land price can be estimated by the land price model formula. However, in the land price model formula, the land price formation factor and the appraisal price are variable-converted, and there is a hidden layer (in the multiple regression model, it is a composite variable), so the grounds for land price estimation by the land price model formula will be explained. Is difficult.

In order to explain the land price estimation by the land price model formula, a road price ratio graph and a road price ratio table are prepared.

The road price ratio graph and the road price ratio table are tables for visually presenting the relationship between land price formation factors and appraisal land prices in the land price model formula.

Since the land price model has a relationship between n inputs and one output, it is not possible to visually represent the relationship between all input variables and output variables. Therefore, by changing the value of one or two of the n land price formation factors and fixing the other land price formation factors, the state of the output change of the neural network is visually displayed. .

The land price formation factor data is represented by x _i (i = 1, 2,
..., n). Among the n data, the variables x _j and x _k
Focusing on, respectively which, x _j1, x _j2, ..., is changed from _{x jN, x k1, x k2} , ..., and one that changes the x _kM.

Thus,

(Equation 23) Generate data for

An estimated price is obtained for each of the data obtained above. When the estimated prices are summarized in a two-dimensional table format,

(Equation 24) Becomes

Description will be made with reference to an example of setting the road price ratio graph creating parameter shown in FIG. Average values of width, station distance, station area, city gas, land use, floor area ratio
5.27, 948.2, 1.15, 0.29, 6.03, and 193.8 shall be the reference values for creating the road price comparison graph (the average value is automatically set by the system). Focusing on the width and the station distance, these are changed to 5, 6, 7,..., 100, 200,.

When the data of the road value comparison table is created in this way, a visual display is performed using the function of the spreadsheet software. An example of a road price ratio graph is shown in FIG.

4. Automatic correction of land price formation factor data In the land price model automatically generated from the analysis data, an error remains between the appraisal price and the estimated price. That is, even if the weight between the units is learned, the sum E (v, w) of the square errors calculated from the N pattern vectors does not become zero.

As a method of matching the estimated price of the land price model formula with the actual appraisal price, there are: 1. Addition of land price formation factor data 2. Correction of land price formation factor data 3. Change of appraisal price.

Regarding 1., the error may be assigned to another variable, but it is difficult to find a variable having sufficient explanatory power as a variable of land price formation factor data. For item 2, the land price formation factor data includes objective data (distance from the station) and subjective data (land use characteristics). For the latter, different values may be set for each appraiser. sell. For item 3, the appraisal price will not be changed unless there is a special reason. Adopt 2 for automatic correction processing, correct the subjective data, and automatically correct the estimated price of the land price model formula to match the appraisal price.

The land price model formula obtained by generating the land price model formula by the neural network is

(Equation 25) Can be expressed as

The above equation is a function having n variables as input and one variable as output. Assuming that m variables to be automatically corrected (m <n) and the other variables are constants,

(Equation 26) Is obtained. This function uses the variable x _i (~ ant) (i
= 1, 2,..., M) as variables.

The square error of one pattern vector is given by e =
The variable x _i (〜ant) is corrected so as to make ｚ (z−d) ² and minimize e.

[0106]

[Equation 27]

Here, γ is a small value, and the value of x _i (〜ant) at the time of the t-th repetition calculation is x ^(t) _i (〜ant).

By this repetitive calculation, the variable x _i (〜ant) (i = 1, 2, 2) is set so that the square error e becomes the minimum value.
.., M) can be calculated, and the output result of the land price model formula can be matched with the appraisal price.

[0109] de / dx _i can be calculated in the following manner.

[0110]

[Equation 28]

The calculation algorithm is as follows. 1.t = 0 2.i = 1 3.x ^{(t + 1)} _i (~ ant) = x ^(t) _i (~ ant) + Δx _i (~
Ant) 4. While i <m, calculate i = i + 1 5.2 squared error e. 6. If e <ε, end 7. If e> ε, go to 2. as t = t + 1

5. Sensitivity analysis of land price model A sensitivity analysis function is provided. The model equation of multiple regression analysis describes the relationship between land price formation factors and appraisal land prices.

(Equation 29) Therefore, which of the n land price formation factors is the major factor in land price determination is determined by each a _i
(I = 1, 2,..., N).

In the model formula using the neural network, there is a hidden layer between the input layer and the output layer, and it cannot be determined only by the weight coefficient between units.

First, a gradient vector ▽ Z (〜ant) in the space subjected to variable conversion is defined, and then a gradient vector ▽ Z in the space before variable conversion is defined.

By using this gradient vector ▽ Z, it is possible to analyze the sensitivity of each local value forming factor.

[0116]

[Equation 30] here

(Equation 31) It is.

In the above equation, since v _ji and w _k are known, the gradient vector ▽ Z (▽ ant) can be calculated.

Before the variable conversion, the gradient vector ▽ Z is

(Equation 32) Can be defined as

As is clear from the relationship between x _i and x _i (〜ant) (see FIGS. 3 and 4), dx _i (〜ant) / dx _i
Cannot be calculated mathematically. Therefore,

[Equation 33] Approximately, it will be calculated by the following equation.

The gradient vector ▽ Z can be calculated from the equations (56), (57) and (58). That is, by substituting the land price formation factor data into the above equation, a gradient vector at an arbitrary point can be calculated.

6. Land Price Evaluation System FIG. 13 is a perspective view showing the overall hardware configuration of the land price evaluation system, and FIG. 14 is a block diagram showing the electrical configuration.

The land price evaluation system includes a computer 10. The computer 10 has an internal memory (semiconductor memory or the like) 11 built therein. Display device 1 as input / output device
2, a printer 13 and a keyboard / mouse 14 (the keyboard is shown at 14A and the mouse at 14B) are connected to the computer 10. FD drive as storage device
15, a CD-ROM drive 16 and an HD unit 17 are connected to the computer 10. The FD drive 15 reads data or programs stored in the floppy disk 18 or writes data or programs to the floppy disk 18. CD-
The ROM drive 16 reads a program stored in the CD-ROM 19. The HD unit 17 reads / writes data and programs from / to a hard disk.

As shown in FIGS. 15 to 19, the internal memory 11 has an analysis data variable area, a "cumulative distribution of price" variable area, a "variable distribution of width" variable area, and a "cumulative distribution of station distance" variable. Area, area of other land price formation factor data variables, “normal distribution” variable area, “weight coefficient between input layer and hidden layer” variable area, “weight coefficient between hidden layer and output layer” variable area, The change area of the “weight coefficient between the input layer and the hidden layer this time” variable, the change area of the “weight coefficient between the hidden layer and the output layer this time” variable, the “output result of the neural network” variable area, Variable area such as “Total of squared errors of this time”, “Total of squared errors of previous time”, “Reference data for comparison table” variable area, “1st attention” variable area, “2nd attention” variable area, etc. Data for ratio table "variable Rear, “Roadline ratio table” variable area, “Modification 1” variable area, “Modification 2” variable area, etc., “Learning result” variable area, “Change amount of modification 1” variable area, “Modification target” A “variation amount of 2” variable area is provided.

The land price model formula generation program, the road price ratio table, the road price ratio graph creation program, and the land price formation factor data automatic correction program shown in FIGS. 20 to 24 are stored in the floppy disk 18. Installed on the hard disk. In actual execution of the program, the program is loaded into the internal memory 11. Of course, some or all of these programs may be stored in the CD-ROM 19.

FIG. 20 and FIG. 21 are flow charts of a process of generating a land price model formula by a neural network.

Recorded on a floppy disk,
The analysis data is read into the internal memory 11 and stored in the "analysis data" variable (area) (step 31). According to Example 1 of Table 1, the first line is a “price” variable, the second line is a “width” variable, and the third line is a “station distance” variable. The same applies to the fourth and subsequent lines.

For the “price” variable, the stored data is sorted and the cumulative distribution is calculated. The calculated value of the cumulative distribution is a set of pairs of price data and cumulative frequency data such that the value of the cumulative distribution data is 1 when the price is the minimum value and is equal to the number of data when the price is the maximum value. Dividing this cumulative distribution by the number of data gives the cumulative density function. The result is stored in a variable “cumulative distribution of price” on the internal memory 11 (area). This variable is an array variable with several data areas. Similarly for other variables,
It is stored in a variable such as "cumulative distribution of width" variable (area). On the other hand, a normal distribution with zero mean and one variance (see FIG. 4) is read from the floppy and stored in a “normal distribution” variable (area) on the memory. Variable conversion is performed using the “cumulative distribution of price” variable and the “normal distribution” variable (step 32) (see FIGS. 3 and 4). Each variable after the variable conversion is newly set as a “price” variable, a “width” variable, and a “station distance” variable. These are collectively referred to as “analysis data” variables.

Initial values are set to variables “area” of “weight coefficient v _ij between input layer and hidden layer” and variables (area) of “weight coefficient w _k between hidden layer and output layer” on the memory. Also, "the change amount of the weight coefficient between the input layer and the hidden layer this time"
The variable (area) and the “change amount of the weight coefficient between the hidden layer and the output layer this time” are each set to an initial value of zero (step 33).

The following processing is performed for N records one record at a time from the “analysis data” variable. "Width" variable,
"Station distance" variable (there are other land price formation factor variables, these two are mentioned as representatives), "weight coefficient between input layer and hidden layer" variable, "between hidden layer and output layer" The values of the “weighting factor” variable are substituted into equations (5) and (6) to calculate the output result of the neural network (step
34). The result is stored in the “output result of neural network” variable (area).

“Weight coefficient between input layer and hidden layer” variable,
The amount of change of the “weight coefficient between the hidden layer and the output layer” variable is determined by the following procedure (step 36).

The value of the “price” variable, the value of the “output result of the neural network” variable, and the value of the “variation of the weighting coefficient between the current input layer and the hidden layer” are calculated by the equations (21) and (2).
3), (24), (29), (37), (39), and (41), and calculate the amount of change of the “weight coefficient between input layer and hidden layer” variable. The amount of change is stored in a variable of “the amount of change in weight coefficient between the current input layer and the hidden layer”. Next, the “weight coefficient between input layer and hidden layer” variable is calculated by adding the amount of change to the current value of the “weight coefficient between input layer and hidden layer” variable.

Similarly, the amount of change in the “weight coefficient between the hidden layer and the output layer” variable is calculated, and the amount of change is calculated as the “amount of change in the weight coefficient between the current hidden layer and the output layer” variable (area ). Next, the “weight coefficient between the hidden layer and the output layer” variable is calculated by adding the current value of the “weight coefficient between the hidden layer and the output layer” and the amount of change thereof.

Using the “weight coefficient between input layer and hidden layer” variable and the “weight coefficient between hidden layer and output layer” variable after modification, the “output result of neural network” is calculated. (Area) (step 37).

The value of the “price” variable in the “output result of the neural network” and the “data for analysis” variable is substituted into equation (18), and the sum of the square errors of the N data is calculated. The result is stored in a “sum of the current square error” variable (area), and the previous result is stored in a “sum of the previous square error” variable (area) (step 38).

If the difference between the “sum of the square errors of this time” variable and the “sum of the previous square errors” variable is sufficiently small, it is considered that the convergence has been completed, and the processing is terminated (step 38). If these differences are not sufficiently small, the process continues from step 34.

When the convergence has occurred, the variables “cumulative distribution of price”, “cumulative distribution of width”, “cumulative distribution of station distance”, “weight coefficient between input layer and hidden layer”, “hidden layer” Each value of the "weight factor between the output layer" variable is output to the floppy (step 40). The above is called the land price model formula.

The values of the “price” variable and the “output result of the neural network” variable are inversely transformed, and the results are output to the floppy (step 41). The method of the inverse transformation is as follows. Since the value of the “price” variable obtained on the cumulative distribution function in FIG. 4 is the value of the probability density (horizontal axis), the value of the cumulative distribution (value on the vertical axis) is obtained from the “normal distribution” variable. Next, using the obtained cumulative distribution value and the cumulative distribution curve of “price”, the value is converted into an actual “price” (see FIG. 10). The data in Fig. 10 is the "cumulative distribution of prices" on the memory.
Since this is stored in a variable (area), this "price cumulative distribution" variable is used. With the above procedure, the price is calculated for one record of the “data for road price ratio table” variable. In addition, the correlation coefficient between the “price” variable and the “neural network output result” variable is calculated and output to the floppy. This is called a learning result (see FIG. 6).

FIG. 22 is a flow chart of a process of preparing a road value comparison table and a comparison graph.

The reference data on the floppy disk is read (step 51). The reference data for this comparison table is:
The state of the estimated price accompanying the change of the land price formation factor data is shown in 3
This is a parameter for creating a three-dimensional graph for dimensional representation. In this graph, when two variables of the land price formation factor data are taken on the x-axis and the y-axis, and the other variables are fixed, it is estimated with the change of the values on the x-axis and the y-axis. The value of the land price (calculated using the above) is plotted on the z-axis. The average value of the land price formation factor data is calculated in a process of generating a land price model formula by a neural network, and is set in advance as a reference value of reference data for creating a comparison table on a floppy disk. The user of the system can create a three-dimensional graph by selecting two variables from the land price formation factor data and giving their coordinate axes. In the example of FIG. 11, focusing on “width” and “station distance” as two variables, the user of the system assigns the respective coordinate values to 5, 6, 7,..., 16, and 100, 200, 300,
…, It is assumed that 1500 is set. The user can load the floppy into the computer and input and set these values using a normal program (for example, spreadsheet software).

The system uses, as reference values from the reference table creation reference data on the floppy disk, a “price reference” variable, a “width reference” variable, a “station distance reference” variable, etc. in the memory (“comparison table creation reference data”). Variable area). next,
As the two variables of the land price formation factor data, “width 1” is used as the coordinate value of the “target 1” variable 5, 6, 7,.
Is read into memory. Read "100, 200, 300, ..., 1500" into the memory with "station distance" as the coordinate value for the "attention 2" variable.

A “price-based” variable, a “width-based” variable,
Using the “station distance reference” variable, the “attention 1” variable, and the “attention 2” variable, the data is expanded into “data for road line price ratio table” variables on the memory. This variable has an array structure as shown in equation (44).

The land price model formula output to the floppy is read. First, variables such as a "cumulative distribution of price" variable, a "cumulative distribution of width" variable, a "cumulative distribution of station distance" variable, and the like are read. Next, the weight coefficients between the units are read into the “weight coefficient between the input layer and the hidden layer” variable and the “weight coefficient between the hidden layer and the output layer” variables on the memory. Finally, the "normal distribution" variable is read (step 53).

An estimated price is calculated by the following procedure for each record of the “data for road price ratio table” variable (step 5).
Four). Variable conversion is performed on the “price” data of one record from the “cumulative distribution of prices” variable and the “normal distribution” variable (see FIGS. 3 and 4). Similarly, variable conversion is performed for “station distance” data and “station area” data. The values after the variable conversion and the values of the “weight coefficient between the input layer and the hidden layer” variable and the “weight coefficient between the hidden layer and the output layer” are expressed by Equation (5).
Substitute in (6) to calculate “land price”. The obtained “land price” is inversely converted by the method described in step 41. This is repeated for n records.

The obtained n prices are stored in the memory by the formula (4
The data is converted to the format of 5) and output to the floppy disk as a "road price comparison table" (step 55).

By executing the spreadsheet software and reading the "road value ratio table" on the floppy disk and displaying it as a three-dimensional contour graph using the graph display function, FIG. 12 is obtained.

FIGS. 23 and 24 are flow charts of the automatic correction processing of the land price formation factor data.

The variable of the land price formation factor data to be automatically corrected is input by a variable number. In the example of FIG. 2, when 3 and 5 are input, it means that “living environment” and “roadside” are selected as correction targets in the land price formation factor data.
These are referred to as “modification target 1” variables and “modification target 2” variables.

The analysis data (eg, FIG. 2) is read from the floppy and stored as a “analysis data” variable on the memory. Also, the learning result (see FIG. 6) output to the floppy is read and stored in the “price” variable and the “neural network output result” variable in the memory.
Together, these are used as a “learning result” variable.

The land price model equation output to the floppy is read in the same manner as in step 53 of FIG. 22 (step 63).

Variables are converted one by one for n records of the “analysis data” variable (step 64), and the variables after the variable conversion are set as “width” variables, “station distance” variables, etc. (step 65). . The n records of the “learning result” variable are also converted one by one into a “price” variable and a “neural network output result” variable. The square error between the value of the converted "price" variable and the value of the "output result of the neural network" variable is calculated (step 66).
If this error is smaller than a small value ε (step 67),
Return to step 64 for processing the next record. If this error is larger than a small value ε, the following automatic correction calculation is performed.

Assuming that i = 1, the change amount of the i-th variable is calculated (steps 68 and 69). In other words, "analysis data"
The value of the variable and the value of the variable to be modified 1
Substitute in 2) and calculate the amount of change in the “Modification 1” variable. This value is referred to as a “change amount of the correction target 1” variable. Similarly, the value of the “change amount of correction target 2” variable is obtained. That is, 1 is added to i, and the processing of step 69 is repeated until i reaches m (steps 70 and 71).

The value obtained by substituting the value of the “correction target 1” variable and the value of the “variation amount of the correction target 1” into equation (48) is defined as the value of the “correction target 1” variable. Similarly, the value of the “modification target 1” variable is obtained. In the example of FIG. 2, the “living environment” variable and the “railway” variable in the “analysis data” variable have been respectively corrected. "Analysis data" variable, "correction target 1" variable, "correction target 2" variable value, "weight coefficient between input layer and hidden layer" variable, "weight coefficient between hidden layer and output layer" The value of the “” variable is substituted into equations (5) and (6) to calculate a “output result of the neural network” variable (step 70).

A square error is calculated from the value of the “output result of the neural network” variable obtained in step 72 and the value of the “price” variable (step 73). If this square error is smaller than the minute value ε (step 74), the step
Return to 64 to process the next record. If this error is larger than the small value ε, the process returns to step 68.

When the processing for all records is completed, the “analysis data” variable and the “learning result” variable are inversely transformed and each is output to the floppy (step 75,
76, 77).

[Brief description of the drawings]

FIG. 1 shows an example of data for analysis.

FIG. 2 shows another example of analysis data.

FIG. 3 is a graph showing an example of a cumulative distribution function of width.

FIG. 4 is a graph showing an example of a cumulative distribution function of a normal distribution.

FIG. 5 shows the structure of a neural network.

FIG. 6 shows an output example of a learning result.

FIG. 7 shows an example of cumulative distribution function data (width).

FIG. 8 shows an example of a weight coefficient of a neural network.

FIG. 9 shows an example of cumulative distribution function data (appraisal price).

FIG. 10 is a graph showing an example of a cumulative distribution function of appraisal prices.

FIG. 11 shows an example of setting parameters for creating a road price ratio graph.

FIG. 12 shows an example of a road value comparison graph.

FIG. 13 is a perspective view showing the appearance of the land evaluation system.

FIG. 14 is a block diagram showing an electrical configuration of the land evaluation system.

FIG. 15 shows the contents of an internal memory.

FIG. 16 shows the contents of an internal memory.

FIG. 17 shows the contents of an internal memory.

FIG. 18 shows the contents of an internal memory.

FIG. 19 shows the contents of an internal memory.

FIG. 20 is a flowchart showing a processing procedure according to a land price model formula generation program using a neural network.

FIG. 21 is a flowchart showing a processing procedure according to a land price model formula generation program using a neural network.

FIG. 22 is a flowchart showing a processing procedure according to a road price ratio table and a road price ratio graph creation program.

FIG. 23 is a flowchart showing a processing procedure according to an automatic correction program for land price formation factor data.

FIG. 24 is a flowchart showing a processing procedure according to an automatic correction program for land price formation factor data.

[Explanation of symbols]

 10 Computer 11 Internal memory 12 Display device 15 FD drive

Claims (15)

[Claims] An input data of a plurality of records including a set of a plurality of land price formation factor data and appraisal land price data corresponding to the land price formation factor data is taken and stored in a storage device, and the storage device is provided with a neural network. The initial value of the weighting factor is set, the land price formation factor data of a plurality of records stored in the storage device is input to the neural network, and learning is performed on the neural network using the appraised land value data as a teacher signal until the solution converges. A method of automatically generating a land price model formula that outputs a set of weighting factors of the neural network obtained by learning as a land price model formula. 2. Input data of a plurality of records, each of which is a set of a plurality of land price formation factor data and corresponding appraisal land price data, are variable-converted according to a predetermined function, and the data obtained by the variable conversion is converted. 2. The method of claim 1, wherein the method is used to perform neural network learning. 3. The land price model formula is obtained from data or formulas for variable conversion, a set of weighting factors obtained by learning using the input data subjected to variable conversion, and data or formulas for variable reverse conversion. The method of claim 2, wherein the method comprises: 4. Inputting land price formation factor data to a neural network, obtaining an output of the neural network,
4. The method according to claim 3, wherein the obtained output is inversely transformed to obtain data representing an estimated land price. 5. A land price ratio in which one or more items of one or more of the plurality of land price forming factor items are discretely changed in one direction and reference data is fixed for other items. Create the table creation data and store it in the storage device. According to the land price model formula for calculating the estimated land price, the estimated land price data for all sets of land price formation factor data included in the data for creating land price comparison table Calculating a land price ratio table and outputting the land price ratio table data including the data for preparing the land price ratio table and the calculated estimated land price data. 6. The method according to claim 5, wherein a land price comparison graph is displayed based on the land price comparison table data. 7. A land price ratio in which one or more items of one or more of the plurality of land price formation factor items are discretely changed in one direction and reference data is fixed for other items. 3. Data for creating a table and storing it in a storage device, and converting the created data for creating a land price ratio table into variables according to a predetermined function.
Calculate the estimated land price data for all the sets of land price formation factor data included in the converted land price comparison table creation data according to the land price model formula for calculating the estimated land price generated according to the method described in A land price ratio table creation method for inverting the land price ratio table creation data and the calculated estimated land price data and outputting the converted data as land price ratio table data. 8. A plurality of land price formation factor data and input data of a plurality of records each consisting of a set of appraisal land price data corresponding to the land price formation factor data are fetched and stored in a storage device. Specify the items to be modified among the items, calculate the estimated land prices for each set of land price formation factor data according to the land price model formula for calculating the estimated land prices for each record, and add the calculated estimated land prices and land price formation factor items To reduce the difference from the corresponding appraisal land price,
A method of correcting the land price formation factor by correcting the data of the specified correction target item described above. 9. The method according to claim 8, wherein the correction is not performed or the correction is finished if the difference is within a predetermined value. 10. A means for fetching input data of a plurality of records comprising a set of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data, and storing the input data in a storage device, and a neural storage device. Means for setting the initial value of the network weighting factor, learning the land price formation factor data of multiple records stored in the storage device as input to the neural network, and using the appraisal land price data as the teacher signal until the solution converges to the neural network And a means for outputting a set of weighting coefficients of the neural network obtained by learning as a land price model formula. 11. A land price ratio in which one or more of one or more of the plurality of land price formation factor items in a set is changed in one direction in a discrete manner, and the reference data is fixed for other items. A means for creating table creation data and storing it in a storage device, and in accordance with a land price model formula for calculating an estimated land price, an estimated land price is calculated for all sets of land price formation factor data included in the data for creating a land price comparison table. An apparatus for preparing a land price ratio table, comprising: means for calculating data; and means for outputting land price ratio table data including the data for preparing the land price ratio table and the calculated estimated land price data. 12. A means for fetching input data of a plurality of records consisting of a set of a plurality of land price formation factor data and appraisal land price data corresponding to the plurality of land price formation factor data and storing the input data in a storage device, and a plurality of land price formation factor data. Means for designating the items to be corrected among the factor items, means for calculating the estimated land price for each set of land price formation factor data in accordance with the land price model formula for calculating the estimated land price for each record, and the calculated estimated land price and land price An apparatus for correcting a land price formation factor, comprising: means for correcting data of the specified correction target item so as to reduce a difference from an appraisal land price corresponding to the formation factor item. 13. Input data of a plurality of records comprising a set of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data is stored in a storage device, and the neural network is stored in the storage device. The initial value of the weighting factor is set, the land price formation factor data of a plurality of records stored in the storage device is input to the neural network, and learning is performed on the neural network using the appraised land value data as a teacher signal until the solution converges. In addition, a medium storing a program for automatically generating a land price model formula for controlling a computer so as to output a set of weighting factors of the neural network obtained by learning as a land price model formula. 14. A land price ratio in which one or more items of one or more of the plurality of land price forming factor items are discretely changed in one direction and reference data is fixed for other items. Create the table creation data and store it in the storage device. According to the land price model formula for calculating the estimated land price, the estimated land price data for all sets of land price formation factor data included in the data for creating land price comparison table And recorded a land price ratio table creation program for controlling a computer to output the land price ratio table data including the above-mentioned land price ratio table creation data and the calculated estimated land price data. Medium. 15. Input data of a plurality of records comprising a set of a plurality of land price formation factor data and appraisal land price data corresponding to these land price formation factor data are stored and stored in a storage device. Items to be corrected among the items are designated, and the estimated land price for each set of land price formation factor data is calculated for each record according to the land price model formula for calculating the estimated land price, and the calculated estimated land price and land price formation factor A medium storing a program for correcting a land price formation factor for controlling a computer to correct the data of the specified correction target item so as to reduce a difference from a corresponding appraisal land price.