JP2003058579A - Method for optimizing design/blending - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for efficiently and easily calculating an optimum value of factor data to produce optional characteristic data by using a method of a neural network and learning a mapping relation between a factor group and a characteristic group of experimental data of design/blending, etc. SOLUTION: A learning data table in which each experimental data in blending/design is united in a tabular format subjected to attribute division between a factor and an attribute is prepared, each column of the learning data table is selected, the each column is allocated to an input layer or output layer, the mapping relaton between a factor group and a characteristic group of the experimental data is learned, the learning results is used to calculate an optimum value of factor data to produce optional characteristic data, also the optimization of design/blending is realized in a GUI environment, and the neural network is learned and the optimization of design combination for blending is carried out by selecting a function tab on a computer screen.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention learns a mapping relationship between a factor and a characteristic by using a method of a neural network in various product design, compounding design, condition setting of a manufacturing apparatus, etc. On the other hand, the present invention relates to a method for obtaining the optimum value of the factor data that produces it.

[0002]

2. Description of the Related Art Conventionally, multiple regression analysis is known as a method of analyzing experimental data in various product designs, compounding designs, or setting of conditions for manufacturing equipment. If there are not enough numbers, accuracy cannot be obtained, there are many errors in the analysis of strong non-linear phenomena because the regression formula is applied to at most quadratic, and there is some statistical expertise to master the above method. However, there is a problem that there is no software that can be easily used. Further, in the regression analysis, graphs and tables are often used, but the GUI (graphical user interface) is not easy to understand visually. On the other hand, as a method of analyzing a strong nonlinear phenomenon, a. Multi-input 1-output ,. Output switching characteristics ,. A method of simulating the relationship between factors and characteristics in product design and compounding design is actively used by using a neural network in which neuron models simulating variable synapse weights are connected in a network. This forms a neural network in which the neurons of the input layer, the intermediate layer, and the output layer are connected by synapses, and, for example, the values of the shape and structure of the product are input to the input layer, and the intermediate layer is The product characteristic is output from the output layer via the output layer. When each value of the factor is input after learning, which will be described later, the corresponding characteristic is output.

FIG. 16 is a diagram showing a structural example (hierarchical NN) of the above-mentioned neural network. This neural network has units I ₁ , I ₂ , corresponding to neurons.
.., an input layer consisting of I _p , a number of units M ₁ , M ₂ ,
.., an intermediate layer composed of M _q , and units U ₁ , U ₂ ,.
.., U _{r has a} multi-input / multi-output structure composed of an output layer. In the figure, B ₁ and B ₂ are offset units. In the input layer unit,
For example, design parameters such as product shape and structure, manufacturing conditions, or values of factors such as composition ratios of compositions are used as input values, and the unit of the output layer is, for example, the surface roughness of the product or composition. Characteristic values such as tensile strength, wear resistance, and heat resistance are used as output values. In addition, in general,
For the above-mentioned input / output relationship, a basis function f (x) representing a characteristic of a neuron such as a standardized sigmoid function is used, and output values are 0 to 1. In the above neural network, when each input signal is x _i , each output y _i of the unit of the intermediate layer and the unit of the output layer can be expressed by the following equations (1) and (2).

[Equation 1] However, p The number of synapses on the input side, w _ji ; the load b _j corresponding to the strength of each synaptic connection between layers (i, j); the offset value, and therefore, by inputting each value of the above factors of the input layer, Each value corresponding to the number of product characteristics is output from the unit of the output layer. The weight w _{ji of} each unit is calculated from the error of known prediction data, that is, the actual measurement value that is the data of the characteristic item used for learning and the data of the factor item used for learning by the learning process described later. Learning and correction are performed so that the error between the estimated value and the estimated value is minimized.

Next, the learning processing of the neural network will be described with reference to the flowchart of FIG. First, the learning data and the test data obtained in advance are read (step S50), and each weight w _ji and the offset value b _j in the neural network are set to predetermined values for initialization (step S5).
2). Next, in order to train the neural network using the plurality of input learning data, each weight w
at least one of _ji and offset values b _j is changed little by little obtains the error of each of the intermediate layers and the output layer, to learn so that the error is minimized (step S54). The error E _j (w) can be represented by the following equation (3), where y _j is the output and v _j is the teacher data.

[Equation 2] Then, each load w obtained above_jiAnd offset value b_j
Is updated (rewritten) (step S56), and then the above update
Neural network based on each weighted load and offset value
Test each piece of test data by work
Characteristic data is obtained as a result value (step S58).
Then, the value of the above test result is the criterion for the convergence judgment.
Whether it converged by judging whether it is a value within a certain range
Whether or not to repeat the above process a predetermined number of times
If it is judged positive (step S60),
If so, the learning ends. On the other hand, in the case of negative judgment,
The above process is repeated. This allows you to enter the learning data
Error of each unit in the middle layer and the output layer
Each load w to be the minimum_jiAnd offset value b _jBut
Determined. When the above learning is completed,
Each load w_jiAnd offset value b_jUsing
Enter each value of the causal condition and enter the value that represents the product performance (characteristic).
Change the above factor conditions by outputting as an estimated value
It is possible to simulate changes in characteristics when
It

[0005]

However, although the above neural network method has the advantage that it can be performed without the user having specialized knowledge of statistics, data preparation, learning, and
It was difficult for even a beginner to easily realize the series of flow up to the use for analysis. Further, it is difficult to simulate and optimize the factor condition for generating arbitrary characteristic data only by obtaining the mapping relationship between the factor group and the characteristic group or by estimating the characteristic data for the factor data group. .

The present invention has been made in view of the conventional problems, and learns the mapping relationship between the factor group and the characteristic group of the experimental data such as design / mixture by using the method of the neural network to arbitrarily It is an object of the present invention to provide a method for efficiently and easily obtaining the optimum value of factor data for producing the characteristic data of the above.

[0007]

A method of optimizing a design / compound according to claim 1 of the present invention uses a neural network method to use a factor group and a characteristic group of experimental data in product design, formulation design, etc. When learning the mapping relationship with and designing / compounding optimization, create a learning data table in which the above experimental data is divided into a column of factor items and a column of characteristic items, and each column of this learning data table After selecting and assigning it to the input layer or the output layer to learn the above mapping relation, the optimization method is used to find the optimum value of the factor data that creates it for any characteristic data. In addition to the neural network configured to allocate the factor data group to the input layer and the characteristic data group to the output layer, the characteristic data group to the input layer and the factor data group to the output layer. Structure to allocate Since can build a neural network, for any characteristic data, it an optimum value of factor data can be determined accurately to produce, it is possible to improve the efficiency of various design work and accuracy.

According to a second aspect of the present invention, there is provided a method for optimizing a design / compounding method, wherein when learning the mapping relationship, a column of factor items and a column of characteristic items of the learning data table are set as input layer items and output items. By providing means for setting either a layer item or an item to be ignored, selection and setting of factor items and characteristic items to be learned can be performed easily and efficiently. The method of optimizing the design and composition according to claim 3 is characterized in that a Newton method or a combination method is used as the optimization method.

Further, in the method of optimizing the design and composition according to the fourth aspect, in order to improve the efficiency of optimization, when obtaining the optimum factor data, the factor values are set to the initial value, the maximum value and the minimum value. A value is set, and the factor value is increased or decreased between the initial value and the maximum value or the minimum value. The design / compound optimization method according to claim 5 sets the estimated value obtained last time to a new target value of the characteristic,
This is characterized in that the optimization process is performed again to obtain the optimum factor data for obtaining the desired characteristics, which makes it possible to improve the efficiency of the optimization. The design / compound optimization method according to claim 6 is:
It is characterized by setting the factor value that produced the estimated value obtained last time to the initial value of a new factor value, performing optimization processing again, and obtaining the optimum factor data. This makes it possible to further improve the efficiency of optimization.

Further, the designing / compounding optimizing method according to claim 7 is characterized in that each of the characteristic groups is lightly weighted when the optimum factor data is obtained. As a result, the convergence of optimization can be improved. The method of optimizing the design and composition according to claim 8 is characterized in that an allowable variation amount is added to each of the characteristic groups when obtaining the optimum factor data. Can further improve the convergence. The design / compound optimization method according to claim 9 sets a plurality of optimum frequencies defined by a convergence determination condition,
When obtaining the optimum factor data, one of the above-mentioned optimum frequencies is selected for optimization, whereby the degree of optimization according to the required accuracy can be determined. Since it can be set, it is possible to efficiently perform optimization. The design / compound optimization method according to claim 10 is:
It is characterized in that each of the factor groups is given a slight severity when determining the optimum factor data.
It is possible to further improve the convergence in the optimization.

The designing / compounding optimizing method according to claim 11 is characterized in that a function tab is selected on a computer screen to perform learning and optimization of the neural network. By this, G
Since the optimization can be performed in the UI environment, it becomes possible to easily and efficiently optimize the design and formulation.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing an outline of a personal computer for carrying out the optimization of the design and formulation according to the present embodiment. In this example, a graphical user using a personal computer 10
In the interface (GUI) environment, learn the product / mixture using the neural network method,
Using the above learning result, optimization is performed for any characteristic data to find the optimum value of the factor data that creates it. In the figure, 11 is a personal computer 10.
Is a display body, 12 is a computer main body, and 13a and 13b are input means keyboard and mouse, respectively. As shown in FIG. 2, the computer main body 12 includes a CPU 14 that performs arithmetic processing according to a processing program described later, a ROM 15, a RAM 16 and a memory 17 that are storage means for storing data, programs, calculation results, etc., and the main body 12 and input / output. An I / O board 18 for exchanging data and the like with a device and a bus 19 connected so that data and commands can be input and output.
It has and. Note that the above data is the keyboard 13
It may be directly input to the computer main body 12 from a,
The data FD may be input via the FDU 12F built in the personal computer 10.

Next, the optimization method of this embodiment will be described with reference to the flowchart of FIG. First,
The personal computer 10 is started up, the design / compounding simulation system program is started, and the initial screen G0 as shown in FIG. 4 is displayed on the display screen of the display 11. Then, the learning data table column g1 displayed above the initial screen G0 is clicked to call the data table creation screen G1 as shown in FIG. 5 to create the learning data table (step S10). In this example, the learning data table is executed on Excel, which is a spreadsheet application software installed in a normal personal computer. Specifically, after inputting the number of factors, the number of characteristics, and the number of data in the data creation field g11 of the data table creation screen G1, the table creation button B11 is clicked to call the Excel book, and the figure is displayed on this Excel book. As shown in FIG. 6, a learning data table 10T in which data values are arranged in each cell having columns of factor conditions and characteristics used for design / compounding simulation and rows of experimental data is created. At this time, the experimental data may be input by the keyboard 13a, or as described later, the file operation column g7 in FIG. 4 may be clicked to read out the learning data table previously created in the Excel book (FIG. 12
reference). Note that calculations such as spreadsheets and graph creation are
VBA (Visual Ba), which is a programming language of Excel
sic Applications), but if the number of learning data groups is large, it is a binary file D containing a compiled function set inside.
By using LL (Dynamic Link Library),
The processing speed can be increased.

Next, the input / output setting column g2 is clicked to call the input / output setting screen G2 as shown in FIGS. 7 (a) and 7 (b) and displayed in the respective assignment columns g21, g22, ..., G29. The respective columns of the factor items and the characteristic items of the learned data table 10T are sequentially set to any of the items of the input layer, the items of the output layer, or the items to be ignored (set to default values) (step S12). ). As a result, it is possible to easily select and set factor items and characteristic items to be learned, so that efficient learning can be performed. In the first learning, normally, all factor items are assigned to the input layer and all characteristic items are assigned to the output layer. afterwards,
The learning field g3 is selected, and learning of the neural network, that is, learning of the mapping relationship between the factor group and the characteristic group is performed using the experimental data such as the above-mentioned design / mixture.

When the learning field g3 is selected, a learning screen G3 as shown in FIG. 8 is displayed, and the learning parameters and the learning method can be set using this screen. First, the parameter setting field g of the neural network
Set the parameter values such as the number of hidden layer nodes, the maximum value (limit) of the number of repeated calculations, the convergence determination limit, the slope of the dicuoid shown in 31, and the learning order (step S14). And the mapping relationship of the characteristic group are learned, and the weight w _ji and the offset value b _j are obtained (step S16). In the present invention, for the purpose of improving learning efficiency, it is possible to select either a normalized value (0 to 1) or a value in an actual unit as the value of the convergence determination limit, and In addition to the usual learning parameters,
The correction ratio η of the middle layer, the load correction ratio η ′ of the output layer, the moment coefficient that is the correction ratio to the immediately preceding load correction amount used when correcting the load, the amount of shaking of the saturated node when learning is saturated, The saturation check interval and the number of moving averages of saturation determination can be set as learning parameters.

In general, the learning end is determined by the difference between the measured value that is the teacher data and the estimated value that is output from the output layer node and that is calculated using the weight w _ji and the offset value b _j to be determined. It is performed depending on whether or not the value is equal to or less than a preset convergence determination limit value. As a unit of this value, a value on the real scale is more than a normalized value (0 to 1) which is usually used. It may be easy to handle. Also,
When the neural network model is assigned to each phenomenon of various units, it may be more convenient to use the above-mentioned normalized value. Therefore, as in this example, it is possible to improve learning efficiency by making it possible to select either a normalized value or a value in an actual unit as the value of the convergence determination limit. .

In the learning of the neural network, the magnitude of the load is repeatedly corrected in the order of the load between the intermediate layer and the output layer and the load between the input layer and the intermediate layer.
Since it is experientially effective to correct the amount of correction calculated by 100% rather than 100%, in this example, as described above, in the middle layer, The modification ratio η and the output layer weight modification ratio η ′ are set as learning parameters, whereby the learning convergence is accelerated. In addition, since the above-mentioned load correction is repeated thousands of times and tens of thousands of times, when correcting,
As described above, by setting the moment coefficient, which is the modification ratio to the immediately preceding load modification amount, as a learning parameter and adding some percentage of the immediately previous modification amount, learning can be accelerated and, Divergence can be suppressed.

By the way, in the course of learning, the output y of each node is given by the following equation (4), where x is the input.
It is calculated by the sigmoid function shown in.

[Equation 3] Here, if the input x is too large or too small to some extent, the value of y changes to 0 or 1 even if the value of x changes to some extent, resulting in a saturated state. Therefore, as described above, it is possible to speed up the convergence of learning by making it possible to set the fluctuation amount of the saturation node when the learning is saturated, the saturation check interval, and the moving average number of times of the saturation determination as learning parameters. .

Further, the learning efficiency may be improved by randomly changing the pattern learning order rather than by executing the pattern learning order fixedly. Therefore, in the present invention, the pattern learning order column g33 is provided for learning. , It is possible to select whether to change the learning order of the pattern at random or to execute it while keeping it fixed, and when changing the learning order of the pattern at random, determine the change timing by the number of times, It is possible to select whether to execute even when the node output is saturated. If there is a constraint that the learning order should be fixed, the learning can be performed by clicking "Fix in table order" in the pattern learning order column g33.

In learning the neural network, the factor group data and the characteristic group data in the learning data table 10T described above are read, the learning data is read using these values, and the neural network of the neural network is read using a plurality of input learning data. The respective errors of the intermediate layer and the output layer are obtained by changing the respective _weights w _ji and the offset values b _j , and learning is performed so that the errors are minimized. Then, after updating the respective loads w _{j i} and the offset values b _j obtained above, each of the test data is tested by a neural network using the updated respective loads w _ji and the offset values b _j The characteristic data (estimated value) as the value of is obtained. In addition, the state of convergence of learning is
It is sequentially displayed in the convergence state display field g34 of the learning screen G3. Next, it is judged whether or not the test result has converged by judging whether or not the value of the test result is within a predetermined range which is a criterion for the convergence judgment, or whether or not the above processing is repeated a predetermined number of times, and an affirmative judgment is made. In the case of, this routine is ended. On the other hand, in the case of a negative determination, the above processing is repeated to calculate each load w _ji and the offset value b _j so that the error of each unit of the intermediate layer and the output layer is minimized. After the learning is completed, the parameters of the neural network used for the learning can be stored by clicking the system parameter storage field g37 on the learning screen G3. It should be noted that in this example, as described later, since the continuous learning is possible, it is desirable to store the above parameters.

Further, in this example, by clicking the learning degree verification column g4 after the learning is completed, the learning degree verification screen G4 as shown in FIG. The graph of the relationship between the actual measurement value and the estimated value is displayed so that the learning degree can be verified (step S18). At this time, the correlation coefficient between the measured value and the estimated value is obtained and displayed simultaneously. As a result, the user can surely grasp the learning degree of the neural network. Note that, in FIG. 9, as an example, the correlation between the actual measurement value and the estimated value regarding the characteristic 1 is displayed in the graph column g41. However, by changing the numerical value in the characteristic item number column g42, another characteristic for which learning is completed is performed. It goes without saying that items can also be displayed.

Further, in this example, since the continuous learning is possible, when the continuous learning is performed, the learning column g3 is selected and the learning screen G3 shown in FIG. 8 is returned to the continuous learning condition column g3.
Select 2 continuous learning. At that time, a predetermined characteristic item is designated in the characteristic item number column g35 of the learning screen G3, and the learning parameter is changed, the same value as the previous time is used, or the default value is used. You can As a result, when using the same value as last time,
It is not necessary to perform learning from the beginning for the same learning parameter, and even when the learning parameter is changed, the learning result can be changed by referring to the previous result, so that the learning efficiency can be improved. After the learning parameters are reset, the learning restart column g38 can be selected to continue the learning. Further, the continuous learning is not necessarily performed after the learning degree is verified, and may be performed when the learning is interrupted midway due to a slow convergence or the like. At that time, as described above, it is possible to continue the learning with the same learning parameter as the previous one, or to change the learning parameter, so that the learning can be efficiently performed.

After the above processing is completed and the learning of the neural network is sufficiently performed, the network structure, that is, the weight and the offset value are stored in the memory 17. After that, click the characteristic estimation field g5 to
A characteristic estimation screen G5 as shown in 0 is displayed, and factors and characteristics are estimated (step S20). In the present invention, the characteristic estimation column g51 or the factor estimation column g of the characteristic estimation screen G5 described above.
By selecting 52, both of the conventional estimation calculation of characteristic data for the factor data group and the estimation calculation of factor data for the characteristic data group can be performed. Then, using the above learning result, the factor data for producing arbitrary characteristic data is estimated and its optimum value is calculated (step S2).
2) Find the optimum combination of factor data in design and formulation. As the optimization method, a well-known optimization method such as a one-variable Newton method, a multivariable Newton method, or a combination method can be used.

In the case of the Newton method, after setting factor items as variables, an estimation calculation at the current location is performed to obtain an error amount with respect to a target characteristic value (hereinafter referred to as a target value), and further, , The variable is, for example, 1/100 of the change range
An estimation calculation is performed when the value is changed by about 0 to obtain an error amount with respect to the target value, a differential coefficient is calculated numerically, and the variable is changed so that the error amount with respect to the target value becomes zero. By repeating such an operation until the error amount becomes equal to or less than the predetermined value, it becomes possible to numerically calculate the optimum values of the plurality of factor variables for realizing the target value. In addition, in the combination method, a combination is determined for each of a plurality of factors by division, and a characteristic estimation calculation is performed for each combination to obtain an optimal factor combination for achieving a predetermined target value. be able to. These can be effectively realized by using them in combination with estimating approximate values by a neural network, because of constraints on calculation time and convergence.

The optimization according to the present invention will be described below. First, click the factor optimization column g6, and
The factor optimization screen G6 as shown in (1) is displayed, and the optimization 1 button g61 or the optimization 2 button g62 of this factor optimization screen G6 is selected. As will be described later, “optimization 2” is applied to optimization in the case where there are a plurality of stable points with respect to changes in factors and it is easy to converge to a position other than the optimum value. Optimization is performed in "Optimization 1". When performing optimization, the initial value, the maximum value (MAX), and the minimum value (MI) of factor data are stored in the respective columns of the optimization factor column g63.
N) is set, the target value of the characteristic data is set in the characteristic column g64, and the optimum factor data for generating the characteristic data set to the target value between the maximum value and the minimum value of the set factor data is set. Find the value. The initial value and the target value can be increased / decreased within the range between the set maximum value and minimum value by using the spin buttons indicated by the upper and lower arrows in FIG.

In this example, by clicking the estimated value button B61, the characteristic value is estimated and calculated by the neural network using the factor value that is currently set, and the characteristic value can be displayed in the estimated value column g65. I have to. Also,
By clicking the optimum value button B62, the factor value can be estimated and calculated by the neural network using the currently set characteristic value and displayed in the optimum value column g66. Further, in this example, the estimated value copy button B63 is clicked to copy the obtained estimated value to the target value in the characteristic column g64 to set it as a new target value, or to copy the optimum value button B64. It is also possible to click and copy the obtained optimum value to the initial value of the optimization factor column g63 to set a new initial value. Further, the difference amount between the estimated value and the characteristic value obtained by the above estimation is calculated and displayed in the error amount column g67, or the difference amount between the optimum value and the factor value obtained by the above estimation is calculated. , Can be displayed in the error amount column g68. Therefore, the value copied to the target value in the characteristic column g64 is set as a new target value while referring to the difference amount displayed in the error amount columns g67 and g68, and the initial value in the optimization factor column g63 is set. Set the optimum value copied to as a new initial value,
By performing the optimization calculation again, the optimum factor condition for producing the desired characteristic can be efficiently obtained. After the optimization, the file column g7 is selected to call the file operation screen G7 as shown in FIG. 12, and the learning data table used for the neural network and the above result are saved in the Excel book.

Next, an optimization method by "optimization 2" will be described. When the optimization 2 button g62 is selected, a division optimization screen G8 as shown in FIG. 13 is displayed. The division optimization screen G8 is
The optimum value of the factor data that creates it is determined more strictly, and it has a division number designation column g81 for designating the number of divisions of the region to be optimized, and also has a characteristic column g82, an estimated value column g83, and an error amount column g84. In, the target value, the estimated value, and the error amount of one specified characteristic are displayed when one characteristic item is selectively set to a high priority. In optimization 2, when the number of divisions is designated, the characteristic value is estimated and calculated by the neural network for each area of division and displayed in the estimated value column g83 for each area. Specifically, FIG.
As shown in, the characteristic value at the stable point of the error amount of the characteristic value in each divided area is set as an estimated value, which is displayed in the estimated value display field g85, and when there is no stable point in the area,
"Not Found" is displayed. This makes it possible to find the true convergence point even when there are a plurality of stable points that easily converge with respect to the change in the factor, and thus it is possible to reliably obtain the optimum value of the factor data for the target characteristic.

In the present invention, the following functions are further added in order to more efficiently find the optimum factor data for producing arbitrary characteristic data. (1) Function of assigning a seriousness (priority) to each of the characteristic groups Specifically, the priority is set by numbers for the n characteristic groups. For example, the highest priority is 1, and the priority is 10. A characteristic group having a fluctuation allowable amount of 0 (which does not allow fluctuation), which will be described later, is a characteristic group having a priority of 0. (2) Function of adding allowable fluctuation amount to each of the characteristic groups The allowable fluctuation amount is given in a graph for the n characteristic groups. Specifically, when an ID number is specified, a graph can be specified with the maximum value and the minimum value of the characteristic group on the vertical axis and the generated random number on the horizontal axis. For example, in the case of generating random numbers in the range of 0 to 1 with a substantially uniform probability, if an allowable fluctuation graph as shown in FIG. 15A is given, the value of the characteristic value corresponding to the random value p is P. Therefore, as shown in the correspondence table of FIG. 15B, by setting the priority of this characteristic item to be high according to the above (1), the probability that the width of the generated random value p becomes the peak value becomes high. Corresponds to. Priority 0
Means that the random number value is p = 0.5, that is, the characteristic value is fixed to the peak value. (3) Function to quantify the optimality It is expressed by numerical values for the conditions for convergence determination when obtaining the optimal solution by the relaxation method that performs repeated calculation such as Newton's method.
Specifically, the optimum frequency is set to 100% when the convergence condition is the strictest, and the optimum frequency is associated with a smaller value as the convergence condition is loosened. (4) Function of assigning severity (priority) to each of the factor characteristic groups Specifically, the priority is set by a numeral for m factor groups. For example, the highest priority is 1, and the priority is 10. The optimum solution is obtained by changing the factors having a high priority preferentially. The factor of priority 10 is considered to be a fixed value and is not optimized.

A method of obtaining optimum factor data using the above functions will be described. First, through the learning described above, a plurality of input factor groups X as shown in the following equation (5) are obtained.
The functional relationship that maps the pattern of P to the pattern of the plurality of output factor groups Y is obtained. Y _j = f _ij (X _j ) ... (5) where i = 1 to m, j = 1 to n This functional relationship can be explicitly given,
By applying a neural network as in the present invention, the following method can be applied even when the form of the equation is unknown. Next, Y _i is selected according to the priority designated by the degree of lightness of the characteristic group, and a predetermined variation amount designated by the straight line of the graph showing the relationship between the priority and the allowable variation amount in FIG. 15A is given. . At this time, the characteristic item designated as the priority 0 is fixed to the given value and the fluctuation is not allowed. With respect to the characteristic group, the higher the priority is, the more the generated random number range is limited to the vicinity of the peak value. Therefore, the higher the priority of the data, the higher the probability of the peak value. Then, for example, iterative calculation applying the multivariable Newton's method determines X _j that satisfies the above equation (5). At this time, when the convergence limit of the solution of the Newton method is set to an extremely strict value, the optimum degree of the obtained solution is close to 100%, and the optimum degree of the solution is reduced as the convergence limit is loosened. go. In this way, the degree of optimization can be set according to the required accuracy, so that the optimization can be performed efficiently. In this way, in the Newton method, when calculating the modification amount of a variable, by changing the coefficient of the variable according to the degree of lightness and weight, it becomes possible to preferentially bring the variable with high priority closer to the optimal solution. Therefore, the optimization can be efficiently performed, and the convergence of the optimization can be improved.

[0030]

As described above, according to the present invention,
Create a learning data table that summarizes each experimental data in formulation and design in a table format that attribute is classified into factors and characteristics, select each column of the above learning data table, assign it to the input layer or output layer, and perform the above experiment By learning the mapping relationship between the factor group and the characteristic group of the data and using the above learning results, the optimum value of the factor data that produces it is obtained for any characteristic data. A suitable factor condition for obtaining can be easily obtained, and the optimization process can be efficiently performed in various design works. Furthermore,
Since the optimization of the above design and formulation is realized in the GUI environment, by selecting the function tab on the computer screen, the learning of the neural network and the optimization of the design formulation can be performed easily and efficiently. Can be done on a regular basis.

[Brief description of drawings]

FIG. 1 is a diagram showing an outline of a personal computer for carrying out the optimization of design and formulation of the present invention.

FIG. 2 is a block diagram showing a configuration of a computer main body.

FIG. 3 is a flowchart showing an optimization method of design / blending according to the present embodiment.

FIG. 4 is a diagram showing an initial screen for implementing design / compound optimization according to the present embodiment.

FIG. 5 is a diagram showing a learning data table creation screen according to the present embodiment.

FIG. 6 is a diagram showing an example of a learning data table according to the present embodiment.

FIG. 7 is a diagram showing an input / output setting screen according to the present embodiment.

FIG. 8 is a diagram showing a learning screen according to the present embodiment.

FIG. 9 is a diagram showing a learning degree verification screen according to the present embodiment.

FIG. 10 is a diagram showing a characteristic estimation screen according to the present embodiment.

FIG. 11 is a diagram showing a factor optimization screen according to the present embodiment.

FIG. 12 is a diagram showing a file operation screen according to the present embodiment.

FIG. 13 is a diagram showing a division optimization screen according to the present embodiment.

FIG. 14 is a diagram showing a relationship between factor data and an error amount of a characteristic value.

FIG. 15 is a diagram showing an allowable fluctuation graph according to the present embodiment.

FIG. 16 is a diagram showing a configuration example of a neural network.

FIG. 17 is a flowchart showing an outline of a learning method of a neural network.

[Explanation of symbols]

10 personal computers, 11 displays,
12 computer main body, 12F FDU, 13a keyboard, 13b mouse, 14 CPU, 15 RO
M, 16 RAM, 17 memory, 18 input / output device, 19 bus.

Claims (11)

[Claims] 1. The above experimental data is used when a mapping relationship between a factor group and a characteristic group of experimental data in product design, mixture design, etc. is learned by using a neural network method to optimize the design / mixture. Create a learning data table in which is divided into a column of factor items and a column of characteristic items, select each column of this learning data table, assign it to the input layer or output layer, and learn the above mapping relationship, then A method for optimizing design and composition, characterized in that the optimum value of factor data for creating arbitrary characteristic data is obtained by using the optimization method. 2. When learning the mapping relationship, the row of factor items and the row of characteristic items of the learning data table are set to either input layer items, output layer items, or items to be ignored. The method for optimizing the design / compounding according to claim 1, further comprising: 3. A Newton method or a combination method is used as the optimization method.
Alternatively, the method of optimizing the design and composition according to claim 2. 4. When obtaining the optimum factor data, an initial value, a maximum value, and a minimum value are set for the factor value, and the factor value is increased / decreased from the initial value to the maximum value and the minimum value. The method for optimizing the design / compounding according to any one of claims 1 to 3, wherein 5. The previously obtained estimated value is set as a new target value of the characteristic, the optimization process is performed again, and the optimum factor data for obtaining the desired characteristic is obtained. The method of optimizing the design / compounding according to any one of claims 1 to 4. 6. The optimum factor data is obtained by setting the factor value that produced the previously obtained estimated value as an initial value of a new factor value and performing optimization processing again. The method for optimizing the design / compounding according to any one of claims 1 to 5. 7. The optimization of the design / compounding according to any one of claims 1 to 6, wherein each of the characteristic groups is given a degree of lightness when obtaining the optimum factor data. Method. 8. The optimum design / compounding method according to claim 1, wherein an allowable variation amount is added to each of the characteristic groups when the optimum factor data is obtained. Method. 9. When setting a plurality of optimum frequencies defined by a convergence judgment condition and obtaining optimum factor data,
9. The method for optimizing a design / compounding according to claim 1, wherein any one of the optimum frequencies is selected for optimization. 10. The optimization of the design / compounding according to any one of claims 1 to 9, characterized in that each of the factor groups is assigned a degree of lightness when obtaining the optimum factor data. Method. 11. The method according to claim 1, wherein a function tab is selected on a computer screen to perform learning and optimization of the neural network.
The method for optimizing the design and composition according to any one of 1.