JP2014164336A - Model analysis device, model analysis method and model analysis program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an operation confirmation system for confirming a change in an estimate value associated with a change in a conditional probability table of a Bayesian network model.SOLUTION: A model analysis device comprises: a simulation data generation unit 102 that generates multiple pieces of simulation data 126 to be input into a defect number estimation model; an estimate value calculation unit 103 that changes a conditional probability table by inputting the simulation data 126 into the defect number estimation model, calculates an estimate value 127 by inputting a model input value 123 into the defect number estimation model having the changed conditional probability table, accumulates information having the calculated estimate value 127 associated with the simulation data 126 input into the defect number estimation model, and outputs the information as estimate value information 130; and a graph generation unit 104 that generates a graph 128 on the basis of the estimate value information 130.

Description

  The present invention relates to a model analysis apparatus, a model analysis method, and a model analysis program. In particular, the present invention relates to analysis of Bayesian network models.

  The Bayesian network model is composed of a network diagram and a conditional probability table. The network diagram is a diagram in which the target phenomenon is a node, the nodes are connected by arrows, and the causal relationship is represented by a directed acyclic graph. A conditional probability table (Conditional / Probability / Table) exists for each node in the network diagram. The combination of this figure and the table is called a Bayesian network model.

  FIG. 1 is an example of a constructed Bayesian network model. X1 to X5 are events (nodes), and arrows indicate their causal relationship. Each node has a conditional probability table. P (X4 / X2) is a numerical example, and each node follows a binomial distribution with values of 0 and 1. When X4 is 0, the probability of X2 is 0.8 (the probability of 1 is 0.2). Hereinafter, the conditional probability table of each node is collectively referred to as a “conditional probability table”.

The operation using the Bayesian network model consists of two steps. First, a conditional probability table optimized by “Step 1 (learning step)” is generated, and then the estimated value is calculated by “Step 2 (estimated value calculating step)”. calculate.
(1) Step 1 (learning step):
An optimized conditional probability table is generated using actual data (past actual measurement data). (2) Step 2 (estimated value calculation step):
By inputting values to some of the nodes in the network diagram, estimated values for other nodes are calculated. However, the calculation result of the estimated value varies depending on the input value and the value in the conditional probability table (the network diagram does not change).

The above two steps will be specifically described using an example.
(Example) Egg model Consider a model (egg model) that estimates the number of eggs a chicken lays the next day from the weather and temperature of the previous day. It is assumed that there are three nodes in the egg model: “previous day weather, previous day temperature, number of eggs laid by next day chicken”.
(1) Step 1 (learning step):
For example, the weather of the previous day, the temperature of the previous day, and the number of eggs of the next day are measured for 10 days. The measurement data for 10 days is given to the model as performance data, and an optimal conditional probability table is generated for estimating the number of eggs laid by the chicken on the 11th day.
(2) Step 2 (estimated value calculation step):
On the 11th day, an estimated value of “the number of eggs laid by the chicken the next day” is obtained by inputting the value of two nodes “the weather of the previous day, the temperature of the previous day”. That is, if the weather and temperature values measured the previous day are input to the egg model, an estimated value of the number of eggs laid by the chicken the next day can be obtained in advance.
However, due to the difference in the measured data for 10 days used in Step 1, even if the weather and temperature values input on the 11th day are the same, the estimated value of the number of eggs laid by the chicken changes.

Here, the operation confirming means of the Bayesian network model refers to confirming means for how the estimated value calculated by numerical input to the model varies.
The conventional Bayesian network model operation confirmation means has made it possible to confirm the change of the estimated value by comparing the estimated values calculated when the input value is increased or decreased (for example, Patent Document 1).
Japanese Patent Application Laid-Open No. 2004-228561 provides means for changing the input value in the estimated value calculating step and confirming the change of the calculated estimated value.

JP 2011-81451 A

As described above, if the conditional probability table generated in Step 1 is different in the Bayesian network model, the estimated value changes even if the input value is constant in Step 2.
However, the conventional Bayesian network model operation confirmation means (hereinafter also referred to as analysis means) can only confirm and analyze the estimated value change that occurs when the input value at Step 2 is changed. Since it is a means, the subject that the change of the conditional probability table produced | generated by Step1 cannot be considered occurred.

  The present invention has been made to solve the above-described problems, for example, and is an operation confirmation method / model analysis method for confirming a change in an estimated value due to a change in a conditional probability table with respect to a Bayesian network model. The purpose is to provide.

The model analysis apparatus according to the present invention is a Bayesian network model that targets a plurality of nodes including an input target node that is an input target to which an input value is input and an estimation target node that is an estimation target. In a model analysis apparatus for analyzing a Bayesian network model including a conditional probability table that represents a dependency relationship of nodes by a conditional probability,
An input value storage unit for storing the input value to be input to the input target node in a storage device;
A simulation data generation unit for generating a plurality of simulation data input to the Bayesian network model;
A probability table changing unit that inputs each simulation data of the plurality of simulation data generated by the simulation data generating unit to the Bayesian network model and changes the conditional probability table;
The input value stored in the input value storage unit is input to the input target node of the Bayesian network model including the conditional probability table changed by the probability table changing unit, and the estimated value of the estimation target node And an analysis target information generating unit that accumulates estimated value information in which the calculated estimated value is associated with the simulation data input to the Bayesian network model and stores the estimated value information as analysis target information in a storage device. It is characterized by that.

  According to the model analysis apparatus of the present invention, the probability table changing unit inputs each simulation data of the plurality of simulation data to the Bayesian network model to change the conditional probability table, and includes the changed conditional probability table. Calculates an estimated value by inputting a preset input value to the input target node of the network model, accumulates estimated value information that associates the calculated estimated value with the simulation data input to the Bayesian network model, and is analyzed Since the information is generated, it is possible to analyze the change in the estimated value accompanying the change of the conditional probability table.

It is an example of the constructed Bayesian network model. 3 is a network diagram of a defect number estimation model that is an operation check target of the operation check apparatus 100 according to Embodiment 1. FIG. It is the figure which showed the specific utilization procedure at the time of SW development of the defect number estimation model used as the operation confirmation object of the operation confirmation apparatus 100 which concerns on Embodiment 1. FIG. 2 is a functional block diagram of an operation check apparatus 100 according to Embodiment 1. FIG. 3 is a diagram illustrating an example of a hardware configuration of an operation check apparatus 100 according to Embodiment 1. FIG. It is a figure which shows each structure of the data 120 which concerns on Embodiment 1, and the division | segmentation distribution allocation table 129, (a) Basic distribution 121 for simulation data, (b) Distribution division information 122 for simulation data, (c) Model input The value 123, (d) an example of the graph setting information 124, and (k) an example of the post-division distribution assignment table 129. It is a figure which shows the structure of distribution 125 (e) after the division | segmentation for simulation data which concerns on Embodiment 1. FIG. It is a figure which shows an example (f) of the simulation data 126 which concerns on Embodiment 1. FIG. 6 is a diagram illustrating an example of an estimated value 127 calculated by an estimated value calculating unit 103 according to Embodiment 1. FIG. 6 is a flowchart showing an operation confirmation method (distribution information generation process, simulation data creation process) of the operation confirmation apparatus 300 according to the first embodiment. 5 is a flowchart showing an operation confirmation method (estimated value calculation process, graph creation process) of the operation confirmation apparatus 300 according to the first embodiment. (A) It is a figure for demonstrating a definition of the basic distribution 121 for simulation data. FIG. 6 is an image diagram showing a mechanism for dividing the simulation data basic distribution 121 according to the first embodiment. FIG. 6 is an image diagram showing a mechanism for dividing the simulation data basic distribution 121 according to the first embodiment. FIG. 6 is an image diagram showing a mechanism for dividing the simulation data basic distribution 121 according to the first embodiment. 4 is an image diagram showing a mechanism for creating a simulation data set and simulation data 126 according to Embodiment 1. FIG. 4 is an image diagram showing a mechanism for creating a simulation data set and simulation data 126 according to Embodiment 1. FIG. It is a figure which shows an example of the (h) 1st graph 128h which concerns on Embodiment 1. FIG. It is a figure which shows the example of DIR of (i) 2nd graph 128i which concerns on Embodiment 1. FIG. It is a figure which shows the example of DRE of (i) 2nd graph 128i which concerns on Embodiment 1. FIG. 6 is a table showing a change rate of each estimated value of another simulation data set with respect to a simulation data set serving as a reference according to the first embodiment. It is a figure which shows the example of DIR of (j) 3rd graph 128j which concerns on Embodiment 1. FIG. It is a figure which shows the example of DRE of the (j) 3rd graph 128j which concerns on Embodiment 1. FIG.

Embodiment 1 FIG.
First, a defect number estimation model using a Bayesian network model that is an operation check target (analysis target) of the operation check device 100 (model analysis device) according to the present embodiment will be described.

The outline of the defect number estimation model is as follows (a) to (c).
(A) Purpose:
During SW (software) development, an estimate of numerical values (defect mixture density, defect removal rate) related to the final number of defects at the end of SW development is obtained.

(B) How to use:
INPUT (estimate) the values (defect mixture density, defect removal rate) that cannot be known by inputting (input) the values that can be measured during SW development (review man-hour density, review indication density) into the model. .

(C) Details:
FIG. 2 is a network diagram of a defect number estimation model that is an operation check target of the operation check apparatus 100 according to the present embodiment. This model consists of six nodes (elements).
During development, since DIR (defect contamination density) and DRE (defect removal rate), which are the estimation target nodes of this model, cannot be known, ED (review man-hour density) and DR (review indication density) that can be measured. ) To estimate. ED (review man-hour density) and DR (review indication density) are examples of input target nodes to which input values are input.

  When the development is completed, ED, DR and SQ (quality of specifications) and RQ (quality of review process) are FIXed to create a full set of data. When this data is used as performance data and the model conditional probability table is regenerated, the defect number estimation model is optimized for the next development.

Further, (c) the detailed description will be continued.
FIG. 3 is a diagram illustrating a specific use procedure during SW development of the defect number estimation model to be an operation check target of the operation check apparatus 100 according to the present embodiment.
At the start of SW development, a conditional probability table optimized by giving actual data is generated (Step 1). Next, an INPUT value observed during development is input to the model to obtain an estimated value (Step 2). When the development is completed, the values of all 6 nodes including the value calculated as the estimated value are finalized. By adding these as actual data and executing Step 1 again, the values have been optimized for the next development. A conditional probability table is generated, and an estimated value (estimated value and standard deviation) is calculated in Step 2. Thereafter, the above process is repeated according to development.

  The above-mentioned “defect number estimation model” is used to input “defect mixing density” and “defect removal rate” that cannot be measured by inputting “review man-hour density” and “review indication density” that can be measured during software development. This model is intended to calculate the estimated value.

Details of each node in the “defect number estimation model” are shown below. Hereinafter, the following abbreviations in parentheses are used for each node name.
(A) Quality of specifications: Specifications / Quality (SQ);
Distribution name: binomial distribution;
・ Role on the model: Organizational specification creation level;
・ Characteristics: Evaluated from past performance at the time of software development. Expresses the quality of the specification before the review. Numeric value is optional (user defined).

(B) Quality of the review process: Review Quality (RQ);
Distribution name: binomial distribution;
・ Model role: Organizational review level;
・ Characteristics: Evaluated from past performance at the time of software development. Represents the quality of the organization's review process. Numeric value is optional (user defined).

(C) Review man-hour density: Effort Density (ED);
-Distribution name: normal distribution;
-Role on the model: measured value (input value);
・ Property: Measurable at the time of software development. Review time (hour, day, month) divided by development scale. The man-hour may be an arbitrary unit. Development scale refers to the amount of document creation (number of pages). Takes a value greater than zero.

(D) Review indication density: Defect / Review (DR);
-Distribution name: normal distribution;
-Role on the model: measured value (input value);
・ Characteristic: Measurable during software development, the number of reviews pointed out divided by the development scale. Development scale refers to the amount of document creation (number of pages). Takes a value greater than zero.

(E) Defect contamination density: Defect / Injection / Rate (DIR);
-Distribution name: normal distribution;
・ Role on model: Estimated value;
• Property: Value that cannot be measured during software development. The number of defects mixed in is divided by the development scale, and the number of defects mixed is the final number of defects found after the end of software development and cannot be measured. Development scale refers to the amount of document creation (number of pages). Takes a value greater than zero.

(F) Defect removal rate: Defect / Removal / Efficiency (DRE); Distribution name: Normal distribution;
・ Role on model: Estimated value;
• Property: Value that cannot be measured during software development. The defect removal number is divided by the defect mixture number, and the defect removal number represents the number of defect corrections. The number of defects mixed is the final number of defects found after the end of software development and cannot be measured. Takes a value greater than zero.

When using the above defect number estimation model at the software development site, create a model diagram on the tools of a Bayesian network construction support system (eg, BeyoNet, HuginExpert, etc.) A probability table can be automatically generated.
This is the end of the description of the defect number estimation model.

  Next, an outline of the operation confirmation method of the operation confirmation apparatus 100 according to the present embodiment will be described.

FIG. 4 is a functional block diagram of the operation check apparatus 100 according to the present embodiment.
As illustrated in FIG. 4, the operation check apparatus 100 includes a distribution information generation unit 101, a simulation data creation unit 102, an estimated value calculation unit 103, and a graph creation unit 104.

  In the operation confirmation method of the operation confirmation apparatus 100, for example, the above-described defect number estimation model is constructed on a tool such as a Bayesian network construction support system, and is used as an operation confirmation target model. Then, without directly manipulating the conditional probability table of this defect number estimation model, it automatically changes arbitrarily, and for each changed conditional probability table, the estimated value obtained when a fixed value is input is compared. .

As shown in FIG. 4, in the operation check method of the operation check apparatus 100, data 120 created in advance by the user is used.
Data 120 includes simulation data basic distribution 121, simulation data distribution division information 122, model input value 123, and graph setting information 124.

  In FIG. 4, the distribution information generation unit 101 reads the simulation data basic distribution 121 and the simulation data distribution division information 122, and generates a simulation data post-division distribution 125 that is data necessary to create the simulation data 126. Has a function to calculate.

  The simulation data creation unit 102 has a function of generating a plurality of simulation data 126 in cooperation with statistical processing software (R or the like) using the simulation data post-partition distribution 125 and the post-partition distribution allocation table 129. .

  The distribution information generation unit 101 and the simulation data generation unit 102 are an example of a simulation data generation unit that generates a plurality of simulation data input to the Bayesian network model.

  The operation check device 100 changes the conditional probability table by inputting a plurality of simulation data, and analyzes the change in the estimated value accompanying the change. Hereinafter, the simulation data may be abbreviated as SD.

  The estimated value calculation unit 103 uses the simulation data 126 and the model input value 123 created by the simulation data creation unit 102, cooperates with the Bayesian network dedicated software, and the number of defects assembled in advance on the software. It has a function of calculating the estimated value 127 by operating the estimation model.

The estimated value calculation unit 103 is an example of a probability table changing unit that inputs each simulation data of the plurality of simulation data 126 to a Bayesian network model and changes the conditional probability table.
In addition, an input value is input to an input target node of the Bayesian network model including the modified conditional probability table to calculate an estimated value, and the calculated estimated value 127 corresponds to the simulation data 126 input to the Bayesian network model. It is an example of the analysis object information generation part which produces | generates the attached estimated value information 130 as analysis object information.

The graph creation unit 104 has a function of creating a graph 128 from the estimated value 127 using statistical processing software, Excel, or the like.
The graph 128 includes a plurality of types such as a first graph 128h, a second graph 128i, and a third graph 128j.

Based on the estimated value information 130 (analysis target information), the graph creation unit 104 analyzes a change in the estimated value 107 accompanying a change in the conditional probability table by inputting a plurality of simulation data into the Bayesian network model. It is an example of a part.
In addition, the graph creation unit 104 displays changes in the value of the estimation target node (estimation value 107) accompanying the change of the conditional probability table based on the estimation value information 130 (analysis target information) in the first graph 128h and the second graph. It is graphed (visualized) as 128i, the third graph 128j, and the like.

  FIG. 5 is a diagram illustrating an example of a hardware configuration of the operation check apparatus 100 according to the present embodiment.

  In FIG. 5, the operation check apparatus 100 is a computer, and includes an LCD 901 (Liquid / Crystal / Display), a keyboard 902 (K / B), a mouse 903, an FDD 904 (Flexible / Disk / Drive), and a CDD 905 (Compact / Disk / Drive). ) And a hardware device such as a printer 906. These hardware devices are connected by cables and signal lines. Instead of the LCD 901, a CRT (Cathode / Ray / Tube) or other display device may be used. Instead of the mouse 903, a touch panel, a touch pad, a trackball, a pen tablet, or other pointing devices may be used.

  The operation check apparatus 100 includes a CPU 911 (Central Processing Unit) that executes a program. The CPU 911 is an example of a processing device 191. The CPU 911 includes a ROM 913 (Read / Only / Memory), a RAM 914 (Random / Access / Memory), a communication board 915, an LCD 901, a keyboard 902, a mouse 903, an FDD 904, a CDD 905, a printer 906, and an HDD 920 (Hard / Disk) via a bus 912. Connected with Drive) to control these hardware devices. Instead of the HDD 920, a flash memory, an optical disk device, a memory card reader / writer, or other recording medium may be used.

  The RAM 914 is an example of a volatile memory. The ROM 913, the FDD 904, the CDD 905, and the HDD 920 are examples of nonvolatile memories. These are examples of the storage device 192. The communication board 915, the keyboard 902, the mouse 903, the FDD 904, and the CDD 905 are examples of the input device 193. The communication board 915, the LCD 901, and the printer 906 are examples of the output device 194.

  The communication board 915 is connected to a LAN (Local / Area / Network) or the like. The communication board 915 is not limited to a LAN, but is an IP-VPN (Internet, Protocol, Private, Network), a wide area LAN, an ATM (Asynchronous / Transfer / Mode) network, a WAN (Wide / Area / Network), or the Internet. It does not matter if it is connected to. LAN, WAN, and the Internet are examples of networks.

  The HDD 920 stores an operating system 921 (OS), a window system 922, a program group 923, and a file group 924. The programs in the program group 923 are executed by the CPU 911, the operating system 921, and the window system 922. The program group 923 includes programs that execute the functions described as “˜units” in the description of the present embodiment. The program is read and executed by the CPU 911. The file group 924 includes data, information, and signal values described as “˜data”, “˜information”, “˜ID (identifier)”, “˜flag”, and “˜result” in the description of this embodiment. And variable values and parameters are included as items of “˜file”, “˜database”, and “˜table”. The “˜file”, “˜database”, and “˜table” are stored in a recording medium such as the RAM 914 or the HDD 920. Data, information, signal values, variable values, and parameters stored in a recording medium such as the RAM 914 and the HDD 920 are read out to the main memory and the cache memory by the CPU 911 via a read / write circuit, and extracted, searched, referenced, compared, and calculated. It is used for processing (operation) of the CPU 911 such as calculation, control, output, printing, and display. During the processing of the CPU 911 such as extraction, search, reference, comparison, calculation, calculation, control, output, printing, and display, data, information, signal values, variable values, and parameters are temporarily stored in the main memory, cache memory, and buffer memory. Remembered.

  The arrows in the block diagrams and flowcharts used in the description of this embodiment mainly indicate input / output of data and signals. Data and signals are recorded in memory such as RAM 914, FDD904 flexible disk (FD), CDD905 compact disk (CD), HDD920 magnetic disk, optical disk, DVD (Digital Versatile Disc), or other recording media Is done. Data and signals are transmitted by a bus 912, a signal line, a cable, or other transmission media.

  In the description of the present embodiment, what is described as “to part” may be “to circuit”, “to device”, “to device”, and “to step”, “to process”, “to”. ~ Procedure "," ~ process ". That is, what is described as “˜unit” may be realized by firmware stored in the ROM 913. Alternatively, what is described as “˜unit” may be realized only by software, or only by hardware such as an element, a device, a board, and wiring. Alternatively, what is described as “to part” may be realized by a combination of software and hardware, or a combination of software, hardware and firmware. Firmware and software are stored as programs in a recording medium such as a flexible disk, a compact disk, a magnetic disk, an optical disk, and a DVD. The program is read by the CPU 911 and executed by the CPU 911. That is, the program causes the computer to function as “to part” described in the description of the present embodiment. Or a program makes a computer perform the procedure and method of "-part" described by description of this Embodiment.

  Next, the configuration of each data in the operation confirmation method of the operation confirmation apparatus 100 will be described. The format of each data is not limited, but in the following example, each column is handled as a set of data in a table format in which column names are determined for each column.

  FIG. 6 is a diagram showing each configuration of the data 120 and the post-partition distribution allocation table 129 according to the present embodiment. (A) Simulation data basic distribution 121, (b) Simulation data distribution division information 122, ( c) model input value 123; (d) an example of the graph setting information 124; and (k) an example of a post-partition distribution assignment table 129. FIG. 7 is a diagram showing a configuration of (e) the distribution after distribution 125 for simulation data according to the present embodiment.

The simulation data basic distribution 121 describes general realistic numerical values for the “defect number estimation model” obtained from literatures and survey results. It becomes a general distribution value corresponding to each node.
The simulation data basic distribution 121 shown in FIG. 6A is an example of distribution information.

The ED, DR, DIR, and DRE items are representative metrics often used in general SW development. Here, the distribution is assumed assuming values that may be measured (or estimated) in general SW development in the world.
In addition, since SQ and RQ are arbitrarily defined by the user, there is no specific assumed value. Therefore, here, it is assumed that success is good = 1, failure is bad = 0, and the probability distribution is B (10, 0.7).

FIG. 12 is a diagram for explaining the definition of (a) the basic distribution 121 for simulation data.
The basic distribution 121 for simulation data will be described with reference to FIGS.

The breakdown of each column in the simulation data basic distribution 121 shown in FIGS. 6A and 12 is as follows.
“NodeName” is a node name of the Bayesian network model, and all node names are listed.
The node of the normal distribution (Normal) is assumed to have a random variable according to N (μ, σ ² ) when the average value μ is the standard deviation σ. A binomial node is assumed to follow B (n, P), where n is the number of trials and P is the success probability.
In a specific example, numerical values are set for SQ, RQ, ED, DR, DIR, and DRE that are all nodes of the defect count estimation model.

“Type” is a distribution name and is defined for common use in the system. As shown in FIG. 6A, binomial is a binomial distribution and normal is a normal distribution.
In “Distribution”, a distribution value corresponding to the distribution name is entered.
In the specific example, it was set as the description according to the general statistical notation.

In “Value1”, a value corresponding to the distribution name is entered. When the type of the node is binomial, the name of each event is entered. Each event follows the Distribution.
In a specific example, the probability of good is 7 out of 10 trials, and the probability of bad is 3 out of 10 trials.
If Type is normal, this field is not used, so entry is not necessary.

A value corresponding to the distribution name is entered in “Value2”. If Type is binomial, enter the value of each event. The entries are described in the same order as Value1.
In a specific example, good is 1 and bad is 0. If Type is normal, this field is not used, so entry is not necessary.

In “Value3”, a value corresponding to the distribution name is entered. When Type is binomial, it is a standard deviation with respect to the average value of the distribution. This value corresponds to an average fluctuation range that can be practically considered.
In the specific example, since the value (standard deviation σ) is 0.1, the distribution value is set to B (10, 0.5) in an organization with a low probability of good. As a result, the average value is twice that of SD.
In addition, B (10, 0.9) is assumed in an organization having a high probability of good. As a result, the average value is doubled by SD.
If Type is normal, this field is not used, so entry is not necessary.

Next, the simulation data distribution division information 122 shown in FIG. 6B will be described.
Information related to the number of simulation data sets is set in the distribution division information 122 for simulation data.
The number of simulation data sets is a number that determines how many simulation data sets are generated by using a plurality of simulation data as one simulation data set. That is, when the number of simulation data sets is 3, three simulation data sets including a plurality of simulation data are generated.
The breakdown of each column is as follows.

  In the “Name” item, four items of “Dataset”, “Normal Distribution Data Interval”, “Max Data”, and “Graph 1 Interval” are set.

In “Dataset”, the number of generated simulation data sets (an example of the number of divisions) is set. Numeric values that can be specified are defined in the system. This is to correspond to the contents of the post-division distribution allocation table 129 generated in advance in the system.
In the specific example, it is “3”, and the following three types of simulation data sets (data) are generated.
(A) SD1;
(B) SD2;
(C) SD3.

In “Normal Distribution Data Interval”, the definition of the data range of the normal distribution is defined.
In the specific example, “95” is set. This means that the data range of the normal distribution uses the confidence interval from the lower 95% to the upper 95% as the data range.

In “MaxData”, the number of data in one simulation data set is defined.
In the specific example, “40” is set, which means that each simulation data set of SD1 to SD3 is composed of 40 simulation data.

In “Graph 1 Interval”, designation of how many pieces of simulation data in the simulation data set are to be set as a data set is set. This is information for creating a data set obtained by dividing the generated simulation data set in order to create a graph 128 described later.
In the specific example, “10” is set. This means that, starting from 1, one data set is generated for every 10 pieces of data.

For example, data is generated as follows.
(A) SD1-data1, SD1-data10, SD1-data20, SD1-data30, SD1-data40;
(B) SD2-data1, SD2-data10, SD2-data20, SD2-data30, SD2-data40;
(C) SD3-data1, SD3-data10, SD3-data20, SD3-data30, SD3-data40.

  As shown in FIG. 6B, a value corresponding to the content described in the “Name” column is entered in the “Value” item.

Next, the model input value 123 shown in FIG. 6C will be described. In the model input value 123, an input value to the “defect number estimation model” node is set.
As shown in FIG. 6C, the breakdown of each column is as follows.
In the “InputNodeName” field, a node name for inputting an input value is set.
In the specific example, “ED” and “DR” are entered.
A value (input value) corresponding to the node name is set in the “InputValue” field.
In the specific example, “0.63” is entered for “ED” and “0.83” is entered for “DR”.
In the operation check method, the model input value 123 (input value) is stored in advance in the storage device by the input value storage unit, thereby making the input value constant.

Next, the graph setting information 124 shown in FIG. In the graph setting information 124, a value used for graphing the estimated value is set.
As shown in FIG. 6 (d), the breakdown of each column is as follows.
In “SimulationDataName”, a simulation data set name is written. In a specific example, SD1 to SD3 are entered.
In “BaseLine”, designation of a simulation data set as a comparison reference is set. In the specific example, “1” is set to any one of SD1 to SD3, and 0 is set otherwise. Here, since SD2 is set to 1, SD2 is set as a comparison reference.
In “XAxisOrder”, designation of the memory notation order of the X axis (simulation data set name) of the graph is set.
In the specific example, it is expressed in the order of SD1, SD2, SD3 on the X axis of the graph.

  6 (a) to 6 (d) have been described above, these pieces of information are prepared by the user arbitrarily setting contents, described in advance, and stored in the storage device. The description of FIG. 6K will be described later.

  Next, the simulation data distribution 125 shown in FIG. 7 will be described. The post-division distribution for simulation data 125 is a data distribution for generating a plurality of simulation data. The numerical value is a value calculated by processing information of the basic distribution 121 for simulation data.

  As shown in FIG. 7, the breakdown of the distribution 125 for the simulation data is “NodeName”, “Type”, “Distribution”, “Value1”, and “Value2”. These contents are the same as “NodeName”, “Type”, “Distribution”, “Value1”, and “Value2” of the basic distribution 121 for simulation data. However, the column “Value3” does not exist.

The contents of the basic distribution 121 for simulation data are divided (here, divided into three), and the obtained data distribution is a distribution 125 after simulation data division.
Specifically, the distribution information generation unit 101 generates SQ1 to SQ3 from SQ of the simulation data basic distribution 121, generates RQ1 to RQ3 from RQ, and generates ED1 to ED3 from ED. The same applies to DR, DIR, and DRE.
The post-division distribution 125 for simulation data is automatically generated in the system (for example, in the operation check apparatus 300).

  Next, returning to FIG. 6, the post-division distribution allocation table 129 shown in FIG. 6 (k) will be described. The post-division distribution assignment table 129 shown in FIG. 6 (k) is information prepared in advance on the system (for example, the operation check apparatus 300) and stored in the storage device by the post-division distribution assignment table storage unit.

The post-division distribution allocation table 129 is a table that defines what data sets are to be created for the combination of the distributions for post-division 125 for simulation data, which is information for each node shown in FIG.
The post-division distribution allocation table 129 is defined in advance in the system so as to correspond to the number that can be designated by “Dataset” of the distribution division information for simulation data 122 defined in the system. This setting value can be arbitrarily created according to the purpose from the purpose of the model and the nature of the node.

In the specific example, a table for data division into three corresponding to the number “3” designated by “Dataset” of the distribution division information 122 for simulation data is shown. For example, in SD1, a combination of six nodes SQ, RQ, ED, DR, DIR, and DRE is “1, 1, 1, 1, 3, 1”. Therefore, SD1 is generated by extracting the combination of “SQ1, RQ1, ED1, DR1, DIR3, DRE1” from the distribution 125 for simulation data in FIG. 7E.
Here, SD1 is set based on knowledge so that the data of the organization of the software development organization is low, SD2 is the organization of the medium level, and SD3 is the organization of the high level.

FIG. 8 is a diagram showing an example of a simulation data set including (f) simulation data 126 according to the present embodiment.
The simulation data 126 is simulation data used for the “defect number estimation model”. One record is one simulation data 126, and each data is generated for each node in accordance with the simulation data post-partition distribution 125 and the post-partition distribution assignment table 129.
Here, for the generation of the simulation data 126, for example, general statistical software such as “R” is used.

  The specific example shows a case where a simulation data set (data 20) of SD1 is generated.

FIG. 9 is a diagram showing an example of the estimated value information 130 (estimated value 127) calculated by the estimated value calculating unit 103 according to the present embodiment.
The estimated value 127 is an estimated value obtained by giving the simulation data 126 created instead of the actual data at Step 1 to the “defect number estimating model” and inputting the model input value 123 at Step 2. A plurality of estimated values 127 are generated. One estimated value 127 is output for one of the generated simulation data 126, and is output by the number of generated simulation data 126.

FIG. 10 is a flowchart showing an operation confirmation method (distribution information generation process, simulation data creation process) of the operation confirmation apparatus 300 according to the present embodiment. FIG. 11 is a flowchart showing an operation confirmation method (estimated value calculation process, graph creation process) of the operation confirmation apparatus 300 according to the present embodiment. The operation confirmation method of the operation confirmation apparatus 300 will be described with reference to FIGS. 10 and 11.
First, the overall flow of the operation confirmation method of the operation confirmation apparatus 300 will be described.

<Distribution information generation process>
In S100, the distribution information generation unit 101 reads the simulation data basic distribution 121 (FIG. 6A) and the simulation data distribution division information 122 (FIG. 6B).
In S101, the distribution information generation unit 101 creates a simulation data post-division distribution 125 (FIG. 7E) based on the read simulation data basic distribution 121 and simulation data distribution division information 122.

<Simulation data creation process>
In S <b> 102, the simulation data creation unit 102 reads the simulation data post-partition distribution 125 and the post-partition distribution assignment table 129.
In S <b> 103, the simulation data creation unit 102 creates a plurality of simulation data 126 based on the read simulation data post-partition distribution 125 and the post-partition distribution assignment table 129. It is assumed that n pieces of simulation data 126 are created. That is, the simulation data creating unit 102 creates n pieces of simulation data from the first to the nth (n is an integer of 2 or more).

<Estimated value calculation process>
In S <b> 104, the estimated value calculation unit 103 reads the model input value 123.
In S105 to S106, the estimated value calculation unit 103 sequentially reads n pieces of simulation data 126 one by one. The simulation data to be processed is i-th (i is an integer from 1 to n).
First, step 1 is executed on the i-th simulation data using the defect number estimation model to calculate a conditional probability table. Next, the read model input value 123 is input, step 2 is executed, and the i-th estimated value 127 is calculated.
As shown in S105 of FIG. 10, the process is repeated under the conditions of i = 0, 1 <n, i ++, the first to nth simulation data 126 is processed, and the corresponding first to nth estimated values are processed. 127 is calculated.

  Here, for example, when each of the simulation data sets SD1, SD2, and SD3 has 40 simulation data 126, n = 120.

<Graph creation processing>
In S107, the graph creation unit 104 reads all n of the estimated values 127 and the graph setting information 124.
In S108, the graph creation unit 104 creates the first graph 128h, the second graph 128i, and the third graph 128j.

  This is the end of the description of the overall flow of the flowcharts shown in FIGS.

  Next, details of distribution information generation processing, simulation data creation processing, estimated value calculation processing, and graph creation processing of the operation confirmation method of the operation confirmation apparatus 300 according to the present embodiment will be described.

First, (1) distribution information generation processing: S100 to S101 will be described.
The distribution information generation unit 101 reads the simulation data basic distribution 121 (FIG. 6A) and the simulation data distribution division information 122 (FIG. 6B), and distributes the simulation data distribution 125 (FIG. 7B). e)).
INPUT:
(A) Basic distribution 121 for simulation data;
(B) Distribution division information 122 for simulation data.
OUTPUT:
(E) Distribution after simulation 125 for simulation data.

The distribution information generation unit 101 divides the distribution of the basic distribution 121 for simulation data based on the information set in “Dataset” of the distribution division information 122 for simulation data (b).
In the specific example, since “Dataset” is “3”, (a) the distribution of the basic distribution 121 for simulation data is divided into three.

  The distribution information generation unit 101 is an example of a reference distribution setting unit that sets a simulation data basic distribution 121 (reference distribution, node-specific reference distribution). Further, the simulation data basic distribution 121 (reference distribution) is divided into “Dataset” values (predetermined number of divisions) of the distribution division information 122 for simulation data (predetermined number of divisions), and a distribution 125 ( It is an example of the distribution division part which produces | generates the distribution after a division | segmentation of a division | segmentation number, and the distribution after a division | segmentation according to a node of a division | segmentation number.

  FIG. 13, FIG. 14, and FIG. 15 are image diagrams showing a mechanism for dividing the simulation data basic distribution 121 according to the present embodiment.

The distribution information generation unit 101 uses the value of “Normal Distribution Data Interval” of the distribution division information 122 for simulation data (b) to calculate the data range of the basic distribution 121 for simulation data (a).
However, all the nodes of the “defect number estimation model” are data having values of 0 or more, but (a) the simulation data basic distribution 121 is a result of simple distribution calculation performed on the measurement data. The data distribution range of ED, DR, and DIR (the range from the lower 95% point to the upper 95% point of the basic distribution) is a negative value.
In this case, the range is adjusted by setting the lower limit (lower 95% point) of the data range to 0, and the adjusted range is divided into three as shown in FIG.

As illustrated in FIG. 13, the distribution information generation unit 101 (a) divides the simulation data basic distribution 121 into three.
As shown in “divided image (normal distribution)” in FIG. 14 and FIG. 15, the distribution information generation unit 101 distributes each of the three divided distribution ranges (group 1, group 2, group 3).

  The distribution information generation unit 101 outputs a calculation result obtained by further distributing the three-divided distribution range as (e) simulation data distribution 125 (see FIG. 7E). The distribution after division of each node is as shown in this table ((e) distribution after distribution 125 for simulation data).

Next, (2) simulation data creation processing: S102 to S103 will be described.
16 and 17 are image diagrams showing a mechanism for creating a simulation data set and simulation data 126 according to the present embodiment.

The simulation data creation unit 102 reads the simulation data post-partition distribution 125 and the post-partition distribution assignment table 129 and creates a plurality of simulation data 126.
INPUT:
(E) a post-division distribution 125 for simulation data;
(K) The post-partition distribution allocation table 129.
OUTPUT:
(F) Simulation data 126 (arbitrary n pieces are created based on “DataSet”).

As shown in FIG. 16, the simulation data creation unit 102 compares the read (k) post-partition distribution allocation table 129 and (e) the post-partition distribution 125 for simulation data, and the distribution contents of simulation data from SD1 to SD3. To decide.
For example, for SD1, (k) DIR is set to 3 in the post-partition distribution assignment table 129, and 1 is set otherwise. Therefore, the simulation data 126 to be generated is (e) data generated from a combination (distribution set) of distributions “SQ1, RQ1, ED1, DR1, DIR3, DRE1” of the distribution for distribution 125 for simulation data (6 nodes). Full set of data)

  Next, the simulation data creation unit 102 determines the number of data in the simulation data set to be created from the “MaxData” value and the “GraphInterval” value of the distribution division information 122 for simulation data.

Here, the value of “MaxData” is “40” and the value of “GraphInterval” is “10”. This indicates that the maximum is 40 and the data count interval for each file is 10.
Files of 1, 10, 20, 30, and 40 data are created for SD1 to SD3, respectively. Note that the initial value of the number of data is defined in advance in the system. Here, since it is 1, it is 1 to 40.

As shown in FIG. 17, the simulation data creation unit 102 creates simulation data 126 using statistical software or the like. The list of simulation data created is as follows. However, in FIG. 17, only 13 pieces of simulation data 126 are shown.
(A) SD1-data1, SD1-data10, SD1-data20, SD1-data30, SD1-data40;
(B) SD2-data1, SD2-data10, SD2-data20, SD2-data30, SD2-data40;
(C) SD3-data1, SD3-data10, SD3-data20, SD3-data30, SD3-data40.

  Note that the simulation data generation itself is created using statistical software (R (21) or the like). An example of the created data is as shown in (f) simulation data 126 (example of SD1-data20) shown in FIG.

  The simulation data creation unit 102 obtains the “MaxData” value (predetermined in the simulation data distribution division information 122) from each of SD1, SD2, and SD3 of the distribution 125 for simulation data (distribution after division). This is an example of a simulation data creation unit that generates simulation data 126 of (number of data) and generates simulation data 126 for each of SD1, SD2, and SD3 (simulation data of the number of data for each distribution after division).

Next, (3) estimated value calculation processing: S104 to S106 will be described.
The estimated value calculation unit 103 reads (f) simulation data 126 and (c) model input value 123 and calculates estimated value information 130.
INPUT:
(F) simulation data 126;
(C) Model input value 123.
OUTPUT:
(G) Estimated value information 130.

A Bayesian network tool is used to calculate the estimated value by the estimated value calculation unit 103. A defect number estimation model is assembled on the tool in advance.
First, the estimated value calculating unit 103 optimizes the conditional probability table of the defect number estimation model on the tool by sequentially giving simulation data 126 one by one to the defect number estimation model on the tool at Step 1. .
The estimated value calculation unit 103 is an example of a probability table optimization unit that optimizes a conditional probability table by inputting each simulation data 126 of a plurality (n) of simulation data 126 to a defect number estimation model.

Next, the estimated value calculation unit 103 inputs the value specified as the (c) model input value 123 in Step 2 and calculates one set of estimated values 127.
The estimated value calculation unit 103 inputs the model input value 123 to the input target nodes ED and DR, and uses the defect number estimation model including the optimized conditional probability table, and the estimated target nodes DIR and DRE. It is an example of the estimated value output part which estimates the value of and outputs as the estimated value 127.

  For example, the estimated value calculation unit 103 gives SD1-data1 to the defect number estimation model, inputs ED = 0.63, DR = 0.83, and calculates the values of DIR and DER as the estimated value 127. The estimated value calculation unit 103 repeats the above-described process, and calculates the estimated value 127 for all the simulation data 126.

  The estimated value calculation unit 103 accumulates information in which each simulation data 126 of the plurality (n pieces) of simulation data 126 input to the defect count estimation model and the corresponding estimated value 127 are associated with each other, thereby estimating the estimated value information. It is an example of the analysis object information storage part memorize | stored in a memory | storage device as 130 (refer FIG. 8) (analysis object information).

Next, (4) graph creation processing: S107 to S108 will be described.
The graph creation unit 104 reads all n estimated values 127 (estimated value information 130) and the graph setting information 124, and creates a first graph 128h, a second graph 128i, and a third graph 128j.
INPUT:
(G) estimated value information 130;
(D) Graph setting information 124.
OUTPUT:
(H) first graph 128h (graph 1);
(I) the second graph 128i (graph 2);
(J) The third graph 128j (graph 3).

  FIG. 18 is a diagram illustrating an example of (h) the first graph 128h according to the first embodiment. FIG. 19 is a diagram showing an example of DIR of (i) the second graph 128i according to the first embodiment. FIG. 20 is a diagram illustrating an example of DRE of (i) the second graph 128i according to the first embodiment. FIG. 21 is a table showing a change rate of each estimated value of another simulation data set with respect to the simulation data set serving as a reference according to the first embodiment. FIG. 22 is a diagram showing an example of DIR of (j) the third graph 128j according to the first embodiment. FIG. 23 is a diagram illustrating an example of DRE of (j) the third graph 128j according to the first embodiment.

The graph creation unit 104 uses (g) the estimated value information 130 to (h) the first graph 128h (graph 1), (i) the second graph 128i (graph 2), (j) the third graph 128j ( Create graph 3).
(H) The first graph 128h (graph 1) is a graph visualizing the relationship between the number of simulation data and model accuracy (see FIG. 18).
(I) The second graph 128i (graph 2) is a graph visualizing the difference in the estimated values changed due to the difference in the simulation data (see FIGS. 19 and 20).
(J) The third graph 128j (graph 3) is a graph visualizing the relationship between the estimated values changed by the simulation data (see FIGS. 22 and 23).

  When (i) the second graph 128i (graph 2) is created, (d) “XAxisOrder” of the graph setting information 124 is used to arrange the X-axis simulation data sets (SD1, SD2, SD3). decide.

In addition, (h) when creating the first graph 128h (graph 1), (d) the value of “BaseLibne” in the graph setting information 124 is used, and each estimated value is changed based on the simulation data to which 1 is input. The rate is calculated and displayed on graph 1.
In the specific example, since SD2 is a reference (see FIG. 6D), numerical values are calculated as shown in FIG. 21 and displayed on the graph 1 (see FIG. 18).

  Next, each graph created by the graph creation unit 104 will be described with reference to FIGS.

With reference to FIG. 18, (h) the first graph 128h (graph 1) will be described.
(G) A graph created by using the estimated value information 130, and a graph for checking the positioning of the estimated value with respect to the reference value at a time for each simulation data set (SD1, SD2 (reference), SD3). is there. This graph makes it possible to confirm how the estimated value and its confidence interval (estimated value ± 2σ) change as the number of data to be given increases.

19 and 20, (i) the second graph 128i (graph 2) will be described.
(G) A graph created using the estimated value information 130, and is a graph for checking the value of the estimated value itself for the simulation data set (SD1, SD2, SD3). If this graph is used, the difference of the estimated value for every simulation data set (SD1, SD2, SD3) can be confirmed.

(J) The third graph 128j (graph 3) will be described with reference to FIGS.
(G) It is a graph created using the information of the estimated value information 130, and is a graph for checking the value of the estimated value itself for each simulation data. This graph makes it possible to confirm how the estimated value and its confidence interval (estimated value ± 2σ) change as the number of data to be given increases.

  This is the end of the description of the functions and operations of the operation check apparatus 100 according to the present embodiment.

  Hereinafter, the functions and operations of the operation check device 100 according to the present embodiment will be described in other words and the effects will be described.

The operation check device 100 according to the present embodiment creates simulation data 126 according to a predefined distribution (simulation data basic distribution 121) instead of the actual data in order to arbitrarily change the conditional probability table. And give to the model.
The operation check apparatus 100 further creates a plurality of distributions within the range of the simulation data basic distribution 121. The distribution created in this way is the distribution 125 for simulation data. Each of the plurality of created distributions is set as a distribution of the data set (SD1, SD2, SD3).

  In creating the distribution of the data set, the operation checking apparatus 100 defines the granularity, and creates a distribution in the data range of the basic distribution of each node by the number of the granularities. “Granularity” corresponds to the value of “Dataset” in the distribution division information 122 for simulation data described above.

  The operation checking apparatus 100 calculates the estimated value 127 for each distribution unit of the data set. As for the normal distribution node, since the data range cannot be limited as it is, the effective range for the distribution, such as the upper 95% point to the lower 95%, is arbitrarily determined to be a range. This range uses the value of “Normal Distribution Data Interval” in the distribution division information 122 for simulation data (b).

  In the operation check apparatus 100, the divided distribution created in this way is assigned to each node (e) a divided distribution 125 for simulation data is created, and data is generated for each node according to the distribution. Thereafter, the generated data are combined according to the post-division distribution assignment table 129 to form a simulation data set. When these simulation data sets are given to the model, different conditional probability tables are automatically generated. At this time, a fixed value defined as a reference value is input and an estimated value is calculated. Keep the input value constant and eliminate factors other than the effects of the conditional probability table. Visualize the calculated estimate and evaluate the results.

As described above, according to the operation checking apparatus 100 according to the present embodiment, it is possible to calculate an estimation result for simulation data and create a graph comparing the results. As a result, it is possible to obtain an indication of the model use effect even when no actual data is obtained.
In addition, by comparing the graph with the estimated value obtained using the “defect number estimation model” at the time of actual software development and the value finally obtained as actual data, Since it is possible to know whether the organization is close to the simulation data, the organization can be evaluated.

For example, (h) it can be seen from graph 1 that the DIR is lowered and the DRE is increased in the organization (SD3) having the development ability compared with the organization (SD2) having the medium software development ability. On the other hand, it can be seen that DIR increased and DRE decreased in the organization (SD1) with insufficient development ability. (I) From the graph 2, an approximate estimated value when used in each organization is known.
Therefore, since the approximate values of other organizations can be found from the information in (h) Graph 1 and (i) Graph 2, it is possible to check your organization by checking against the graph when values are actually obtained in your organization. You can check the level of the level and compare with other organizations.
Also, if you raise the level of development, you can know how much it will be.
(J) From the graph 3, it is possible to know how much performance data can be prepared in each organization and how much model accuracy can be expected.

  In particular, there is an effect that simulation data can be created by processing of (1) distribution information generation function and (2) simulation data creation function. In addition, (4) in the process of the graph creation function, there is an effect that the calculated estimated value is visualized in a form according to the purpose.

In the description of the first embodiment, the “distribution information generation unit”, “simulation data generation unit”, “estimated value calculation unit”, “graph generation unit”, etc. constitute the operation check device 300 as independent function blocks. ing. However, the present invention is not limited to this. For example, the “distribution information generation unit” and the “simulation data creation unit” are realized by one functional block, and the “estimated value calculation unit” and the “graph creation unit” have one function. It may be realized by a block. Alternatively, the operation check apparatus 300 may be configured by any combination of these functional blocks.
In addition, this invention is not limited to this embodiment, A various change is possible as needed.

DESCRIPTION OF SYMBOLS 100 Operation confirmation apparatus, 101 Distribution information generation part, 102 Simulation data creation part, 103 Estimated value calculation part, 104 Graph creation part, 120 data, 121 Basic distribution for simulation data, 122 Distribution division information for simulation data, 123
Model input value, 124 graph setting information, 125 simulation data divided distribution, 126 simulation data, 127 estimated value, 128 graph, 128h first graph, 128i second graph, 128j third graph, 129 divided distribution allocation table , 130 Estimated value information, 901 LCD, 902 keyboard, 903 mouse, 904
FDD, 905 CDD, 906 printer, 911 CPU, 912 bus, 913
ROM, 914 RAM, 915 communication board, 920 HDD, 921 operating system, 922 window system, 923 program group, 924 file group.

Claims (9)

A Bayesian network model that targets a plurality of nodes including an input target node that is an input target to which an input value is input and an estimation target node that is an estimation target, and the dependency probability of the plurality of nodes is conditional probability In a model analyzer for analyzing a Bayesian network model including a conditional probability table expressed by
An input value storage unit for storing the input value to be input to the input target node in a storage device;
A simulation data generation unit for generating a plurality of simulation data input to the Bayesian network model;
A probability table changing unit that inputs each simulation data of the plurality of simulation data generated by the simulation data generating unit to the Bayesian network model and changes the conditional probability table;
The input value stored in the input value storage unit is input to the input target node of the Bayesian network model including the conditional probability table changed by the probability table changing unit, and the estimated value of the estimation target node And an analysis target information generating unit that accumulates estimated value information in which the calculated estimated value is associated with the simulation data input to the Bayesian network model and stores the estimated value information as analysis target information in a storage device. A model analyzer characterized by that. The simulation data generation unit
Generate n simulation data from the first to n (n is an integer of 2 or more),
The probability table changing unit is
Each time the n pieces of simulation data generated by the simulation data generation unit are sequentially read, and the i-th simulation data is read (i is an integer between 1 and n), the i-th simulation data is converted into the Bayesian network model. To change the conditional probability table,
The analysis target information generation unit
By outputting the i-th estimated value information in which the calculated estimated value is associated with the i-th simulation data, n pieces of the estimated value information are output, and the n pieces of the estimated value information that have been output The model analysis apparatus according to claim 1, wherein the information is stored in the storage device as the analysis target information. The simulation data generation unit
A reference distribution setting unit for setting a reference distribution used for generating the plurality of simulation data;
A distribution dividing unit that divides the reference distribution set by the reference distribution setting unit into a predetermined number of divisions and generates a distribution after division of the number of divisions;
A simulation data creating unit that generates simulation data of a predetermined number of data from each of the divided distributions of the division number and generates simulation data of the data number for each of the divided distributions of the division number; The model analysis apparatus according to claim 1, wherein: The simulation data to be input to the Bayesian network model is composed of simulation data of the number of data generated for each divided distribution of the number of divisions,
The analysis target information generation unit
For each simulation data of the number of data generated for each distribution after the division, the estimated value information in which the calculated estimated value and the simulation data input to the Bayesian network model are associated is accumulated and divided. The model analysis apparatus according to claim 3, wherein each model is generated as post-distribution analysis target information. The reference distribution has a node-specific reference distribution that is a distribution for each node of the plurality of nodes.
The distribution divider is
Dividing the node-specific reference distribution into the number of divisions, generating the divided number of divided distributions per node to obtain the divided number of divided distributions,
The simulation data creation unit
A post-partition distribution allocation information storage unit that pre-stores post-partition distribution allocation information that defines a combination of the post-partition distribution by node of the number of partitions for each post-partition distribution of the number of divisions;
Based on the post-division distribution allocation information stored by the post-division distribution allocation information storage unit, for each post-division distribution of the division number, generate a distribution set that combines the post-division distribution by node of the division number, The model analysis apparatus according to claim 4, wherein simulation data of the number of data is generated from the generated distribution set. The model analyzer further includes:
Based on the analysis target information stored by the analysis target information generation unit, a change in the value of the estimation target node accompanying a change in the conditional probability table by inputting the plurality of simulation data to the Bayesian network model The model analysis apparatus according to claim 1, further comprising a model analysis unit that analyzes The model analysis unit
The model according to claim 6, wherein a change in the value of the estimation target node accompanying a change in the conditional probability table is visualized based on the analysis target information stored by the analysis target information generation unit. Analysis equipment. A Bayesian network model that targets a plurality of nodes including an input target node that is an input target to which an input value is input and an estimation target node that is an estimation target, and the dependency probability of the plurality of nodes is conditional probability In the model analysis method of the model analyzer for analyzing the Bayesian network model including the conditional probability table represented by
A simulation data generation unit generates a plurality of simulation data input to the Bayesian network model,
A probability table changing unit inputs each simulation data of the plurality of simulation data generated by the simulation data generating unit to the Bayesian network model to change the conditional probability table,
The analysis target information generation unit inputs a predetermined input value to the input target node of the Bayesian network model including the conditional probability table changed by the probability table change unit, and estimates the estimation target node A model analysis method characterized by calculating a value and accumulating estimated value information in which the calculated estimated value is associated with the simulation data input to the Bayesian network model to generate analysis target information. A Bayesian network model that targets a plurality of nodes including an input target node that is an input target to which an input value is input and an estimation target node that is an estimation target, and the dependency probability of the plurality of nodes is conditional probability In the model analysis program of the model analyzer that analyzes the Bayesian network model including the conditional probability table expressed by
A simulation data generation unit that generates a plurality of simulation data input to the Bayesian network model; and
A probability table changing unit that inputs each simulation data of the plurality of simulation data generated by the simulation data generating process to the Bayesian network model and changes the conditional probability table;
The analysis target information generation unit inputs a predetermined input value to the input target node of the Bayesian network model including the conditional probability table changed by the probability table change process, and estimates the estimation target node. An analysis object information generation process for calculating the value and storing the estimated value information in which the calculated estimated value and the simulation data input to the Bayesian network model are associated to generate the analysis object information A model analysis program that is executed by a model analyzer.

Images