JP2012123749A - Sensibility estimation device, sensibility estimation method, and sensibility estimation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To estimate user's sensibility highly precisely and efficiently.SOLUTION: A sensibility estimation device which estimates a sensibility during and after use of a sensibility evaluation object by a user has a preference characteristic pattern extraction means which classifies for each preference characteristic pattern acquiring a predetermined overall evaluation on a plurality of sensibility fields, using a hierarchical neural network skeleton learning method configured beforehand.

Description

  The present invention relates to a sensitivity estimation device, a sensitivity estimation method, and a sensitivity estimation program, and more particularly, to a sensitivity estimation device, a sensitivity estimation method, and a sensitivity estimation program for accurately and efficiently estimating a user's sensitivity.

  Conventionally, in the sensitivity evaluation such as usability in cosmetics and pharmaceuticals (skin external preparations) applied to the skin, for example, the evaluation at the previous stage of applying cosmetics etc. to the skin etc., for example, viscosity of the base, rolling friction, etc. The base was evaluated. In addition, for the evaluation during and after application of cosmetics on the skin, etc., for example, using a preset sensitivity word (for example, refreshing, moist, sticky, penetration feeling, etc.) A consumer test, etc. was conducted, and the sensitivity was evaluated based on the results.

  Note that, as described above, a technique is disclosed in which sensitivity evaluation analysis is performed based on evaluation data by a learning type neural network based on evaluation data created from answer items collected by a questionnaire (for example, Patent Document 1). reference.).

Japanese Patent Laid-Open No. 5-108602

  By the way, there are individual differences in preference and sensitivity for sensibility. For example, in general consumers, there are users who can understand even a slight difference in feeling of use, users who do not feel a slight difference in feeling of use, and the like. However, the individual differences and the like in the prior art described above are usually classified only by attributes (such as age), and therefore, the sensitivity is not related to the individual differences, and appropriate adjustment has not been performed. Therefore, some measures are required to improve the accuracy of sensitivity evaluation.

  In addition, it is considered that there are multiple levels of sensibility. For example, among the levels such as “shallow sensibility” or “deep sensibility”, for example, there is an individual difference in how to feel one deep sensibility. ”And“ deep sensitivity ”need to be elucidated. Note that “shallow sensibility” refers to, for example, sensibility (eg, moist, smooth) that is directly felt by the skin's sensilla, and “deep sensibility” refers to, for example, complex judgment (eg, penetrating sensation) , And the like). However, conventionally, a large number of questionnaire items are only analyzed side by side, and this also requires some measures to improve analysis accuracy.

  In addition, as an examination item necessary for accurately quantifying Kansei evaluation, for example, extracting individual differences in Kansei evaluation, reflecting Kansei hierarchical structure, extracting Kansei items that can be converted into base physical properties, etc. However, in the conventional questionnaire evaluation method, there is no questionnaire method that can appropriately collect the above-described contents. The above-described hierarchical structure of sensibilities indicates a form in which a plurality of sensibilities such as “shallow sensibility” and “deep sensibility” are hierarchically structured.

  That is, in the past, since many questionnaire items were analyzed side by side, evaluations associated with a hierarchical structure such as “shallow sensitivity” and “deep sensitivity” could not be reflected.

  In addition, since the correlation between the physical properties of the base and the sensibility has not been found in the past, the precept development has to rely on the experience and intuition of the person in charge, and the user's sensibility can be estimated with high accuracy and efficiency. There was no way to do this.

  The present invention has been made in view of the above-described problems, and provides a sensitivity estimation device, a sensitivity estimation method, and a sensitivity estimation program for estimating the sensitivity of a user (customer) with high accuracy and efficiency. With the goal.

  In order to solve the above-described problems, the present invention employs means for solving the problems having the following characteristics.

  The invention described in claim 1 uses a skeleton learning method of a hierarchical neural network that is set in advance in a sensitivity estimation device that estimates the sensitivity during and after use of a sensitivity evaluation object used by a user. It is characterized by having a preference characteristic pattern extracting means for classifying each preference characteristic pattern for obtaining a predetermined comprehensive evaluation for a plurality of sensitivity items.

  According to the first aspect of the present invention, by using a neural network that is a non-linear type, it is possible to effectively make an inference on sensitivity without a physical model. In addition, it is possible to estimate the sensitivity with high accuracy by reflecting the individual difference for each user in the input sensitivity evaluation that affects the comprehensive evaluation.

  The invention described in claim 2 is an input means for inputting base material property information for the sensitivity evaluation object, and a hierarchy for hierarchizing the sensitivity of the user to the sensitivity evaluation object by using the plurality of sensitivity items. Based on the structure estimation means, the base physical property information obtained by the input means, and the hierarchical structure obtained by the hierarchical structure estimation means, to the sensitivity evaluation object for the classification obtained by the preference characteristic pattern extraction means And a sensibility estimation means for estimating the sensibility.

  According to the second aspect of the present invention, the sensitivity of the user who is a general consumer can be estimated with high accuracy and efficiency from the base material property information.

  According to a third aspect of the present invention, there is provided a sensibility evaluation unit that evaluates a sensibility by associating a sensitivity estimation result obtained by the sensitivity estimation unit with a preset evaluation index.

  According to the third aspect of the present invention, the evaluation result can be easily grasped by evaluating the user's sensibility by using numerical values or indexes such as predetermined characters.

  According to a fourth aspect of the present invention, at least one of the content extracted by the preference characteristic pattern extraction unit, the estimation result obtained by the hierarchical structure estimation unit, and the sensitivity estimation result obtained by the sensitivity estimation unit is provided. It has the screen generation means which produces | generates the screen for displaying on a display means, It is characterized by the above-mentioned.

  According to the invention of claim 4, various execution contents can be easily grasped.

  The invention described in claim 5 uses a skeleton learning method of a hierarchical neural network set in advance in a sensitivity estimation method for estimating sensitivity during use and after use of a sensitivity evaluation object used by a user. And a preference characteristic pattern extracting step for classifying each preference characteristic pattern for obtaining a predetermined comprehensive evaluation for a plurality of sensitivity items.

  According to the fifth aspect of the invention, by using a neural network that is a non-linear type, it is possible to effectively make an inference on sensitivity without a physical model. In addition, it is possible to estimate the sensitivity with high accuracy by reflecting the individual difference for each user in the input sensitivity evaluation that affects the comprehensive evaluation.

  The invention described in claim 6 includes an input step of inputting base material property information for the sensitivity evaluation object, and a hierarchy for hierarchizing the sensitivity of the user to the sensitivity evaluation object using the plurality of sensitivity items. Based on the structure estimation step, the base physical property information obtained by the input means, and the hierarchical structure obtained by the hierarchical structure estimation step, to the sensitivity evaluation object for the classification obtained by the preference characteristic pattern extraction step A susceptibility estimation step for estimating the sensibility.

  According to the sixth aspect of the present invention, the sensitivity of the user who is a general consumer can be estimated with high accuracy and efficiency from the base material property information.

  The invention described in claim 7 includes a sensitivity evaluation step of evaluating sensitivity by associating a sensitivity estimation result obtained in the sensitivity estimation step with a preset evaluation index.

  According to the seventh aspect of the present invention, the evaluation result can be easily grasped by evaluating the user's sensibility by using numerical values or indexes such as predetermined characters.

  In the invention described in claim 8, at least one of the content extracted by the preference characteristic pattern extraction step, the estimation result obtained by the hierarchical structure estimation step, and the sensitivity estimation result obtained by the sensitivity estimation step is provided. A screen generation step for generating a screen to be displayed on the display means is provided.

  According to the eighth aspect of the invention, various execution contents can be easily grasped.

  The invention described in claim 9 is a sensitivity estimation program that causes a computer to function as the sensitivity estimation device according to any one of claims 1 to 4.

  According to the ninth aspect of the invention, by using a neural network that is a non-linear type, it is possible to effectively make an inference on sensitivity without a physical model. In addition, it is possible to estimate the sensitivity with high accuracy by reflecting the individual difference for each user in the input sensitivity evaluation that affects the comprehensive evaluation. Also, by installing the program, the sensitivity estimation process according to the present invention can be easily realized by a general-purpose personal computer or the like.

  According to the present invention, the sensitivity of a user (customer) can be estimated with high accuracy and efficiency.

It is a figure which shows an example of a function structure of the sensitivity estimation apparatus in this embodiment. It is a figure which shows an example of the hardware constitutions which can implement | achieve the sensitivity estimation process in this embodiment. It is a flowchart which shows an example of the sensitivity estimation processing procedure after the sensitivity evaluation target object application in this embodiment. It is a figure for demonstrating the outline | summary of the sensitivity estimation process in this embodiment. It is a figure for demonstrating a viewpoint required for the sensitivity evaluation in this embodiment. It is a figure for demonstrating the correlation method using the hierarchical neural network in this embodiment. It is a figure for demonstrating concretely the classification in this embodiment. It is a figure which shows an example of the result classified. FIG. 9 is a table showing the classification results shown in FIG. 8. It is a figure for demonstrating the hierarchical structure estimation method in this embodiment. It is a figure which shows an example of an analysis result. An example of a sensitivity item that affects the permeation feeling in the upper two groups is shown. It is a figure for demonstrating the correlation of a base physical property and shallow sensitivity. It is a figure which shows an example of the relationship between a base physical-property value and a sensitivity evaluation value. It is a figure which shows an example of the production | generation screen in this embodiment.

<About the present invention>
The present invention pre-analyzes prescriptions, physical properties of products, and questionnaire results for users (general consumers) in sensitivity evaluation such as usability in sensitivity evaluation objects such as cosmetics and pharmaceuticals (skin external preparations) applied to the skin. By this, the system which estimates the sensitivity when a user uses the sensitivity evaluation object is provided.

  At this time, in the present invention, for example, in order to estimate the relationship between the physical property value and the sensibility using nonlinear processing, the usability (sensitivity) that the user feels from the physical properties of the prescription / sensitivity evaluation target product using a neural network or the like. ).

  Note that there are usually large individual differences in sensitivity evaluation for tactile sensation. Therefore, in the present invention, users are classified according to preference characteristics by, for example, a group classification method using a skeleton learning method, and are used for sensitivity estimation in association with a hierarchical structure of emotions.

  Note that the above-described preference characteristics include, for example, users who have a slight difference in feeling of use and those who do not feel a difference in feeling of use. Perform group classification according to Thereby, sensitivity estimation accuracy can be improved reflecting individual differences.

  In the present invention, the above-described hierarchical structure of sensitivity is reflected in the evaluation. Specifically, a plurality of different sensitivity items (for example, “shallow sensitivity”, “deep sensitivity”, etc.) are hierarchically related as the hierarchical structure of sensitivity. In addition, since there is individual difference in how to feel one “deep sensibility”, as described above, the hierarchical structure of sensibilities is clarified for each preference characteristic, and “shallow sensibility” and “deep sensibility” are clarified. Further, according to the present invention, the sensitivity when the user uses the sensitivity evaluation object can be estimated from the physical property information of the base. Note that a plurality of sensitivity items in the present invention can be set in advance according to, for example, a sensitivity evaluation object used by the user. The sensitivity items are not limited to names such as “shallow sensitivity” and “deep sensitivity”, and can be arbitrarily set using predetermined numerical values, parameters, and the like.

  DESCRIPTION OF EMBODIMENTS Embodiments of a sensitivity estimation device, a sensitivity estimation method, and a sensitivity estimation program according to the present invention will be described below with reference to the drawings. In the embodiment described below, cosmetics applied to the skin are used as an example of the sensitivity evaluation target, but the present invention is not limited to this.

<Sensitivity estimation device: functional configuration example>
A functional configuration example of the sensitivity estimation apparatus according to the present embodiment will be described with reference to the drawings. FIG. 1 is a diagram illustrating an example of a functional configuration of the sensitivity estimation apparatus according to the present embodiment. The sensitivity estimation device 10 shown in FIG. 1 includes an input means 11, an output means 12, a storage means 13, a preference characteristic pattern extraction means 14, a hierarchical structure estimation means 15, a sensitivity estimation means 16, and a sensitivity evaluation means 17. And a screen generation means 18, a transmission / reception means 19, and a control means 20.

  The input means 11 inputs preset physical property information, or a preference characteristic pattern extraction instruction for extracting a preference characteristic pattern based on a predetermined condition or the like for a user who uses cosmetics or the like, or estimates a hierarchical structure Input of various instructions such as a hierarchical structure estimation instruction, a sensitivity estimation instruction for estimating sensitivity, a sensitivity evaluation instruction for finally evaluating sensitivity, a screen generation instruction, and a transmission / reception instruction is accepted. Note that the input unit 11 includes, for example, a keyboard and a pointing device such as a mouse.

  The base physical property information in the present embodiment is, for example, a viscosity or a contact angle if it is a static characteristic, and is a rolling friction or a surface friction if it is a dynamic characteristic.

  In addition, for rolling friction, it is possible to input the respective physical property values with respect to predetermined conditions such as the average value after finishing and the peak value when drying, and the friction coefficient is also the average value after finishing and when drying Each physical property value can be input with respect to a predetermined condition such as a peak value of 1 and an average value immediately after coating. In addition, the base physical property information value input by the input unit 11 may be a value obtained from an existing measuring device or the like, or may be manually input by a person who performs a sensitivity estimation process.

  The input means 11 can also input various data of the physical property values of the base material by an external device or the like, or input learning data of a set neural network to be described later.

  The output unit 12 displays / outputs the content input by the input unit 11 and the content executed based on the input content. The output unit 12 includes a display, a speaker, and the like. Further, the output unit 12 may have a function such as a printer. In that case, the screen content generated by the screen generation unit 18 is printed on a print medium such as paper and presented to the user or the like. You can also

  The storage means 13 is a variety of preset physical property information, various neural network data used in the present embodiment, preference characteristic pattern extraction results, hierarchical structure estimation results, sensitivity estimation results, sensitivity evaluation results, Accumulate various data such as generated screens.

  The storage unit 13 also stores the above-described various information acquired by the transmission / reception unit 19. Further, the storage means 13 can read out various data stored as required.

  The preference characteristic pattern extraction unit 14 performs group classification of a plurality of users according to individual differences (preference characteristics) of sensitivity set in advance. Specifically, the preference characteristic pattern extraction unit 14 classifies the user according to the user's preference characteristic pattern, and, for example, classifies the user according to the hierarchical structure of “shallow sensibility” and “deep sensibility”, or the relevance of the sensitivity evaluation and the comprehensive evaluation. Classify. Details of the classification method in the preference characteristic pattern extraction unit 14 will be described later.

  The hierarchical structure estimation means 15 stratifies the relationship between preset sensitivity items obtained when, for example, cosmetics or the like are applied to the user, and the comprehensive evaluation result by a plurality of layers. Specifically, the hierarchical structure estimation means 15 uses, for example, a hierarchical neural network or the like to construct a neural network having a plurality of sensitivity items immediately after application as input nodes and an overall evaluation point as an output node. To do.

  Thus, by elucidating the hierarchical structure of sensibilities for each preference characteristic pattern, for example, the structure of multiple hierarchies such as “shallow sensibility” and “deep sensibility” can be clarified. It is possible to clarify individual differences in how to feel “deep sensitivity” and improve sensitivity evaluation accuracy. In this embodiment, for example, the process of human sensitivity evaluation can be simply expressed by using only two layers such as “shallow sensitivity” and “deep sensitivity” described above. The number of hierarchies in is not limited to this. Details of the hierarchical structure estimation method in the hierarchical structure estimation means 15 will be described later.

  The sensitivity estimation means 16 estimates sensitivity from physical property values. For example, “shallow sensibility” is usually directly felt by a sensory organ of the skin, such as moist and smooth, and thus can be easily converted into a base material property value. Therefore, the sensitivity estimation means 16 estimates the correlation between the “shallow sensitivity” and the physical property value as the base physical property information based on preset conditions such as the result obtained by the hierarchical structure estimation means 15. Thus, “shallow sensibility” can be determined by estimating the sensibility hierarchical structure.

  This makes it possible to relate “shallow sensibility” and “base physical property value”, so that it is easy to utilize for prescription development, for example, when conducting a questionnaire, it is possible to reduce questionnaire items, A highly sensitive questionnaire can be proposed. The sensitivity estimation method in the sensitivity estimation means 16 will be described later.

  The sensitivity evaluation means 17 comprehensively evaluates the sensitivity by matching the sensitivity estimation result obtained by the sensitivity estimation means 16 with a preset evaluation index. It should be noted that the sensibility evaluation means 17 gives an evaluation result as one of a plurality of preset comprehensive evaluation item indexes such as “I like / dislike this cosmetic”, “Good / bad usability”, etc. May be selected, or may be the result of quantification based on a preset index such as “favorability 95%”, “favorability level 8 (10-step evaluation)”, etc. .

  The screen generation means 18 generates various screens such as data input necessary for execution by each functional configuration of the sensitivity estimation apparatus 10 in the present embodiment, output of execution results, notification at the time of error, and the generated screens. Is output to the output means 12 such as a display. Specifically, the screen generation unit 18 includes the content of the preference characteristic pattern obtained by the preference characteristic pattern extraction unit 14, the estimation result obtained by the hierarchical structure estimation unit 15, the sensitivity estimation result obtained by the sensitivity estimation unit 16, and the sensitivity. A screen for displaying at least one of the sensitivity evaluation results obtained by the evaluation unit 17 on the output unit 12 is generated.

  The transmission / reception means 19 inputs various data necessary for completion estimation processing in the present embodiment from an external device connected by a communication network or the like, and outputs various execution results executed by the sensitivity estimation device 10 to the external device. be able to.

  The control means 18 controls the entire components of the sensitivity estimation device 10. Specifically, the control means 18 is based on, for example, various instruction information from the input means 11 by the user or the like, such as preference characteristic pattern extraction processing, hierarchical structure estimation processing, sensitivity estimation processing, sensitivity evaluation processing, screen generation processing, Each control such as transmission / reception processing is performed.

<Sensitivity estimation device 10: hardware configuration>
Here, in the sensitivity estimation apparatus 10 described above, an execution program (sensitivity estimation program) that can cause a computer to execute each function is generated, and the execution program is installed in, for example, a general-purpose personal computer or server. The sensitivity estimation process in the present invention can be realized. Here, a hardware configuration example of a computer capable of realizing the sensitivity estimation process in the present embodiment will be described with reference to the drawings.

  FIG. 2 is a diagram illustrating an example of a hardware configuration capable of realizing the sensitivity estimation process according to the present embodiment. 2 includes an input device 31, an output device 32, a drive device 33, an auxiliary storage device 34, a memory device 35, a CPU (Central Processing Unit) 36 that performs various controls, and a network connection device. 37, and these are connected to each other by a system bus B.

  The input device 31 has a pointing device such as a keyboard and a mouse operated by a user or the like, and inputs various operation signals such as execution of a program from the user or the like. In addition, the input device 31 is obtained from an external device connected to the network connection device 37 or the like via a communication network, and includes various physical property information already measured by the external device or the like, learning data used in the neural network, and the like. Data can also be entered.

  The output device 32 has a display for displaying various windows and data necessary for operating the computer main body for performing the processing according to the present invention, and displays the program execution progress and results by the control program of the CPU 36. can do. Further, the output device 32 can print the above processing result or the like on a print medium such as paper and present it to the user or the like.

  Here, the execution program installed in the computer main body in the present invention is provided by, for example, a portable recording medium 38 such as a USB (Universal Serial Bus) memory, a CD-ROM, or a DVD. The recording medium 38 on which the program is recorded can be set in the drive device 33, and the execution program included in the recording medium 38 is installed in the auxiliary storage device 34 from the recording medium 38 via the drive device 33.

  The auxiliary storage device 34 is a storage means such as a hard disk, and can store an execution program in the present invention, a control program provided in a computer, and the like, and can perform input / output as necessary.

  The memory device 35 stores an execution program read from the auxiliary storage device 34 by the CPU 36. The memory device 35 includes a ROM (Read Only Memory), a RAM (Random Access Memory), and the like.

  The CPU 36 controls processing of the entire computer, such as various operations and input / output of data with each hardware component, based on a control program such as an OS (Operating System) and an execution program stored in the memory device 35. Thus, each processing can be realized. Various information necessary during the execution of the program can be acquired from the auxiliary storage device 34, and the execution result and the like can also be stored.

  The network connection device 37 acquires an execution program from another terminal connected to the communication network by connecting to a communication network or the like, or an execution result obtained by executing the program or an execution in the present invention The program itself can be provided to other terminals.

  The network connection device 37 can also acquire various data used in the physical property values of the base already measured by the external device connected to the communication network, the learned neural network, and the like.

  With the hardware configuration as described above, the sensitivity estimation process according to the present invention can be executed. Also, by installing the program, the sensitivity estimation process according to the present invention can be easily realized by a general-purpose personal computer or the like.

<Sensibility estimation processing procedure>
Next, the sensitivity estimation processing procedure in this embodiment will be described. FIG. 3 is a flowchart showing an example of a sensitivity estimation processing procedure after application of the sensitivity evaluation object in the present embodiment. In the following description, cosmetics are used as an example of the sensitivity evaluation target, but the present invention is not limited to this.

  In the sensitivity estimation processing procedure shown in FIG. 3, first, sensitivity evaluation information of a plurality of users for cosmetics whose sensitivity is estimated is input (S01), and a preference characteristic pattern is extracted from the input sensitivity evaluation information (S02). Further, the relationship between preset sensitivity items obtained when cosmetics are applied to the user and the overall evaluation result is hierarchized by a plurality of hierarchies, and a hierarchical structure is estimated for the sensitivity (S03).

  In addition, base physical property information is input (S04). Note that the base material property information may be used by reading out only information necessary for sensitivity estimation from the base material property information already stored in the storage unit 13 or the like.

  Next, using the hierarchical structure estimated in the process of S03, the sensitivity to the base physical property information obtained by the process of S04 is estimated (S05), and the sensitivity is evaluated based on the estimated result (S06).

  Next, a screen for displaying the results (estimation / evaluation results) obtained by the above-described processes on the output means 12 such as a display is generated (S07), and the generated screen is output (S08).

  Here, it is determined whether or not to end the sensitivity estimation process (S09), and when it does not end (NO in S09), the process returns to S01 to input other sensitivity evaluation information, base physical property information, etc. Process. In the process of S09, when the process is terminated based on an end instruction from the user or the like (YES in S09), the sensitivity estimation process is terminated. In the above process, the order of the processes of S01 to S03 and the process of S04 may be switched.

  By the processing procedure described above, the user's sensitivity can be estimated with high accuracy and efficiency. Therefore, the sensibility expected from users who have relied on the experience and intuition of researchers at the time of product development can be estimated. In addition, usability can be estimated without conducting a consumer questionnaire, so it is possible to efficiently estimate sensitivity, and using the result can shorten the time for final product development. . In addition, users can be classified according to taste characteristics, and the taste characteristics of the target layer can be known. Furthermore, the usability of cosmetics that satisfy “deep sensibility” that has been difficult to design can be designed using the physical properties of the base as an index.

<Specific example of sensitivity estimation using this embodiment>
Next, a specific example of sensitivity estimation using the above-described embodiment will be described with reference to the drawings. FIG. 4 is a diagram for explaining the outline of the sensitivity estimation process in the present embodiment. In the present embodiment, for example, the relationship between base physical properties and sensibilities as shown in FIG. 4 is systematized for each group in which users are classified according to preference characteristic patterns.

  For example, FIG. 4 shows the relationship in four items of “base physical properties”, “shallow sensitivity”, “deep sensitivity”, and “overall evaluation” as an example. That is, in the present embodiment, the relationship between each of the base physical properties consisting of one level is expressed using sensitivity items consisting of a plurality of levels (in the example of FIG. 4, “shallow sensitivity” and “deep sensitivity”). Estimate and finally make a comprehensive evaluation.

  The “base physical properties” include, for example, static properties such as “viscosity” and “contact angle”, and dynamic properties such as “rolling friction” and “surface friction”, and “shallow sensitivity” includes , “Feel fresh”, “moist”, “sticky”, “smooth”, etc. “deep sensibility” includes “penetration”, “feel”, etc. , "I like this cosmetic!" In addition, a circle (“◯”) in FIG. 4 indicates an intermediate node between items.

  In this way, by associating the above-mentioned items, it is possible to accurately and efficiently estimate the sensibility such as usability and effect feeling expected from the user who relied on the experience and intuition of the prescription developer. Become.

  In addition, about the specific content in each item, it is not limited to the content mentioned above. Further, as the estimation method using the relevance to the sensibility in which a plurality of elements are complicated as described above, for example, a neural network analysis method or the like can be used.

<Neural network used in this embodiment>
Here, the neural network used in this embodiment will be described. A neural network is a tool that can elucidate the correlation between input and output having a complicated relationship. In addition, as a type of neural network applicable in the present embodiment, for example, there is a hierarchical neural network or the like as a combination of elements, and a learning method includes back pro vacation or the like.

  In addition, the neural network has two major characteristics of “learning ability” and “nonlinearity”. As “learning ability”, a network pattern representing the correlation between input and output can be automatically formed. For example, an input and an output whose correlation is known are prepared in advance, and the weight is automatically corrected so that the output is calculated from the input to form a network pattern. In addition, the sensitivity evaluation is estimated by inputting one or more base material property values in this embodiment as input data for the network pattern.

  In addition, since neural networks are “non-linear”, no preconditions are necessary for data distribution, etc. (in the case of linear, data distribution such as Gaussian distribution is required), and it is possible to deal with relationships that are difficult to formulate by learning. (In the linear case, it must be expressed as a physical formula). Therefore, the neural network method is effective for sensitivity analysis without a physical model.

  Here, FIG. 5 is a figure for demonstrating a viewpoint required for the sensitivity evaluation in this embodiment. In the present embodiment, as shown in FIG. 5A, based on the physical property values (viscosity, friction, etc.) of the prescription, a user characteristic is classified for each preference characteristic to extract a preference characteristic pattern (FIG. 5A- ( 1)), the hierarchical structure of sensitivity is estimated (FIGS. 5A to 5), and the sensitivity is estimated (evaluated) based on the results (FIGS. 5A to 3).

  Specifically, as shown in FIG. 5 (b), a cosmetic containing a predetermined base property value is applied to the skin surface of the user's skin or the like, and the result is divided into a plurality of classified users (FIG. 5 (b)-(1)), shallow sensibility and deep sensibility are hierarchically estimated (FIG. 5 (b)-(2)), and sensitivity estimation is performed from the result.

  In this embodiment, in FIGS. 5A and 5B described above, (1) is a user classification (preference characteristic pattern extraction) method by preference characteristic, (2) is a Kansei hierarchical structuring method, ( In 3), as shown in FIG. 4 described above, the above-described neural network is used for the method of estimating the sensitivity from the physical property value.

<Preference Character Pattern Extraction Method in Preference Character Pattern Extraction Unit 14>
Next, a preference characteristic pattern extraction method in the above-described preference characteristic pattern extraction unit 14 will be described. In this embodiment, a user who is a general consumer can understand a slight difference in the feeling of use, a person who does not feel a slight difference in the feeling of use, a person who prefers moistness to the feeling of use, a person who likes refreshment, There are individual differences in preference and sensitivity to sensitivity, such as those who need high moisture and those who need a refreshing feeling, and group classification according to individual differences (preference characteristics) of sensitivity to improve analysis accuracy To do.

  Specifically, group classification is performed for each preference characteristic pattern preset by the preference characteristic pattern extraction unit 14. For example, when classifying according to the hierarchical structure of preset sensibilities, users who feel “moist” and “thorough” as “shallow sensibility” and “penetration sensation” as “deep sensibility” in one group Classification as follows. Further, for example, when classifying by sensitivity pattern that contributes to the overall preference, the user who feels “moist” and “thorough” as “shallow sensibility” and “whether or not?” As the overall preference is 1 Classification as one group.

  As described above, in this embodiment, for example, users are classified according to the hierarchical structure of “shallow sensibility” and “deep sensibility”, or users are classified according to the relevance between the sensitivity evaluation and the comprehensive evaluation, thereby extracting the preference characteristic pattern. To do. Since the sensitivity is non-linear, this embodiment uses a technique using a hierarchical neural network, for example.

<Neural Network Learning Method for Preference Character Pattern Extraction>
Here, the learning method of the neural network used for the preference characteristic pattern extraction of the present embodiment will be specifically described.

  Normally, the weight parameter is adjusted so that (the square sum of learning errors) is minimized during learning of the neural network. Note that the learning error means, for example, a difference between an estimated value by a neural network and a teacher value (teacher data).

  In the conventional method described above, there is a possibility that the weight parameter is adjusted so as to minimize the sum of squares of the learning error by using even the input node that does not affect the output node. As a result, for example, if the weight parameter is expressed by a graphical network structure, the network structure becomes complicated when the magnitude of the value is expressed by the line thickness, and it is difficult to interpret the influence of each input node on the output node. Become. That is, the conventional method may not be suitable for the purpose of finding the correlation between the input node and the output node.

  On the other hand, by using the skeleton learning method in the present embodiment, the network can be adjusted by adjusting the weight parameter so that “(the sum of squares of learning error) + (complexity of network structure)” is minimized during learning. The structure can be simplified.

  By using the skeleton learning method, it is possible to express with a simple network structure, and it becomes easy to understand the direction of the influence of each input node on the output node and the input node having a large influence on the output node. This results in high learning accuracy of learning data that matches the correlation pattern even in the case of a simple network structure. Here, the data with low learning accuracy is regarded as data that should be expressed in a pattern other than the correlation pattern, and the learning data is matched with the pattern with the set learning threshold as a boundary (data group). Can be divided into data groups.

<Classification method into multiple data groups using skeleton learning method>
Here, an example of a classification method into a plurality of data groups using the skeleton learning method will be described. As the classification method, for example, an evaluation function (Σ (Xi−Xi ^* ) ² , Xi is a teacher value, Xi ^* is an estimated value of the neural network) of the sum of deviations in the neural network, and a joint dispersion degree (C · ΣW _jk , C is a joint distribution weighting factor, W _jk is a skeleton learning evaluation function E = Σ (Xi−Xi ^* ) ^Σ + C · ΣW _jk with the addition of a weight parameter value from node j to node k) For a plurality of input nodes and a large number of data derived from the factors and having a result of being a teacher value, the coupling variance weighting coefficient C in the evaluation function E is “0 <C <1”. The range value is set and the skeleton learning is performed.

  Next, it sorts into the data group which has the said teacher value which falls within a setting error range with respect to the said estimated value obtained in the learning process mentioned above, and the data group of the said teacher value which remove | deviates from the said setting error range. Next, with respect to the teacher value data group that is out of the set error range sorted in this sorting step, the value of the range of “0 <C <1” is again set as the joint dispersion weight coefficient C in the evaluation function E. And repeating the process of performing the skeleton learning, the data is classified into a plurality of data groups having a common pattern with the result of the factor.

  For the data group having the teacher value outside the setting error range classified in the sorting step described above, the coupling dispersion weight coefficient C in the evaluation function E is again in the range of “0 <C <1”. The learning step of setting the value and performing the skeleton learning may be repeated until all the teacher values are within the set error range in the learning step.

In other words, the above-described method uses an evaluation function E = Σ (Xi−Xi ^* ) ² + C · ΣW _jk, which is obtained by adding a degree of joint dispersion to the evaluation function of the sum of deviations in the conventional neural network, On the other hand, skeleton learning is performed by setting a value in a range of “0 <C <1”. In other words, in the skeleton learning, an optimization process is performed with an emphasis on making the joint dispersion degree smaller than the sum of deviations. As a result, the weak connection between the nodes in the neural network can be eliminated and only a stronger connection can be left.

  On the other hand, according to the skeleton learning described above, since it is important to make the neural network structure simple by reducing the degree of joint dispersion, the deviation sum becomes a large value. For this reason, there are many data that are estimated values having a large difference from the teacher value (that is, low accuracy).

  Therefore, in each learning, a data group having a teacher value that falls within the setting error range with respect to the obtained estimated value is classified into a data group having a teacher value that falls outside the setting error range, and further from the setting error range. By repeating the skeleton learning again for the data group with the teacher value deviating and classifying the data group into a plurality of data, each of the classified data groups is based on the connection state between the lost nodes. Relationship patterns between specific input nodes (factors) and teacher values (results) included in the data group can be acquired, and evaluation can be performed by classifying into a plurality of groups having a common pattern.

  Here, not all of the monitors that responded to the questionnaire have the same preference pattern. Therefore, by applying the above-described method to the questionnaire analysis, a plurality of preference patterns are extracted, and the consumer's Kansei analysis can be performed. Thereby, for example, in the product development process, the analysis becomes more precise and the estimation accuracy can be improved as compared with the conventional method which has been treated as one consumer preference so far.

  Here, FIG. 6 is a diagram for explaining an association method using the hierarchical neural network in the present embodiment. FIG. 7 is a diagram for specifically explaining classification according to the present embodiment.

  6A shows 18 items (sensitivity items) immediately after application as input nodes, and FIG. 6B shows a comprehensive evaluation by a neural network using the input nodes shown in FIG. 6A. It shows until it is done. That is, FIG. 6B shows the relationship between the sensitivity items at the time of application and the comprehensive evaluation results.

  As shown in FIG. 6A, as the sensitivity items at the time of application, for example, “(1) I like the color of lotion: visual”, “(2) I like the smell of lotion: smell”, “(3 ) Strong scent of lotion: olfaction "," (4) Lotion is cool "," (5) Lotion is thick "," (6) Lotion is mellow "," (7) “There is a lot of lotion”, “(8) Moisture lotion”, “(9) Moisture lotion”, “(9) Moisture lotion”, “(10) Moisture lotion”, “(11) Lotion softens”, “(12) Lotion spreads well on skin”, “(13) Skin penetrates well”, “(14) Skin blends well”, “(15) Skin "(16) Lotion is refreshing", "(17) Lotion is moist", "(18) Lotion is sticky" It is a V "and the like. In the present invention, the contents of the 18 items described above are not limited to this, and the number of items is not limited to this.

  Further, as shown in FIG. 6B, as a learning operation using the neural network in the present embodiment, first, the test result is learned by the neural network (FIG. 6B- (1)), and then The learning data is traced (FIG. 6 (b)-(2)). Here, for example, the case data is separated when the estimation error is a predetermined value (for example, ± 5%) (FIG. 6 (b)-(3)). Next, learning is performed by performing the processing of FIGS. 6B to 6A described above using each case data after separation.

  Further, for example, as shown in FIG. 7A, among the data cases of the learning data obtained by the processing as described above, when there are three learning data cases of preferences A to C, the largest preference A Extract cases. Further, grouping is performed depending on whether the learning data is a “pattern case” or “pattern case” that is set in advance.

Note that the above-mentioned pattern, for example, the _{X n} as a measurement value, _{"Y = a 1 * X 1 +} a 2 * X 2 + a 3 * X 3 + ··· a n * X n '(coefficient: _{a n} is It can be expressed that the relationship can be expressed by the same value. More specifically, in the above-described function “Y = a ₁ * X ₁ + a ₂ * X ₂ + a ₃ * X ₃ +... A _n * X _n (X _i is an input value, Y is an output value)” Even if the form of the function is the same (the structure of the neural network is the same), if the combination of the coefficients a _i is different, the output value Y will be different even if the same set of X _i is given. That is, the same combination of coefficients a _i is treated as the same pattern, and the different combinations are treated as different patterns.

  As shown in FIG. 7B, the above-described classification is performed with respect to all panels with respective patterns (for example, pattern A, pattern B, pattern C, etc.), thereby performing classification (for example, Group1). , Group 2 etc.). As described above, by classifying users, it is possible to extract individual differences in the sensitivity evaluation of the input that affects the overall evaluation. Note that the contents in this embodiment as shown in FIG. 6 and the like are generated by the screen generation means 18 and the like and displayed by the output means 12 and the like.

  Here, FIG. 8 is a diagram illustrating an example of the classified result. FIG. 9 is a table showing the classification results shown in FIG. In the example shown in FIG. 8, the result (Group_1 to 12) obtained by classifying a plurality of users into 12 groups is shown. FIG. 8 shows the number of cases (number of questionnaires) of the learning data at the head and the branch points (Branch_1 to 15) on each route, and the groups at the finally grouped points (Group_1 to 12). The number of cases of the corresponding data (number of questionnaires) is shown (total 1850).

  In addition, FIG. 9 shows the number of case data and the ratio thereof with respect to the group numbers grouped in correspondence with FIG. Note that the classified results in the present embodiment as shown in FIGS. 8 and 9 are generated as a screen by the screen generator 18 and displayed by the output unit 12 and the like.

  In the present embodiment, by performing the above-described processing, for example, as shown in FIG. 9, a total of 1842 learning cases can be classified into 12 groups according to preference characteristics. Note that “out of group” 8 in FIG. 9 means that there are 8 cases that do not belong to any group.

  In the present embodiment, as shown in FIG. 9, grouping that occupies approximately 80% of all cases with five groups of group numbers 1, 6, 9, 11, and 12 can be performed, and other preset ratios are possible. A total of eight groups can be performed except for groups of less than (less than 2% in the example of FIG. 9).

<Hierarchical structure estimation method in the hierarchical structure estimation means 15>
Next, the hierarchical structure estimation method in the hierarchical structure estimation means 15 described above will be specifically described. FIG. 10 is a diagram for explaining the hierarchical structure estimation method in the present embodiment.

  In the present embodiment, there is an individual difference in how to feel one “deep sensibility”, and it is necessary to clarify “shallow sensibility” and “deep sensibility”. Therefore, as described above, the hierarchical structure of sensibilities according to user preference characteristics And clarify the relationship between “shallow sensibility” and “deep sensibility”.

  Specifically, in order to elucidate the hierarchical structure of sensibilities, a model that predicts “penetration sensation” from “shallow sensibility” described above as an example is examined. As an example, in this embodiment, a sensitivity evaluation having a high influence on “penetration feeling” is extracted. Specifically, for example, in the case of a hierarchical structure, when extracting “penetration” that is deep sensitivity from “moist” and “thoroughness” of “shallow sensibility”, “freshness” and “ Sensitivity evaluation is increased when extracting “penetration”, which is a deep sensibility.

  That is, as shown in FIG. 10, the neural network is used to input sensitivity evaluation points other than the penetrating sensation and extract the Kansei evaluation points related to the penetrating sensation. That is, a sensitivity evaluation having a high influence on “penetration” is extracted. For example, a weight parameter analysis method or a sensitivity analysis method can be used as an influence analysis method in sensitivity evaluation.

Here, the weight parameter analysis method, for example, usually has a plurality of routes (combinations of arrows between nodes) to reach the output node starting from each input node, but the route between the nodes. The weight parameter value (P _ij ) is set on (arrow), the weight parameter value obtained as a result of learning is multiplied by the weight parameter value on all arrows from the input node to the output node, and the value This is an analysis method for comparing the magnitude of the influence of each input node on the output node according to the size of.

  The sensitivity analysis method is a method that changes the value of a specific input node among multiple input nodes from the minimum value to the maximum value (the values of other input nodes are fixed), and how much the value of the output node changes. , And by examining how it changes, the sensitivity of the input node value to the output node is analyzed.

  Here, FIG. 11 is a diagram illustrating an example of the analysis result. In the example of FIG. 11, the influence of the input node on the output node is analyzed by weight parameter analysis and sensitivity analysis from the neural network obtained by the skeleton learning method for Group 1 described above.

  FIG. 11 shows the top five evaluation results in each of the above-described methods. In the weight parameter analysis, the first place is “Thorough”, the second place is “Chilly”, and the third place is “Color”. Like, “No. 4” is “mellow”, No. 5 is “moist”, and in sensitivity analysis, No. 1 is “I like fragrance”, No. 2 is “fresh”, No. 3 is “sticky” The fourth place is “moist” and the fifth place is “cool”.

  In addition, among the top five evaluation results in each method shown in FIG. 11, the overall evaluation of the portion indicated by bold letters decreases as the sensitivity evaluation score increases.

  In this case, for example, the analysis is performed in the top two groups having a large number of people classified according to the preference characteristic for the comprehensive evaluation among the 12 groups classified as shown in FIG. FIG. 12 shows an example of a sensitivity item that affects the penetration feeling in the top two groups. In FIG. 12, the frequency for each sensitivity evaluation item set in advance for the two groups of Group 1 and Group 6 using the weight parameter analysis is shown. In addition, in the frequency | count with respect to each sensitivity evaluation item shown in FIG. 12, plus shows a preferable numerical value, minus shows a numerical value which is not preferable, and 0 has shown no influence.

  As shown in FIG. 12, the sensitivity evaluation items that affect the penetrating sensation are different depending on the group. More specifically, for example, in Group 1, the permeation feeling is determined based on whether or not it “soaks well into the skin” or “the lotion is not refreshed”. On the other hand, in Group 6, it can be seen that the permeation feeling is determined by whether or not “the lotion is not rich”, “the lotion is not cool”, “well fits well with the skin” or the like.

  That is, in the present embodiment, the feature difference of each group can be shown from the result obtained by the above-described weight parameter analysis or the like. Therefore, as shown in FIG. 12, it is possible to grasp the preference characteristics of each group by analyzing the influence degree for each classified group. In the present embodiment, this difference in preference characteristics can be applied to group classification.

<Sensitivity estimation method in sensitivity estimation means 16>
Next, the sensitivity estimation method in the sensitivity estimation means 16 in this embodiment will be specifically described. In the sensitivity estimation means 16, the sensitivity is non-linear in order to find the relationship (correlation) between the “base property value” and the “shallow sensitivity”, for example, and in this embodiment, the sensitivity estimation is performed using a neural network. .

  FIG. 13 is a diagram for explaining the correlation between the physical properties of the base and the shallow sensitivity. As shown in FIG. 13, in this embodiment, “(average value after finishing) rolling friction” and “(peak value at the time of drying) rolling for each classified group (for example, Group 1 in FIG. 13). Bases such as “friction”, “(average value after finishing) friction coefficient”, “(peak value upon drying) friction coefficient”, “(average value immediately after application) friction coefficient”,... Using the neural network, the relationship of “freshness”, which is a shallow sensibility, is extracted from the physical property values.

  Here, FIG. 14 is a diagram illustrating an example of the relationship between the base property value and the sensitivity evaluation value. The example of FIG. 14 shows an example in which the relationship between the base material property value and the sensitivity evaluation value is learned by using the input node as the base material property value and the output node as the sensitivity evaluation value. Further, the structure of FIG. 14A shows an example of the preference characteristic Group1, and the structure of FIG. 14B shows an example of the preference characteristic Group6.

  14A and 14B, A is the viscosity (mPa · s), B is the surface tension (mN / m), C is the coefficient of friction (C1: average value immediately after application, C2: Peak value upon drying, C3: average value after finishing), D indicates rolling friction (D1: peak value upon drying, D2: average value after finishing).

  14A and 14B, the thickness of the route (arrow) between the nodes indicates the magnitude of the weight parameter. Further, the solid line indicates a positive weight parameter value, and the dotted line indicates a negative weight parameter value. With the graphic display as shown in FIG. 14, it is possible to visually grasp which input node item has an influence on the output node item (how it acts on the plus side / minus side). Note that such a screen is generated by the screen generation means 18 described above and output by the output means 12.

  Thereby, the relationship between a base physical property and shallow sensitivity can be estimated, and a user's sensitivity can be estimated with high precision and efficiency from base property information using these estimation results.

<Screen example>
Here, an example of a screen generated by the screen generation unit 18 in the present embodiment will be described with reference to the drawings. FIG. 15 is a diagram illustrating an example of a generation screen in the present embodiment. In the screen generation unit 18 in the present embodiment, for example, at least of the content extracted by the preference characteristic pattern extraction unit 14, the estimation result obtained by the hierarchical structure estimation unit 15, and the sensitivity estimation result obtained by the sensitivity estimation unit 16. A screen for displaying one on the display means is generated.

  The screen 40 shown in FIG. 15 has a first input area 41-1 for inputting the components A to X and a second input area 41-2 for inputting the predicted physical properties of the physical properties A to X. By inputting data such as a predetermined numerical value for the sensitivity evaluation target for performing sensitivity estimation to each input area and executing the above-described sensitivity estimation processing, for example, the sensitivity estimation results of sensitivities A to N on the screen 40, Evaluation prediction results and the like can be displayed as numerical values, characters, and the like. Note that the screen 40 shown in FIG. 15 is an example, and the layout and items generated on the screen are not limited to this in the present invention.

  Further, the screen generated by the screen generation means 18 is not limited to this in the present invention. For example, the above-described screens of FIGS. Or the like.

  As described above, according to the present embodiment, the user's sensitivity can be estimated with high accuracy and efficiency from the base material property information. This makes it possible to estimate the sensibility expected of users who have relied on the experience and intuition of researchers at the time of product development. Moreover, since it is possible to estimate the sensibility without having to conduct a consumer questionnaire, the time for product development can be shortened. Further, since the users can be classified according to the preference characteristics, the target target taste characteristics can be grasped with high accuracy. Furthermore, it is possible to design the usability of a cosmetic that satisfies “deep sensibility” that has been difficult to design, using the physical properties of the base as an index.

  The sensitivity evaluation target applicable to the invention of the present application is not limited to the above-described cosmetics and pharmaceuticals, but includes, for example, five human senses (tactile sense, smell sense, taste sense, visual sense, hearing sense) such as food and drink, daily necessities, and clothing items. This includes a wide range of products that can be evaluated for sensibilities that can be obtained through In this case, the neural network described above selects the learning data according to each sensitivity evaluation object and performs sensitivity estimation.

  As described above, according to the present invention, the sensitivity of the user can be estimated with high accuracy and efficiency from the base material property information. Thereby, for example, it is possible to accurately and efficiently estimate the sensibility expected of a user who has relied on the experience and intuition of a researcher at the time of product development based on the present invention. In addition, for example, usability can be estimated without conducting a consumer questionnaire, so that product development time can be reduced. Further, by classifying users according to preference characteristics, it is possible to grasp the preference characteristics of the target layer to be obtained with high accuracy and efficiency. Furthermore, the usability of the sensitivity evaluation object that satisfies the “deep sensitivity” that has been difficult to design can be designed using the base physical properties as an index.

  The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to such specific embodiments, and various modifications, within the scope of the gist of the present invention described in the claims, It can be changed.

DESCRIPTION OF SYMBOLS 10 Kansei estimation apparatus 11 Input means 12 Output means 13 Storage means 14 Preference characteristic pattern extraction means 15 Hierarchical structure estimation means 16 Kansei estimation means 17 Kansei evaluation means 18 Screen generation means 19 Transmission / reception means 20 Control means 31 Input apparatus 32 Output apparatus 33 Drive Device 34 Auxiliary storage device 35 Memory device 36 CPU
37 Network connection device 38 Recording medium 40 Screen 41 Input area

Claims (9)

In the sensitivity estimation device that estimates the sensitivity during and after use of the sensitivity evaluation object used by the user,
Kansei estimation apparatus comprising preference characteristic pattern extraction means for classifying each preference characteristic pattern for obtaining a predetermined comprehensive evaluation for a plurality of sensitivity items using a skeleton learning method of a preset hierarchical neural network . Input means for inputting base material property information for the sensitivity evaluation object;
Hierarchical structure estimating means for hierarchizing the sensitivity of the user to the sensitivity evaluation object using the plurality of sensitivity items;
Based on the base physical property information obtained by the input means and the hierarchical structure obtained by the hierarchical structure estimating means, the sensitivity to the sensitivity evaluation object for the classification obtained by the preference characteristic pattern extracting means is estimated. The sensitivity estimation device according to claim 1, further comprising sensitivity estimation means.   The sensitivity estimation device according to claim 2, further comprising sensitivity evaluation means for evaluating sensitivity by matching a sensitivity estimation result obtained by the sensitivity estimation means with a preset evaluation index.   Generates a screen for displaying at least one of the content extracted by the preference characteristic pattern extraction unit, the estimation result obtained by the hierarchical structure estimation unit, and the sensitivity estimation result obtained by the sensitivity estimation unit on the display unit The sensitivity estimation apparatus according to claim 2, further comprising a screen generation unit configured to perform the screen generation. In the sensitivity estimation method for estimating the sensitivity during and after use of the sensitivity evaluation object used by the user,
A sensitivity estimation method comprising a preference characteristic pattern extraction step for performing classification for each preference characteristic pattern to obtain a predetermined comprehensive evaluation for a plurality of sensitivity items using a skeleton learning method of a preset hierarchical neural network . An input step of inputting base material property information for the sensitivity evaluation object;
A hierarchical structure estimation step of hierarchizing the sensitivity of the user to the sensitivity evaluation object using the plurality of sensitivity items;
Based on the base physical property information obtained by the input means and the hierarchical structure obtained by the hierarchical structure estimation step, the sensitivity to the sensitivity evaluation object for the classification obtained by the preference characteristic pattern extraction step is estimated. The sensitivity estimation method according to claim 5, further comprising a sensitivity estimation step.   The sensitivity estimation method according to claim 6, further comprising a sensitivity evaluation step of evaluating sensitivity by associating a sensitivity estimation result obtained in the sensitivity estimation step with a preset evaluation index.   Generate a screen for displaying at least one of the content extracted in the preference characteristic pattern extraction step, the estimation result obtained in the hierarchical structure estimation step, and the sensitivity estimation result obtained in the sensitivity estimation step on the display means The method for estimating a sensibility according to claim 6, further comprising a step of generating a screen.   A sensitivity estimation program for causing a computer to function as the sensitivity estimation device according to any one of claims 1 to 4.

Images