JP2015204084A - Apparatus, method and program for generating rule using data mining, and care support system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a data analysis apparatus configured to generate a rule or acquire knowledge by repeating: selecting two parent rules from a collation set extracted from a set of rules including a conclusion section which may have a continuous value; creating two child rules from the two parent rules by use of an AT value; and adding the created child rules to the rule set, a data analysis method, a care support system, or knowledge acquisition method thereof.SOLUTION: A rule generation apparatus includes: storage means for storing a rule set and a collation set; collation set generation means which generates the collation set from the rule set; and GA application means which updates rules constituting the rule set, on the basis of the generated collation set, by use of genetic algorithm. The rule generation apparatus repeats modifying data to be selected by the collation set generation means from multiple pieces of data, and adding two child rules created by the GA application means to the rule set, to update rules constituting the rule set.

Description

  The present invention relates to a technique for generating a rule or acquiring knowledge by analyzing a plurality of data.

  There is known a data mining technique for generating rules that match the contents of data from a plurality of data by data analysis using a statistical method or the like.

  In addition, with the informatization of society, data generated in various scenes is accumulated, and a method for acquiring knowledge (rules) from a huge amount of accumulated data is required. For example, as a care support method in a care home, there is a need to create a care plan (knowledge) for care support based on a plurality of data representing the daily behavior or state of a care recipient.

  As such a rule acquisition technique, a learning classifier system (LCS) has been proposed. This LCS is an adaptation system that incorporates the concepts of both learning and evolution, which are adaptation strategies to living environments. Specifically, the LCS learns a rule (classifier) composed of a condition part and a conclusion part (action part) by a reinforcement learning method and uses a genetic algorithm (GA) to make a rule. It is a system that generates rules (classifiers) that can be adapted to a plurality of data (environmental states) by evolving (classifiers). Of these LCSs, a system that makes a highly accurate rule (classifier) an evolution target is called an XCS (Accuracy-based Learning System). (For example, Non-Patent Documents 1 and 2).

  In addition, an LCS in which a technique called Attribute Treacking / Feedback is introduced has been proposed (Non-patent Document 2). In the conventional LCS, one data includes a condition data element corresponding to the rule condition part and a conclusion data element corresponding to the rule conclusion part. Then, calculate the evaluation value indicating the importance of each conditional data element in the individual data, that is, how much the conditional data element contributes to the accuracy of the rule that matches the data. (Attribute Tracking) Using the calculated evaluation value, a rule having higher compatibility is generated in crossover or mutation in GA (Attribute Feedback).

On the other hand, from the viewpoint of the content of the rule to be acquired, in order to acquire the exception rule (special rule), the case data (data set) that is the target of the exception rule using the condition part of the general rule (generalized rule) A technique for separating and acquiring exception rules based on separated case data has been proposed (Patent Document 1).

JP 11-338701 A

4 Nakata, “Promotion of generalization of learning classifier system by individualization”, Society of Instrument and Control Engineers, Transactions of Society of Instrument and Control Engineers, 47 (11), 581-590, November 30, 2011 Two people from Urbanowicz, “Instance-Linked Attribute Tracking and Feedback for Michigan-Style Supervised Learning Classifier Systems”, Association for Computing Machinery, GECCO '12 Proceedings of the fourteenth international conference on Genetic and evolutionary computation conference, Pages 927-934, 2012 July 7,

  In the technique disclosed in the above document, the data element for conclusion is not binary (for example, whether or not a certain illness has been obtained) but data indicating a continuous degree such as some degree or score (for example, sleep It is not assumed that a rule is generated from data indicating the degree of depth.

  In addition, in the technologies disclosed in Non-Patent Documents 1 and 2, knowledge may not be acquired (a rule may be generated) when there is conflicting data or special date data such as a birthday. .

  Furthermore, with the technique disclosed in Patent Document 1, there are cases where the generalization rule and the special rule cannot be generated (acquired) at the same time. In the technique disclosed in Patent Document 1, for example, a special rule may not be generated using a plurality of data sets that constitute a generalized rule.

  The present invention has been made under such circumstances, and it is a first object of the present invention to provide a technique for appropriately generating a rule even if the conclusion data element is data indicating a continuous level. And

  A second object of the present invention is to provide a technique that enables simultaneous generation of general-purpose rules and special rules.

  Furthermore, a third object of the present invention is to provide a technique that makes it possible to simultaneously generate a plurality of rules having different values in the conclusion part.

The rule generation device according to the first aspect of the present invention includes:
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A data storage unit for storing a plurality of data including at least a conclusion data data element that can take one of the result values composed of continuous values representing a predetermined result under a condition to be performed;
One or more condition rule elements representing rule application conditions corresponding to the condition data element, wherein each of the condition rule elements is any of the two or more types of state values and the condition value conditions It represents one of the wildcard values indicating goodness, a rule element for conclusion representing the conclusion when the condition is satisfied as a continuous value, and a characteristic evaluation value of the rule including the accuracy of the rule A rule set storage unit for storing a set of rules including at least a rule element for characteristic evaluation;
A rule extraction unit that reads out the data from the data storage unit and extracts from the rule set storage unit the rule having the value of the condition rule element to be compared with the state value of the condition data element of the read data When,
A rule strengthening unit that updates the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so that the compatibility with the read data is higher;
Among the extracted rules, a rule classification unit that classifies a rule that satisfies a predetermined condition in which a continuous value of the rule element for conclusion is applicable to a plurality of values;
A feature extraction model generated for each piece of data such that a rule classified by the rule classification unit exists, including a model element corresponding to the conditional data element of the data, each of the model elements A feature extraction model storage unit that stores a model having a feature evaluation value that represents how much the corresponding conditional data element contributes to the accuracy of the rule;
Regarding the feature extraction model corresponding to the data read by the rule extraction unit, the condition rule elements of the rule classified by the rule classification unit having the state value are used for the characteristic evaluation. The model element is updated so that the characteristic evaluation value of the corresponding model element becomes higher as the evaluation value representing the accuracy of the rule included in the rule element is larger. For the wild card value, the corresponding model is updated. A feature extraction model update unit that updates the model element so that the feature evaluation value of the element is low;
A rule selection unit that selects two rules as parent rules from the classified rules in a given method;
As shown in the following (1) and (2), a rule crossing unit that generates two child rules by exchanging values of the condition rule elements corresponding to the selected two parent rules; ,
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and a rule addition unit that stores by adding child rules to perform the replacement when necessary in the rule set storage unit,
A series of processes by the rule extraction unit, the rule strengthening unit, the rule classification unit, the feature extraction model update unit, the rule selection unit, the rule crossing unit, and the rule addition unit are stored in the data storage unit. This is repeated for a plurality of data.

Moreover, the rule generation method according to the first aspect of the present invention includes:
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A rule application corresponding to the condition data element from a plurality of data, including at least a conclusion data data element that can take one of the result values consisting of continuous values representing a predetermined result under a given condition One or more condition rule elements representing the condition of the condition, wherein each of the condition rule elements is any one of the two or more types of state values and the wild card value representing the condition of the condition value At least a rule element for conclusion that represents a conclusion when the condition is satisfied as a continuous value, and a rule element for characteristic evaluation that represents a characteristic evaluation value of the rule including the accuracy of the rule A rule generation method for generating Lumpur,
Extracting the rule having the value of the condition rule element to be compared with the state value of the condition data element of one of the plurality of data from the rule set storage unit that stores the set of rules And steps to
Updating the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so as to be more compatible with the read data;
Among the extracted rules, a step of classifying a rule that satisfies a predetermined condition in which a continuous value of the conclusion rule element can correspond to a plurality of values;
A feature extraction model generated for each of the data such that the classified rule exists, including a model element corresponding to the conditional data element of the data, each of the model elements corresponding to the corresponding data element Among the models having feature evaluation values representing how much the conditional data element contributes to the accuracy of the rule, the feature extraction model corresponding to the one data is included in the classified rule. Among the condition rule elements, those having the state value have a higher feature evaluation value of the corresponding model element as the evaluation value representing the accuracy of the rule included in the characteristic evaluation rule element is larger. The model element is updated as described above, and for the wild card value, the model element is updated so that the feature evaluation value of the corresponding model element is lowered. And updating the,
Selecting two rules to be parent rules from the classified rules in a given way;
As shown in the following (1) and (2), generating two child rules by replacing values of the condition rule elements corresponding to the two selected parent rules;
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and a step of storing by adding child rules for performing said replacement as required in the rule set storage unit,
A series of processes in each step is repeatedly performed for a plurality of data among the plurality of data.

  A rule generation program according to the first aspect of the present invention causes a computer to execute the rule generation method.

The rule generation device according to the second aspect of the present invention provides:
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A data storage unit that stores a plurality of data including at least a conclusion data data element that can take a result value representing a predetermined result under a state of being performed;
One or more condition rule elements representing rule application conditions corresponding to the condition data element, wherein each of the condition rule elements is any of the two or more types of state values and the condition value conditions Represents one of the wildcard values representing good, a rule element for conclusion representing a conclusion when the condition is satisfied, and a characteristic evaluation value of the rule including accuracy and versatility of the rule A rule set storage unit for storing a set of rules including at least a rule element for characteristic evaluation;
A rule extraction unit that reads out the data from the data storage unit and extracts from the rule set storage unit the rule having the value of the condition rule element to be compared with the state value of the condition data element of the read data When,
A rule strengthening unit that updates the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so that the compatibility with the read data is higher;
Among the extracted rules, a rule classification unit that classifies a rule in which the value of the rule element for conclusion satisfies a predetermined condition;
A feature extraction model generated for each piece of data such that a rule classified by the rule classification unit exists, including a model element corresponding to the conditional data element of the data, each of the model elements A feature extraction model storage unit that stores a model having a feature evaluation value that represents how much the corresponding conditional data element contributes to the accuracy of the rule;
Regarding the feature extraction model corresponding to the data read by the rule extraction unit, the condition rule elements of the rule classified by the rule classification unit having the state value are used for the characteristic evaluation. The model element is updated so that the characteristic evaluation value of the corresponding model element becomes higher as the evaluation value representing the accuracy of the rule included in the rule element is larger. For the wild card value, the corresponding model is updated. A feature extraction model update unit that updates the model element so that the feature evaluation value of the element is low;
A rule selection unit that selects two rules as parent rules from the classified rules in a given method;
As shown in the following (1) and (2), a rule crossing unit that generates two child rules by exchanging values of the condition rule elements corresponding to the selected two parent rules; ,
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and versatility evaluation value update unit that performs a process of changing the evaluation value representing the versatility of the rules to perform the replacement when necessary with a predetermined probability,
A rule addition unit for adding and storing the child rules in the rule set storage unit;
A series of processes by the rule extraction unit, the rule strengthening unit, the rule classification unit, the feature extraction model update unit, the rule selection unit, the rule crossing unit, the versatility evaluation value update unit, and the rule addition unit Is repeatedly performed for a plurality of data stored in the data storage unit.

The rule generation method according to the second aspect of the present invention includes:
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value One or more data representing a rule application condition corresponding to the condition data element from a plurality of data, including at least a conclusion data data element that can take a result value representing a predetermined result A condition rule element, wherein each of the condition rule elements can take one of the two or more types of state values and a wildcard value representing any of the condition value conditions; A rule generation method for generating a rule including at least a rule element for conclusion that represents a conclusion when the condition is satisfied and a rule element for characteristic evaluation that represents a characteristic evaluation value of the rule including the accuracy of the rule There is,
Extracting the rule having the value of the condition rule element to be compared with the state value of the condition data element of one of the plurality of data from the rule set storage unit that stores the set of rules And steps to
Updating the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so as to be more compatible with the read data;
Among the extracted rules, classifying a rule in which a value of the rule element for conclusion satisfies a predetermined condition;
A feature extraction model generated for each of the data such that the classified rule exists, including a model element corresponding to the conditional data element of the data, each of the model elements corresponding to the corresponding data element Among the models having feature evaluation values representing how much the conditional data element contributes to the accuracy of the rule, the feature extraction model corresponding to the one data is included in the classified rule. Among the condition rule elements, those having the state value have a higher feature evaluation value of the corresponding model element as the evaluation value representing the accuracy of the rule included in the characteristic evaluation rule element is larger. The model element is updated as described above, and for the wild card value, the model element is updated so that the feature evaluation value of the corresponding model element is lowered. And updating the,
Selecting two rules to be parent rules from the classified rules in a given way;
As shown in the following (1) and (2), generating two child rules by replacing values of the condition rule elements corresponding to the two selected parent rules;
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and performing the process of changing the evaluation value representing the versatility of the rules to perform the replacement when necessary with a predetermined probability,
Adding and storing the child rules in the rule set storage unit,
A series of processes in each step is repeatedly performed for a plurality of data among the plurality of data.

  A rule generation program according to a second aspect of the present invention causes a computer to execute the rule generation method.

The rule generation device according to the third aspect of the present invention is:
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A data storage unit that stores a plurality of data including at least a conclusion data data element that can take a result value representing a predetermined result under a state of being performed;
One or more condition rule elements representing rule application conditions corresponding to the condition data element, wherein each of the condition rule elements is any of the two or more types of state values and the condition value conditions Represents one of the wildcard values representing good, a rule element for conclusion representing a conclusion when the condition is satisfied, and a characteristic evaluation value of the rule including accuracy and versatility of the rule A rule set storage unit for storing a set of rules including at least a rule element for characteristic evaluation;
A rule extraction unit that reads out the data from the data storage unit and extracts from the rule set storage unit the rule having the value of the condition rule element to be compared with the state value of the condition data element of the read data When,
A rule strengthening unit that updates the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so that the compatibility with the read data is higher;
Among the extracted rules, a rule classification unit that classifies a value of the conclusion rule element into a twelfth rule set that satisfies a predetermined condition and a second rule set that does not satisfy the predetermined condition;
For each of the first and second rule sets, a feature extraction model generated for each piece of data such that a rule exists in each rule set, corresponding to the condition data element of the data A feature extraction model storage unit that stores a model having a feature evaluation value representing how much the corresponding conditional data element contributes to the accuracy of the rule. When,
Regarding the feature extraction model corresponding to the data read by the rule extraction unit, the condition rule elements of the rule classified by the rule classification unit having the state value are used for the characteristic evaluation. The model element is updated so that the characteristic evaluation value of the corresponding model element becomes higher as the evaluation value representing the accuracy of the rule included in the rule element is larger. For the wild card value, the corresponding model is updated. A feature extraction model update unit that updates the model element so that the feature evaluation value of the element is low;
A rule selection unit that selects two rules as parent rules from the classified rules in a given method;
As shown in the following (1) and (2), a rule crossing unit that generates two child rules by exchanging values of the condition rule elements corresponding to the selected two parent rules; ,
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and a rule addition unit that stores by adding child rules to perform the replacement when necessary in the rule set storage unit,
A series of processes by the rule extraction unit, the rule strengthening unit, the rule classification unit, the feature extraction model update unit, the rule selection unit, the rule crossing unit, and the rule addition unit are stored in the data storage unit. This is repeated for a plurality of data.

The rule generation method according to the third aspect of the present invention includes:
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value One or more data representing a rule application condition corresponding to the condition data element from a plurality of data, including at least a conclusion data data element that can take a result value representing a predetermined result A condition rule element, wherein each of the condition rule elements can take one of the two or more types of state values and a wildcard value representing any of the condition value conditions; A rule generation method for generating a rule including at least a rule element for conclusion that represents a conclusion when the condition is satisfied and a rule element for characteristic evaluation that represents a characteristic evaluation value of the rule including the accuracy of the rule There is,
Extracting the rule having the value of the condition rule element to be compared with the state value of the condition data element of one of the plurality of data from the rule set storage unit that stores the set of rules And steps to
Updating the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so as to be more compatible with the read data;
Classifying the extracted rules into a twelfth rule set in which the value of the rule element for conclusion satisfies a predetermined condition and a second rule set that does not satisfy the predetermined condition;
For each of the first and second rule sets, a feature extraction model generated for each piece of data such that the classified rule exists, and a model element corresponding to the conditional data element of the data Each of the model elements corresponds to the one piece of data among models having feature evaluation values representing how much the corresponding conditional data element contributes to the accuracy of the rule. In the feature extraction model, the condition rule elements of the classified rules having the state value have a large evaluation value indicating the accuracy of the rule included in the characteristic evaluation rule element The model element is updated so that the feature evaluation value of the corresponding model element becomes higher, and for the wild card value, the corresponding model is updated. A step of characterization value of the unit to update the model elements to be lower,
Selecting two rules to be parent rules from the classified rules in a given way;
As shown in the following (1) and (2), generating two child rules by replacing values of the condition rule elements corresponding to the two selected parent rules;
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and performing the process of changing the evaluation value representing the versatility of the rules to perform the replacement when necessary with a predetermined probability,
Adding and storing the child rules in the rule set storage unit,
A series of processes in each step is repeatedly performed for a plurality of data among the plurality of data.

A rule generation program according to a third aspect of the present invention causes a computer to execute the rule generation method.

  According to the first aspect of the present invention, even if the conclusion data element is data indicating a continuous level, it is appropriate to classify a rule whose value can take any of a plurality of desired values. Can be generated.

  Further, according to the second aspect of the present invention, the general-purpose evaluation value is probabilistically updated, thereby generating a general-purpose rule and a special rule that are generated at the same time. Appropriate general-purpose rules and special rules can be generated simultaneously.

Furthermore, according to the third aspect of the present invention, the rule range elements for the conclusion are classified into two rule sets having different value ranges, and optimization is performed for each rule set. It becomes possible to generate a plurality of different rules at the same time.

It is explanatory drawing explaining an example of the data used by the rule production | generation apparatus which concerns on the 1st Embodiment of this invention, a rule, and a feature extraction model (AT). It is a schematic functional block diagram explaining an example of the rule production | generation apparatus which concerns on the 1st Embodiment of this invention. It is explanatory drawing explaining an example of the data set analyzed and the rule to produce | generate with the rule production | generation apparatus which concerns on the 1st Embodiment of this invention. It is explanatory drawing explaining an example of the rule produced | generated with the rule production | generation apparatus which concerns on the 1st Embodiment of this invention. It is explanatory drawing explaining an example of the update process of AT value in the rule production | generation apparatus which concerns on the 1st Embodiment of this invention. It is explanatory drawing explaining an example of the process (crossover) using the genetic algorithm in the rule production | generation apparatus which concerns on the 1st Embodiment of this invention. It is explanatory drawing explaining an example of the process (mutation | mutation) using the genetic algorithm in the rule production | generation apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart figure explaining an example (extension of continuous value reward) of a rule generation method concerning a 1st embodiment of the present invention. It is a flowchart figure explaining an example (genetic algorithm) of the rule production | generation method concerning the 1st Embodiment of this invention. It is a flowchart figure explaining the other example (generalization rule and special rule) of the rule production | generation method concerning the 1st Embodiment of this invention. It is a flowchart figure explaining the other example (genetic algorithm) of the rule production | generation method concerning the 1st Embodiment of this invention. It is a flowchart figure explaining the other example (handling of contradiction data) of the rule production | generation method concerning the 1st Embodiment of this invention. It is a flowchart figure explaining the other example (genetic algorithm) of the rule production | generation method concerning the 1st Embodiment of this invention. It is a schematic system diagram explaining an example of the care support system concerning a 2nd embodiment of the present invention. It is explanatory drawing explaining an example of the acquisition result (data set) of the data acquisition means of the care support system which concerns on the 2nd Embodiment of this invention. It is explanatory drawing explaining an example of the sleep stage (evaluation value) used with the care support system which concerns on the 2nd Embodiment of this invention. It is explanatory drawing explaining an example of the analysis result (The general knowledge of special sleep and insomnia, special knowledge) analyzed with the care support system concerning a 2nd embodiment of the present invention. It is explanatory drawing explaining the update method of a probability model in cGA. It is explanatory drawing explaining the update method of a probability model in DcGA. It is explanatory drawing explaining the expression by a suboptimal solution and a probability model in DcGA. It is explanatory drawing explaining filtering of a probability model in DcGA. It is explanatory drawing explaining the inclination of the heart rate of each test subject, and the example of grouping. It is explanatory drawing explaining the sleep stage estimation by the filter of DcGA (for 1 day). It is a schematic block diagram explaining an example of a sleep stage estimation apparatus.

  The present invention will be described using a rule generation device according to a non-limiting exemplary embodiment with reference to the accompanying drawings. The present invention generates a rule (knowledge, knowledge, etc.) based on a plurality of data other than the rule generation device described below, and values that can be taken by data elements corresponding to the conclusion part of the rule are continuous. Any rule generator (device, equipment, unit, system, etc.) can be used as long as it generates a rule from a target data set. Note that “the value that can be taken by the data element corresponding to the conclusion part of the rule is continuous” means that the value that the data element can take is binary, such as “normal” or “abnormal”. It means that it indicates a continuous level, such as the depth of sleep or some score.

In the following description, the same or corresponding devices, parts, or members described in all the attached drawings are denoted by the same or corresponding reference numerals, and redundant description is omitted. Also, the drawings are not intended to show a limited relationship between devices, parts or members unless specifically described. Accordingly, specific correlations can be determined by one skilled in the art in light of the following non-limiting embodiments.

The present invention will be described in the following order using a rule generation device according to an embodiment of the present invention.

1. Rule generation device (first embodiment)
2. Rule generation method (first embodiment)
3. Nursing care support system (second embodiment)
3-1. Configuration and operation of care support system 3-2 Example of sleep stage estimation method Program and recording medium recording the program


[1. Rule generation device (first embodiment)]
The present invention will be described using the rule generation device according to the first embodiment. In the following description, the rule generation device is adapted to a plurality of data by applying a genetic algorithm using a feature extraction model (hereinafter referred to as “AT”) by Attribute Tracking to a set of rules. A device that generates rules (acquires knowledge).

  As shown in FIG. 1A, data (DT) is a set of data elements composed of the first data element (or data item, attribute, etc.) d1 to the nth data element dn. In the present embodiment, the data is a condition data element for deriving a rule condition (hereinafter, referred to as “state value”) that is composed of a value (hereinafter referred to as “state value”) that indicates the characteristics of the state (activity, action, etc.) from which data is acquired. For example, it is composed of d1 to d4 in FIG. 1A and a conclusion data element (for example, d5 in FIG. 1A) for deriving the conclusion of the rule to be generated.

  The condition data element is composed of the first to (n-1) th data elements. That is, the target state from n−1 viewpoints is included as information. In this embodiment, n = 5. As shown in FIG. 1A, the conditional data element is a 4-bit bit string from data element d1 to data element d4. Here, the “state value” of each condition data element can take a binary value (0, 1) in the present embodiment. It should be noted that the “state value” that can be used in the present invention may be a multivalue representing three or more states.

  The conclusion data element is the nth data element. In the present embodiment, as shown in FIG. 1A, the data element d5. The value of the conclusion data element d5 can be a continuous value. Here, in the present embodiment, the conclusion data element d5 is an actual measurement value P (result value) indicating a state at the time of data acquisition.

  For example, in the data of FIG. 1A, when the condition d1 is “0”, d2 is “1”, d3 is “0”, and d4 is “1”, the actual measurement value P (d5) is “0.15”. It means that it was.

  Note that data (data elements) that can be used by the rule generation device is not limited to the data DT shown in FIG. In other words, the data of the present invention can be any number of conditional or conclusion data elements, possible values of each condition data element, etc., as long as the conclusion data element can take a continuous value. can do.

  As shown in FIG. 1B, a rule (RL) is a set of rule elements composed of the first rule element (or rule item, attribute, etc.) r1 to n + 3 rule element rn + 3. . In this embodiment, the rule includes a condition rule element corresponding to the condition in the rule (for example, r1 to r4 in FIG. 1B) and a rule element for conclusion corresponding to the conclusion in the rule (for example, FIG. 1B). r5) and characteristic evaluation rule elements (for example, r6 to r8 in FIG. 1B) representing rule characteristics such as rule accuracy and versatility (speciality).

  The condition rule elements are composed of the first to (n-1) th rule elements, and each rule element corresponds to the first to the (n-1) th condition data element of the data (DT). In this embodiment, n = 5. As shown in FIG. 1B, the condition rule elements are composed of rule elements r1 to r4. Here, each condition rule element indicates either the same value as the condition data element, that is, in the present embodiment, any of the two (0, 1) state values, or any of these two state values. Three values “#” (wildcard value) can be taken. That is, “#” used for the rule element for a condition indicates that it does not contribute to the property constituting the rule or the dominant characteristic of the rule. When the “state value” of the condition data element takes a multivalue representing a state of 3 or more, the condition rule element can take those multivalues or “#”.

  The conclusion rule element is the nth rule element and corresponds to the conclusion data element which is the nth data element of the data (DT). In this embodiment, as shown in FIG. 1B, it is the rule element r5 and can take a continuous value similarly to the conclusion data element d5. In this embodiment, an evaluation value p (r5) is used as the conclusion rule element.

  Here, the evaluation value p (r5) is a value corresponding to the actual measurement value P (data element d5) of the data DT, and the actual measurement value P under the condition satisfying each value set in the condition rule element. It is a predicted value.

  The rule element for characteristic evaluation includes n + 1 to n + 3 rule elements. In this embodiment, it is composed of three rule elements: an error value ε (r6), a fitness F (r7), and an allowable error (university evaluation value) ε0 (r8).

  Here, the error value ε (r6) is an evaluation value representing a difference between the actual measurement value P of the data DT and the evaluation value p (r5) used for updating (strengthening) the fitness F (r7). . The allowable error ε0 (r8) means the “width” of the value of the rule element for conclusion allowed in the rule. In other words, the rule is more versatile (less specific) as the value of the tolerance ε0 (r8) is larger, and the rule is more specific (lower versatility) as the value of the tolerance ε0 (r8) is smaller. is there. The conformity F (r7) indicates the degree of conformity or reliability (accuracy) of the rule. The fitness F (r7) is determined based on the error value ε (r8) and the allowable error ε0 (r6). Specifically, if the error value ε (r8) exceeds the allowable error ε0 (r6), the value of the fitness F (r7) is determined to be small.

  Note that the rules (rule elements) generated by the rule generation device are not limited to the rules RL shown in FIG. In other words, the rules of the present invention can have any configuration as long as they are composed of rule elements corresponding to condition data elements and conclusion data elements, and characteristic evaluation rule elements.

  As shown in FIG. 1C, the AT is a set of AT elements composed of the first AT element (or item, attribute, etc.) a1 to the (n-1) th AT element an-1. Each AT element in one AT corresponds to each of the condition data elements of the data DT, and further to each of the condition rule elements of the rule RL corresponding to each condition data element. Further, there are as many ATs as the number of data DTs. The value of each AT element represents how much the condition data element of the corresponding data DT contributes to the conformity F of the rule RL to be collated with the data. In the present embodiment, AT is a variable within a range from “0” (small contribution) to “1” (large contribution). The AT is calculated based on a rule constituting a collation set (block B04 in FIG. 2 in the present embodiment) described later. In the example of FIG. 1, the value “0.9” of the AT element a1 is the same as the condition rule element r1 (FIG. 1B) of the corresponding rule RL, and the condition data element d1 ( This acts in the direction of setting “0”, which is the value of FIG. On the other hand, the value “0.0” of the AT element a3 (FIG. 1C) is the value of the condition rule element r3 (FIG. 1B) of the corresponding rule RL and the condition data of the corresponding data DT. It does not act in the direction of “0”, which is the value of the element d3 (FIG. 1A), but acts in the direction of any value “#” indicating that it does not contribute to rule determination.

  Note that the AT (AT element) used by the rule generation device is not limited to the AT shown in FIG. That is, the AT of the present invention can have any configuration as long as it is configured from the AT element corresponding to the condition data element and the condition rule element.

  Each configuration of the rule generation device will be described with reference to FIGS. 2, 3, and 4. FIG. 2 is a schematic functional block diagram of the rule generation device according to the first embodiment of the present invention. FIG. 3 is an explanatory diagram illustrating an example of a data set used in the rule generation device and an example of a rule to be generated. FIG. 4 is an explanatory diagram illustrating an example of a rule generated by the rule generation device according to the first embodiment of the present invention.

  The function of the rule generation device shown in FIG. 2 is an example, and the function of the rule generation device that can use the present invention is not limited to that shown in FIG. The data sets and rules shown in FIG. 3 are examples, and the data sets and rules according to the present invention are not limited to those shown in FIG. The analysis result of the rule generation device shown in FIG. 4 is an example, and the analysis result that can be analyzed using the present invention is not limited to that shown in FIG.

  As shown in FIG. 2, the rule generation device according to the present invention is a storage means for storing data (for example, FIG. 1 (a)), rules (for example, FIG. 1 (b)) and AT (for example, FIG. 1 (c)). 10 and a set of rules to be collated with data (hereinafter referred to as “collation set”) from a plurality of rule sets (hereinafter referred to as “rule set”), and the rule elements and characteristics for conclusion of the extracted rules It further includes a collation set generation unit 20 that updates the rule element for evaluation and classifies the updated rule, and a GA application unit 30 that updates a rule constituting the rule set using a genetic algorithm. Note that the rule generation device may further include an acquisition unit that acquires a data set (data, data elements, and the like) and an interface unit that inputs and outputs information and the like outside the rule generation device.

  As shown in FIG. 3, the rule generation device according to the present invention analyzes a data set DS composed of k pieces of data DT shown on the horizontal axis in the drawing. Here, the data DT is evaluated by the actual measurement value P (vertical axis in the figure). Further, the rule generation device generates rules (RLg1, RLs1, RLg2, and RLs2 in the drawing) from the plurality of data DT. That is, as shown in FIG. 4, the rule generation device is the eighth rule from the first rule element r1 as in the example of the rules RLg1, RLs1, RLg2, and RLs2 shown in FIGS. 4 (a) to 4 (d). A rule composed of the rule element r8 is generated. Further, the rule generation device “#” represents data that does not contribute to a common property or dominant feature as in the example of the rules RLg1, RLs1, RLg2, and RLs2 shown in FIGS. 4 (a) to 4 (d). It prescribes as

  Hereinafter, each configuration of the rule generation device according to the present invention will be described in detail with reference to FIG.

  The storage means 10 is means for storing data (Raw data, measured values, etc.) and a set of rules (rule set, collation set, positive rule set described later). As shown in FIG. 2, the storage means 10 stores a data storage unit 11 for storing a data set composed of a plurality of data, a rule set storage unit 12 for storing a rule set (not shown), and an AT value. And an AT storage unit 13. The rule set storage unit 12 includes a collation set storage unit 12r that stores a collation set. In addition, each structure of the memory | storage means 10 can use flash memory which can memorize | store electronic data, semiconductor memories, such as RAM and ROM, a memory card, HDD (Hard Disc Drive), and other well-known techniques.

  The data storage unit 11 stores a data set composed of a plurality of data for data composed of a plurality of data elements (FIG. 1A).

  The rule set storage unit 12 stores a rule set composed of a plurality of rules for a rule composed of a plurality of rule elements (FIG. 1B). In this embodiment, the rule set storage unit 12 stores a plurality of rules in which random values are set in advance for each rule element as initial values of the rule set.

  The collation set storage unit 12r stores a collation set. In this embodiment, the collation set storage unit 12r stores a rule in the rule set storage unit 12 by adding a flag M to the rule extracted by the collation set generation unit 20 (rule extraction unit 21 described later). To do. The collation set storage unit 12r may store the rules extracted by the collation set generation unit 20 (rule extraction unit 21) in a storage area different from the original rule set without using the flag M. Good.

  Further, the collation set storage unit 12r assigns a flag P to the rule stored in the rule set storage unit 12 in order to specify the Positive rule set classified by the collation set generation unit 20 (rule classification unit 23 described later). To do. The collation set storage unit 12r stores the positive rule set classified by the collation set generation unit 20 (rule classification unit 23) in a storage area different from the original rule set and collation set without using the flag P. You may make it do.

  The AT storage unit 13 stores a set of ATs composed of a plurality of ATs for ATs composed of a plurality of AT elements (FIG. 1C). In this embodiment, the AT storage unit 13 stores a plurality of ATs corresponding to a plurality of data in the data storage unit 11. In the present embodiment, the AT storage unit 13 stores an AT corresponding to the Positive rule set classified by the collation set generation unit 20 (rule classification unit 23).

  The collation set generation means 20 is a means for generating a collation set from the rule set. As shown in FIG. 2, the collation set generation means 20 extracts a rule from the rule set (generates a collation set) 21 and updates (or strengthens) the rule element of the extracted rule (collation set). And a rule classification unit 23 that classifies the extracted rules (collation set). The collation set generation unit 20 may be configured to use information processing resources of a control unit (such as a controller) that is preinstalled in the rule generation device.

  The rule extraction unit 21 extracts rules from the rule set. The rule extraction unit 21 selects and reads one data among a plurality of data stored in the data storage unit 11 (storage unit 10), and selects a rule to be checked against the selected one data. A rule set stored in (storage means 10) is extracted to generate a collation set. That is, the rule extraction unit 21 extracts a rule having a condition rule element to be matched with the condition data element of the selected data from the rule set. Specifically, when the data in FIG. 1A is selected, the value of the condition data elements d1 to d4 is “0101”, and therefore, the rule having “0101” as the condition rule elements r1 to r4. In addition, rules having “# 101”, “01 ##”, etc. using “#” described above collate with this data.

The rule strengthening unit 22 uses the conclusion data element (d5 in FIG. 1A) of the selected data, that is, the evaluation value that is the rule element for conclusion of the rule (rule of collation set) extracted using the actual measurement value P. p (r5 in FIG. 1 (b)), error value ε (r6 in FIG. 1 (b)) and rule F (r7 in FIG. 1 (b)) which are characteristic evaluation rule elements are updated. Specifically, the rule strengthening unit 22 updates using the following formula.

p (j + 1) = p (j) + β × (P−p (j)) (1)

ε (j + 1) = ε (j) + β × (| P−p (j) | −ε (j)) (2)

F (j + 1) = F (j) + β × (κ−F (j)) (3)

Here, the above equation indicates that p (j), ε (j), and F (j) are updated to p (j + 1), ε (j + 1), and F (j + 1). β is a predetermined coefficient. Also, κ is calculated by the following equation, for example.

κ = 1 (when ε (j + 1) <ε0) (4.1)
κ = α (ε (j + 1) / ε0) ^−ν (when ε (j + 1) ≧ ε0) (4.2)

That is, in this embodiment, the rule strengthening unit 22 updates the evaluation value p using (1). Next, ε is updated using (2), κ is calculated by substituting ε after the update into (4.1) or (4.2), and the calculated κ is substituted into (3). The fitness F is updated accordingly.

  The rule classification unit 23 classifies the rules (collation set) updated (strengthened) by the rule strengthening unit 22. In the present embodiment, the rule classification unit 23 extracts a Positive rule set from the collation set according to the value of the rule element for conclusion. Here, the Positive rule set is a set of rules that lead to a positive conclusion. That is, the Positive rule set is composed of rules in which the value of the evaluation value p is larger than a predetermined threshold value. Specifically, in this embodiment, the rule classification unit 23 first calculates an average value of the actual measurement values P (data elements d5) of a plurality of data stored in the data storage unit 11 (storage unit 10). . Next, the rule classification unit 23 uses the calculated average value as a threshold value and compares the rule evaluation value p (rule element r5) included in the collation set. At this time, the rule classification unit 23 classifies the rule having the evaluation value p larger than the average value into the positive rule set.

  The GA application unit 30 is a unit that updates a rule constituting the rule set based on the collation set generated by the collation set generation unit 20 using a genetic algorithm (hereinafter referred to as “GA”). . Here, GA (genetic algorithm) refers to each of a plurality of data (corresponding to a rule in this embodiment) as an individual having a gene, and the shape of the gene (data (data ( This is a calculation method that finds individuals with high fitness for the environment (problem) by changing the content of the rule).

  As shown in FIG. 2, the GA application unit 30 includes an AT update unit 31 that updates the AT, and a rule selection unit 32 that selects two parent rules from the rules classified by the rule classification unit 23 (collation set generation unit 20). And a rule crossing unit 33c that creates two child rules from two parent rules by a crossover operation, and a rule addition unit 34a that adds the created child rules to the rule set. Further, the GA application unit 30 includes a rule mutation unit 33v that performs a mutation operation on the child rule created by the rule crossing unit 33c, a rule selection unit 34d that deletes (淘汰) a rule from the rule set, and / or An allowable error updating unit 34r that updates the allowable error ε0 may be further included. The GA application unit 30 may be configured to use information processing resources of a control unit (such as a controller) that is pre-installed in the rule generation device.

  The AT update unit 31 updates the AT stored in the AT storage unit 13 (storage unit 10). The AT update unit 31 updates the AT based on the rules in the Positive rule set. As described above, the value of each AT element represents how much the condition data element of the corresponding data DT contributes to the matching degree F of the rule RL to be collated with the data. Therefore, when the value of a certain condition data element collates with the value of “#” of the corresponding condition rule element, the AT update unit 31 determines the matching degree F of the matched rule. Judge that it does not contribute to. On the other hand, when matching with a specific value of the corresponding rule element (here, “0” or “1”), the conditional data element contributes according to the value of the matching degree F of the matched rule. Judge that you are doing. That is, it is determined that the condition data element contributes to the determination of an accurate rule, so that it matches with the rule with high adaptation F.

  An example of the AT update operation will be described with reference to FIG. FIG. 5 is an example when there are two rules in the collation set. The AT update unit 31 according to the present invention can update the AT value in the same manner as the following procedure even when there is one rule in the collation set or when there are three or more rules.

  First, the AT update unit 31 uses the rule element r7 (compliance F) of two rules in the collation set shown in FIG. 5A to determine each value of the rule elements (r1 to r4) as shown in FIG. Convert as shown in (b). Specifically, when the condition rule element is “1” or “0”, the AT update unit 31 sets the AT element corresponding to the rule element as the value of the fitness F. In addition, when the rule element is “#”, the AT update unit 31 sets the AT element corresponding to the rule element to “0”.

  Next, the AT update unit 31 performs addition for each rule element for the two cases shown in FIG. FIG. 5C shows the addition result.

  Next, the AT update unit 31 normalizes the addition result (AT (C) in FIG. 5C). FIG. 5D shows the normalization result. Specifically, the AT updating unit 31 converts the value of the maximum AT element to “1” in order to correct the absolute value of each element to “1” or less, and values of other elements Are converted at a similar conversion rate (here, each value is divided by 1.2).

  Thereafter, the AT updating unit 31 updates the AT (for example, FIG. 1C) using each normalized value (AT (R) in FIG. 5D). Specifically, for each element, the AT update unit 31 averages each value after normalization (AT (R) in FIG. 5D) and the AT before update (for example, FIG. 1C). And the calculated average value is set as a new AT value (AT (A) in FIG. 5E). When there is no AT before update (for example, when data corresponding to the AT is processed for the first time), the AT update unit 31 sets each value after normalization (AT (R) in FIG. 5D). ) As a new AT.

  The rule selection unit 32 selects two parent rules from the rules classified by the rule classification unit 23 (collation set generation means 20). In the present embodiment, the rule selection unit 32 selects two parent rules from the classified rules (Positive rule set) based on the evaluation value p (r5 in FIG. 1B). As a specific selection method, a known method in a genetic algorithm such as roulette selection, ranking selection, tournament selection, or the like can be used. At this time, the selected parent rule is not deleted from the rule set.

  The rule crossing unit 33c uses the AT to create two child rules from the two parent rules according to a given probability. Specifically, from the copy of the two parent rules selected by the rule selection unit 32 (Offspring 1 in FIG. 6C, Offspring 2 in FIG. 6D), two child rules (Best, Create Worst) in FIG. Here, the rule crossing unit 33c replaces the rule elements of the two copies so that important features are aggregated in the best child rule (FIG. 6E).

  Specifically, the rule crossing unit 33c sets the AT value to a predetermined threshold value so that the value of the corresponding condition data element becomes the value of the condition rule element when the value of the AT element is larger than the predetermined threshold value. In the following cases, an attempt is made to replace the data elements of the two copies so that the value of the corresponding condition data element is “#”. Here, if the two copies do not have a data element that allows such replacement, the replacement is not performed. In the example of FIG. 6, since the values of the AT elements a1 and a2 are larger than a predetermined threshold (for example, “0.5”), the value “0” of the data element d1 is set as the first rule element for each Offspring condition. The offspring1 having the value “1” of the data element d2 is an important feature in the second rule element for the condition. On the other hand, since the AT values of a3 and a4 are equal to or less than a predetermined threshold, “#” is an important feature for the third and fourth conditional rule elements. However, since the third rule element of each Offspring is “0”, it remains “0” (even if replaced). In the fourth rule element, Offspring2 has a value of “#”. Therefore, in this example, the second and fourth rule elements of Offspring (FIGS. 6 (c) and 6 (d)) are crossed, and the child rule in which the important features of FIG. A child rule in which unimportant rules of (f) are aggregated is created. Further, the rule crossing unit 33c sets the average value of the allowable errors ε0 of the two copies as the allowable error ε0 of the two child rules.

  The rule mutation unit 33v mutates the child rule created by the rule crossing unit 33c. The rule mutation unit 33v mutates the child rule OffspringA (FIG. 7C) to, for example, the child rule VariationAv (FIG. 7D) with a predetermined probability μ. Specifically, for example, when performing the mutation, the rule mutation unit 33v sets the value of the AT element of the corresponding AT (FIG. 7B) as the probability μ for each rule rule element of the child rule. To determine whether to change the value. Here, when the value of the condition rule element is to be changed, the rule mutation unit 33v sets the value of the condition rule element to the corresponding data if the value of the AT element is larger than a predetermined threshold value. The value is changed to the value of the data element in FIG. 7A (second rule element in FIG. 7D). On the other hand, if the value of the AT element is equal to or less than the predetermined threshold, the value of the rule element for the condition is changed to “#” (the fourth rule element in FIG. 7D). FIGS. 7E and 7F are examples of mutation of another child rule. As shown in the figure, the values are changed in the third and fourth rule elements in the same manner as in the previous example.

  The rule adding unit 34a adds the two child rules created by the rule crossing unit 33c to the rule set stored in the rule set storage unit 12 (storage means 10). The rule adding unit 34a may add one of the two created child rules to the rule set. The rule adding unit 34a may add a child rule mutated by the rule mutation unit 33v to the rule set.

  When the rule adding unit 34a adds a child rule to the rule set, the rule hook unit 34d deletes (淘汰) the rule from the rule set. The rule hook 34d deletes, for example, a rule that is not used for analysis, a rule with a small evaluation value p, a rule with a small fitness F, or any other rule that corresponds to a predetermined condition.

  The allowable error update unit 34r mutates (updates) the allowable error ε0 with a given probability. For example, the allowable error updating unit 34r can update the allowable error ε0 before or after generating the child rule by the rule crossing unit 33c, or before or after mutating the child rule by the rule mutation unit 33v. Further, the rules to be processed by the allowable error updating unit 34r may be all of the rules selected by the rule classification unit 23 (even in this case, the allowable error ε0 is selective with the given probability. Only the rule selected by the rule selection unit 32 may be used.

  In the present embodiment, the allowable error update unit 34r updates the allowable error ε0 using the following equation.

(Equation 6)
ε0 (g + 1) = ε0 (g) × α
Here, the above equation indicates that ε0 (g) is updated to ε0 (g + 1) and F (g + 1). Α is a coefficient representing the degree of update. The allowable error updating unit 34r can use a value larger than “1” or a value smaller than “1” for α. Alternatively, α may be, for example, 0 <α <1.5 or 0 <α <1.0).

  In the present embodiment, the allowable error ε0 indicates a range of a region of rules to be generated (for example, RLg1, RLs1, RLg2, and RLs2 illustrated in FIG. 3). Further, the allowable error ε0 is, for example, a range of RLg1 (ε0 = 0.08) shown in FIG. 4A or a value shown in FIG. 4C as a value equal to or less than the first predetermined value used in the generalization rule. The range of RLg2 (ε0 = 0.06) may be used. The allowable error ε0 is, for example, a range of RLs1 (ε0 = 0.01) shown in FIG. 4B as a value equal to or smaller than a second predetermined value smaller than a first predetermined value used in a special rule (described later), or The range of RLs2 shown in FIG. 4D may be set (ε0 = 0.01).

  Here, the generalization rule is a rule applicable to a relatively large number of data (for example, a plurality of days). A special rule is a rule applicable to a relatively small number of data (for example, a specific day). The rule generation device generates a generalized rule as a rule that characterizes a daily state, for example, at the time of data analysis in a later-described care support system. In addition, the rule generation device generates a special rule as a rule that characterizes an extraordinary state (birthday, going out date, event date, etc.).

  The rule generation device according to the present invention generally operates as follows. That is, one data is selected from the data set DS stored in the storage unit 10 (data storage unit 11), and the storage unit 10 (rule set storage unit) is selected using the collation set generation unit 20 (rule extraction unit 21). In step 12), a collation set that is a set of rules to be collated with one data is extracted from the rule sets stored in advance. Further, the rule generation device uses the collation set generation means 20 (rule strengthening unit 22), and the rule element for conclusion (evaluation value p) of the rules constituting the extracted collation set and part of the rule elements for characteristic evaluation ( The error value ε and the fitness F) are updated (strengthened). Furthermore, the rule generation device extracts a positive rule set from the updated collation set using the collation set generation unit 20 (rule classification unit 23).

  Thereafter, the rule generation device uses the GA application unit 30 (AT update unit 31) to update the AT value stored in the storage unit 10 (AT storage unit 13) based on the rule elements of the Positive rule set. . In addition, the rule generation device uses the GA application unit 30 (rule selection unit 32) to select two parent rules from the Positive rule set. Next, the rule generation device generates two child rules based on the selected two parent rules using the GA application unit 30 (rule crossing unit 33c). Here, the GA applying unit 30 (rule crossing unit 33c) generates two child rules using the AT value. At this time, one of the two child rules is generated so that the values of the condition rule elements representing important features are aggregated. Next, the rule generation device uses the GA application unit 30 (rule addition unit 34a) to add the generated two child rules to the rule set stored in the storage unit 10 (rule set storage unit 12).

  Further, the rule generation device may mutate the two generated child rules by using the GA application unit 30 (rule mutation unit 33v). Further, the rule generation device may delete (淘汰) a desired rule from the rule set when adding the generated two child rules to the rule set using the GA applying means 30 (rule rule unit 34d). Good. The rule generation device may update the allowable error ε0 using the GA applying unit 30 (allowable error updating unit 34r).

[2. Rule generation method (first embodiment)]
An example of the operation of the rule generation device (rule generation method) according to the first embodiment of the present invention will be described with reference to FIGS. Here, FIG. 8 is a flowchart for explaining an example of the rule generation method according to the present embodiment (expansion of continuous value reward). FIG. 9 is a flowchart for explaining an example (GA) of the rule generation method according to the present embodiment. The operation of the rule generation device shown in FIGS. 8 and 9 is an example, and the operation of the rule generation device according to the present invention is not limited to that shown in FIGS.

  As shown in FIG. 8, in step S800, the rule generation apparatus first initializes a rule set [P] in order to start an analysis operation. Specifically, the rule generation device stores a plurality of rules (for example, FIG. 1B) set at random within a range of values that can be taken by each rule element in the rule set storage unit 12 (block B03 in FIG. 2). Store (rule set storage step). After that, the rule generation device goes to step S801.

  In step S801, the rule generation device uses the rule extraction unit 21 (block B02 in FIG. 2) to select one data from a plurality of data stored in the data storage unit 11 (block B01 in FIG. 2). . Thereafter, the rule generation device proceeds to step S802. Note that the rule generation device may select, for example, data having a new recording date and time or new data in order from among a plurality of data. For example, the rule generation device may select data from a desired date and time among a plurality of data.

  In step S802, the rule generation device uses the rule extraction unit 21 to extract the collation set [M] (rule extraction step). Specifically, the rule generation device extracts a rule to be matched with the data from the stored rule set [P] stored in step S800 based on the one data selected in step S801. After that, the rule generation device goes to step S803.

  In step S803, the rule generation device updates (strengthens) the evaluation value p, the error value ε, and the fitness F using the rule strengthening unit 22 (block B05 in FIG. 2) (rule strengthening step). Specifically, the rule generation device reinforces the evaluation value p, the error value ε, and the fitness F using (1) to (4.2) described above. Thereafter, the rule generation device proceeds to step S804.

  In step S804, the rule generation device classifies the collation set [M] using the rule classification unit 23 (block B06 in FIG. 2) (rule classification step). Specifically, the rule generation device compares the evaluation value p strengthened in step S803 with the average value of the actual measurement values P of a plurality of data stored in the data storage unit 11, thereby obtaining the evaluation value p. A rule larger than the average value is classified into a Positive rule set [Mp]. After that, the rule generation device goes to step S805.

  In step S805, the rule generation device uses the AT update unit 31 (block B07 in FIG. 2) to update the AT value stored in the AT storage unit 13 (block B08 in FIG. 2) according to the above-described procedure. . After that, the rule generation device goes to step S806.

  In step S806, the rule generation device applies GA to the Positive rule set [Mp] using the GA application unit 30 (GA application step).

  Specifically, the rule selection unit 32 (rule generation device) selects two parent rules from the classified positive rule set [Mp] in step S901 of FIG. Next, in step S902, the rule crossing unit 33c (block B10a in FIG. 2) creates two child rules from the two parent rules based on the AT value by the above-described procedure (rule crossing step). Next, the rule mutation unit 33v (block B10b in FIG. 2) mutates the generated child rule by the above-described procedure in step S903. Thereafter, the rule adding unit 34a (block B11a in FIG. 2) adds two child rules (step S902) or a mutated child rule (step S903) to the rule set [P] in step S904 (rule adding step). ). At this time, rule ruler 34d (rule generation device) deletes two rules having a small fitness F from set [P]. Note that the rule generation device may proceed to step S904 without performing step S903 (mutation).

  Thereafter, the rule generation device returns to Step S807 in FIG.

  In step S807, the rule generation device determines whether to end the analysis operation. Specifically, for example, when generating a generalized rule from k pieces of data, the rule generation device counts repeated calculations and analyzes all of the k pieces of data, or performs a predetermined number of repeated calculations. The analysis operation may be determined to end.

  If it is determined that the analysis operation is to be terminated, the rule generation device proceeds to END in the drawing and ends the analysis operation. If it is determined that the analysis operation is not terminated, the rule generation device returns to step S801 and repeats the operations of steps S801 to S807 described above.

  As a result of this iterative process, the rules included in the rule set are optimized. That is, in the case of an appropriate rule, the conclusion rule element is updated to a value according to the value of the raw data stored in the data storage unit 11 and the matching degree F of the characteristic evaluation rule element. Is getting higher. In addition, the rule element for a condition is appropriately applied to a state value or wildcard value having high importance based on an AT reflecting a high degree of fitness by the rule crossing unit 33c and the rule mutation unit 33v as necessary. Updated. As a result of the optimization, for example, a rule having a high fitness F in the final rule set and a conclusion rule element having a desired value can be acquired as an appropriate rule.

  As described above, even if the rule element for conclusion has a continuous value, the operation of the rule generation device according to the first embodiment (rule generation method) is a positive rule set whose value satisfies a predetermined condition. Since classification is performed and optimization by GA using AT is performed for this Positive rule set, an appropriate rule is generated.

(Modification 1 of rule generation method)
A modification example 1 of the operation (rule generation method) of the rule generation device according to the first embodiment of the present invention will be described with reference to FIGS. 10 and 11. Here, FIG. 10 is a flowchart illustrating a first modification of the rule generation method according to the present embodiment (simultaneous acquisition of special rules and general rules). FIG. 11 is a flowchart for explaining the present modification (GA).

  As shown in FIG. 10, steps S1001 to S1006 performed by the rule generation device according to the present modification are the same as the operation of the rule generation device according to the first embodiment (FIG. 8).

  Here, in step S1006 of FIG. 10, the rule generation device according to the present modification is the same as the operation of the rule generation device according to the first embodiment (steps S901 to S901 in FIG. 9). This is the same as step S903). In step S1104 in FIG. 11, the allowable error ε0 is updated using the allowable error (ε0) update unit 34r (block B11c in FIG. 2). The rule generation device according to this modification is the same as the operation (FIG. 8) of the rule generation device according to the first embodiment except that the allowable error (ε0) is updated. Omitted.

  As described above, the operation (rule generation method) of the rule generation device according to Modification 1 can obtain the same effects as the operation of the rule generation device according to the first embodiment (FIG. 8). Further, in the operation (rule generation method) of the rule generation device according to the modified example 1, by updating the allowable error ε0, it is possible to generate both a special rule with a small allowable error and a generalized rule with a large allowable error.

  For example, when the value of ε0 is updated in a decreasing direction, the value of the fitness F is hardly increased by the subsequent iterative process (see equations (3), (4.1), and (4.2)). As a result, since the AT value does not easily increase, the number of condition rule elements increases by “#” due to crossover or mutation. Then, since a large amount of data is collated, the error ε between the evaluation value p of the conclusion rule element and the result value of the conclusion data element increases, and the fitness F decreases. If this happens, it will be inappropriate as a rule and will be easily deceived. On the other hand, even if the value of ε0 is small, if there is some data in which the result value of the conclusion data element is within the range of the evaluation value p ± ε0, the above-described flow does not occur and the fitness F is It is kept high and appropriate special rules are generated.

  On the other hand, when the value of ε0 is updated in the increasing direction, the value of the fitness F is easily increased by the subsequent iterative process (see equations (3), (4.1), and (4.2)). As a result, the AT value tends to increase, and the condition value of the condition rule element increases to “0”, “1”, etc. due to crossover or mutation. Then, only the specific data tends to be collated, the error ε between the evaluation value p of the conclusion rule element and the result value of the conclusion data element is reduced, and the fitness F is increased. Thereby, an appropriate generalization rule is generated.

  In the example of FIG. 3, the rules RLg1 and RLg2 having a wide “width” in the vertical axis direction are generalization rules, and the rules RLs1 and RLs2 having a narrow “width” are special rules.

(Modification 2 of the rule generation method)
Modification 2 (handling of contradictory data) of the rule generation device (rule generation method) according to the first embodiment of the present invention will be described.

  In the present modification, the storage unit 10 stores the negative rule set together with the positive rule set.

  Specifically, the collation set storage unit 12r is stored in the rule set storage unit 12 in order to identify the Positive rule set and the Negative rule set classified by the collation set generation unit 20 (rule classification unit 23 described later). A flag P or N is assigned to the existing rule. The collation set storage unit 12r uses the positive rule set and the negative rule set classified by the collation set generation unit 20 (rule classification unit 23) without using the flags P and N as the original rule set and collation set. You may make it memorize | store in another memory area.

  Further, in the present modification, the AT storage unit 13 stores ATs respectively corresponding to the Positive rule set and the Negative rule set classified by the collation set generation unit 20 (rule classification unit 23).

  Then, the rule classification unit 23 classifies the collation set into a Positive rule set and a Negative rule set according to the value of the rule element for conclusion. Here, the Positive rule set is a set of rules that lead to a positive conclusion. That is, the Positive rule set is composed of rules in which the value of the evaluation value p is larger than a predetermined threshold value. The Negative rule set is a set of rules that lead to a negative conclusion. That is, the Negative rule set is configured by rules in which the value of the evaluation value p is equal to or less than a predetermined threshold value. Specifically, in this modification, the rule classification unit 23 first calculates an average value of the actual measurement values P (data elements d5) of a plurality of data stored in the data storage unit 11 (storage means 10). . Next, the rule classification unit 23 uses the calculated average value as a threshold value and compares the rule evaluation value p (rule element r5) included in the collation set. At this time, the rule classification unit 23 classifies a rule having an evaluation value p larger than the average value into a Positive rule set, and classifies a rule having an evaluation value p equal to or less than the average value into a Negative rule set.

  FIG. 12 is a flowchart for explaining a rule generation method according to the second modification. FIG. 13 is a flowchart for explaining the second modification (GA).

  As shown in FIG. 12, steps S1201 to S1203 performed in the rule generation device according to this modification are the same as the operations of the rule generation device according to the first embodiment (steps S801 to S803 in FIG. 8). It is.

  In step S1204, the rule generation device classifies the collation set [M] using the rule classification unit 23 based on the evaluation value p (rule classification step). Specifically, in this modification, the rule generation device compares the evaluation value P strengthened in step S1203 with the average value of the actual measurement values P of a plurality of data stored in the data storage unit 11. The rules with the evaluation value p larger than the average value are classified into the Positive rule set [Mp], and the rules with the evaluation value p smaller than the average value are classified into the Negative rule set [Mn]. Thereafter, the rule generation device proceeds to Step S1205A and Step S1205B.

  In step S1205A and step S1206A, the rule generation device uses GA (FIG. 9) in the same manner as the operation of the rule generation device according to the first embodiment (FIG. 8) (step S805 and step S806). Add the generated child rule to the rule set. In step S1205B and step S1206B, the rule generation device selects a parent rule from the classified negative rule set [Mn] using GA (FIG. 13), and applies GA to the parent rule. After that, the rule generation device goes to step S1207. That is, in the rule generation device according to this modification, the category of the conclusion part of the rule is different between the A side step (Step S1205A and Step S1206A) and the B side step (Step S1205B and Step S1206B) (for example, shallow sleep). Two types of rules can be generated at the same time. Note that FIG. 13 is the same as step S1205A and step S1206A except that two parent rules are selected from the negative rule set [Mn], and thus the description thereof is omitted.

  Next, in step S1207, the rule generation device determines whether to end the analysis operation in the same manner as in the operation of the rule generation device according to the first embodiment (FIG. 8) (step S807). If it is determined that the analysis operation is to be terminated, the rule generation device proceeds to END in the drawing and ends the analysis operation. If it is determined that the analysis operation is not terminated, the rule generation device returns to step S1201 and repeats the operations of steps S1201 to S1207 described above.

  As described above, the operation (rule generation method) of the rule generation device according to Modification 2 can obtain the same effects as the operation of the rule generation device according to the first embodiment (FIG. 8). In addition, the operation (rule generation method) of the rule generation device according to the second modification is different from the rule set (a plurality of rules) in the category of the rule conclusion part (for example, light sleep and deep sleep). Can be generated.

  In the example of FIG. 3, for example, the rule RLg1 (or RLs1) in the high value range and the rule RLg2 (or RLs2) in the low value range are generated simultaneously. .

  In addition, if the combination of the first and second modifications, that is, the allowable error update unit 34r is provided and both the positive rule set and the negative rule set are classified, the rules RLg1, RLs1, RLg2, and RLs2 are generated simultaneously. can do.

Furthermore, even when the conclusion rule element does not take a continuous value but takes a binary value, the generalized rule and the special rule can be generated at the same time if the aspect of the first modification is adopted. Moreover, if it is set as the aspect of the modification 2, two types of rules from which the category of the conclusion part of a rule differs can be produced | generated simultaneously.

[3. Care support system (second embodiment)]
A care support system according to the second embodiment of the present invention will be described with reference to FIGS. 14 to 16. Here, FIG. 14 is a schematic system diagram illustrating an example of a care support system according to the present embodiment. FIG. 15 is an explanatory diagram illustrating an example of an acquisition result (data set) of the data acquisition unit of the care support system. FIG. 16 is an explanatory diagram illustrating an example of a sleep stage (evaluation value) used in the care support system.

The care support system shown in FIG. 14 and the like is an example, and the care support system according to the present invention is not limited to the one shown in FIG. Further, in the following description, since the care support system (second embodiment) includes the configuration of the above-described rule generation device (first embodiment), different portions will be mainly described.

[3-1 Configuration and operation of care support system]
The care support system according to the present embodiment includes a data acquisition unit that acquires a data set, and the [1. Rule generation device (first embodiment)]. In other words, the care support system according to the present embodiment acquires the daily data of the person to receive care using the data acquisition means, and receives care based on the acquired data set (for example, data for one month). Analyzes the situation of the target person and the content of care support to acquire knowledge about care support. Thereby, for example, a care recipient in a care support method such as a care home using knowledge acquired by the care support system according to the present embodiment (rules analyzed by the rule generation device according to the first embodiment). A care plan for care support can be created based on the past behavior or state (a plurality of data).

Specifically, as shown in FIG. 14, the care support system is used for care support of a care target person (for example, an elderly person) in a care welfare facility. The care support system may be used by, for example, a caregiver, a care manager, a care planner, a caregiver support agent, and the like.

For example, as shown in FIGS. 14 and 15, the care support system uses (i) a pleasant sleep level indicating the depth of sleep as an evaluation value based on a plurality of (for example, one month) data (data set). , Acquire knowledge about care support. Further, the care support system simultaneously acquires (ii) a general rule (generalized rule) and a special rule that lead to deep (or shallow) sleep based on the data set.

As shown in FIG. 15, in the care support system, the data acquired by the data acquisition means includes (i) care activity elements performed on a certain day for a care recipient (breakfast, exercise, bathing, sleep, health status, etc. ) And (ii) a sleep stage that means the depth of sleep (degree of sleep) of the day. For example, the sleep stage may be evaluated by a degree LV as shown in FIG. The sleep stage shown in FIG. 16 means deeper sleep as the degree LV (degree of sleep) is higher. Here, the set of care activity elements corresponds to the condition data element in the first embodiment, and the sleep stage corresponds to the conclusion data element.

[3-1-1 Rules obtained by the care support system]
Examples of rules (knowledge) obtained using the care support system according to the present embodiment are shown below.

(A) Rules that lead to deep sleep and can be used for multiple days (deep sleep-general rules, such as RLg1 in FIG. 3)
(B) Rules that lead to deep sleep and can be used only on specific days (deep sleep-special rules such as RLs1 in FIG. 3)
(C) Rules that lead to shallow sleep and can be used for multiple days (shallow sleep-general rules, such as RLg2 in FIG. 3)
(D) Rules that lead to shallow sleep and can be used only on specific days (shallow sleep-special rules such as RLs2 in FIG. 3)

Specifically, the rules used in the care support system are a set of care activity elements that are condition rule elements and a predicted sleepiness level p (corresponding to the evaluation value p in the first embodiment) that is a rule element for conclusion. , A rule conformance f that is a rule element for characteristic evaluation (corresponding to the conformity F of the first embodiment), and an allowable error εo of the predicted sleepiness (corresponding to the tolerance ε0 of the first embodiment) Is done. Further, the rule used in the care support system can constitute a general-purpose condition by incorporating “#” indicating an arbitrary value into the input data (a set of care activities).

  The predicted sleepiness level p is a value obtained by predicting the sleepiness level of the input data when the rule is used. That is, the user of the care support system according to the present embodiment predicts the degree of sleepiness using the acquired knowledge (rule). Further, the user of the care support system according to the present embodiment recognizes that the rule (content of care support) leads to deep sleep when the predicted sleepiness p is a high value, for example, and the predicted sleepiness p is low. If it is a value, it can be recognized that the rule (contents of care support) leads to shallow sleep.

  The fitness f means the reliability (correctness) of the predicted sleepiness level. When the error between the predicted sleepiness level and the actually measured value is large, the fitness level decreases, which means that the rule is incorrect.

  The allowable error ε0 of the predicted sleepiness level means an allowable range of errors when calculating the fitness level. When the allowable error is large, a general rule (generalized rule) that has a large error but can be used for a plurality of data. Also, when the tolerance is small, a special rule that is very accurate but can use only a small number of data.

  In the care support system, for example, when the rule condition part is “00 ##”, the estimated predicted value is p, and the allowable error is ε0, the set of care activity elements on a certain day is “0001”, “0010”, “ When the value is “0000” or “0011”, it is predicted that the night's sleep comfort level is p ± ε0.

In the present embodiment, in the care support system, the rule [A] in the collation set [M] that is equal to or higher than the average sleepiness level p of the rule set [P] is “Positive Rule Set”, and is less than the average sleepiness level p. A certain rule [nA] is referred to as “Negative Rule Set”.

[Examples of knowledge (rules) acquired by the 3-1-2 care support system]
FIG. 17 shows an example of knowledge acquired by the care support system (general knowledge and special knowledge for sleep and insomnia).

  As shown in FIG. 17, in the care support system, for example, the following four knowledge ("(a) pleasant sleep (general knowledge)", "(b) pleasant sleep (special knowledge)", "(c) insomnia (general knowledge)" And “(d) Insomnia (special knowledge)”). An example of rule interpretation is shown below.

    (A1) When “bathing in the afternoon” is performed, deep sleep is predicted that night.

    (A2) When “rehabilitation” is not performed, deep sleep is predicted that night.

    (B) When performing “morning bathing” and “morning rehabilitation”, deep sleep is predicted that night.

    (C) When “morning rehabilitation” is performed and “bath” is not performed, light sleep is predicted that night.

    (D) When “tea”, “bathing”, and “rehabilitation” are not performed, light sleep is predicted that night.

The person who provides care support realizes care support with a high degree of satisfaction by creating a menu of care support using the knowledge (rules) of (a1), (a2), and (b) to (d) above. .


[3-2 Example of sleep stage estimation method]
As an example of a method for estimating the sleep stage (predicted sleepiness level p) of the data acquisition means of the care support system according to the second embodiment of the present invention, a method proposed by the present applicant in Japanese Patent Application No. 2013-123257 is described. explain. Note that the sleep stage (predicted sleepiness level p) estimation method that can be used in the present invention is not limited to the sleep stage estimation method described below.

  The sleep stage estimation method according to the present example is based on the steps of extracting frequency band data from the biological data acquired by the biological data acquisition unit using a filter that transmits a desired frequency band, and the extracted data. Determining a sleep stage at the time of obtaining biometric data.

  In this example, the filter is a filter generated by a database-based compact genetic algorithm (hereinafter referred to as “DcGA”) using a data set that is an aggregate of a plurality of filter individuals. The filter includes, for example, a plurality of filters that transmit various frequency bands, and an evaluation value for a result of determining a sleep stage based on data extracted from past biological data using the filter.

The sleep stage estimation method according to this example will be specifically described below.

[3-2-1 Generation Method of Bandpass Filter for Sleep Stage Estimation]
In the sleep stage estimation method of this example, a bandpass filter that extracts a frequency component of a heart rate suitable for estimation is generated using DcGA improved by a learning method based on a genetic algorithm. Here, the band pass filter is expressed by a 14-bit bit string in this example. Each bit represents a frequency (in order, 1/16200 second, 1/8100 second, 1/4050 second, 1/16 second, 1/8 second, 1/4 second, 1/2 second). Each bit means that the frequency represented by that bit is used for estimation when the value is "1", and the frequency represented by that bit is used for estimation when the value is "0". It means not.

(A) cGA (Compact Genetic Algorithm)
cGA is a genetic algorithm (GA) using a probability model. The target data is binary (0, 1) binary data. The probability indicated by the probability model is a probability that a bit can be “1”. Each data corresponds to a locus. Note that the DcGA used in this example is a technique obtained by extending cGA.

(mechanism)
(I) In cGA generation, cGA generates an individual according to the probability of the probability model. Individuals are composed of genes that have data areas and sequences according to the problem. Moreover, the position (locus) of the gene that holds the information is determined. cGA discovers individuals to be applied to the environment (problems) by capturing the genes biologically and changing the shape of the genes through selection, crossover and mutation operations. The probability of the cGA probability model represents the probability that the bit is “1”. For example, if the probability model element is 1.0, the corresponding locus element is “1”. In addition, cGA produces | generates an individual from a probability model by performing the same operation with respect to all the gene loci.

      (Ii) In the cGA, individuals a and b are generated from a probability model (population probability B01 in the figure) in an updating method (for example, FIG. 18) (B03). Moreover, cGA calculates | requires an evaluation value using an evaluation function with respect to two produced | generated individuals. Here, in cGA, an individual with a high evaluation value is “winner” and a low individual is “loser” (B04). Moreover, cGA compares the gene of the individual who is "loser" based on the individual who became "winner". cGA compares the elements (0, 1) of each locus as a gene comparison, and checks whether they are different or the same. When the elements are different from each other, 1) When the element of “winner” is “1”, cGA is added to the corresponding element of the probability model by the update rate. cGA 2) When the element of “winner” is “0”, the update rate is subtracted from the corresponding element of the probability model (B02). Note that cGA does nothing if the elements are the same. The update rate is a constant determined in advance. The update rate is obtained by dividing the upper limit number of population from “1”. After updating, the cGA repeats updating of the probability model by generating an individual from the probability model again and performing comparison.

(algorithm)
In cGA, first, 1) the probability element of each locus in the probability model (P Vector: P [i]) is initialized to 0.5. Next, 2) Two individuals are generated according to the value of each probability element of the probability model. Next, 3) the two generated individuals are evaluated by the evaluation function, and the evaluation values are compared. Here, 4) the individual having a high evaluation value is “winner” and the individual having a low evaluation value is “loser”, and each locus is compared. Thereafter, the same processing is repeated until all the elements of P converge, and the converged P is obtained.

(B) DcGA (Database-based Compact Genetic Algorithm)
In general, cGA can be applied to the problem that a correct evaluation function (evaluation value) can be designed correctly. However, in the sleep stage estimation, the tendency of the biometric data frequently changes due to changes in physical condition and the like, and thus there are cases where the evaluation function cannot be designed. Therefore, in this example, DcGA, which is a learning method that extends cGA, is used.

  FIG. 19 shows an outline of the DcGA algorithm.

  DcGA handles given data as a child individual (B01 in the figure). In other words, DcGA treats given data (solution extracted from the solution space) as a child individual, thereby avoiding evaluation based on a solution with a biased solution structure, and a solution with a plurality of evaluation values (B02) (sub-optimal solution). (FIG. 20A)) can be searched (B03). Furthermore, DcGA can acquire a model expressing features common to the suboptimal solution by using a probability model (B04).

(mechanism)
(I) In the probability model in DcGA, the probability model is used in the same manner as in cGA. Here, DcGA differs from cGA in that it expresses not only the optimal solution but also the characteristics of the sub-optimal solution. For example, as shown in FIG. 20A, when the individual groups a, b, c, and d are suboptimal solutions, the genetic structures of the individuals are different. DcGA can express a common gene by a real value by using a probability model. The second bit is “1”, the fourth bit is “0”, and the first and third bits are “0” or “1”. At this time, as shown in FIG. 20B, the probability model indicates values of “1”, “0”, and “0.5”, respectively, and expresses gene characteristics. Further, DcGA indicates the probability that the probability handled in the probability model is “1”, similar to cGA.

      (Ii) DcGA uses a data set instead of the environment handled by cGA. That is, DcGA prevents the convergence of an erroneous local solution even when the evaluation value fluctuates due to the uncertain evaluation function, and acquires a suboptimal solution. Specifically, DcGA first uses a data set composed of data evaluated in advance. Next, DcGA avoids convergence to an incorrect local solution by selecting data randomly from the data set. As described above, DcGA can reflect not only the optimal solution but also the characteristics of the sub-optimal solution in the probability model by using the data set.

(algorithm)
First, DcGA evaluates each data (individual) in advance and generates a data set. Here, an evaluation value is associated with each data in the data set. Next, DcGA sets 2) the probability element of each locus of the probability model (P Vector: P [i]) to 0.5 (initial value). In addition, 3) DcGA is randomly selected from the data set. Furthermore, DcGA 4) compares the evaluation values of the two selected individuals. Next, DcGA 5) makes individuals with high evaluation values “winner” and low individuals “loser”, and compares each locus. Thereafter, DcGA repeats the same process until all the elements of 6) P converge. Thereby, DcGA can obtain | require converged P as a solution.

(C) Application of DcGA to sleep stage estimation method (filtering and selection of probability model)
A method of filtering (binary data) a probability model with real values to apply this example to the sleep stage estimation problem, and a method of selecting a probability model (solution extraction) based on the filtered probability model And will be explained. In filtering the probability model, in order to apply the probability model (P Vector) to the sleep stage estimation problem, numerical value filtering (binary data conversion) is performed according to the following conditions.

Condition 1: “0” when the probability model value is 0.25 or less
Condition 2: “1” when the value of the probability model is 0.75 or more
Condition 3: “#” when the value of the probability model is other than Condition 1 and Condition 2.
“#” In condition 3 indicates that “0” or “1” can be arbitrarily selected. FIG. 21A shows an example of a probability model, and FIG. 21B shows an example of filtering.

(Application to sleep stage estimation)
When DcGA is applied to the generation of a bandpass filter for sleep stage estimation, first, 1) the subject's biological data is acquired by an unconstrained air mattress type biosensor, and fast Fourier transform is applied to Generate a path filter. Next, 2) an inverse fast Fourier transform is applied from the generated multiband pass filter to extract a set of components in a certain frequency band and intermediate frequency band. At this time, 3) Applying the discrete Fourier transform to the extracted frequency band component set, obtaining the estimated evaluation value by converting the sleep stage, and generating a data set combining the multiband pass filter and the evaluation value To do. 4) Using DcGA, update the probability model by selecting data from the data set. 5) The updated probability model is converted to binary data, and the probability model is converted to a multiband pass filter.

  Here, the filter generated in 1) corresponds to each individual in the data set. That is, the above 1) to 3) correspond to processing for evaluating each data in advance and generating a data set. Further, the calculation of the estimated evaluation value in the above 3) is performed by assuming that the output data of (i) 2) above represents the temporal transition of the sleep stage, and the sleep stage (wake / REM sleep / Non-REM sleep).

  Next, (ii) using the heartbeat and body motion data, the sleep stage at each time point is estimated separately. Specifically, the heartbeat average and body motion standard deviation are obtained from the heartbeat and body motion data obtained from the beginning of sleep to a certain time i, and the body motion standard deviation for one minute from the time i is obtained. Next, an average of heartbeats from time i to 10 minutes before is obtained, and the generated body movements are distinguished from each other: a) body movements that occur when the sleep stage changes, and b) body movements that occur during the REM sleep period.

  Specifically, “Condition 1: Standard deviation of body movements overnight <standard deviation of body movements for 1 minute” and “Condition 2: average of heartbeats overnight <average of heartbeats of 10 minutes” are used. And judge.

  Thereafter, 4) the sleep stage is estimated according to the following determination method.

    Determination 1: When conditions 1 and 2 are satisfied, it is determined that the REM sleep period.

    Determination 2: When only condition 1 is satisfied, it is determined that the body movement occurred outside the REM sleep period.

    Determination 3: When the conditions 1 and 2 are not satisfied, it is determined as non-REM sleep.

  For example, the sleep stages obtained at each time point are compared and scored by +1 if both sleep stages match, and -1 if they do not match, and estimates the sum of the scores at each time point. It may be calculated as a value.

(Group by heart rate transition slope)
In DcGA, sleep-specific grouping may be considered. By performing grouping in consideration of the inclination of the heart rate transition, the accuracy of sleep stage estimation can be improved. FIG. 22 shows an example of the heart rate inclination and grouping of each subject. Here, the grouping is, for example, two groups of “I” (Increase) having a positive slope value and “D” (Decrease) having a negative slope value.

(Example of filter generation for sleep stage estimation)
DcGA can be applied to sleep stage estimation in, for example, four methods. As four methods, (a) a filter (for example, FIG. 23) based on DcGA (for one day) is used to obtain a multiband pass filter from the learning data for one day. (B) A multiband pass filter is obtained from learning data for a plurality of days by a filter based on DcGA (for a plurality of days). (C) A multiband pass filter is obtained from learning data for one day obtained by grouping filters based on DcGA (grouping for one day). Here, the probability model is updated for each of the “I” group and the “D” group. (D) A multiband pass filter is obtained from learning data for a plurality of days obtained by grouping filters based on DcGA (grouping for a plurality of days). Here, the probability model is updated for each of the “I” group and the “D” group.

[3-2-2 Sleep stage estimation device]
FIG. 24 is a block diagram illustrating an example of the sleep stage estimation device 100E according to the present example.

  As illustrated in FIG. 24, the sleep stage estimation device 100E includes a biological data acquisition unit 102, a biological data recording unit 103, a biological data storage unit 104, a filter generation unit 105, a past individual storage unit 106, a filter storage unit 107, and a specification. A frequency band extraction unit 108 and a sleep stage determination unit 109 are included.

  The biometric data acquisition unit 102 acquires data representing heartbeat and body movement from a subject using a pressure sensor or the like. What can acquire these data, without restraining a test subject is preferable. The sleep stage estimation apparatus 100E can use, for example, an air mattress type pressure sensor (a sensor that is placed under the mattress and senses heart rate, respiratory rate, and body movement in a non-invasive non-restrained state) as the biological data acquisition unit 102. .

  The biometric data recording unit 103 receives the data acquired by the biometric data acquisition unit 102 via a network, for example, and writes it in the biometric data storage unit 104. The biometric data storage unit 104 is a memory that stores data acquired by the biometric data acquisition unit 102.

  The filter generation unit 105 uses the above-described (applying DcGA to sleep stage estimation method) from the biometric data in the biometric data storage unit 104 and the data set of the past individual in the past individual storage unit 106, A filter used for estimating the sleep stage is generated and written in the filter storage unit 107.

  The past individual storage unit 106 is a memory that stores an individual data set generated in the past by the filter generation unit 105. The filter storage unit 107 is a memory that stores the filter generated by the filter generation unit 105.

  The specific frequency band extraction unit 108 receives the biometric data from the biometric data acquisition unit 1 in real time via a network, for example, and uses the filter stored in the filter storage unit 7 to perform the specific frequency band from the biometric data by fast Fourier transform. The data of the set is extracted. Here, the biometric data used is data belonging to a predetermined time window, for example, the latest one hour.

  The sleep stage determination unit 109 applies inverse fast Fourier transform to the data extracted by the specific frequency band extraction unit 8, and determines the sleep stage at that time.

  The biometric data recording unit 103, the biometric data storage unit 104, the filter generation unit 105, the past individual storage unit 106, the filter storage unit 107, the specific frequency band extraction unit 108, and the sleep stage determination unit 109 are mounted on a computer. Also good. Moreover, you may make it the biometric data recording part 103, the filter production | generation part 105, the specific frequency band extraction part 108, and the sleep stage determination part 109 function as each process part by running a computer program.

  When the sleep stage estimation device 100E is used for real-time monitoring of the sleep stages of the nursing facility residents every night, at the time of the morning of a certain day (day n), the night of day n-1 to day n The biometric data up to the morning is stored in the biometric data storage unit 104. At this time, the past individual storage unit 106 stores, for example, an individual data set generated by the filter generation unit 105 in the daytime from the (n-7) th day to the (n-1) th day. The filter storage unit 107 stores a filter generated by the filter generation unit 105 in the daytime of (n-1) th day. That is, in the monitoring of the sleep stage from the night of the (n-1) th day to the morning of the nth day, the filter generated in the daytime of the (n-1) th day is used in the specific frequency band extracting unit 108.

  In the daytime of the nth day, the filter generation unit 105 reads out the biological data from the night of the (n-1) th day to the morning of the nth day from the biological data storage unit 104, generates a new individual data set, Write to the past individual storage unit 106. At this time, the filter generation unit 105 deletes the data set of the individual generated in the daytime of the oldest n-7th day. The filter generation unit 105 further uses the individual data set generated during the daytime from the (n-6) th day to the (n-1) th day, and the individual data set generated during the daytime of the current day (the nth day). A filter is generated by DcGA and written to the filter storage unit 107 (the filter of the previous day is replaced with a new filter).

Then, from the night of the nth day to the morning of the (n + 1) th day, processing by the specific frequency band extracting unit 108 is performed using the filter generated in the daytime of the nth day stored in the filter storage unit 107, The sleep stage determination unit 109 determines the sleep stage in real time.

[3-2-3 Other example of sleep stage estimation method]
The care support system according to the present invention may use the following method as a technique for estimating the sleep stage.

(1) Rechtscha_en, A. and Kales, A. (Eds): A Manual of Standardized Terminology, Techniques and Scaring System for Sleep Stage of Human Subjects, Pulbic Health Service USGovernment Printing O_ce (1968))
Put a special instrument on the subject sleeping in the bed, acquire EEG (electroencephalogram), EMG (electromyogram), EOG (eye movement) data, and determine the sleep stage based on doctor's expertise and experience To do.

(2) Measure body data (heartbeat, respiration, body movement) without wearing a device directly, and obtain data that can be approximated with temporal transition data of the sleep stage obtained by polysomnography from the measured data (Japanese Patent Laid-Open No. 2003-079587)
For example, the sleep stage is determined from the obtained data using an unconstrained air mattress type sensor capable of measuring body data (heart rate, respiration, body movement).

(3) Sleep Stage Estimation By Evolutionary Computation (Matsushima, H., Hirose, K., Hattori, K., Sato, H., and Takadama, K.) (Using Heartbeat Data and Body-Movement, Proceeding of The 15th Asia Pacific Symposium on Intelligent and Evolutionary Systems, pp. 103-110, 2011)
The sleep stage is estimated by extracting characteristic frequency components from the heart rate data for one day obtained from the biosensor. Specifically, first, heart rate and body motion data are acquired from a subject using a pressure sensor. Next, FFT is applied to these data, and the obtained heart rate is determined. Here, the sleep stage is estimated in detail using the heartbeat and body motion data. Generally, when the observed heart rate increases and is irregular, it is the REM sleep period, and when the heart rate decreases, the non-REM sleep period is reached. In addition, when body motion occurs intensively, it is a REM sleep period, and when the magnitude and frequency of body motion decrease, it becomes a non-REM sleep period. Furthermore, fluctuations in body motion data are also observed when the sleep stage changes during sleep. For example, when moving from non-REM sleep with deep sleep to non-REM sleep with low sleep, or when shifting from Non-REM sleep to REM sleep, large body movement occurs. The sleep stage is estimated based on the tendency of heartbeat and body movement.


[4. Program and recording medium recording the program]
A program according to an embodiment of the present invention causes a computer to function as the rule generation device. That is, the computer executes any one of the above rule generation methods. In other words, the rule generation device can be a computer including a CPU, a memory, an external storage device, an input / output device, and the like, and this program is stored in the external storage device and executed.

The present invention may be a computer-readable recording medium that records the above program. The recording medium on which the program is recorded includes a flexible disk (FD), a CD-ROM (Compact Disk-ROM), a CD-R (CD Recordable), a DVD (Digital Versatile Disk), and other computer-readable media, and A semiconductor memory such as a flash memory, a RAM, and a ROM, a memory card, an HDD (Hard Disc Drive), and other computer-readable devices can be used. In addition, the recording medium on which the program is recorded may be a program (so-called difference file) installed in an existing apparatus via a network or the like.

The present invention is not limited to the above-described data elements, data and data sets, information related to the care support system, and information related to the sleep stage estimation device, for example, information communication, electronics, industry, medical care, biotechnology, agriculture, security, or others. In this field, it can be used for analyzing a plurality of data (analysis, identification, etc.).

10: Storage unit 11: Data storage unit 12: Rule set storage unit 12r: Collation set storage unit 13: AT storage unit 20: Collation set generation unit 21: Rule extraction unit 22: Rule strengthening unit 23: Rule classification unit 30: GA Application means 31: AT update unit 32: rule selection unit 33c: rule crossing unit 33v: rule mutation unit 34a: rule addition unit 34d: rule saddle unit 34r: allowable error update unit 100: rule generation device 100E: sleep stage estimation device 102: Biometric data acquisition unit (data acquisition means)
DESCRIPTION OF SYMBOLS 103: Biometric data recording part 104: Biometric data memory | storage part 105: Filter production | generation part 106: Past part individual memory | storage part 107: Filter memory | storage part 108: Specific frequency band extraction part 109: Sleep stage determination part an: AT element (n is a natural number) )
dn: Data element (n is a natural number)
rn: Rule element (n is a natural number)
AT: AT value DT: Data DS: Data set RL: Rule (knowledge, etc.)
RLg1, RLg2: General rules (general knowledge, etc.)
RLs1, RLs1: Special rules (special knowledge, etc.)
RLgc1, RLgc2: Parent rule (current generation rule)
RLgn1, RLgn2: Child rules (next generation rules)
P: Actual measurement value p: Evaluation value F: Goodness of fit ε: Error value ε0: Allowable error μ: Predetermined probability [P]: Rule set [M]: Matching set [Mp]: Positive rule set of matching set [Mn] : Negative rule set of collation set

Claims (9)

One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A data storage unit for storing a plurality of data including at least a conclusion data data element that can take one of the result values composed of continuous values representing a predetermined result under a condition to be performed;
One or more condition rule elements representing rule application conditions corresponding to the condition data element, wherein each of the condition rule elements is any of the two or more types of state values and the condition value conditions It represents one of the wildcard values indicating goodness, a rule element for conclusion representing the conclusion when the condition is satisfied as a continuous value, and a characteristic evaluation value of the rule including the accuracy of the rule A rule set storage unit for storing a set of rules including at least a rule element for characteristic evaluation;
A rule extraction unit that reads out the data from the data storage unit and extracts from the rule set storage unit the rule having the value of the condition rule element to be compared with the state value of the condition data element of the read data When,
A rule strengthening unit that updates the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so that the compatibility with the read data is higher;
Among the extracted rules, a rule classification unit that classifies a rule that satisfies a predetermined condition in which a continuous value of the rule element for conclusion is applicable to a plurality of values;
A feature extraction model generated for each piece of data such that a rule classified by the rule classification unit exists, including a model element corresponding to the conditional data element of the data, each of the model elements A feature extraction model storage unit that stores a model having a feature evaluation value that represents how much the corresponding conditional data element contributes to the accuracy of the rule;
Regarding the feature extraction model corresponding to the data read by the rule extraction unit, the condition rule elements of the rule classified by the rule classification unit having the state value are used for the characteristic evaluation. The model element is updated so that the characteristic evaluation value of the corresponding model element becomes higher as the evaluation value representing the accuracy of the rule included in the rule element is larger. For the wild card value, the corresponding model is updated. A feature extraction model update unit that updates the model element so that the feature evaluation value of the element is low;
A rule selection unit that selects two rules as parent rules from the classified rules in a given method;
As shown in the following (1) and (2), a rule crossing unit that generates two child rules by exchanging values of the condition rule elements corresponding to the selected two parent rules; ,
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and a rule addition unit that stores by adding child rules to perform the replacement when necessary in the rule set storage unit,
A series of processes by the rule extraction unit, the rule strengthening unit, the rule classification unit, the feature extraction model update unit, the rule selection unit, the rule crossing unit, and the rule addition unit are stored in the data storage unit. A rule generation device characterized by repeatedly performing a plurality of data. The characteristic evaluation rule element includes a characteristic evaluation value representing versatility of the rule,
2. The versatility evaluation value update unit that performs a process of changing an evaluation value representing versatility of the rule with a predetermined probability as a processing unit that performs the iterative process. Rule generator.   The said rule classification | category part is classified into the 1st rule set which satisfy | fills the said predetermined condition, and the 2nd rule set which does not satisfy | fill the said predetermined condition, The Claim 1 or 2 characterized by the above-mentioned. Rule generator.   When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the value of the rule element for the condition of the generated child rule Is changed to the value of the condition data element of the data, and the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is low enough not to satisfy a predetermined condition A rule mutation unit that performs a process of changing a value of the conditional rule element of the generated child rule to the wildcard value with a given probability, as a processing unit that performs the repetition process The rule generation device according to any one of claims 1 to 3, wherein Data acquisition means for acquiring the data set;
The rule generation device according to any one of claims 1 to 4,
Including
The data acquisition means acquires, as the data, data on a daily basis for a subject who provides care support,
The data storage unit stores the data of a plurality of days as the data set.
Nursing care support system characterized by that.   6. The daily data includes, as condition data elements, information related to a behavior pattern, health condition, and / or care support content of a target person who supports care. Care support system.   The care support system according to claim 5 or 6, wherein the result value is a value related to a depth of sleep. One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A rule application corresponding to the condition data element from a plurality of data, including at least a conclusion data data element that can take one of the result values consisting of continuous values representing a predetermined result under a given condition One or more condition rule elements representing the condition of the condition, wherein each of the condition rule elements is any one of the two or more types of state values and the wild card value representing the condition of the condition value At least a rule element for conclusion that represents a conclusion when the condition is satisfied as a continuous value, and a rule element for characteristic evaluation that represents a characteristic evaluation value of the rule including the accuracy of the rule A rule generation method for generating Lumpur,
Extracting the rule having the value of the condition rule element to be compared with the state value of the condition data element of one of the plurality of data from the rule set storage unit that stores the set of rules And steps to
Updating the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so as to be more compatible with the read data;
Among the extracted rules, a step of classifying a rule that satisfies a predetermined condition in which a continuous value of the conclusion rule element can correspond to a plurality of values;
A feature extraction model generated for each of the data such that the classified rule exists, including a model element corresponding to the conditional data element of the data, each of the model elements corresponding to the corresponding data element Among the models having feature evaluation values representing how much the conditional data element contributes to the accuracy of the rule, the feature extraction model corresponding to the one data is included in the classified rule. Among the condition rule elements, those having the state value have a higher feature evaluation value of the corresponding model element as the evaluation value representing the accuracy of the rule included in the characteristic evaluation rule element is larger. The model element is updated as described above, and for the wild card value, the model element is updated so that the feature evaluation value of the corresponding model element is lowered. And updating the,
Selecting two rules to be parent rules from the classified rules in a given way;
As shown in the following (1) and (2), generating two child rules by replacing values of the condition rule elements corresponding to the two selected parent rules;
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, and a step of storing by adding child rules for performing said replacement as required in the rule set storage unit,
A rule generation method characterized by repeatedly performing a series of processes in each step for a plurality of data among the plurality of data. On the computer,
One or more conditional data elements representing a target state, each of the conditional data elements being capable of taking one of two or more types of specific state values and represented by the state value A rule application corresponding to the condition data element from a plurality of data, including at least a conclusion data data element that can take one of the result values consisting of continuous values representing a predetermined result under a given condition One or more condition rule elements representing the condition of the condition, wherein each of the condition rule elements is any one of the two or more types of state values and the wild card value representing the condition of the condition value At least a rule element for conclusion that represents a conclusion when the condition is satisfied as a continuous value, and a rule element for characteristic evaluation that represents a characteristic evaluation value of the rule including the accuracy of the rule A rule generation method for generating Lumpur,
Extracting the rule having the value of the condition rule element to be compared with the state value of the condition data element of one of the plurality of data from the rule set storage unit that stores the set of rules And steps to
Updating the value of the rule element for conclusion of the extracted rule and the characteristic evaluation value of the rule element for characteristic evaluation so as to be more compatible with the read data;
Among the extracted rules, a step of classifying a rule that satisfies a predetermined condition in which a continuous value of the conclusion rule element can correspond to a plurality of values;
A feature extraction model generated for each of the data such that the classified rule exists, including a model element corresponding to the conditional data element of the data, each of the model elements corresponding to the corresponding data element Among the models having feature evaluation values representing how much the conditional data element contributes to the accuracy of the rule, the feature extraction model corresponding to the one data is included in the classified rule. Among the condition rule elements, those having the state value have a higher feature evaluation value of the corresponding model element as the evaluation value representing the accuracy of the rule included in the characteristic evaluation rule element is larger. The model element is updated as described above, and for the wild card value, the model element is updated so that the feature evaluation value of the corresponding model element is lowered. And updating the,
Selecting two rules to be parent rules from the classified rules in a given way;
As shown in the following (1) and (2), generating two child rules by replacing values of the condition rule elements corresponding to the two selected parent rules;
(1) When the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit is high enough to satisfy a predetermined condition, the parent rule corresponding to the model element If any of the condition rule elements matches the state value of the condition data element of the data corresponding to the model element, the value is the value of the corresponding condition rule element of one of the child rules. (2) To the extent that the feature evaluation value of the model element of the feature extraction model corresponding to the data read by the rule extraction unit does not satisfy the predetermined condition If any of the condition rule elements of the parent rule corresponding to the model element has the wildcard value, the condition rule corresponding to the one child rule is determined. As the value of the model element is the wildcard value, to execute the steps of storing by adding child rules for performing said replacement as required in the rule set storage unit,
A rule generation program that repeatedly executes a series of processes in each step for a plurality of data among the plurality of data.

Images