JPH0736700A - Estimated rule automatic generating method and device therefor - Google Patents

Abstract

PURPOSE:To provide a user original estimated rule automatic generating method and device of a time series data estimating device. CONSTITUTION:This method and device therefor is composed of a data control function 207 of the input/output of data, a fuzzy inference function 208 comparing data with the time series fluctuation pattern defined by the antecedent part of a fuzzy rule and calculating the adaptability of the fuzzy rule and a fuzzy rule generation function 209 evaluating how much time series data fluctuates by an average value, a standard deviation and a histogram, etc., after the time for the sampled time series data and the time for the extracted time series data and setting the parameter of the succeeding part of a new fuzzy rule when the value of the time series data has a tendencey to rise/fall.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a prediction rule automatic generation method and apparatus in a system for predicting numerical values,
For example, the present invention relates to a method and apparatus for automatically generating a price prediction rule in a system that predicts market prices of stocks and exchange rates.

[0002]

2. Description of the Related Art As a conventional technique relating to a rule automatic generation method and device in a system for predicting a numerical value by using fuzzy inference, there is, for example, Japanese Patent Laid-Open No. 4-364536, and there is an output of fuzzy inference and input / output data Based on the above, while adjusting the parameters of the fuzzy inference rule by the parameter adjustment unit, the inference error calculation unit is used to examine the inference error and the change amount of the inference error. It is proposed to obtain fuzzy rules automatically by generating fuzzy inference rules. In addition, JP-A-4-36904
There is 0, the data collected by the process data collection unit is processed by the data processing unit and sent to the process database generation unit, and the database generated by this generation unit is sent to the fuzzy clustered database generation unit together with the fuzzy evaluation function. Then, the state space satisfying the fuzzy evaluation function is detected, and the fuzzy clustered database generated by the generator is sent to the fuzzy rule generator and the membership function generator, where the fuzzy rules and membership functions are obtained. Is proposed. Further, in Japanese Patent Laid-Open No. 372045/1992, a fuzzy inference prediction unit obtains a prediction value by referring to a knowledge base, a comparator obtains a difference between an actual measurement value and a prediction value, and a knowledge base correction unit obtains this prediction error. Based on this, we propose to modify the center of gravity of the consequent membership function of the knowledge base and the weight of the rule.

[0003]

PROBLEM TO BE SOLVED BY THE INVENTION Automatic generation of fuzzy inference rules from each application while dynamically generating fuzzy inference rules and adjusting the parameter values thereof to automatically obtain fuzzy inference rules, fuzzy rules It is known that the membership function can be automatically generated, and the center of gravity of the consequent membership function of the knowledge base and the weight of the rule can be modified based on the prediction error.

However, the conventional technique has the following problems.

(1) Even if different people use the apparatus, the fuzzy rules generated are the same. You need to create your own fuzzy rules for your system. For example, different stock dealers and traders have different know-how in forecasting time series data of stocks and foreign exchange. We have our own know-how to keep up with other dealers and traders. However, in the conventional fuzzy rule generation methods, the same fuzzy rule is generated for all of them, and it is not possible to use unique know-how.

(2) Since the center of gravity value of the membership function and the weight of the rule in the consequent part are corrected by comparing the error between the measured value and the predicted value, the measured value is easily affected by irregular data. . In particular, the actual measured values of time series data such as stocks and exchanges often include large disturbances due to the movement of the world.

An object of the present invention is to provide a method and apparatus for automatically generating a prediction rule unique to a user of a time series data prediction apparatus.

Another object of the present invention is to provide a method for correcting a consequent part parameter which is less likely to be affected by irregular data disturbance.

Other objects of the present invention can be easily guessed by those skilled in the art from the present specification and the accompanying drawings.

[0010]

[Means for Solving the Problems] The above object is to discover a combination of user-specific fuzzy rules registered in advance by the user, which is important for time series data prediction, and define the combination as a rule antecedent part. This is accomplished by a predictive rule generation method and apparatus that generates rules. In particular,
It is realized by an apparatus characterized by the following method.

(1) Data Management / Processing Unit The data management / processing unit inputs data in real time, stores the data in a time-series data file, and retrieves the previously stored data as needed. If processing such as a moving average line is required when predicting time-series data, it is processed by the data management and processing functions and the data is passed to the fuzzy inference unit.

(2) Fuzzy reasoning unit The fuzzy reasoning unit compares the data passed from the data management / processing function with the time-series fluctuation pattern defined in the antecedent part of the fuzzy rule, and determines the degree of conformity of the fuzzy rule. To calculate. The fuzzy rules are stored in advance in a fuzzy rule file, and the fuzzy rules are searched and retrieved as needed. The synthesis result of the consequent part of the fuzzy rule is calculated from the goodness of fit of each rule.

(3) Rule Generation Unit The fuzzy rule generation unit symbolizes the time series data into a numerical sequence. Specifically, the time series data is represented by a numerical sequence in which the fuzzy rule name that matches a certain time and the matching degree of the rule are arranged. Using the symbolized fuzzy rule name and the numerical sequence of its fitness as a sample, the time series data for each other time is symbolized, and the degree of similarity between the sample and the numerical sequence at another time is calculated. The time of the time series data having a degree of similarity larger than a preset value is extracted. Evaluate how much the time-series data fluctuates after the time of the sample time-series data and the time of the extracted time-series data with an average value, a standard deviation, a histogram, or the like. After the evaluation result, when the value of the time-series data tends to rise / fall, a parameter of the consequent part of the new fuzzy rule is set based on the histogram.

[0014]

According to the present invention, it is possible to discover a combination of user-specific fuzzy rules registered in advance that is important for time series data prediction, and generate a new rule by defining the combination as a rule antecedent part. It allows you to create your own fuzzy rules. By shaving both ends of the measured value data, it is possible to accurately capture the dense part of the measured value data and set it as a consequent part parameter, so it is possible to avoid disturbance of irregular data To

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT A first embodiment of the present invention will be described below with reference to FIGS.
Will be explained.

FIG. 2 shows the structure of the predictive rule automatic generation apparatus of this embodiment. A computer 201 executes a program for automatically generating rules for predicting time series data. Two
Reference numeral 02 is a display device. Reference numerals 203 and 204 are a keyboard and a mouse as input devices for information necessary for automatically creating a prediction rule. A data file 205 stores time series data. A fuzzy rule file 206 stores the fuzzy rules. Although the files 205 and 206 are separate files for convenience of explanation,
The number of storage devices may be one.

The computer 201 has nine functions for time series data prediction and automatic prediction rule generation. Each function is achieved by a program executed by the computer 201. The program is roughly divided into four parts (modules)
Composed of. A data management unit 207 that stores, retrieves, and manages time-series data, a fuzzy inference unit 208 that calculates a goodness of fit of rules and synthesizes a membership function of a consequent part,
Fuzzy rule generation unit 20 for generating fuzzy rules
9. An execution control unit 210 that controls the execution of each of these units.

The data management unit 207 realizes the following functions. The first is a function 211 for inputting data in real time, storing the data in the time-series data file 205, and taking out previously stored data as needed. When processing such as a moving average line is required when predicting time series data, the data management and processing function 211
And passes the data to the rule antecedent part conformance calculation function 212.

The fuzzy inference unit 208 realizes the following functions. The first is a function 212 that compares the data passed from the data management / processing function with the time-series fluctuation pattern defined in the antecedent part of the fuzzy rule to calculate the degree of conformity of the fuzzy rule. The fuzzy rules are stored in advance in a fuzzy rule file, and the function 212 retrieves and extracts the fuzzy rules as needed. Second is function 2
This is a function 213 for calculating the result of synthesis of the consequent part of the fuzzy rule from the goodness of fit of each rule calculated in 12.

The fuzzy rule generator 209 realizes the following functions. The first is a function 214 for symbolizing time series data into a numerical sequence. Specifically, the time series data is represented by a numerical sequence in which the fuzzy rule name that matches a certain time and the matching degree of the rule are arranged. Second, using the fuzzy rule name and the numerical sequence that is the degree of conformity symbolized by the function 214 as a sample, the time series data for each other time is symbolized, and the degree of similarity between the sample and the numerical sequence at another time is determined. Calculation function 2
It is 15. The third is a function 216 for extracting the time of the time series data having a degree of similarity larger than a preset value. The fourth is the time of the sample time series data and the function 21
This is a function 217 that evaluates how much the time-series data has changed after the time of the time-series data extracted in 6 with an average value, a standard deviation, a histogram, and the like. Fifth, function 2
This is a function 218 for setting the parameter of the consequent part of a new fuzzy rule based on the histogram created by the function 217 when the value of the time series data tends to rise / fall as a result of evaluation in 17.

The execution control unit 210 centrally controls the execution of each of the above units.

The details of each function and the operation description will be described below.

FIG. 3A shows a fuzzy rule file 2
It is a block diagram of the record 31 of the fuzzy rule antecedent part information stored in 06. The fuzzy rule antecedent information of the record 31 includes seven items. They are,
Rule number 31 used as a search key for record 31
1, time series data 312, relative time point 313, comparison operator 314, time series data name 315, relative time point 31
6 is a fuzzy value 317 which is a value for expressing the membership function by a numerical value.

FIG. 3B shows the fuzzy rule file 2.
It is a block diagram of the record 32 of the fuzzy rule consequent part information stored in 06. The fuzzy rule consequent part information of the record 32 includes three items. They are,
Rule number 32 used as a search key for record 32
1, a predicted value 322 as a change value of time series data when the rule is matched, and a predicted hit rate 323.

FIG. 4 is a structural diagram of a record 41 of time series data information stored in the time series data file 205. They are the date 41 that represents the time of the time series data.
1 and the time series data value 412.

FIG. 5 is a structural diagram of a record 51 of time series data processing information stored in the time series data file 205. 511 is a date representing the time of the processing time series data, 512 is a 3-day moving average line represented by MA3, and 512 is a 6-day moving average line represented by MA6. Although omitted in FIG. 5,
Hereinafter, the 9-day moving average line is represented by MA9, and the 12-day moving average line is represented by MA12.

FIG. 1 schematically shows a predictive rule automatic generation method in which the flow of functions realized by a program in the predictive rule automatic generating apparatus of the present embodiment is schematically shown along the flow of processing of the execution control unit 210. It is a plot chart.

Step 1: In time unit of time series data,
The time-series data is compared with the time-series data fluctuation pattern defined in advance in the antecedent part of the fuzzy rule to calculate the goodness of fit of the fuzzy rule. A numerical sequence is created by combining the fuzzy rule names that match here and the matching degrees of the rules. This numerical sequence serves as a symbol representing the time series data at that time. Perform this task in whole or in a specific range.

Step 2: Time series data at an arbitrary time is used as a sample. The similarity between the numerical sequence of goodness of fit as a sample and the numerical sequence of goodness of fit of time series data at another time is calculated. Details will be described later.

Step 3: Extract the time of the time series data having the degree of similarity equal to or more than a preset threshold value.

Step 4: Evaluate the time series data fluctuation value after the extracted time.

Step 5: If, as a result of the evaluation, the changed values have variations, then the processing of Step 2 is performed. If the variation is small and the change in the data tends to rise or fall, then the process of step 6 is performed.
It is a well-known fact to those skilled in the statistical analysis whether the variation in the price fluctuation value is small or large, and any variation determination method may be applied to the present invention, and therefore the description thereof is omitted.

Step 6: The parameters of the consequent part are created from the changing tendency of the data. Details will be described later.

Step 7: A rule name that matches the time-series data as a sample is registered in the antecedent part, and the parameter created in step 6 is defined in the antecedent part to make a new rule.
It is stored in the fuzzy rule file 206.

The above is the outline of the predictive rule automatic generation method according to the present invention.

Next, a method of predicting time series data using a prediction rule will be described in detail with reference to the flowchart shown in FIG.

Step 11: Data management / processing section 207
To time-series data from a certain time point to the current time point in units of hours, minutes, etc. to the fuzzy inference unit 208.

Step 12: Using the data input in step 11, the suitability of all or specific fuzzy rules stored in the fuzzy rule file 206 is calculated.

Step 13: Based on the goodness of fit calculated in Step 12, a time series data fluctuation predicted value predefined in the consequent part of the fuzzy rule is synthesized, and from the sum of the fluctuation predicted value and the current time series data value. , Calculate the value of the time-series data at the next point in time.

FIG. 7 shows an example of rules used in this time series data prediction method.

FIG. 7A shows a rule name 701, rule antecedent parts 702 to 708, and rule a consequent part 70.
9 to 711. Antecedent section 702 of rule a
Reference numeral 708 indicates a three-day moving average line (MA3) of time series data 702, reference time points 703 and 706, a comparison operator 704, and a six-day moving average line (MA6) of time series data.
Is represented by 705, and the slope of the membership function is represented by a numerical value 708. Consequent part 709 to 71 of rule a
1 is composed of a lower limit value 709 of the expected time-series data fluctuation value, an upper limit value 710 of the expected time-series data fluctuation value, and its hit rate 711. Fig. 7 shows the consequent part of the fuzzy rules.
As shown in (b), the time series data fluctuation prediction value may be one.

A specific method of calculating the degree of conformity of fuzzy rules will be described with reference to FIG.

FIG. 8A shows a date 801 which is the time of the time series data, a 3-day moving average line 802, and a reference time point 8 of the time.
It is composed of 03. FIG. 8B shows the condition clauses 804 to 809 of the antecedent part of the fuzzy rule and the membership function 81.
0 to 814. The conditional clauses 804 to 809 in the antecedent part of the fuzzy rule are the 3-day moving average (MA3) 8
04, 804 showing the time point of the data, 805, the comparison operator 806, 808 showing the time point of the data of the 3-day moving average line (MA3) 807, 807, showing the membership function 80
It is composed of 9. Membership functions 810-814
Indicates the vertical axis 810 indicating the grade, the horizontal axis 811 indicating the time-series data value, and the conditional section 804 to the antecedent part of the fuzzy rule.
Membership functions 812, 807 and 8 obtained from 809
It is composed of the values 813 and 804 obtained from 08 and the value 814 obtained from 805.

Conditional clause 804 of the antecedent part of the fuzzy rule
808 means the 3-day moving average (MA3)
Is equal to or more than the 3-day moving average line one time point before the reference time point. In this example, the value 81
Since 4 is larger, the conformance of the conditional clauses 804 to 809 in the antecedent part of the fuzzy rule is 1.0. One or more of the conditional clauses 804 to 809 in the antecedent part of the fuzzy rule are used as the conditional clause of the rule, and the logical product of the conditional clauses is the fitness of the rule.

The synthetic calculation of the consequent part of the fuzzy rule with the goodness of fit greater than 0 is easy, for example, according to the method described in Japanese Patent Application No. 4-301724 “Defuzzification Method” of the applicant's patent application. Since it can be carried out, the description is omitted.

Next, a method of calculating the degree of similarity of digitized time series data will be described in detail with reference to the flowchart shown in FIG.

Step 21: Extract a fuzzy rule suitable for the time-series data used as the sample. The suitability of the rule is arranged and expressed by a numerical sequence. FIG. 10A shows an example of this numerical sequence. FIG. 10A shows a rule name 1001,
It is composed of a goodness of fit 1002. In this example, the fuzzy rule that fits the time series data at the time of sampling is
Only the rule a, the rule b, and the rule have the matching degrees of 0.8, 1.0, and 1.0.

Step 22: Calculate the goodness of fit of the fuzzy rule extracted in step 21 at another time point of the sample. If the example of FIG. 10A is used, the suitability of the rules with respect to the rules a, b, and c is calculated.

Step 23: Calculate the degree of similarity between the numerical value sequence of the goodness of fit at each time point obtained in step 22 and the numerical value sequence of the goodness of fit of the rule with respect to the time-series data sampled in step 21. An equation for calculating the similarity is shown in FIG. FIG. 10B shows the similarity 1003, the average value 1004 of the goodness of fit that matches the sample time series data, the variance 1005 of the goodness of fit that matches the sample time series data, and the similarity to the sample time series data. The average value 1006 of the goodness of fit that is suitable for the time series data for calculating the degree, the variance 1007 of the goodness of fit that is suitable for the time series data for calculating the similarity to the time series data of the sample, Goodness of fit 1
008, and the number of fuzzy rules 1009 having a goodness of fit with respect to the time-series data as a sample is larger than 0. FIG. 10C shows an example of the degree of similarity. Figure 10
(C) shows the date 1010 as the time of the time series data,
It is composed of the similarity 1011 at each time point.

Next, the method of calculating the parameters of the consequent part from the time series data fluctuation value after the similar time series data fluctuation pattern will be described in detail with reference to the flowchart shown in FIG.

Step 31: A histogram is created with time-series data fluctuation values after time-series data whose similarity at each time point is equal to or greater than a preset value. FIG. 12A shows an example of the histogram. FIG. 12A shows a histogram frequency 1201, a time series data variation value 1202,
It is composed of a histogram 1203.

Step 32: Connect each vertex of the histogram with a line segment. An example is shown in FIG. Reference numeral 1204 indicates a line segment connecting the vertices.

Step 33: Calculate the area of the portion surrounded by the vertical line and the line segment of the histogram. An example is shown in FIG. 1205 is the name of each area. 1206 is the area surrounded by the line segment 1205 and the horizontal axis 1202.

Step 34: The size relationship between the area reduced by one block at the left end and the area reduced by one block at the right end is checked, and the larger one is used in the next process. FIG. 13 shows a calculation example. In FIG. 13, the leftmost 1 block of the area 1206 is deleted 1207, the rightmost 1 block of the area 1206 is deleted 1208, the leftmost 1 block of the area 1208 is deleted 1209, and the rightmost 1 block of the area 1208 is deleted 1.
It is composed of 210. Area 1206 to area 120
7, 1208 were calculated, leaving a large area 1208. The areas 1209 and 1210 were calculated from the area 1208, leaving the large area 1209.

Step 35: Repeat Step 34,
When the remaining area becomes smaller than the preset value, the repetition of step 34 is ended.

Step 36: The values at both ends of the area remaining by the processing up to step 35 are set to the upper limit value and the lower limit value of the consequent part of the fuzzy rule, and the area is set to the predictive hit probability of the consequent part. . An example is shown in FIG. FIG. 14 shows the lower limit 140
1, an upper limit 1402, and an area 1403 of 4 blocks. Area calculated as a result of calculation up to step 35 is 14
03 remains, and the upper limit 1402 and lower limit 140 of this area
1 is the predicted value, and the area 1403 is the predictive hit ratio. Here, the predicted value is obtained as a binary value such as an upper limit value and a lower limit value, but there is also a method of obtaining one predicted value by averaging the upper limit value and the lower limit value.

Finally, the features of the rules generated by this method and the inference method will be described in detail with reference to FIG.
FIG. 15 shows the names 1501-1503 of rules defined by the user or generated by this method, step 3.
It is composed of consequent parts 1504 to 1506 of the rule obtained in 6. FIG. 15 shows the rules created in steps 1 to 7. Rules that match the sample time are rule a1501, rule b1502, and rule c1503, respectively.
Is. The inference is a method in which first the fitness of rule a, rule b, and rule c is calculated, and then the logical sum of three rules is calculated.

[0058]

According to the present invention, a user-registered fuzzy rule, which is unique to the user, is found to be important in time series data prediction, and a new rule is generated by defining the combination in the rule antecedent part. Therefore, there is an effect that a user-specific fuzzy rule can be created. By scraping both ends of the measured data,
Since the dense portion of the actual measurement value data can be accurately captured and set as the consequent part parameter, there is an effect that the disturbance of irregular data is less likely to occur.

[Brief description of drawings]

FIG. 1 is a flowchart showing a basic operation according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a computer hardware configuration and each program applied in the computer.

FIG. 3 is a configuration diagram of a fuzzy rule antecedent part and consequent part information record.

FIG. 4 is a configuration diagram of a time series data information record.

FIG. 5 is a configuration diagram of a time series data processing information record.

FIG. 6 is a flowchart showing a time series data prediction method.

FIG. 7 is a diagram of an example of rules created and registered by a user with an editor.

FIG. 8 is a diagram illustrating calculation of a membership function of a rule and fitness.

FIG. 9 is a flowchart showing a process of calculating the degree of similarity of time-series data that has been digitized.

FIG. 10 is a diagram showing a calculation process of similarity.

FIG. 11 is a flowchart showing a process of calculating parameters of a consequent part from time series fluctuation values after similar time series fluctuation patterns.

FIG. 12 is a diagram showing a process of calculating parameters of a consequent part from time series fluctuation values after similar time series fluctuation patterns. (A) It is a histogram showing the variation of the value of the time series data. (B) It is the figure which connected the vertex of the histogram created by (a) with the line segment. (C) is a diagram in which the area of the portion surrounded by the vertical line of the histogram and the line created in (b) is named.

FIG. 13 is a diagram showing a process of calculating parameters of a consequent part from time-series fluctuation values after similar time-series fluctuation patterns. This is a visual representation of the calculation process on the way.

FIG. 14 is a diagram showing a process of calculating parameters of a consequent part from time-series variation values after similar time-series variation patterns. FIG. 7 is a diagram of the final result.

FIG. 15 is an example of a rule created by a rule generation unit.

[Explanation of symbols]

201 ... Calculator, 202 ... Display, 203 ... Keyboard, 204 ... Mouse, 205 ... Time series data file, 206 ... Fuzzy rule file, 207 ... Data management program, 208 ... Fuzzy inference program, 2
09 ... Fuzzy rule generation unit, 210 ... Execution control unit

Claims (4)

[Claims] 1. A step of symbolizing time-series data, a step of calculating the respective similarities of the symbolized time-series data, and a step of calculating time-series data of the time-series data having a degree of similarity equal to or greater than a threshold value. An automatic prediction rule generation method comprising: a step of evaluating fluctuations in time series data; and a step of setting parameters of a consequent part of a fuzzy rule from a fluctuation distribution of time series data. 2. A step of creating a histogram using the variation values of time-series data, a step of connecting each vertex of the histogram with a line segment, and a step of calculating an area of a portion surrounded by a vertical line and a line segment of the histogram. , A step of comparing the area reduced by one block on the left end with the size of the area reduced by one block on the right end, and adopting the larger one in the next processing, a step of evaluating the size of the area in the step, and an area 2. The method for automatically generating a prediction rule according to claim 1, further comprising the step of setting both ends and the area value in the consequent part of the fuzzy rule. 3. A means for symbolizing time-series data, a means for calculating respective similarities of the symbolized time-series data, and a time-series data after the time of similarity of the time-series data having a certain threshold value or more. An automatic prediction rule generation device comprising means for evaluating fluctuations in time-series data and means for setting parameters of a consequent part of a fuzzy rule from fluctuation distributions of time-series data. 4. A means for creating a histogram by the variation value of time-series data, a means for connecting each vertex of the histogram with a line segment, and a means for calculating the area of a portion surrounded by the vertical line and the line segment of the histogram. , A means for comparing the area reduced by one block on the left end and the area reduced by one block on the right end and adopting the larger one for the next processing, and means for evaluating the size of the area by the means, 4. The predictive rule automatic generation device according to claim 3, further comprising means for setting both ends and the area value in the consequent part of the fuzzy rule.