Rule generation system and method of generating rule

• US 5,802,509 A
• Filed: 07/11/1997
• Issued: 09/01/1998
• Est. Priority Date: 09/21/1993
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Ichimori, Toshihide, Maeda, Akira, Ashida, Hitoshi, Hirai, Chiaki, Takahashi, Yori
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Shah, Sanjiv
• Application Number:
  US08/893,422
• Time in Patent Office: 417 Days
• Field of Search: 395/75, 395/51, 395/72, 395/68, 395/22, 364/157, 364/474.23
• US Class Current: 706/59
• CPC Class Codes: G06N 5/048 Fuzzy inferencing

Abstract

A rule generation apparatus includes a label presenter in which when producing from training data including a set of specific values related to input and output variables rules representing input/output relationships between the input and output variables, numeric data of the training data is converted into categorical data expressed by symbols to generate an instance table, an RI device for extracting rules from the instance table, and a rule converter for converting the extracted rules into fuzzy rules. When the training data is divided to be distributively stored in a plurality of server processors, label assignment is conducted by each server processor such that a client processor later combines instance tables with each other to achieve rule induction and conversion.

Claims

1. A rule generating system for generating rules from training data including a set of specific values related to input and output variables, the rules representing input/output relationships between the input and output variables, the training data including both numeric data and symbol data, comprising:

division method defining means for defining a division method indicating a number of intervals into which a domain is divided and categorical data expressed by symbols corresponding to the respective intervals for the variables corresponding to the numeric data in the training data;

label presenting means for converting all the numeric data of the training data into the categorical data in accordance with the defined division method, results of the conversion producing an instance table, and for generating division information indicating dividing positions of the domains for the respective numeric data;

rule induction means for extracting rules from the instance table, the rules being expressed by the categorical data;

membership function generating means for generating a membership function using the division information generated by the label presenting means; and

fuzzy rule synthesizing means for synthesizing fuzzy rules from the rules inducted by the rule induction means and the membership function generated by the membership function generating means;

wherein the fuzzy rules synthesized by the fuzzy rule synthesizing means outputs inference results in the categorical data form.

2. A rule generation system according to claim 1, wherein the label presenting means divides for each of the variables a definition domain of the variable at predetermined points into several intervals, assigns labels respectively to the intervals, and converts numeric data of the training data into categorical data corresponding to an interval to which the numeric data belongs.

3. A rule generation system according to claim 2, wherein the label presenting means decides in the domain division the dividing points to produce the intervals such that the intervals include an equal number of data.

4. A rule generation system according to claim 2, wherein the label presenting means obtains a mean value and a value of variance of the numeric data which belong to numeric items for each of the variables, thereby achieving the domain division.

5. A rule generation system according to claim 2, wherein the label presenting means divides the definition domain at points inputted by the user.

6. A rule generation system according to claim 2, wherein the label presenting means divides the definition domain at points deviding the domain equally.

7. A rule generation system according to claim 1, wherein the label presenting means conducts the categorical data conversion using dividing methods which are mutually different for groups, each of the groups including a plurality of variables having a numeric attributes in the training data.

8. A rule generation system according to claim 1, wherein the rule induction means uses in the rule induction, in place of a number of instances, a sum of weights representing reliability of respective instances.

9. A rule generating system for generating rules from training data including a set of specific values related to input and output variables, the rules representing input/output relationships between the input and output variables, the training data including both numeric data and symbol data, comprising:

division method defining means for defining a division method indicating a number of intervals into which a domain is divided and categorical data expressed by symbols corresponding to the respective intervals for the variables corresponding to the numeric data in the training data;

label presenting means for converting all the numeric data of the training data into the categorical data in accordance with the defined division method, results of the conversion producing an instance table, and for generating division information indicating dividing positions of the domains for the respective numeric data;

rule induction means for extracting rules from the instance table, the rules being expressed by the categorical data;

membership function generating means for generating a membership function using the division information generated by the label presenting means; and

fuzzy rule synthesizing means for synthesizing fuzzy rules from the rules inducted by the rule induction means and the membership function generated by the membership function generating means;

wherein the fuzzy rules synthesized by the fuzzy rule synthesizing means outputs inference results in the categorical data form;

wherein the label presenting means divides for each of the variables a definition domain of the variable at predetermined points into several intervals, assigns labels respectively to the intervals, and converts numeric data of the training data into categorical data corresponding to an interval to which the numeric data belongs;

wherein the label presenting means divides the definition domain at points inputted by the user; and

wherein the label presenting means includes;

means for displaying a histogram of the numeric data for each of the variables;

input means for operating a pointer superimposed onto the presented histogram; and

means for indicating decision of the dividing points by the pointer.

10. A rule generating apparatus for generating rules from training data including a set of specific values related to input and output variables, the rules representing input/output relationships between the input and output variables, the training data including both numeric data and symbol data, comprising:

a client processor; and

a plurality of server processors connected to the client processor;

the client processor comprising;

division method defining means for defining a division method indicating a number of intervals into which a domain is divided and categorical data expressed by symbols corresponding to the respective intervals for the variables corresponding to the numeric data in the training data;

label presenting means for converting all the numeric data of the training data into the categorical data in accordance with the defined division method, results of the conversion producing a label code correspondence table, and for generating division information indicating dividing positions of the domains for the respective numeric data;

transmitting means for transmitting the label code correspondence table to the server processors;

means for combining instance tables sent from the server processors;

a rule induction device for extracting rules from each of the instance tables, the rules being expressed by the categorical data;

membership function generating means for generating a membership function using the division information generated by the label presenting means; and

fuzzy rule synthesizing means for synthesizing fuzzy rules from the rules inducted by the rule induction means and the membership function generated by the membership function generating means, the fuzzy rules synthesized by the fuzzy rule synthesizing means outputs inference results in the categorical data form; and

each of the server processors including;

a data base for storing therein a portion of the training data;

label presenting means for converting the numeric data and symbol data of the training data into the label codes and thereby producing an instance table; and

transmitting means for transmitting the instance table to the client processor.

11. A rule generation apparatus according to claim 10, wherein the label presenting means of the respective plural server processors achieve simultaneous processing.

12. A rule generation apparatus according to claim 10, wherein the label code correspondence table specifies, for a data item having a value of a real number or an integer, correspondences between an upper limit value and a lower limit value and a label code and specifies for a data item having a value in the form of a character string conversion of the character string.

13. A rule generating method for generating rules from training data including a set of specific values related to input and output variables, the rules representing input/output relationships between the input and output variables, the training data including both numeric data and symbol data, comprising the steps of:

defining a division method indicating a number of intervals into which a domain is divided and categorical data expressed by symbols corresponding to the respective intervals for the variables corresponding to the numeric data in the training data;

converting all the numeric data of the training data into the categorical data in accordance with the division method set forth in the defining step, results of the conversion producing an instance table;

generating division information indicating dividing positions of the domains for the respective numeric data;

extracting rules from the instance table, the rules being expressed by the categorical data;

generating a membership function using the information generated in the converting step; and

synthesizing fuzzy rules from the extracted rules and the generated membership function, the fuzzy rules containing inference results in the categorical data form.

14. A rule generating method according to claim 13,

wherein the synthesizing step includes:

transforming the extracted rules expressed by the categorical data into a fuzzy rule format; and

wherein the step of generating a membership function includes;

assigning the membership function to a respective label of the rule expressed by the categorical data.