JPS62208133A - Inference system - Google Patents

Abstract

PURPOSE:To reduce the volume of calculation to obtain an inference result within a short period and to reduce the storage capacity to required minimum by executing the inference directly when the same experience as the inference did not exist in the past, and automatically storing only an effective inference result. CONSTITUTION:An inference mechanism part executes real inference, a retrieving mechanism pat retrieves data registered in an experience table and a detecting mechanism part detects the degree of jam of the inference time and the inference mechanism and that of an experience table detecting time and the retrieving mechanism. A control table is provided with a decision table, an experience table, an environment table, and a working table. When an inference is applied, necessary decision is executed and whether the actual inference is to be executed or the inference result experiences in the past is to be utilized is determined. At the time of execution of the inference, necessary decision is executed and the inference result or the like is additionally registered in the experience table in accordance with the decided result.

Description

[Detailed Description of the Invention] [Industrial Application Field] This invention is applicable to cases where it is necessary to perform inference of the same content multiple times, the inference time required for the inference is long, and the amount of calculation is large. Since this method aims to shorten the time required to obtain an inference result and reduce the amount of calculation, it can be used in a system that obtains a result by inference and is used by a large number of users who meet these conditions.

[Conventional technology]

Conventional techniques perform inference every time an inference is given, so inferences that require a long time require a long inference time each time they are performed, and many calculations are required for inference. In addition, in conventional technology, decisions are made strategically during execution based on information such as the inference execution time and the current environment, to use past inference experience, or to decide whether to memorize inference results. There was no.

[Problem that the invention seeks to solve]

In conventional technology, even if the same pattern of inference has been processed in the past, the same inference is performed every time, and the same inference time and calculation amount are required each time. For this reason, when performing inferences that take a long inference time, it takes a long time each time to obtain an inference result, and calculations necessary for inference are also required each time. For this reason, when providing processing that requires advanced inference to a large number of users, the response time and total 37. There was a problem in terms of quantity. This invention solves these problems.

[Means for solving problems]

Given an inference, determine whether it is possible to obtain an inference result in a shorter time by directly inferring it, or whether it is faster to obtain an inference result by using inference results experienced in the past. This method uses a method that allows the user to make judgments at different times and obtain inference results in a short period of time based on the judgments made. If the person has not had the same experience as the inference in the past, the inference is made directly, the result of the inference is determined whether it is worth remembering, and only those that are worth remembering are automatically memorized.

[Effect]

Compared to performing inference every time, which requires a long inference time,
If you perform the inference once, the results of the previous inference can be used from the second time onward, so you can obtain the inference result in a shorter time than actually performing the inference, and the amount of calculation required for inference is also reduced. Also, when inference results are stored, only those that are worth recording are stored, so only the minimum necessary storage capacity is required.

[Embodiments of the invention]

Examples of the present invention are shown below in Figures 1, 2, 3, and 4.
This will be explained in detail using figures. FIG. 1 is a diagram showing the configuration of functions necessary for performing inference using this method. Figure 1 (a
) is a diagram showing the functional components necessary for performing inference using this method, which includes an inference mechanism section, an experience table search mechanism section, a detection mechanism section, and a management table. The inference mechanism section is for performing actual inference, and the experience table search mechanism section is for searching data registered in the experience table. The detection mechanism section is for detecting the inference time, the degree of congestion of the inference mechanism, the detection time of the experience table, and the degree of congestion of its search mechanism. Management tables include registration table, judgment table, experience table,
It consists of an environment table and a work table, and is used to store information necessary to determine how to perform inference. FIG. 1(b) shows an outline of the operation of the inference device. When an inference is given, the necessary judgment is made and the inference mechanism part decides whether to perform the actual inference or to search for data in the experience table and use inference results experienced in the past. If you make inferences, make the necessary judgments;
Accordingly, inference results and the like are additionally registered in the experience table.

FIG. 2 shows the contents of the management table. Second
Figure (a) shows a registration table, and the first item contains numbers and inference names (hereinafter referred to as inference IDs) for identifying attributes such as inference types (hereinafter referred to as inference types). The second term is for specifying that each inference of that type is to be processed by directly performing the actual inference every time. The third term is the average value (hereinafter referred to as the standard inference time) of the inference time of the inference performed in the past for that type of inference (hereinafter referred to as the standard inference time) converted into the inference time in the congestion situation that the inference mechanism unit uses as the standard. It is used to remember the average inference time. The fourth term is for specifying how many times the standard inference average time in the third term is to be averaged. This table is managed using inference IDs, and there is only one data set (record) having the same inference ■D.

FIG. 2(b) shows a determination table, which is used to determine the reference average inference time of the third term of the registration table. The first term is used to store the guess IID. Second
The term is the number of times the inference ID is used to find the standard inference average time.
This is to remember whether the inference was made. The third term is for storing the sum total of the standard inference times of the inferences performed the number of times indicated in the second term in order to obtain the standard inference average time. This table is managed using inference IDs, and there is only one record having the same inference ID.

FIG. 2(C) is an experience table for registering inference results determined to be necessary among inferences performed in the past, and the first item is for registering an inference ID. The second term is for registering input data for the inference (input data may be singular or plural). The third term is for registering the inference result (the inference result may be singular or plural). This table does not have multiple entries with the same content for all three items. The above three types of tables are composed of zero or more records, and the number of records and their contents change as necessary by executing the inference process.

For unused records, a value indicating unregistered is given to each item of the record so that it can be known that the records are unused.

FIG. 2(d) is an environment table for registering the current status of the inference device, etc., and the first item is for registering the congestion status of the inference mechanism section. The second term is for registering the congestion status of the search mechanism section that searches the experience table. The third term is the search time (converted to the value when the search mechanism section that searches the experience table is in a standard crowded situation).
This is for registering the average value (hereinafter referred to as the standard search average time). The fourth item is used to register this value for use when there is no specification in the fourth item of the registration table. The fifth term is for registering how many times the average is to be calculated when calculating the reference search average time in the third term of the environment table. The environment table is a table containing only Lou codes.

The values of the first term and the second term are always given by the detection mechanism section. It is assumed that values are given to the fourth and fifth terms at the time of startup. Also, an appropriate value is given to the third term at the time of startup.

Figure 2 (e) shows the work table, which is used in the process of determining the standard average search time in the third term of the environment table, and the first term is used to calculate the standard average search time that will be calculated from now on. This is for registering how many search times have been inspected, and the second term is for registering the sum total of the reference search times for which inspections have been conducted as many times as the first term. The work table is a table containing only Lou codes. At startup, zero is given to the first and second terms.

The contents of each table in FIG. 2 change as necessary during the inference execution process, but apart from this, the values can be changed externally as necessary. Also, if there is something that has already been created, make it possible to import it. This allows you to forcibly change part of the processing flow, continue to use information that has already been accumulated, or use information obtained from other systems.

Next, a method for performing inference will be explained using FIGS. 3 and 4. The above-mentioned inference ID can be given to the inference so that attributes of the inference can be identified as necessary. If only inference is performed unconditionally every time, an inference ID is not given to the inference, but if it is desired to have the inference decide how to perform the actual inference after performing some kind of evaluation, an inference ID is given to the inference. If there is an inference ID when an inference is given, its value is made known. If the given inference has an inference ID, the inference ID
and memorize the input data for that inference (. Before performing inference processing, for inferences that do not seem to require the use of empirical data at the moment, add a record for each inference ID to the registration table for the corresponding inference. This addition is done by registering the inference ID in the first section of the record and adding it in the second section.
Instructions for making inferences are registered in the section. Also, for cases where the value of the fourth term of the record in the registration table, which is the number of times the standard inference average time is averaged, must be given a value different from the value of the fourth term of the environment table. Give the necessary values to the first and fourth terms of the record. At this time, registration is made so that two or more records with the same inference ID do not exist.

When an inference is given, it is determined whether it has an inference ID, and if it does not have an inference ID, an inference is made, the inference result is obtained, and the process ends. If it has an inference ID, it stores the inference ID and the manual data of the inference, and checks the registration table based on the inference ID to see if there is a record with the same inference ID. If the same record is found, the position of the corresponding record is recorded and the contents of the record are memorized. Identify whether the content of the second section of the corresponding record (hereinafter referred to as a registered record) instructs inference, and if instructs inference, performs the inference, obtains the inference result, and ends the process. . If there is no inference instruction in the second term, check the third term to see if a value is registered. Third
If the inference time is registered in the item, check the congestion status of the inference mechanism section in the first item of the environment table, and based on this, set the inference time obtained from the third item of the registered record to the inference time in the current congestion state. Corrected value (hereinafter referred to as estimated inference time)
get. In addition, the standard average search time in the 3rd item of the environment table is corrected to the value of the search time under the current congestion situation by checking the current congestion status of the search mechanism section in the experience table in the 2nd item of the environment table (hereinafter referred to as (called the estimated search time). The corrected estimated inference time and estimated search time obtained in this manner are compared. If the estimated inference time is smaller than or equal to the estimated search time, inference is performed, an inference result is obtained, and the process ends. If the estimated inference time is greater than the estimated search time, a search is made to see if there is any inference whose inference ID and input data are exactly the same as the given inference among the data stored in the experience table. If there is, the inference result and the time required for the search registered in the third section of that record in the experience table are obtained. Search time and environment table 2nd
The standard search time is determined based on the congestion status of the term search mechanism. Next, 1 is added to the value of the number of searches in the first term of the work table, and the calculated standard search time is added to the value of the second term. If the value of the first term of the work table is smaller than the value of the fifth term of the environment table, the process is terminated by using the obtained inference result as the result of the given inference.

If the value of the first term of the work table is equal to or greater than the value of the fifth term of the environment table, calculate the value of the second term of the work table divided by the value of the first term. is registered in the third term of the environment table, and at the same time, the values of the first and second terms of the work table are set to zero, and the inference result obtained by the search is regarded as the result of the given inference, and the process is terminated. If no matching result is found as a result of searching the experience table, inference is performed and the inference result and inference time are obtained. Next, from the obtained inference time, calculate the standard inference time as described above based on the first term of the environment table, compare this with the value of the third term of the environment table, and calculate the value of the standard inference time. If it is larger, a record with the inference ID, inference input data, and inference result is additionally registered in the experience table, and if the standard inference time is smaller than or equal to the third term of the environment table, it is not registered in the experience table. do not have. Then, the process ends with the result obtained by the inference as the inference result.

If the inference ID of a given inference is registered in the registration table, but the inference time in the third term of the registration table is not registered, check the judgment table to see if there is one with the same inference ID. Remember record position. If not, a new record is additionally registered in the determination table and the position of that record is stored (the corresponding records in these determination tables are hereinafter referred to as determination records). In this additional registration, the inference ID is registered in the first term, and zero is registered in the values of the second and third terms. The entry point for the process of additionally registering in this determination table is defined as point (PI). Next, perform the given inference and obtain the inference result and inference time. Next, judgment registration processing is performed. This determination registration process will be explained using FIG. 4. From the inference time obtained by inference, the standard inference time is determined based on the value of the first term of the environment table as described above. The value of the calculated standard inference time is added to the value of the third term of the stored judgment record, and 1 is added to the value of the second term.

After this, the value of the second term of the judgment record is compared with the value of the fourth term of the stored registration record, and the second term of the judgment record is
If the value of the term is smaller than the fourth term of the registered record being processed, the inference result is taken as the result of the given inference and the process is terminated. If the value of the second term of the judgment record is the same as or greater than the value of the fourth term of the registration record, the value obtained by dividing the value of the third term of the judgment record by the value of the second term is registered in the third term of the registration record. , delete the corresponding record having the inference ■D currently being processed from the determination table. Next, the calculated standard inference average time is compared with the value of the third term of the environment table, and if the value of the standard inference time is greater, that inference 10. Register additional records with inference input data and inference results. If the reference inference time is smaller than or equal to the third term of the environment table, no additional registration is performed in the experience table. The process ends with the result obtained by the inference as the result of the given inference. The above is the determination registration process.

The explanation of the process will be continued using FIG. 3 again. If the inference ID of the given inference is not registered in the registration table, additional registration is performed in the registration table (hereinafter, this will be referred to as a registration record). In this case, registration is performed by registering the inference ID in the first term, the value indicating that no registration has been performed in the second and third terms, and the value of the fourth term of the environment table in the fourth term. Next, the processing after the above-mentioned point P1 is performed.

By repeating the above, the standard inference average time of the experienced inferences is stored in the registration table, and the experience table can be searched to see if the inference result can be obtained in a shorter time if the actual inference is performed based on this. It becomes possible to judge whether inference results can be obtained in a shorter time by using past inference results. In addition, the experience table accumulates information such as inference results only for inferences that can be obtained in a shorter time by using past inference results than by directly inferring, and if the same inference is given again, It becomes possible to use the accumulated inference results. As a result, if the inference time is longer than the experience table search time, and if the same pattern of inference has been experienced in the past, the inference result can be obtained in a shorter time than the actual inference by searching for that result. Obtainable. Furthermore, the amount of calculation required for inference to obtain an inference result can be reduced.

〔Effect of the invention〕

By using this invention, when an inference that requires a long inference time is given, if there is empirical data of inferences made in the past using the same inference, it can be used to make the actual inference. Inference results can be obtained in a short time, and the amount of calculation required to obtain inference results can be reduced. This makes it possible to shorten the time required to obtain inference results when processing processes that require complex inferences used by a large number of users, and because the amount of calculation required for inference is reduced, it is possible to perform more advanced inferences. It is now possible to perform processing. In addition, since experience data for inference is automatically accumulated only for necessary data, the storage capacity and search time for experience data can be reduced.

4. Brief Description of the Figures Figure 1 is a diagram for explaining the configuration of functions necessary to perform the present inference method. FIG. 2 is a diagram for explaining a method of storing information necessary to perform the present inference method. FIG. 3 and FIG. 4 are diagrams for explaining the procedure of this inference method.

Claims (1)

[Claims] The inference results obtained through inference are automatically determined based on the time required for inference, stored information, and the execution environment of the device performing the inference, and the inference results are presented in a form that can be reused only for those that are needed. This stored inference result is stored based on the stored information and the execution environment of the device that performs the inference and the device that retrieves the stored inference result when the same inference is given next time. By automatically determining whether to use the inference or perform direct inference, and obtaining the inference result using a method based on this judgment, the time required to obtain the inference result can be shortened and the amount of calculations required for inference can be reduced. An inference method characterized by measurement.