JPH02108127A - Inference control method for expert system - Google Patents

Abstract

PURPOSE:To define and input a multiple knowledge base and to perform inference control at low cost by providing an execution table of simple structure which represents multiple knowledge and providing a processing part which interprets and executes the execution table. CONSTITUTION:The rule base 3-1, fact base 3-2, and function of the multiple knowledge base are divided and registered in the execution table 4, the registered function is executed, and inference processing is performed based upon the production rule of the multiple knowledge base corresponding to the rule base 3-1 and fact base 3-2. Namely, the execution table 4 is referred to by a multiple knowledge inference execution control part 5 and the multiple knowledge base corresponding to a subordinate problem is referred to the execution table 4, inputted, and executed by a production system 1. Consequently, the definition, input, and inference control and processing of the multiple knowledge base are performed efficiently at a high speed.

Description

[Detailed Description of the Invention] (Industrial Application Field) This invention relates to an inference control method for an expert system using a computer, and in particular deals with multiple knowledge bases and large amounts of databases for complex problems such as process design. Concerning necessary inference control methods for expert systems.

[Conventional technology]

Regarding the knowledge representation of expert systems and their inference methods, see Setsuo Osuka, "Knowledge Information Processing" (Ohmsha), p.2.
3 to 39 are introduced in detail.

In general, many expert system development tools are built based on production systems, frame systems, etc.

Figure 5 shows the conventional inference f using a production system.
It is a figure explaining a control method. In the figure (!) is a knowledge base consisting of a production storage unit (1o) and a working storage unit (11), and a production storage unit (10).
stores knowledge expressed in the production rule format of "if condition part) then (action part/conclusion part)".

"01" to "C5" are conditional clauses, and "&" indicates an AND combination of the conditional parts. Moreover, "A1" to "A5" are action nodes.

Further, (11) indicates a working memory section that stores a fact base, which is the known facts of the problem. In this example, “C1”0
3'' and "C5" are facts. (12) is a pattern matching unit that checks whether the conditional part consisting of multiple conditional clauses is satisfied by the facts in the working memory unit (11). (13) is a competitive rule set, which is composed of a plurality of production rules obtained as a result of pattern matching.

(14) is a conflict rule resolution unit for selecting only one rule from the conflict rule set (13) using a conflict resolution algorithm. Furthermore, (15) is an execution unit that executes (action part/conclusion part) of the production rule determined by the conflict rule resolution part (14).

In general, the inference control method using a production system is
For the rules in the production storage (10),
Knowledge is processed by repeating the cognitive-behavior cycle shown below in order from the beginning.

First, in step 1, the production storage (
The condition part of the rule in the rule base in the working memory unit (11) and the pattern matching unit (12
) to perform verification processing [■■]. Then, a production rule that satisfies the condition part in the production storage section (10) is retrieved [■]. In the matching process of the pattern matching unit (12), a plurality of production rules may be established, and a set of these production rules is called a conflict set.

Next, in step 2, a conflict rule resolution unit (14) selects one rule from among the conflict rule sets obtained in step 1 using an algorithm such as the rule that first succeeded in pattern matching. This process is called conflict resolution [■■]. As a result, in the case of the above conventional method, C3->A
Rule 2 is selected.

The processing up to this point is called the cognitive cycle.

Furthermore, in step 3, the action part of the rule selected by the cognitive cycle of step 2, that is, A2
is executed [■]. This is an action cycle, and after this execution returns to step 1 again.

This updates the fact base so that a different rule is chosen for the next cognitive cycle. Thereafter, the cognition-behavior cycle is repeated until there are no more rules to match, that is, until no new facts can be derived.

(Problem to be solved by the invention) Since the conventional inference control method using a production system is configured as described above, the knowledge for problem solving is divided into a rule base in the form of production rules and a fact base in the form of known facts. Since these rule bases and fact bases are stored and processed in one production memory and one working memory, the number of rules and facts becomes large, and the processing time for pattern matching becomes longer than the condition part. The problem was that the number of conditionals increases in combination, depending on the number of conditionals, the variables in the conditional clause, and the number of facts in the working memory.

In addition, in the case of process design, the knowledge base is composed of sub-problems consisting of knowledge of multiple different characteristics, such as selection of machine tools and machining processes, determination of standards for installation, and determination of work order. Complex problems require processing of multiple knowledge bases with different properties, and conventional inference methods using production systems express the knowledge of all sub-problems in a single knowledge base. Therefore, it was difficult to manage knowledge and it was not possible to control reasoning for each subproblem.

As a method that can express and process multiple knowledge for each sub-problem, there is a "blackboard model" in the HEAR5AY-11 system developed at Carnegie Mellon University in the United States to solve speech understanding problems (see the above-mentioned document, p. 30~). 3
3).

The HEAIISAY-II system is an audio-based speech understanding system that separates knowledge into sub-problems consisting of sensor detection, phoneme segmentation, word understanding, word recognition, sentence understanding, etc., and performs knowledge information processing. conduct.

Knowledge for each problem is defined in knowledge sources, and the six individual knowledge sources are configured as a production system, which is controlled to work cooperatively through a common work area called the blackboard.

The system allows forward reasoning as well as backward reasoning based on hypotheses generated by other knowledge sources. HEA monitors the status of the blackboard, determines which level of processing should be prioritized, and plans the activation of knowledge sources.
In R5AY-If, this part is said to be based on human cognitive abilities and thinking methods, but in systems of this type, apart from the control content, there are problems related to the implementation method of the control mechanism. In other words, the method itself has not yet been established, and considering its use, it relies on speech understanding, making it difficult to apply to other problems.

Furthermore, since it is based on a production system,
The problem mentioned above that inference time increases as the number of rule bases and fact bases increases for each subproblem remains unresolved.

Furthermore, ART and ESHELL are used as executable system construction tools derived from IEAR5AY-II.
There are systems such as However, both systems are HE
Because it is based on the AR5AY-II system,
It has the same drawbacks, and in addition, it is expensive in terms of application, and the control mechanism cannot be modified to suit the problem.

This invention was made to solve the above problems,
The purpose of this study is to obtain an expert system inference control method that can efficiently and quickly define, input, and inference control/process multiple knowledge bases.

[Means to solve the problem]

The inference control method for an expert system according to the present invention divides the rule base, fact base, and function of multiple knowledge bases and registers them in an execution table, executes the registered functions, and supports the rule base and fact base. Inference processing is performed based on production rules based on multiple knowledge bases.

[Effect]

The execution table in this invention is referred to by the multiple knowledge inference execution control unit, and the multiple knowledge base corresponding to the sub-problem is input with reference to this execution table and executed by the production system.

〔Example〕

An embodiment of the present invention will be described below with reference to FIGS. 1 to 3. FIG. 1 shows an operation flowchart of the inference control method of this embodiment, FIG. 2 shows a system configuration diagram for executing the inference control method of this embodiment, and FIG. 3 shows a storage state of an execution table. In each of the above figures, the inference control method of the expert system according to the present embodiment includes method data for a plurality of sub-problems stored in a knowledge base (3), a knowledge base for the plurality of sub-problems, and an inference process for the knowledge base. The functions before and after the inference that are executed before and after the execution table (4
) The first process (step 1) of registering in
The pre-inference function (Init-func) (hereinafter referred to as initialization function) registered in the execution table (4) in the process (step 1) is executed by the execution control unit (5), and the registered knowledge base is The production system (1) performs inference processing using the rule base and fact base, and after the inference processing, the above registered post-inference function (End-func)
(hereinafter referred to as a termination function) by the execution control unit (5).

The above production system (1) has an execution table (
4) A production storage unit (1=1) that stores the rule base (set of production rules) of the knowledge base registered in 4), and a working storage unit (1-2) that stores the fact base of the registered knowledge base. and each of the above storage units (1
-1), (1-2) and an inference mechanism (1-3) that performs inference processing based on the fact base.

The knowledge base unit (3) includes a rule base storage unit (3-1) that stores a rule base prepared corresponding to each sub-problem constituting the problem to be inferred, and a rule base storage unit (3-1) that stores a rule base prepared corresponding to each sub-problem constituting the problem to be inferred. This configuration includes a fact base unit (3-2) that stores a fact base prepared by the user.

The execution table (4) registers the name of the rule base and the name of the fact base that are the targets of inference processing among the knowledge bases stored in the knowledge base (3), and is displayed before the start of the inference processing. If there is any information, the message salt and the forward or backward inference mechanism type of the inference mechanism are registered, and the initialization functions and termination functions to be executed before and after the above inference processing are registered. Also, the above execution table (4)
puts two elements (ENJPROC) in the final position of the table
ESS marker and the name of the LISP function to be executed last) are registered as the final entry.

The execution control unit (5) controls various registrations in the execution table (4), processes functions before and after inference registered in the execution table (4), and processes functions before and after inference registered in the execution table (4).
) commands and controls loading, storing, and starting.

Next, the operation of the inference control method for an expert system according to this embodiment based on the above configuration will be explained.

First, in step 1, the execution control unit (5) registers various items in the execution table (4).

This execution table (4) has a structure as shown in FIG. That is, the execution table (4) includes a plurality of registrations corresponding to the plurality of sub-problems stored in the knowledge base section (3). In the example shown in FIG. 3, it is assumed that the order of registration in the execution table (4) corresponds to the order of inference execution. This registration consists of the following three elements and a final entry. These three elements are: first, the message and inference type; second, the knowledge base;
Thirdly, there are functions before and after inference.

First, regarding the registration of the first element, the message and inference type, if there is information that should be displayed before inference starts, the message salt is registered, and the type of knowledge-based inference mechanism defined corresponding to this registration. Register.

Note that the types of the above inference mechanism include forward inference, backward inference, and no inference processing.

Regarding the registration of the knowledge base of the second element, the rule base name and fact base name to be subjected to inference processing are registered.

Regarding the registration of processing before and after inference of the third element above,
If there is a LISP function to be executed before or after inference processing, register the function name.

Note that the final entry is placed at the final position of the table and consists of two elements. One is (END-PROCE
SS), and the other is the name of the LISP function to be executed last.

Next, in step 2, inference control operations for multiple knowledge bases are performed as follows. Basically, inference execution control between multiple knowledge is realized by the execution control unit (5) shown in FIG. 2 interpreting and executing the entries in the execution table (4) in order from the beginning. Inference processing for each knowledge base is performed by activating the production system (1).

First, the execution control unit (5) reads the registered items from the execution table (4) (step 2-1).The execution control unit (5) determines whether the read registered item is the final entry or not. 2) If it is determined that the entry is the final entry, the execution control unit (5) executes the termination function to terminate the multiple knowledge base inference process (step 2-8).

On the other hand, if it is determined in step 2-2 that it is not the final entry, the three elements (
The execution control unit (5) takes out the message, inference type, knowledge base, function) from the execution table (4) (step 2-3). If this extracted element contains a message (see message 1.2... in Figure 3),
Execute and display message output. In addition, the initialization function (In1t-func in Figure 3) that should be executed before inference is
1.2...), the execution control unit (5) executes it (step 2-4).

Furthermore, in step 2-5, the execution control unit (5) refers to the execution table (4), and this execution table (5)
The fact base (see WM-name 1.2... in Figure 3) and the rule base (P name 1. in Figure 3) of the knowledge base specified by the latest entry registered in .
2...) from the knowledge base section (3) and stores it in the production storage section (1) of the production system (1).
1-1). This fact-based load store is stored as an intermediate state in the working memory (1-2), and also stores the inference result of the inference mechanism (1-3), which will be described later, as an intermediate result. The process of step 2-5 is not executed when the production rule is already stored in the production storage unit (1-1) and inference is performed using this production rule, and the process of step 2-5 is not executed.
-6 is executed directly.

After storing the production rule in the production storage section (1-1), the execution control section (5) activates the production system (1) for inference execution (step 2-6). The execution control unit (5) determines forward inference or backward inference based on the inference type registered in the execution table (4) (see R type 1.2... in Fig. 3), and determines whether the inference is forward or backward inference. You will have to start something like that. Note that if there is no inference designation in the above inference type, the process proceeds to the next step without performing inference processing.

The production system (1) that has received the activation command starts the forward or backward reasoning mechanism (1-3) of the reasoning mechanism (1-3).
1-31) or (1-32) executes the inference operation (step 2-7).

Further, after the production system (1) executes the inference, the process returns to step 2-2, and after the determination in step 2-2, the execution control unit (5) executes the termination function and ends (step 2-8).

Furthermore, a case in which the inference processing method is applied to process design will be described in detail. In this process design, it is necessary to handle sub-problems consisting of different knowledge, such as selection of machining processes and machining machines, determination of machining reference planes, determination of mounting plans, and determination of machining procedures.

These sub-problems can be knowledge-processed by registering them in the above-described execution table (4), but if the processing of the sub-problems fails, global backtrack processing across the sub-problems is required. For example, machine tool i4
Assume that A is determined as one candidate, and it is found that there will be a contradiction in machine tool A in the subsequent machining procedure determination problem.

In this case, in the method according to the invention, the execution table (4)
By using the function that is activated after the inference is registered in , it is possible to return to the processing machine selection problem and restart the inference.

In this way, the execution table (4) provided in this invention
is a meta-knowledge archive that controls multiple knowledge of sub-problems, which enables global backtracking across sub-problems.

Next, a case will be described in which the frame system (2) shown in FIG. 2 is operated using the production rule example shown in FIG. 4.

As described above, in the production system (1), as the number of rule bases and fact bases in the knowledge base increases, the inference time increases combinatorially. Therefore, in order to reduce this inference time, a frame access function mechanism section (2-2) is provided.

In this frame system (2), atoms and lists to be processed are stored in an associative list format. For this reason, it is possible to process a certain variable faster by using the frame system (2) than by using the production system (1) to obtain a variable from the fact base in the working memory (1-2) by pattern matching.

The condition part of the production rule stored in the production storage unit (1-1) consists of a plurality of conditional clauses.

In the pattern matching process of the cognitive cycle, the value of a variable is often bound in the first conditional clause, and the attribute value of that variable is often determined in the next conditional clause. If the attribute value of the variable exists in the database, accessing the direct tl frame system (2) from the condition section can reduce the number of pattern matching operations, and the working memory section ( The number of facts in 1-2) can also be reduced.

As a result, it becomes possible to dramatically shorten the inference processing time compared to a system configured only with the production system (1).

Next, a specific example will be explained using the rules showing knowledge of machine tool selection shown in FIG.

In the first conditional clause, the name of the workpiece to be processed is extracted from the fact base by pattern matching.

In the second conditional clause, the frame access function unit (2-2) shown in FIG. 2 is used to extract data on the machine tool set from the production database.

The third and fourth conditional clauses extract one machine tool from the machine tool set, and the subsequent conditional clauses carry out the process of selecting the machine tools.

In the fifth conditional clause, the frame access function mechanism unit (2-2) is used to extract the approximate power work dimensions in the X, Y, and Z axis directions required for the work mechanism from the pattern-matched work name.

Regarding machine tools, use the frame access function mechanism section (2-2) from the machine name, and use the maximum output power X%Y,
The stroke in the Z-axis direction is extracted.

By comparing these data, a machine tool that satisfies the power and size necessary for machining the workpiece is determined.

If such a frame access function mechanism part (2-2)
Without it, pattern matching would have to be used to find the attribute values of the workpiece or machine tool in the fget function part of the fifth conditional clause.

This conditional clause alone requires pattern matching 11 times, and the number of facts in the working memory (1-2) increases by 11.

It can thus be seen that the frame access function mechanism section (2-2) functions effectively in shortening the inference time.

Note that the first argument of the fget function indicates the frame name, the second argument indicates the slot name, and the third argument indicates the facit name.

In this way, rule-based processing is performed in the production system, and database processing is performed in the frame system, so that the database can be directly accessed from the production rule execution part and action part, so the knowledge base Even when the number of systems is large, the inference time, which is a drawback of production systems, can be shortened and high-speed inference processing can be achieved.

Although the above embodiment shows a method of processing the entries in the execution table (4) in order from the beginning, it is also possible to change the control of multiple knowledge based on the inference result of one sub-problem.

Further, in the above embodiment, the case where the production system and the frame system are activated is shown, but it is also possible to use a method of inferring by starting only one of them alone.

Furthermore, the present invention is not limited to process design, but can be applied to other fields as long as the problem consists of multiple knowledge bases.

〔Effect of the invention〕

As explained above, according to the present invention, an execution table with a simple structure that expresses multiple knowledge is provided, and a processing unit that interprets and executes this execution table is provided. This has the effect that control can be realized efficiently and at low cost.

[Brief explanation of drawings]

FIG. 1 is an operation flowchart of an inference control method for an expert system according to an embodiment of the present invention, FIG. 2 is a block diagram of an inference control method of an embodiment of the present invention, and FIG. 3 is an execution table provided in the present invention. FIG. 4 is a program chart showing an example of rules regarding machine tool selection, and FIG. 5 is a configuration diagram of a conventional production system. (1) is the production system, (1-1) is the production storage unit, (1-2) is the working memory unit, (1-3) is the inference mechanism, (1
-31) is a forward reasoning mechanism, (1-32) is a backward reasoning mechanism, (2) is a frame system, (2-1) is a database, (2-2) is a frame access function mechanism, (3) is knowledge Base part, (3-1) is a multiple rule base part, (3-2) is a multiple fact base part, (4) is an execution table, and (5) is an execution control part. In addition, in the figures, the same reference numerals indicate the same or corresponding parts.

Claims (1)

[Claims] A first process of registering a plurality of knowledge bases consisting of fact bases and rule bases and functions before and after inference, which are executed before and after inference, in an execution table for each sub-problem that constitutes the problem to be inferred; The pre-inference function registered in the execution table in step 1 is executed, the production system performs inference processing using the fact base and rule base of the registered knowledge base, and the registered function after inference is executed. The second step, in which the above execution/processing is performed sequentially for each subproblem.
An inference control method for an expert system, comprising the steps of: