JPH01251237A - Inference system - Google Patents

Abstract

PURPOSE:To return a state to a desired inference state and to execute the re-execution therefrom by storing the change history of knowledge of each part of the system generated from the change of the knowledge. CONSTITUTION:When the change is generated in a working memory 8, a token is generated, and this token flows to a rule network 1, and flows into plural net sequences concerned in the network 1. Change information at this time is stored in a change management table 10. Debugging of inference is executed by stopping the inference and inspecting an inference state, based on information stored in the table 10. In that case, when it becomes clear that a fact that a rule being in a state before several rules is started is the cause generating an erroneous state, it is returned to before its state and an erroneous rule is corrected and executed again, and whether the inference as scheduled is executed or not can be confirmed.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention (Field of Industrial Application) The present invention relates to an inference system including an inference engine such as an expert system.

(Conventional technology) In recent years, the development of so-called expert systems that aggregate the knowledge of experts, turn it into a knowledge base, and attempt to function in the same way as experts has begun to be actively carried out, and some of them have been put into practical use. Systems that serve this purpose are also emerging.

Such an expert system consists of a knowledge base that stores fragmented knowledge and an inference engine that interprets it, connects the fragmented knowledge, and derives a conclusion.

For this reason, it is possible to easily realize a different behavior by changing the knowledge, and the system can be expanded easily. However, since the behavior description is not written procedurally, the behavior can be easily followed. The disadvantage is that it is difficult to remove.

Therefore, when debugging during the development of an expert system, there is a need for a means of knowing the effect of changes in knowledge on the system, or a means of searching for knowledge that is causing malfunctions.

For example, the explanation function of inference is known to meet some of these requirements. This traces the reasoning process and tells us what kind of knowledge has been applied and in what order.

However, when debugging such system development, for example, in order to improve the quality of knowledge, it is necessary to go back from a certain state during inference to a few previous states, change the knowledge slightly, and re-execute it. , cannot be handled by the inference explanation function described above.

(Problem to be solved by the invention) In the field of expert system development, when debugging, it is possible to return to several previous states from a certain state during inference, change the knowledge slightly, and re-execute. Development of this function is desired.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide an inference system having a function of returning the state to a desired inference state and re-executing the inference state from there.

[Structure of the Invention] (Means for Solving the Problems) The present invention resides in a system having an inference engine that infers fragmentary knowledge and derives a predetermined conclusion by connecting the inferred knowledge. , stores the history of changes in the knowledge of each part of the system resulting from changes in knowledge, and changes the knowledge of each part of the system in accordance with this knowledge change when knowledge changes occur in any inference state.

(Function) In the present invention, by checking the history of changes in knowledge of each part of the system resulting from changes in knowledge, it is possible to easily return the state to a desired inference state.

In addition, when knowledge changes occur in any inference state, the knowledge of each part of the system is changed in accordance with this knowledge change.
You can rerun it from there.

(Example) An example of the present invention will be described below, but prior to that, the basic principle regarding the expert system will be explained.

An expert system consists of a knowledge base that stores fragmented knowledge and an inference engine that interprets it, connects the fragmented knowledge, and derives a conclusion.

Knowledge in a knowledge base consists of rules and facts, facts are stored in working memory, rules are
As shown in the figure, it is compiled into a network. As shown in FIG. 2, the rule description takes the form of "If conditions 1 to n are satisfied, perform actions 1 to m!".

The inference in the inference engine is based on cognition as shown in Figure 3.
Take action cycles. In the figure, pattern matching means searching for a rule whose conditional part matches the contents of working memory, and conflict resolution means deciding which rule to apply first among the rules that hold at the same time. Action is the execution of the action part of the selected rule.

Then, by taking action, the contents of working memory change, and pattern matching occurs again.

The inference ends when there are no more rules to resolve the conflict through such repetition. Note that the inference is stopped midway through the number of times the rule is executed or when a specified rule is activated.

Now, in the embodiment shown below, the present invention is applied to an expert system using a rule network, and FIG. 1 is a diagram showing the structure of compiled rules according to this embodiment.

In the figure, 1 is a rule network (Rete net), 2 is a comparison node for elements of one condition of each rule,
3 is a join node where two different condition elements within a certain rule are combined, and 4 is a terminal node which is the name of the final rule.

In addition, 5 is a node stack table that points to a token pointer table representing the storage location of tokens (working memory change information) stored in the joint node 3. In the joint node 3, there are cases where tokens come from the left side and accumulate, and when tokens are stored from the right side. Since there are two cases, one in which the data accumulates, and the other in which the data accumulates, this node stack table 5 has pointer areas corresponding to both cases.

6 is a token pointer table in which the table information corresponding to the position pointed to from the node stack table 5 points to the corresponding area of the token stack Boolean.

Note that this token pointer table 6 is
Since multiple tokens are stored in the , link information to those information is also stored.

7 is a token stack table in which the table information corresponding to the position pointed from the token pointer table points to each element of the working memory indicating a group of tokens combined at the connection node 3; 8 is a token stack table created by the action part of the rule. , a working memory in which element units to be modified or deleted are stored.

Here, the operated element unit is sent to the rule network 1 as a token.

Reference numeral 9 is a conflict rule area in which the name of the rule where the token passed in the rule network 1 reaches the terminal node 4 is registered, and the order in which the rule names are registered in this conflict rule area 9 is determined by conflict resolution. They are rearranged and registered in descending order of priority according to the strategy. 10 is a change management table that manages the change history of each area;
Used to restore to previous state.

Next, the operation of this system will be explained.

When a change occurs in working memory 8, a token is generated. This token is passed to rule network 1,
It flows into a plurality of corresponding net series within this rule network 1.

Then, when the test fails at comparison node 2 in a certain net series, the flowing token disappears from the net series.

On the other hand, when all the tests at the comparison node 2 are successful, the token is flowed to the joining node 2, and when it flows from the corresponding direction of the node stack table 5, that is, from the left side, the token flows from the left side, and when it flows from the right side, it flows from the right side. Addition of token pointer table 6 pointed to by pointer/
Deletion processing is performed. At the same time, the change information at this time is registered in the change management table 10.

As shown in FIG. 4, this change management table 10 stores a recognition-behavior cycle number, a change table number, a change index, a change type (addition/modification/deletion), a trigger rule name, and pre-change information as registered information. do.

By the way, as the pattern matching process during inference progresses, the state of the token stack table 7 also changes, so the change information is also stored in the change management table 10.

Further, in the conflict rule area 9, a rule by which the token flows through each net series and finally reaches the terminal node 4 is registered, and this rule is selected and deleted from the conflict rule area 9 in the conflict resolution process. This information is also stored in the change management table 10.

The working memory 8 changes according to the process specified in the action section of the rule selected in the conflict resolution process described above, and the change information is also stored in the change management table 10.

Now, inference debugging is performed by stopping the inference in a certain state and inspecting the inference state based on the information stored in the change management table 10.

In this example, when it is determined that a rule in a previous state was activated that caused the incorrect state, the user returns to the previous state and corrects the incorrect rule and tries again. You can run it and check whether the inference was performed as planned. This is done by putting the rule correction information in the state that created the wrong state out from the change management table 10, and by using the token pointer table 6.
This is because the states of the token stack table 7, conflict rule area 9, and working memory 8 are updated.

Note that the extracted information is deleted from the change management table 10. Thereby, the information in the token pointer table 6, token stack table 7, conflict rule area 9, and working memory 8 can be restored to any inference state while maintaining consistency with the information in the change management table 10.

In this example, as described above, when it is determined that the activation of a rule in a previous state caused an incorrect state, the erroneous rule is corrected by going back to the previous state. Since you can run it again and check whether the inference was performed as planned, knowledge verification becomes much easier.

Note that in such an embodiment, by calculating the difference of changes in each table and managing the difference information, memory space can be saved and execution efficiency can be improved.

Also, the inference becomes very deep and there is a lot of changing data,
If the storage capacity of the change management table is exceeded, a method may be used in which old information is swapped to a secondary storage medium.

Furthermore, if higher processing speed is required, a method may be adopted in which the change management table information is stored directly in the address management 12 and the change memory information. When using virtual memory, the base address of each table may be stored and the storage address may be relocatable.

Furthermore, when tables are managed efficiently, the space management information of each table may also be included as change management table information.

[Effects of the Invention] As explained above, according to the present invention, the history of changes in the knowledge of each part of the system resulting from a change in knowledge is stored, and when a change in knowledge occurs in an arbitrary inference state, Since the knowledge of each part of the system has been changed, the state can be returned to the desired inference state and re-executed from there.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the structure of compiled rules according to the embodiment of the present invention, FIG. 2 is a diagram showing the rule description format in this embodiment, and FIG. 3 is a diagram showing the inference format of the inference engine in this embodiment. The figure shown in FIG. 4 is a diagram showing registration information of the change management table in this embodiment. 1... Rule network, 2... Comparison node, 3
...Connection node, 4...Terminal node, 5...
・Node stack table, 6... Token pointer table, 7... Token stack table, 8...
・Working memory, 9...Conflict rule area, 10
...Change management table. Applicant: Toshiba Corporation Patent Attorney Satoshi Suyama - +If (Condition 1) (Condition n) (Action m) 1 Figure 2 Figure 3 Figure 4

Claims (1)

[Scope of Claims] A system comprising at least an inference engine that infers fragmentary knowledge and derives a predetermined conclusion by connecting the inferred knowledge, comprising: knowledge of each part of the system resulting from a change in the knowledge; An inference system comprising: a storage means for storing a change history of the information; and a change means for changing the knowledge of each part of the system in accordance with the knowledge change when the knowledge is changed in an arbitrary inference state.