JPH0650467B2 - Inference processor - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Object of the Invention (Industrial field of application) The present invention relates to an inference processing method for advancing inference processing by executing production rules, and in particular, rules applicable in the process of inference processing. The present invention relates to an inference processing method that speeds up inference processing by clarifying in advance.

(Prior Art) In a system that executes inference processing by using production rules, a production rule having a condition that matches information about the current state given during the inference processing and a conclusion drawn during the inference process is generated from a large number of production rules. A process of deriving a final conclusion is performed while sequentially extracting and applying these rules. In such processing, it is effective to speed up the inference processing so that the next applicable rule can be extracted at high speed in the inference process.

Conventionally, as a method for quickly extracting a production rule, each condition part and execution part of the production rule are examined, and which of the following conditions is satisfied by the conclusion obtained by the execution of which execution part, that is, which rule is applied? Next, prepare in advance the relationship as to which rule can be applied in the form of a table or network, and at the time of inference, refer to these relationships and select the next applicable rule sequentially. , A method for speeding up inference processing has been proposed (Japanese Patent Laid-Open No. 60-7203).
And JP-A-62-19940).

However, in these methods, since only the applicable relationships between rules are described, only the process of selecting the next applicable rule can be speeded up. Therefore, it is not possible to suggest the direction of global reasoning, for example, whether it is possible to lead from one state to another state, or which rule should be applied before reaching a certain conclusion. It was difficult to speed up inference processing by executing useless rules.

In addition, in the conventional method, even if there is redundancy between the rules registered in the rule storage unit, there is no means for judging it, so the time required for the inference process is increased by performing inference using unnecessarily many rules. There was also a problem that it would be long.

(Problems to be Solved by the Invention) As described above, in the conventional inference processing method, it is possible to speed up the inference processing by not giving a global orientation of the inference or executing useless production rules. There was a problem that I could not do it.

The present invention has been made in view of the above problems, and it is possible to efficiently execute a production rule by giving a global direction of inference processing, and eliminate the waste of inference by eliminating rule redundancy. , It is an object of the present invention to provide an inference processing method capable of speeding up inference processing.

[Structure of the Invention] (Means for Solving the Problem) An inference processing method according to the present invention stores a rule storage unit for storing a plurality of production rules including a pair of a condition unit and an execution unit, and a rule storage unit. A strongly connected component decomposition unit that obtains meta-knowledge described below based on the produced production rule, a meta-knowledge storage unit that stores the meta-knowledge obtained by the strongly connected component decomposition unit, and a memory stored in this meta-knowledge storage unit. It is equipped with an inference unit that extracts production rules based on the acquired meta-knowledge and executes inference processing, and a rule registration determination unit that determines whether to add the rules when registering the rules in the rule storage unit. There is.

Here, the strongly connected component decomposition unit obtains, as a directed graph matrix, a production rule that satisfies the transition condition between all possible states of each variable handled in the production rule stored in the rule storage unit, and further, the directed graph matrix described above. Is decomposed into a strongly connected component and a production rule satisfying the variable condition of each strongly connected component and the transition condition between the strongly connected components is obtained as meta-knowledge indicating which production rule can be applied next to which production rule.

Further, the inference execution unit, based on the information indicating the production rule applicable next to the meta-knowledge stored in the meta-knowledge storage unit, sets the global direction of inference, that is, the current state and the target state. Recognizing the gap, further applying the production rule for approaching the target state, based on this, extracting the next applicable production rule stored in the rule storage unit, inference processing To execute.

Further, the rule registration determination unit determines whether or not the strongly connected component changes due to the addition of the rule when registering the rule in the rule storage unit, and if it does not change, it cancels the registration of the rule. .

Also, instead of this rule registration determination unit, it is determined whether the strongly connected component changes when the rule is deleted among the rules registered in the rule storage unit, and if it does not change, the rule is not changed. It is also possible to provide a rule deletion determination unit that deletes.

(Operation) According to the present invention, the difference between the current state and the target state is recognized by the meta-knowledge indicating which production rule is applicable next to which production rule, and the target state is further recognized. A global orientation of reasoning that measures the application of production rules to approach Then, based on this meta-knowledge, the application order of the production rules until the final conclusion is derived, one production rule stored in the rule storage unit is extracted based on the order, and the inference processing is executed. Therefore, it is possible to prevent unnecessary execution of production rules and speed up inference processing.

In addition, in the present invention, when registering a rule, the rule registration determination unit checks whether or not the strongly connected component has changed before and after the addition of the rule. Then, when there is no change in the strongly connected component, it is determined that the rule is a redundant rule, and the rule is not added. For this reason, redundant rules are not registered, and in the end, useless rules are prevented from being applied in advance, so that inference processing can be speeded up.

In addition, when a rule deletion determination unit is provided instead of the rule registration determination unit, among rules that have already been registered by the rule deletion determination unit, a rule whose strongly connected component does not change even if the rule is deleted is deleted. Therefore, similarly to the above, useless application of rules can be prevented in advance, and inference processing can be speeded up.

Embodiment An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the arrangement of an inference processing apparatus according to this embodiment. That is, this inference processing apparatus includes a rule input unit 11, a rule registration determination unit 12, a rule storage unit 13, a strongly connected component decomposition unit 14, a meta-knowledge storage unit 15, and an inference unit 16.
And the state storage unit 17.

The rule input unit 11 is for registering a production rule in the rule storage unit 13.

When registering the rule in the rule storage unit 13 by the rule input unit 11, the rule registration determination unit 12 determines whether or not a strongly connected component, which will be described later, changes due to the addition of the rule. If the rule is changed without registration, the rule is stored in the rule storage unit 13.
To store.

The rule storage unit 13 stores a plurality of production rules determined by the rule registration determination unit 12 to be registered. Note that these production rules consist of a pair of a condition part and an execution part.

The strongly connected component decomposition unit 14 extracts the production rule stored in the rule storage unit 13 prior to the inference processing by the inference unit 16 to generate meta-knowledge. Here, a production rule satisfying transition conditions between all possible states of each variable handled in the production rule stored in the rule storage unit 13 is obtained as a directed graph matrix, and the directed graph matrix is further connected to the strongly connected component. The production rules that are decomposed and satisfy the variable condition of each strongly connected component and the transition condition between each strongly connected component are obtained as meta-knowledge. This meta-knowledge calculation algorithm will be briefly described later.

The meta-knowledge storage unit 15 stores the meta-knowledge obtained by the strongly connected component decomposition unit 14.

The inference unit 16 sequentially reads out the production rules stored in the rule storage unit 13 and executes the inference process. The inference unit 16 stores the inference process from a given state to another target state as a meta-knowledge memory. A global decision is made based on the meta-knowledge stored in the section 13, and the production rules are executed efficiently.

The state storage unit 17 stores the initial state and the intermediate state of the inference processing in the inference unit 16.

Next, the operation of the inference processing device configured as described above will be described.

Now, it is assumed that the rule storage unit 13 stores the production rules shown by the directed graph in FIG.
Here, nodes a, b, c, ..., I respectively indicate states, arrows connecting the respective states indicate a production rule in which such state transition occurs when the state is applied, and numbers near the arrows indicate rules. Shows the number.

First, prior to the inference process, the strongly connected component decomposition unit 14 generates meta-knowledge from the production rule according to the processes shown in S1 and S2 in FIG.

That is, first, from the transition condition between the nodes based on all the production rules stored in the rule storage unit 13,
The directed graph matrix shown in the figure is generated (step S1).
This effective graph is obtained by connecting the nine nodes a to i with a rule that satisfies the transition condition between these nodes. This matrix describes rule numbers that can transit from node number j to node number k.

In order to obtain this matrix, in step S1, the size is 9
When a matrix (table) M of × 9 is prepared and the state of the node number j satisfies the condition part of the x-th rule and the state of the node number k is obtained as the execution result of the rule, (jk of the matrix M). ) The process of registering the rule number x in the component and registering 0 in the other parts is performed for all the production rules.

Next, the directed graph matrix is decomposed into strongly connected components (step S2). Here, the strongly connected component is a component capable of bidirectionally transitioning between the constituent nodes, and in the example of FIG.
The node (c, d) and the node (g, h, i) are strongly connected components. As a method of strongly connected component decomposition for a directed graph matrix, a simultaneous row and column permutation method is known. In this method, row replacement and column replacement are performed at the same time so that, for example, the lower left part of the matrix becomes 0,
The elements connected on the diagonal of the matrix are strongly connected components. In the example of FIG. 4, a state in which the component on the lower left corner of the diagonal is set to 0 is obtained, and the set of rules gathered on the diagonal line can be a strongly connected component. As a result, six strongly connected components included in the center line of FIG. 5 are obtained. The extracted strongly connected components are sequentially numbered from I to VI,
It is stored in the meta-knowledge storage unit 15. The partial ordering relationships among the strongly connected components are listed in FIG.

Then, the rule registration determination unit 12 determines the strongly connected component decomposition unit 14
The directed graph matrix A of the strongly connected components obtained in step S3 is obtained (step S3). As shown in FIG. 7, the matrix A is an n × n matrix where the number of strongly connected components is n, and the state of the strongly connected component l transits to the state of the strongly connected component m by applying a predetermined rule. (Lm) of matrix A, if possible
It is obtained by performing the process of registering 1 in the component and 0 in the other parts for all production rules.

Next, the rule registration determination unit 12 obtains the matrix A ′ obtained from the obtained matrix A by A ′ = A + A ¹ + A ² + ... + A ⁿ (1). However, here A ⁿ is
The strongly connected components that are in a relation that can be reached by applying the rule n times are displayed. Therefore, for example, A ² indicates a portion that can be reached by applying the rule twice. Therefore the above
By applying the formula (1), all the reachable relations are obtained by applying the rule multiple times. However, in this example, since the number of strongly connected components is n = 6, it is sufficient to calculate up to A ′ = A + A ¹ + A ² + ... + A ⁶ (2). This is because no new achievable relationship can be found even if it is performed further. The matrix A'obtained in this way is shown in FIG.

Here, consider that a rule to be newly registered is input by the rule input unit 11 (step S5). Now, it is assumed that the rule to be newly registered is the rule 5 shown in FIG. The rule registration determination unit 12 creates a matrix B as shown in FIG. 10 showing the possibility of transition between states caused by the new rule 5. This matrix B has a size of n × n similarly to the matrices A and A ′, and 1 is assigned to the portion where the state transition from the strongly connected component 1 to the strongly connected component m is possible by the new registration rule, and 0 to the other portions. It has been registered. In the case of rule 5 shown in FIG. 9, as shown in FIG. 10, a matrix B in which 1 is registered in the portions II → VI and VI → VI and 0 is registered in the other portions is obtained.

Next, the rule registration determination unit 12 compares the obtained matrix B with the matrix A ′, and the matrix A ′ corresponding to the non-zero element of the matrix B.
It is checked whether the component of is zero (step S7).
As a result, if the corresponding matrix A ′ element is non-zero, the registration of the rule is stopped (step S8), and if it is zero, the rule is registered in the rule storage unit 13 (step S9). ).

In the example of FIG. 10, since all the elements of the matrix A ′ (FIG. 8) corresponding to the non-zero elements of the matrix B are non-zero, rule 5 is not registered. This is because even if rule 5 is registered, it is not possible to obtain a combination of strongly connected components that can be newly transitioned. In other words, rule 5 is a redundant rule with respect to already registered rules 1 to 4. That's why. Therefore, in this case, the content of the rule storage unit 13 does not change at all, and the semi-order relation between the strongly connected components is the same as that before the rule 5 is input, as shown in FIG.

On the other hand, considering the case of newly registering the rule 6 as shown in FIG. 12, since all the elements of the matrix A ′ (FIG. 8) corresponding to the non-zero elements of the matrix B are zero,
Rule 6 is registered. This is because by registering rule 6, two partial order relations III <I IV <V are newly obtained, and as shown in FIG. 12, a new strongly connected component I and III are integrated. This is because the strongly connected component I ′ is obtained, in other words, the rule 6 is not redundant with respect to the already registered rules 1 to 4. Therefore, in this case, the rule registration unit 13
Rule 6 is registered in, and the partial order relation between the strongly connected components is also changed as shown in FIG.

When the strongly connected component is changed as described above, this is stored in the meta-knowledge storage unit 15.

In the process of inference, the inference unit 16 determines the global direction of inference based on meta-knowledge, and based on this, the rule storage unit 1 stores the intermediate result of inference in the state storage unit 17.
3 production rules are executed sequentially. For example, suppose that the initial state of inference is e and want to know whether state i can be derived as an inference result from this state, the initial state is included in strongly connected component IV, and the conclusion of inference is the strongly connected component. Included in VI. Therefore, the inference unit 16 searches for a path from the strongly connected component IV to the strongly connected component VI. As a result, the path of IV → V → VI and the path of IV → VI are searched, so it is possible to reach the final goal from the initial state. In that case, rule 4 is applied rather than rule 6. It turns out that it is desirable to do.

As described above, according to the present invention, it is possible to give an overall direction to the inference process, and at the time of newly registering a rule, the redundancy of the rule is judged to determine whether or not the rule can be registered. Therefore, unnecessary rules are not registered, and as a result, inference processing can be speeded up.

In the above embodiment, the rule registration determination unit 12 determines whether registration is possible when newly registering a rule. However, in place of this rule registration determination unit, a rule deletion determination unit is provided, and among rules already registered, Redundant rules may be deleted. In this case, of the rules registered in the rule storage unit 13, it is determined whether or not the strongly connected component changes when the rule is deleted, and if it does not change, the rule is deleted. Just do it.

[Effects of the Invention] As described above, according to the present invention, it is possible to determine the direction of global inference using meta-knowledge, and it is possible to prevent the registration of useless rules. Therefore, the inference processing can be speeded up.

[Brief description of drawings]

1 to 14 are diagrams for explaining an inference processing apparatus according to an embodiment of the present invention. FIG. 1 is a block diagram of the apparatus, and FIG. 2 shows rules stored in a rule storage unit. FIG. 3 is a diagram showing the relationship between states, FIG. 3 is a flow chart showing the processing procedure of the strongly connected component decomposition unit and the rule registration determination unit, FIG. 4 is a diagram showing how strongly connected components are obtained, and FIG. FIG. 7 is a graph showing the obtained strongly connected components and their partial order relations; FIG. 6 is a graph showing the obtained partially connected relations of the strongly connected components;
FIG. 8 is a diagram showing a matrix A obtained by the rule registration determination unit, FIG. 8 is a diagram showing a matrix A ′ obtained by the rule registration determination unit, and FIGS. 9 to 11 are rules when redundant rules are added. FIGS. 12 to 14 are diagrams for explaining the process of the registration determination unit, and FIGS. 12 to 14 are diagrams for explaining the process of the rule registration determination unit when a non-redundant rule is added. 11 ... Rule input section, 12 ... Rule registration determination section, 13 ...
Rule storage unit, 14 ... Strongly connected component decomposition unit, 15 ... Meta-knowledge storage unit, 16 ... Inference unit, 17 ... State storage unit.

Claims (2)

[Claims] 1. A rule storage unit configured to store a plurality of production rules each including a pair of a condition unit and an execution unit, and all possible states of each variable handled in the production rule stored in the rule storage unit. The production rule that obtains a production rule that satisfies the transition condition as a directed graph matrix, further decomposes the directed graph matrix into strongly connected components, and determines which production rule satisfies the variable condition of each strongly connected component and the transition condition between each strongly connected component. Next, a strongly connected component decomposition unit that is obtained as meta knowledge indicating which production rule can be applied, a meta knowledge storage unit that stores the meta knowledge obtained by this strongly connected component decomposition unit, and this meta knowledge storage unit The current state based on the information that indicates the next applicable production rule of the meta knowledge stored in Recognize the gap between the target state and further determine the application of production rules to approach the target state, and conduct a global inference direction, and store it in the rule storage unit based on this. The next applicable production rule is extracted, an inference unit that executes inference processing, and at the time of registering the rule in the rule storage unit, it is determined whether or not the strongly connected component changes due to the addition of the rule, An inference processing device comprising: a rule registration determination unit that cancels registration of the rule if it does not change. 2. In place of the rule registration determination unit, it is determined whether a strongly connected component of the rules registered in the rule storage unit is deleted when the rule is deleted, and if it is not changed. The inference processing apparatus according to claim 1, further comprising a rule deletion determination unit that deletes the rule.