JPH0495132A - Knowledge inference processing device - Google Patents

Abstract

PURPOSE:To efficiently select a effective next rule in each network by setting selection indexes of rules as the next selection objects with respect to each rule. CONSTITUTION:Selection indexes of rules as next selection objects are stored in a storage means (rule network matrix memory) 2 with respect to each rule in a knowledge base, and the selection index of a rule selected out of selection indexes of rules as next selection objects is updated by an updating means based on information indicating whether the condition of the selected rule is satisfied or not. An inference processing part 5 reads out selection indexes of rules as next selection objects of one rule and selects the next rule based on these selection indexes. Thus, the effective next rule in each network is efficiently selected.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a knowledge inference processing device that performs a series of inferences by sequentially selecting rules from a knowledge base having a plurality of rules and interpreting tangent rules.

[Prior Art] This type of knowledge inference processing device includes a knowledge base consisting of a plurality of rules including respective conditions, and an inference processing unit that sequentially extracts rules from this knowledge base. After selecting one rule from the base and determining whether the condition of the selected rule is met, a rule is subsequently selected from the knowledge base and the condition of the rule is determined.

A series of inferences is made by sequentially performing the selection and judgment of such rules.

Here, in order to shorten the inference processing time, it is very effective to improve the efficiency of selecting rules from the knowledge base.For this reason, conventionally, each A rule pointer table storing pointers has been proposed. This pointer table stores pointers for each rule in a predetermined arrangement order, and during inference processing, the pointers for each rule are selected based on the arrangement order, and the rule indicated by the selected pointer is added to the knowledge base. , and use the conditions of this rule as the subject of judgment.

The arrangement order of each pointer in the rule pointer table is updated every time inference is completed, and the arrangement order of pointers pointing to rules that have met the condition with a high probability up to that point goes up, and the arrangement order of pointers that point to rules that have met the condition with a high probability up until then is The array order of pointers pointing to rules that have satisfied the conditions is lowered.

Therefore, during inference, rules that have satisfied the conditions with a high probability are selected preferentially, thereby improving the efficiency of selecting rules from the knowledge base.

[Problem to be solved by the invention] However, in the method using the conventional rule pointer table described above, the arrangement order of each pointer in this pointer table corresponds to the priority order of all rules in the knowledge base. Since the priority order of each rule is specified all at once, when inferences are made for each of multiple specialized fields, the priority order is not necessarily the most appropriate for each specialized field. For example, if a set of rules for reasoning in one specialized field is completely different from a set of rules for reasoning in another specialized field, the arrangement order of pointers in the rule pointer table will be different for both specialized fields. However, both methods may not be fully effective. In other words, when building a network consisting of a series of rules for each specialized field by considering the relationships between rules for each specialized field, it is not possible to improve the rule selection efficiency even by using this rule pointer table. could not.

Additionally, when a rule's conditions are met, each pointer in the rule pointer table must be searched from the beginning according to its arrangement order in order to select the next valid rule. It could not be said that the efficiency was extremely good.

Therefore, the present invention provides a knowledge inference processing device that can construct a network consisting of a series of rules for each specialized field and select the next rule that is effective in each network with good efficiency. The purpose is to

[Means for Solving the Problems] In the present invention, for each rule in the knowledge base, the selection index of each of the rules to be selected subsequent to the one rule is stored in a storage means. Store it and
When a series of inferences is completed, for each rule out of a series of rules that satisfy the selected conditions, the selection index of the selection index of each rule to be selected is The updating means updates the selection index of the selected rule based on whether the condition of the selected rule is satisfied, and the inference processing unit updates each of the rules to be selected following one rule. The selection indices are read from the storage means, and the next rule is selected based on these selection indices.

[Operation] According to the second aspect of the present invention, selection indices are set for each rule to be selected following one rule, and the inference processing section selects the next rule based on these selection indices. I try to do that. Since the selection index of each rule is set corresponding to each rule, the next rule can be selected from any of the rules.

In addition, when a series of inferences is completed, for each rule used in this inference, the rule actually selected from the selection index of each rule that is the target of selection following that one rule. Since the selection index of is updated based on whether the conditions of the selected rule are satisfied, a history of a series of inferences can be kept.

[Embodiments] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram showing an embodiment of a knowledge inference processing device according to the present invention. In the figure, a rule base memory 1 stores a plurality of rules including respective conditions and a pointer table in which pointers for indicating each rule are arranged, and a rule network matrix memory 2 stores a plurality of rules including respective conditions, and a rule network matrix memory 2 stores a pointer table in which pointers for indicating each rule are arranged. In addition, the working memory 3 stores various programs, data, etc. necessary for inference processing. The rule base control unit 4 selects and takes out a predetermined rule from among the rules in the rule base memory 1, provides this rule to the inference processing unit 5, rewrites the contents of these rules, and stores the rule network matrix memory. Rewrite the contents of 2. Every time the inference processing unit 5 receives a rule from the rule base control unit 4, it interprets the rule and executes inference. The external interface unit 6 is an interface between the knowledge inference processing device and the outside, and includes data related to changing the contents of the rule base memory 1, data related to changing the contents of the rule network matrix memory 2, activation instructions to the inference processing unit 5, etc. is input from the outside, or the result of the inference processing by the inference processing unit 5 is output to the outside.

The rule base memory 1 stores a pointer table 21 as shown in FIG. 2. If n rules are stored in the rule base memory 1, the pointer table 21 stores n+1 rules. Stores a pointer.

Here, the first rule stored in the rule base memory 1 is given an identifier 1, the second stored rule is given an identifier 2, and the nth rule is given an identifier n. Given.

In the pointer table 21, pointers indicating each rule are arranged corresponding to each identifier. For this reason, for example, if you want to select the rule stored in No. 1 and given the identifier k, the pointer stored in this pointer table 21 corresponding to the identifier k is read out, and the rule pointed to by this pointer is read out. By accessing the rules, you can get the rules for the identifier.

The rules for this identifier have a configuration as shown in Figure 3, where an identifier k is stored in correspondence with "rule id", a condition of the rule is stored in correspondence with "condition", and cone lus ton is Correspondingly, the conclusion to be executed when the conditions of the rule are met is stored, and the "rule pointer table"
A matrix indicating the priority order of each rule to be selected next when the conditions of the rule are satisfied, that is, a matrix of identifiers of the rules arranged in the priority order is stored.

In addition, the rules for other identifiers are also similar to the configuration shown in Figure 3, with the identifier for each rule, the conditions for each rule, the conclusion to be executed when the conditions for each rule are met, and the results for each rule. A matrix is stored in which the identifiers of the rules to be selected next when the rule conditions are met are arranged in priority order.

Note that the pointer to the rule with identifier O is the pointer that is first selected at the start of the inference process.

The rule with the identifier 0 indicated by this pointer is set in advance as a basic definition for starting the inference process, and is applied only when starting the inference process. Of course, for this rule with identifier O, there is a matrix in which the identifiers of the rules to be selected next are arranged in order of priority.
er table”.

The rule network matrix memory 2 stores a matrix table 41 as shown in FIG. 4, and since the pointer table 21 stores (n+1) pointers, this matrix table 41 has (n+1) It becomes a matrix of X(n+1). The first line corresponds to the rule with identifier 0, and w (0,0
), w (0,1).

-0-w(0,n) is the index of the rule with identifier O, the index of the rule with identifier 1, . . . , the index of the rule with identifier n. Similarly, the 2nd to (n+1)th lines correspond to the rules with identifiers 1 to n, respectively, and these lines include the index of the rule with identifier 0, the index of the rule with identifier 1, ..., the rule with identifier n. The indices of are arranged respectively.

Therefore, the (k+1)th line corresponds to the rule for the identifier, and the (k+1)th line w(k, o), w(k,
1) , −, w (k, n) are the exponents of the rule with identifier 0.

The index of the rule with identifier 1, . . . is the index of the rule with identifier n. Here, these indices w(k, o).

If w (k, 1), ..., w (k, n) are selected in order from the largest value, and each identifier indicating each index is arranged in this selection order, the order in which these identifiers are arranged will depend on the identifiers. Indicates the priority of each rule to be selected next when the rule conditions are met. Therefore, the matrix of these identifiers becomes a matrix of identifiers of each rule indicating the priority of each rule to be selected next, which corresponds to the rule pointer table in the rule for identifiers shown in FIG.

Next, the operation of inference processing performed in such a configuration will be described according to the flowchart shown in FIG.

First, the inference processing unit 5 is connected to the external interface unit 6 from the outside.
When the startup command is received via the inference controller, a command to read the rules at the time of starting the inference is issued to the rule base control unit 4 (step 10).
1). In response to this, the rule base control section 4 uses the pointer table 2 shown in FIG.
1, the pointer to the rule with identifier 0 to be selected at the start of inference is read from the pointer table 21, and the rule with identifier 0 pointed to by this pointer is accessed (step 102).

When the rule base control unit 4 accesses the rule with identifier 0, the “rule pointer ta” of the rule is
In step 103), it is determined whether an identifier to be read is set in the matrix of identifiers of each rule indicating the priority order of each rule corresponding to "ble". Reads out the identifiers of the rules arranged in , that is, the identifier of the rule with the highest priority.For example, if identifier n is arranged first, the rule base control unit 4 reads out identifier n, and then reads out the identifier of the rule with the highest priority. The pointer to the rule indicated by is read from the pointer table 21, and the rule with the identifier n indicated by this pointer is accessed.

Then, the rule base control section 4 hands over the rule with the identifier n to the inference processing section 5 (step 104).

When the inference processing unit 5 receives the rule with the identifier n, it determines whether the condition corresponding to condition of this rule is satisfied (step 105). Here, if this condition is not satisfied, the inference processing unit 5 informs the rule base control unit 4 that the condition of the rule with the identifier n is not satisfied.
to notify. Since the conditions of the rule with the identifier n selected as the next rule following the rule with the identifier 0 are not satisfied, the rule base control unit 4 sets the index of the rule with the identifier n corresponding to the rule with the identifier 0, that is, the rule Select the index w(0,n) of the rule of the identifier n in the first row in the matrix table 41 shown in FIG. 4 in the network matrix memory 2, and select this index w(0,n).
is updated based on the following equation (1).

w (j, j)=w (+, j) −Δ
w-(1) new old However, in the above formula (1), w (j, j> is o
ld represents the index w (0, n) before being updated, w
(1ev, j) represents the updated index w(0, n), ΔW
represents a predetermined value.

Therefore, if the condition of the rule with identifier n selected as the next rule after the rule with identifier 0 is not satisfied, the index w(0, n) of the rule with identifier n corresponding to the rule with identifier 0 is be made smaller.

Thereafter, the process returns to step 103, and the rule base control unit 4 stores the "rule pointer table" of the rule with identifier 0 in the rule base memory 1.
Check that the identifier to be read is set in the matrix of identifiers of each rule corresponding to step 103).
The identifier of Lulu with the highest priority other than , that is, the identifier arranged second in the matrix is read out. For example, if the read identifier is identifier 1, the rule base control unit 4 reads a pointer to the rule indicated by identifier 1 from the pointer table 21, accesses the rule with identifier n indicated by this pointer, The rule with identifier 1 is handed over to the inference processing unit 5 (step 104).

When the inference processing unit 5 receives the rule with the identifier 1, it determines whether the condition corresponding to condition of this rule is satisfied (step 105). Here, if this condition is satisfied, the inference processing unit 5
The conclusion corresponding to onclusion is executed (step 107), and the rule base control unit 4 is notified that the condition of the rule with identifier 1 is satisfied. Since the condition of the rule with identifier 1 selected as the next rule following the rule with identifier 0 is satisfied, the rule base control unit 4 calculates the index of the rule with identifier 1 corresponding to the rule with identifier 0, that is, the rule network matrix. Select the exponent W (0, 1) of the rule of the identifier 1 in the first row in the matrix table 41 shown in FIG. Update based on.

w (j, j) −w (i, j)
+ΔW... (2) new
However, in the above formula (2), w (
i, j) represents the index w (0, 1) before the old update, w
(400w, j) represents the updated index w(0,1), ΔW
represents a predetermined value.

Therefore, if the conditions of the rule with identifier 1 selected as the next rule after the rule with identifier 0 are satisfied, the exponent w(0,1) of the rule with identifier 1 corresponding to the rule with identifier O is increased. Ru.

Thereafter, the process returns to step 103, and since the condition of the rule with identifier 1 is satisfied, the rule base control unit 4 sets the "rule pointer t" of this rule.
It is determined whether or not the identifier to be read is set in the matrix of identifiers of each rule corresponding to "able" (step 103), and if it is set, the identifier of the rule arranged first in the matrix of identifiers of each rule is read. Read out. For example, if the identifier k is arranged first, the rule base control unit 4 reads the pointer to the rule indicated by the identifier k from the pointer table 21, and sends the rule to the identifier indicated by this pointer to the inference processing unit 5. Delivery (step 104).

When the inference processing unit 5 receives the rule for the identifier, it determines whether the condition corresponding to the condition of this rule is satisfied (step 105).

Here, if this condition does not hold, inference processing 5
The rule base control unit 4 is notified that the condition is not satisfied. Since the rule condition for the identifier selected as the next rule following the rule for identifier 1 is not satisfied, the rule base control unit 4 sets the rule index, that is, the matrix, for the identifier corresponding to the rule for identifier 1. The index W (1, k) of the rule is selected as the identifier in the second row of the table 41, and this index w (1, k) is updated based on the above formula (1).
, k).

Furthermore, returning to the process from step 103, the rule pointer table of the rule with identifier 1 is
Make sure that the identifier to be read is set in the identifier matrix of each rule corresponding to step 103).
, reads the second identifier arranged from this matrix, reads the pointer to the rule indicated by this identifier from the pointer table 21, and delivers the rule indicated by this pointer to the inference processing unit 5 (Step 104) , it is determined whether the conditions of this rule are satisfied (step 105).

If the rule conditions are not met in this way, the process from step 103 is repeated, and r of the rule with identifier 1 is
The second, third, etc. identifiers are sequentially read out from the matrix of identifiers of each rule corresponding to "ule pointer table", and it is sequentially determined whether the conditions of each rule indicated by these identifiers are satisfied. Then, each time the condition is not satisfied, the index of each selected identifier in the second row of the matrix table 41 is updated to become smaller based on the above formula (1).

Further, if the condition of the rule for the identifier described earlier is satisfied in step 105, the inference processing unit 5 executes the conclusion corresponding to 'eone lus ton' of the rule (step 107), and also applies the rule to the identifier. The rule base control unit 4 is notified that the rule conditions have been met. Since the rule condition for the identifier selected as the next rule following the rule for identifier 1 is satisfied, the rule base control unit 4 sets the index of the rule for the identifier corresponding to the rule for identifier 1, that is, the matrix table' 4
The index of the rule w(1,
k) and updates this index w(1,k) based on the above equation (2), thereby increasing the index W(1,k).

Furthermore, returning to the process from step 103, since the condition of the rule for the identifier is satisfied, this rule is
It is determined whether the identifier to be read is set in the matrix of identifiers of each rule corresponding to the "rule pointer table" (step 103), and if it is set, the identifier of the rule arranged first is read from the matrix. A pointer to the rule indicated by this identifier is read from the pointer table 21, and the rule indicated by this pointer is delivered to the inference processing unit 5 (Step 104), and it is determined whether the conditions of this rule are satisfied or not. (Step 105).

Therefore, if the conditions of the predetermined rule are satisfied, the process returns to step 103 and the r of the predetermined rule is satisfied.
Read the identifier of the next rule from the matrix of identifiers of each rule corresponding to ule pointer table',
It is then determined whether the conditions of the next rule are satisfied. Then, when the condition is satisfied, the index of the identifier of the law rule arranged in the row corresponding to the predetermined rule in the matrix table 41 is updated to become larger based on the above formula (2). .

By repeating each of the steps 103 to 108 in this manner, a series of inferences is made by sequentially selecting rules from among the rules in the rule base memory 1 and sequentially interpreting these rules. ``Then, the index in the matrix table 41 is updated depending on whether the condition of the selected rule is satisfied or not.

In this inference, one rule -ule point
When all identifiers are sequentially read from the matrix of identifiers of each rule corresponding to ter table, and it is sequentially determined that the conditions of each rule indicated by these identifiers do not hold, and therefore there are no more identifiers to be read from the matrix. In step 1, the rule base control unit 4 determines that the identifier to be read is not set in the matrix.
03), notifies the inference processing unit 5 of this fact. When the inference processing unit 5 receives this notification, it determines that the inference has ended, and notifies the result of this inference to the outside via the external interface unit 6.

At this time, since the series of inferences has been completed, the rule base control unit 4 stores each matrix of identifiers set as 1: rule pointer table for each rule in the rule base memory 1 in each row in the matrix table 41. Update based on the exponent of (
Step 109). That is, matrix table 4
Select each index in the first row of 1 in order from the largest value, arrange each identifier indicating each index in this selection order,
This forms a matrix of each identifier. This nine matrix is the first
This corresponds to the rule indicated by the line 0, that is, the identifier 0, so the rule pointer of this rule
It is stored as a matrix of identifiers of each rule corresponding to table, and the matrix is updated accordingly. Also, each index in the second row in the matrix table 41 is selected in order from the largest value, each identifier indicating each index is arranged in this selection order, and the matrix of each identifier thus formed is created by identifier 1. rule poi of the indicated rule
The identifier of each rule corresponding to the rule table is stored as a matrix, and the matrix is updated accordingly. Thereafter, similarly, the third to nth rows in the matrix table 41 are associated with each identifier matrix included in each rule indicated by each identifier 3 to n, and each matrix is updated accordingly. .

In this way, when making inferences, the index in the matrix table 41 is set so that it increases if the rule conditions are met.
Also, if the condition does not hold, the matrix table 41 is updated so that it becomes smaller.
Since each identifier matrix included in each rule is updated based on the size of the exponent in The rule with the highest probability will be selected preferentially. In addition, because the index of each rule to be selected next is set for each rule, no matter how you construct the network of rules, you can easily infer from one rule to the next. This can be done with great efficiency.

[Effects of the Invention] As explained above, according to the knowledge inference processing device according to the present invention, for each rule, each selection of each rule to be selected subsequent to the one rule is determined. Set indices so that the next rule is selected based on the selection index of each rule that is selected following one rule, and those indices are updated when a series of inferences is completed. I am trying to keep a history of my inferences. Therefore, a network of rules can be constructed in various ways without any difficulty, and the next rule that is effective in each network can be selected with good efficiency.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the knowledge inference processing device according to the present invention, FIG. 2 is a diagram showing the configuration of a pointer table in the rule base memory in the embodiment shown in FIG. 1, and FIG. 1 is a diagram showing the configuration of Lulu in the rule base memory in the embodiment shown in FIG. 1, FIG. 4 is a diagram showing a matrix table in the rule network matrix memory in the embodiment shown in FIG. 1, and FIG. 1 is a flowchart used to explain the operation in the embodiment shown in FIG. 1... Rule base memory, 2... Rule network matrix memory, 3... Working memory,
4. Rule base control unit, 5. Inference processing unit, 6. External interface unit, 21. Pointer table, 41
...Matrix table. Figure 1 (2 problems, Figure 3, continuous use, 1r coffin rule identifier)

Claims (1)

[Scope of Claims] A knowledge base having a plurality of rules including respective conditions, and a process of selecting a rule from this knowledge base and determining whether or not the conditions of this rule are satisfied are sequentially performed. In a knowledge inference processing device comprising an inference processing unit that makes a series of inferences, for each rule in the knowledge base, the one
a storage means for storing a selection index of each of said rules that are to be selected successively after said one rule; Then, the selection index of the selected rule among the selection indices of the rules to be selected subsequent to the one rule is updated based on whether or not the conditions of the selected rule are satisfied. updating means, the inference processing section reads out selection indices of each of the rules to be selected subsequent to one rule from the storage means, and selects the next rule based on these selection indices. A knowledge inference processing device characterized by: