JPH0432927A - Fast inference device - Google Patents

Abstract

PURPOSE:To speed up a fact data deleting process in a discrimination network by collating a rule that an inference engine performs during inferring with fact data by tracking the discrimination network in a rule storage area with the fact data in a real data storage area. CONSTITUTION:The fact data reaching an alpha storage node are stored by a process in the discrimination network when the fact data are added. At this time, the fact data and the position of the alpha storage node are stored in a storage node table 301. Then when a process for deleting the fact data is started, the storage node table 301 is referenced without tracking the discrimination network from the head node and thus the alpha storage node which is stored is accessed directly and deleted. Further, a beta storage node following the alpha storage node is deleted by retrieving the storage contents by tracking only storage nodes until it becomes evident that the nodes are not stored. Consequently, the deletion of the fact data in the discrimination network is speeded up.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a high-speed inference device that performs inference at high speed using knowledge of rules and facts.

Conventional technology Conventionally, inference devices that perform inference by comparing data representing facts (hereinafter referred to as fact data) and rules (hereinafter referred to as rules) have to perform a large number of processes to compare fact data and rules. did not become. Therefore, there is a problem that inference takes too much time when there is a huge amount of rules and factual data or when real-time processing is required.

To solve this problem, a well-known method is to expand the condition part of a rule into a discrimination network and perform high-speed matching. Discrimination networks are based on Charles El Forgy's “A Fast Algorithm for the
Many Patterns / Meni, -Object Patterns
“Matsuchi Proprem” Artificial Intelligence, 1982 (Chales L, Forgy
＋”Rete: AFast Algorithm
for the Many Pattern/M
any 0bject Pattern Match
Problem *Artificial Int
intelligence, Vol. 19. pp.

17-37. 1982. ) is generally known.

Hereinafter, a conventional inference device that performs inference using a discrimination network will be briefly described with reference to FIGS. 10, 11, and 12.

In FIG. 10, an inference device 701 includes an inference engine 2 that performs inference, a rule storage area 3, and a fact data storage area 4.
It is composed of. A rule is composed of a part expressing a condition (condition part) and an execution part that is executed when the condition part is satisfied.

FIG. 11 is a flowchart of inference processing performed by a conventional inference device. When inference is started (step 70), the inference engine first compares the rule condition part with fact data (steps 71-72). Next, step 73 selects one combination with the highest priority from among the combinations of rules and fact data that satisfy the rule conditions in the comparison.
), the processing of the rule execution part is performed (step 74).

It is also possible to delete or add fact data through the processing of the execution unit.

When the rule execution section processing is completed, the process returns to verification (step 71). The cycle of matching and execution of the rule execution unit is repeated until the rules and fact data are no longer matched (steps 72 and 75).

FIG. 12 is a conceptual diagram showing an example of a discrimination network of a conventional inference device.

A discrimination network is one in which rules are converted into a network format convenient for matching.

Verification processing using a discrimination network means that added or deleted fact data follows the discrimination network by processing according to instructions from each node from the top node of the discrimination network. Reaching the end of the discrimination network means that a certain rule condition part has been satisfied.

The discrimination network achieves faster matching than conventional matching due to the following two points. First, common parts of rule condition parts are expressed as common test nodes to avoid redundant condition checks on fact data. Second,
Since the current fact data comparison status is stored in the storage node in the discrimination network, the comparison process only needs to be performed for changes in the fact data.

Each node of the discrimination network can be classified into five types, ie, a head node, a test node, a storage node, a connection node, and a terminal node, depending on the processing performed at the node. The head node is the root node where fact data is first processed. Test node Hiro corresponds to one of the conditions performed within the rule condition section. Fact data that follows a test node is traced to the next node if the value of the fact data is checked and holds true, otherwise it is not traced.

The storage node is a node that stores the arrived fact data. If the factual data being traced is additional,
It is stored in the storage node, and in the case of deletion, it is deleted from the storage node. The storage nodes include an α storage node located between the test node and the combination node, and a β storage node located next to the combination node. At the join node, the relationship between the values of the fact data that has been traced is checked, and if the relationship is established, the fact data are joined and the next node is traced. Therefore, uncombined fact data is stored in the α storage node, and combined fact data is stored in the β storage node. The terminal node is the terminal end of the discrimination network. When there is fact data that has arrived at the terminal node, it means that all of the conditional parts of a rule are satisfied by the fact data that has arrived.

Problems to be Solved by the Invention However, although efficiency generally improves in inference using conventional discrimination networks, efficiency decreases when factual data is frequently added or deleted during inference. There were drawbacks.

Furthermore, in the method of allocating a new storage area each time fact data is created, the capacity of the storage area becomes enormous, resulting in lower memory usage efficiency and lower performance efficiency. Therefore,
There was a problem in that storage area management became complicated.

The first object of the present invention is to speed up the process of deleting factual data within the discrimination network in view of the above-described shortcomings in efficiency of the discrimination network and shortcomings in the storage area management method. The second purpose is to speed up the storage processing of the fact data storage area. Furthermore,
The third purpose is to speed up the process of adding fact data within the discrimination network.

Means for Solving the Problems To achieve this object, the first invention of the present invention provides a rule storage area for storing a discrimination network and a fact data storage area for storing factual data representing facts and inference situations. , an inference engine that performs inference by collating the discrimination network and fact data, and a storage node table that stores the correspondence between the fact data and storage nodes in the discrimination network.

Further, the second invention of the present invention provides a rule storage area for storing a discrimination network, a fact data storage area for storing fact data representing facts and inference situations, and performing inference by collating the discrimination network and fact data. It is configured to include an inference engine and a storage area management unit that manages free areas and used areas of the factual data storage area.

Furthermore, the third invention of the present invention provides a rule storage area for storing a discrimination network, a fact data storage area for storing fact data expressing facts and inference situations, and performing inference by collating the discrimination network and fact data. An inference engine, an inference preprocessing unit that finds a node of a discrimination network to start matching when creating fact data, and an input node table that stores the matching start node for the rule execution unit that creates fact data are provided. It is composed of

Effects of the present invention With the above configuration, in the first invention, the process of matching rules and fact data performed by the inference engine during inference is a process in which the fact data in the fact data storage area follows the discrimination network in the rule storage area. carried out by
When the process of deleting fact data is started, by referring to the storage node table without performing the process of tracing from the first node of the discrimination network, the process of deleting fact data in the discrimination network becomes faster than before. It has the effect of speeding up inference.

Next, in the second invention, the storage area management section secures an area from an unused area and writes the fact data as a used area when fact data is added, and when fact data is deleted. simply changes the area where fact data is stored from a used area to an unused area. In this way, by using unused space, the memory usage efficiency of the factual data storage area is improved, so it is possible to alleviate the constraints on the required storage area for the operating system and hardware that are the environment in which the inference device operates. The efficient use of the storage area has the effect of speeding up the storage processing of the fact data storage area and speeding up the inference.

Furthermore, in the third invention, when the execution part of the rule is executed during inference, additional processing of fact data is performed by performing collation processing from the node stored in the input node table, without tracing from the top node of the discrimination network. Therefore, unnecessary verification processing can be omitted. In this way, there is an effect of speeding up the process of adding fact data within the discrimination network and speeding up the inference.

Example Embodiments of the present invention will be described below with reference to the drawings.

First, a first embodiment of the present invention will be described with reference to FIGS. 1, 2, 3, and 4.

As shown in FIG. 3, the high-speed inference device 1 includes a rule storage area 3 that stores a discrimination network of rules, a fact data storage area 4 that stores fact data representing facts and inference situations, and a discrimination network and facts. It is composed of an inference engine 2 that performs inference by collating data, and a storage node table 301 that stores the correspondence between fact data and storage nodes in the discrimination network.

FIG. 1 is an example of rules. Let me first explain the meaning of the rules.

Rule A101 is that if the farmer is on the left bank and the cabbage is on the left bank, both the farmer and the cabbage will cross the river and move to the right bank. Rule B102 is a rule that if the farmer is on the left bank and the goat is on the left bank, the farmer and the goat will both cross the river and move to the right bank. Rule ClO3 is a rule that if the farmer is on the right bank, he crosses the river and moves to the left bank. For example, the process of moving to the right bank is a process of deleting fact data that the user is on the left bank and adding fact data that the user is on the right bank.

These rules are converted into a discrimination network shown in FIG. 2 and stored in the rule storage area 3 of the high-speed inference device 1.

The inference execution is performed by the process shown in the flowchart shown in FIG. 11, as in the conventional example. The comparison with the fact data is performed by the fact data tracing the network links from the first node 201 as shown in FIG. The test node, which is a node that examines the content of fact data, checks the value of fact data in response to one of the conditions performed in the rule condition part, and if it is true, the fact data is traced to the next node. , otherwise it will not be traced. Nodes 202 to 208 check whether the subject is a farmer and whether the location is on the left bank. When the fact data reaches the terminal node 217, it indicates that the condition part of rule A is satisfied. Terminal node 2
18 and terminal node 219 are the condition parts of rule B and rule C, respectively. The terminal node stores information indicating which rule's condition part is satisfied and the execution part of that rule. After the matching process, one is selected from among the fact data that have reached the terminal node and the matching is established, and the processing is performed by the rule execution unit stored in the terminal node.

Fact data storage area 4 stores fact data indicating that the farmer is on the left bank and fact data indicating that the cabbage is on the left bank. These fact data are compared with the discrimination network in the rule storage area 3 as follows. The fact data that the farmer is on the left bank traces the test node 202 from the first node. There, a check is made to see if the subject is a farmer. This is true because the subject of the fact data is a farmer, and then one of the two nodes 205 and 206, 205, is traced. Test node 205
is valid because it checks whether the location is on the left bank,
The α memory node 209 is traced. When the α storage node is reached, the α storage node pointer is stored in the corresponding location of the storage node table 301.

In this case, the information is stored in the identification number 1 of the fact data that the farmer is on the left bank. This fact data follows the next join node 213, but does not follow the next node because it cannot be joined.

Nodes 206, 203, and 204 are traced, but since no check is established for any of them, this fact data matching process ends.

When similar processing is performed for the fact data indicating that the cabbage is on the left bank, the α storage node 211 is stored in the storage node table. Also, this fact data is the connection node 2]
3 is combined with the factual data indicating that the farmer is on the left bank, and the β storage node 215 is traced to the terminal node 217. This means that all the conditional parts of rule A are satisfied. In this way, the storage node table 301 stores the pointer of the α storage node when fact data is stored in the α storage node during the collation process.

If the execution part of rule A101 is executed here, the fact data that the farmer is on the right bank will be deleted. In the conventional fact data deletion process, a process similar to the addition process follows a discrimination network, and when a storage node is reached, the stored fact data is deleted. However, by referring to the storage node table from the identification number of the fact data to be deleted, it is possible to directly know the α storage node in which the fact data is stored. For example, when deleting the fact data that a farmer is on the left bank, it can be seen that this fact data is stored in the α node 209 by referring to the contents of the storage node table 301. After deleting the fact data from the storage contents of the α storage node, the processing of tracing the lower nodes from the α node, searching the β storage node until no fact data exists, and deleting it is continued. As a result, when additional verification of fact data is performed, the process of deleting fact data in the discrimination network can be considerably speeded up simply by registering it in the storage node table, which has the effect of speeding up the overall inference.

Next, a second embodiment of the present invention will be described with reference to FIGS. 1, 2, 5, 6, and 7.

As shown in FIG. 5, the high-speed inference device 1 includes a rule storage area 3 for storing a discrimination network, a fact data storage area 4 for storing fact data representing facts and inference situations, and a rule storage area 3 for storing a discrimination network and a fact data storage area 4 for storing fact data representing facts and inference situations. It is composed of an inference engine 2 that performs inference by comparison, and a storage area management unit 401 that efficiently manages free areas and used areas of the fact data storage area. The storage area management unit 401 is composed of a queue management table 402 and a badge table 403 (see FIG. 6).

FIG. 1 is an example of rules. Let me first explain the meaning of the rules.

Rule A101 is a rule that if the farmer is on the left bank and the cabbage is on the left bank, both the farmer and the cabbage cross the river and move to the right bank. Rule B102 is a rule that if the farmer is on the left bank and the goat is on the left bank, the farmer and the goat will both cross the river and move to the right bank. Rule ClO3 is a rule that if the farmer is on the right bank, he crosses the river and moves to the left bank. For example, the process of moving to the right bank is a process of deleting fact data that the user is on the left bank and adding fact data that the user is on the right bank.

These rules are converted into a discrimination network shown in FIG. 2 and stored in the rule storage area 3 of the high-speed inference device 1.

The inference execution is performed by the process shown in the flowchart shown in FIG. 11 as in the conventional case.

The comparison with the fact data is performed by the fact data tracing the network links from the head node 201, as shown in FIGS. 2 and 6. The test node, which is a node that examines the contents of the fact data, corresponds to one of the conditions performed in the rule condition part, and if the value of the fact data is checked and is true, the fact data is traced to the next node. It must be so. Nodes 202 to 208 check whether the subject is a farmer and whether the location is on the left bank. When the fact data reaches the terminal node 217, it indicates that the condition part of rule A is satisfied. The terminal nodes 218 and 219 are the condition parts of rule B and rule C, respectively. The terminal node stores information indicating which rule's condition part is satisfied and the execution part of that rule. After the matching process, one is selected from among the fact data that have reached the terminal node and the matching is established, and the processing is performed by the rule execution unit stored in the terminal node.

The fact data is stored in the fact data table 404 and the character string table 4 in the fact data storage area 4, as shown in FIG.
It is stored in 07.

One row in the fact data table 404 is an area for storing one piece of fact data (referred to as a fact data area), a pointer for management is written in the first pointer area 405, and a data area 406 is used to store the fact data. Data content The data can be written and sized appropriately. In the pointer area 405, by writing pointers to the fact data areas, the fact data areas are connected with pointers and managed in a queue format. If the data type of the data area 406 is a numeric value, the numeric value is directly written, but if it is a character string, a pointer to the data entity taken in the character string table 407 is written. The key management table 402 of the storage area management unit 401 uses three areas: an in-use queue, a deletion queue, and a free queue to create a fact data table 404.
The fact data area of is managed as follows (Figure 7 (a)
reference).

In the initial state when the system is started up, all fact data areas are connected to free queues. That is, in the free queue of the queue management table 402, a pointer to the area of the first row of the fact data table 404 is written, and the pointer area 405 of the area of the first row is written.
In this way, all fact data areas are connected, with a pointer to the next line area being written. Next, when fact data is created, the contents of the fact data are written in the fact data area at the head of the free queue, and this area is connected to the busy queue.

When factual data is deleted, the factual data storage area is linked to a deletion queue. When processing additional factual data,
If there is no free queue, a fact data area is secured from the deletion key area.

If nothing is connected to the deletion queue, the fact data area of the free queue is secured and used.

The expression format of factual data and rules representing human knowledge includes
In general, character strings are often used instead of numbers. Additionally, these strings are used repeatedly both within the factual data and within the rules. Therefore, if the same character string is centrally managed in one character string area, instead of taking up a character string area each time a character string appears, it is possible to use the memory area economically.

In the embodiment, by storing the actual character string in a character string table 407 and providing a hash table 403 in the storage area management unit 401, a high-speed character string-to-source chain ring is realized by hashing.

Fact data is created during inference, and character strings included in the fact data are stored in the data area 406 of the fact data table 404.
When storing the data, the following processing is performed.

First, check whether the character string is already stored in the character string table 407 in the hash table 403, and if it is stored, return the pointer of the character string,
If it is not stored, it is stored in the character string table 407 and its pointer is returned (see FIG. 7(b)).

As a result, the area for storing fact data can be used without waste, which is effective in speeding up the storage processing of fact data and speeding up inference.

Next, a third embodiment of the present invention will be described with reference to FIGS. 1, 2, 8, and 9.

As shown in FIGS. 8 and 9, the high-speed inference device 1
A rule storage area 3 that stores a discrimination network, a fact data storage area 4 that stores fact data representing facts and inference situations, an inference engine 2 that performs inference by collating the discrimination network and fact data, and creates fact data. It consists of an inference preprocessing unit 602 that finds the node of the discrimination network at which to start matching, and an input node table 601 that stores the matching start node for the rule execution unit that creates fact data.

FIG. 1 is an example of rules. First, explain the meaning of the rules.

Rule A101 is a rule that if the farmer is on the left bank and the cabbage is on the left bank, both the farmer and the cabbage cross the river and move to the right bank. Rule B102 is a rule that if the farmer is on the left bank and the goat is on the left bank, the farmer and the goat will both cross the river and move to the right bank. Rule ClO3 is a rule that if the farmer is on the right bank, the farmer and her husband cross the river and move to the left bank. For example, the process of moving to the right bank is a process of deleting fact data that the user is on the left bank and adding fact data that the user is on the right bank.

These rules are converted into a discrimination network shown in FIG. 2 and stored in the rule storage area 3 of the high-speed inference device 1.

Before inference, the inference preprocessing unit 602 performs the following processing.

The fact data creation function of each rule execution unit of the terminal nodes 217, 218 and 219 is referred to, and fact data is created. At that time, there are values that are undefined at the moment, such as variables, but are determined only at the time of inference.

Here, we will call it an indefinite value. The fact data is α
Each test node up to the memory node is sequentially checked to see if it holds true. Do not follow the node that checks for an unspecified value. A node that is established by checking the value of a definite value that is not an indefinite value further traces the next node.

As a result of this processing, the factual data either does not reach any node, reaches a test node that performs a comparison of undefined values, or reaches an α memory node. The pointer of the node reached is written into the input node table 601 for each rule and fact data addition function. For example, in the case of the second execution part of rule A, which is a function that adds data about the fact that the farmer is on the right bank, the first node 201 is reached, so this pointer is stored in the input node table.

The inference execution is performed by the process shown in the flowchart shown in FIG. 11 as in the conventional case.

The comparison with the fact data is performed by the fact data tracing the network links from the head node 201, as shown in FIGS. 2 and 9. The test node, which is a node that examines the contents of the fact data, corresponds to one of the conditions performed in the rule condition part, and if the value of the fact data is checked and is true, the fact node traces to the next node. Otherwise I can't follow it. Nodes 202 to 208 check whether the subject is a farmer and whether the location is on the left bank. When the fact data reaches the terminal node 217, it indicates that the condition of rule A is satisfied. In the case of the terminal node 218 or 219, this is the condition part of rule B or rule C, respectively. The terminal node stores information indicating which rule's condition part is satisfied and the execution part of that rule. After the collation process, one is selected from among the fact data arrived at at the terminal node where the collation is established, and is processed by the rule execution unit stored in the terminal node. However, factual data is added during the processing of the rule execution part,
When performing matching with the rule condition part, the input node table 601 is used instead of tracing from the first node of the discrimination network.
Verification processing is performed starting from the node stored in .

With these, it is possible to eliminate wasteful matching processing in the discrimination network and realize faster matching.

The explanation of the embodiments described above can be summarized as follows.

The matching process between rules and fact data performed by the inference engine during inference is performed by a process in which the fact data in the fact data storage area follows the discrimination network in the rule storage area. First, when fact data is added, when the α storage node is reached in the processing within the discrimination network, processing to store the fact data is performed. At this time, this fact data and the position of the α storage node are stored in the storage node table.

Next, when the process to delete this fact data is started, by referring to the memory node table without performing the process of tracing from the first node of the discrimination network, the stored α memory node can be directly accessed and deleted. . As for the β storage nodes below the α storage node, searching for and deleting the storage contents by tracing only the storage nodes is continued until it is determined that the storage contents are not stored in the storage nodes.

In this way, by significantly reducing the process by which fact data follows the discrimination network, the process of deleting fact data in the discrimination network becomes faster than in the past, which has the effect of speeding up inference.

1. The process of matching rules and fact data performed by the inference engine during inference is performed by a process in which the fact data in the fact data storage area follows the discrimination network in the rule storage area. The storage area management section divides the fact data storage area into an unused area and a used area and performs the following management. When factual data is added, an area is secured from the unused area and used as a used area, and the factual data is written therein. When fact data is deleted, the area where the fact data is stored is simply changed from a used area to an unused area. In this way, by using the unused area, the memory usage efficiency of the fact data storage area is improved, so that the constraints on the necessary storage area for the operating system and hardware that are the environment in which the inference device operates can be relaxed. In addition, the efficient use of storage area means faster storage processing of fact data storage area,
It has the effect of speeding up inference.

Furthermore, the process of matching rules and fact data performed by the inference engine is performed by a process in which the fact data in the fact data storage area follows the discrimination network in the rule storage area. Before the inference is executed, the inference preprocessing unit checks all rule execution units to see if there is a process for adding fact data. For each fact data addition process, it is checked to which node in the discrimination network the fact data created by the addition process can be traced, and that node is stored in the input-to-touchful. When executing the rule execution part during inference, additional processing of fact data is performed from the node stored in the input node table without tracing from the top node of the discrimination network, eliminating unnecessary matching processing. be able to. In this way, there is an effect of speeding up the process of adding fact data within the discrimination network and speeding up the inference.

Effects of the Invention As described above, according to the first invention, the process of deleting fact data in a discrimination network can be sped up with almost no additional processing burden, and inference can be sped up. Become.

Further, according to the second invention, since the capacity of the fact data storage area can be reduced, it is possible to alleviate the constraints on the necessary storage area for the O8 and hardware that are the environment in which the inference device operates. Moreover, efficient use of the storage area enables faster storage processing of the fact data storage area and faster inference.

Furthermore, according to the third invention, it is possible to speed up the process of adding fact data within the discrimination network and speed up the inference.

[Brief explanation of the drawing]

FIG. 1 is a conceptual diagram showing an example of rules in an embodiment of a high-speed inference device according to the present invention, FIG. 2 is a conceptual diagram showing a discrimination network in an embodiment of a high-speed inference device according to the present invention, and FIG. FIG. 4 is a block diagram of a high-speed inference device in the first embodiment of the high-speed inference device according to the present invention, and FIG. 5 is a block diagram of the configuration of the high-speed inference device in the first embodiment of the high-speed inference device according to the present invention. , a block diagram of a high-speed inference device in a second embodiment of the high-speed inference device according to the present invention, FIG. 6 is a block diagram of a high-speed inference device in a second embodiment of the high-speed inference device according to the present invention, FIG. (a
) and (b) are flowcharts showing the processing of the storage area management unit in the second embodiment of the high-speed inference device according to the present invention, and FIG. Block diagram of high-speed inference device, No. 9
10 is a block diagram showing a general configuration of a conventional inference device, and FIG. 11 is a block diagram showing a general configuration of a conventional inference device. A flowchart showing the inference processing of the inference device,
FIG. 12 is a conceptual diagram showing an example of a discrimination network of a conventional inference device. 1...High-speed inference device, 2...Inference engine, 3...
Rule storage area, 4...Fact data storage area, 301
. . . Storage node table, 401 . . . Storage area management unit, 402 . . . Queue management table, 403 .
5... Pointer area, 406... Data area, 4
07...Character string table, 601...Input node table, 602...Inference preprocessing unit, 701...Inference device. Name of agent: Patent attorney Shigetaka Awano and one other person

Claims (3)

[Claims] (1) a rule storage area for storing a discrimination network;
A fact data storage area that stores fact data expressing facts and inference situations, an inference engine that performs inference by collating the fact data with a discrimination network, and a storage node that stores the correspondence between fact data and storage nodes in the discrimination network. A high-speed inference device having a table. (2) a rule storage area for storing the discrimination network;
A fact data storage area that stores fact data expressing facts and inference situations, an inference engine that makes inferences by collating the fact data with a discrimination network, and a storage area management unit that manages free space and used area of the fact data storage area. A high-speed inference device having (3) a rule storage area for storing the discrimination network;
A fact data storage area that stores fact data that expresses facts and inference situations, an inference engine that performs inference by collating fact data with a discrimination network, and a pre-inference process that finds the node of the discrimination network that starts collation when creating fact data. A high-speed inference device that has a processing unit and an input node table that stores a matching start node for a rule execution unit that creates fact data.