JPH04123231A - Inference control system in artificial intelligence - Google Patents

Abstract

PURPOSE:To resolve a problem with a simple processing by storing a formed operation program and resolving a program to be resolved with a processing procedure similar to the processing procedure of the stored operation program in accordance with the stored operation program. CONSTITUTION:This inference control system is provided with an inference engine 1 for referring to a problem resolving knowledge base 4 and meta- knowledge base 5 in order to resolve a problem, forming a resolving tree, arranging execution functions positioned on the terminal of the resolving tree in the order of execution to form an operation program indicating a processing procedure for executing problem resolution, and executing the problem resolution in accordance with the processing procedure specified by the operation program. The formed operation program is stored in the knowledge base 4 to be reserved and a problem to be resolved by a processing procedure similar to the processing procedure specified by the operation program is resolved in accordance with the stored operation program. In the case of executing the problem resolution experienced in the past again, the problem resolution can be more easily and rapidly executed without repeating the same inference.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to an inference control method in artificial intelligence, and particularly to an inference control method that solves problems while recognizing the external world and its own state.

[Background Art] Research and development of artificial intelligence is currently being actively conducted, and applications are being considered to control the operation of robots and other machines, or to assist in diagnosis and design.

On the other hand, in fields where artificial intelligence is expected to be applied, the ability to autonomously deal with dynamic environmental changes at any time is often required. In such cases, inference control using artificial intelligence is expected to have the following capabilities.

(1) Problem-solving ability that reflects the recognition results of the external situation Rather than determining the entire course of action in advance using a program, etc., after rough planning, take actions according to the situation at that time while taking into account the surrounding situation. ability to decide.

(2) Ability to interrupt, interrupt, and resume commands at any time. It allows you to receive a command with a higher priority than the command given to the cutting edge later, interrupting the previous command, or after the interrupt command has finished. A running force that resumes execution of a previous command.

(3) Input/output ability using natural language Ability to write commands in natural language and output responses.

By the way, in order to realize the above three abilities, various inference control methods have been researched and developed that solve problems by inference without inputting a problem-solving program in advance, but the conventional inference control methods are Similarly, for five individual problems, the problem was solved by reading problem-solving knowledge and processing methods from the problem-solving knowledge base and meta-knowledge base and making inferences.

[Problems to be Solved by the Invention] However, the above-mentioned inference control method does not use problem-solving knowledge or processing in the same way each time, even when problem solving is repeated many times using the same processing procedure, such as basic operation. Since a problem is solved while reading out a method and making inferences, there is a lot of wasteful processing, and in such cases there has been a desire for an inference control system that solves the same problem through simpler processing.

The inference control method in artificial intelligence of the present invention has been developed in view of the above-mentioned conventional problems, and provides a means for solving the above-mentioned problems with simpler processing when resolving problems experienced in the past. The purpose is to

[Means for Solving the Problems] The invention of item 1 that achieves the above object includes an inference engine as a control means, a self-state recognition unit that monitors its own state, and an external environment that monitors the state of the outside world. a recognition unit, a problem-solving knowledge base storing knowledge rules for problem-solving,
It has a meta-knowledge base that stores information indicating a problem-solving execution method and a data processing method, solves problems according to the information in the problem-solving knowledge base and meta-knowledge base, and uses the self-state recognition unit and external environment recognition In an inference control method in artificial intelligence that recognizes the external world situation using information from a computer, the inference engine analytically calculates a solution to the problem by referring to the problem-solving knowledge base and meta-knowledge base in order to solve the problem. Generate a solution tree shown in the diagram, arrange the execution functions located at the end of the solution tree in the order of execution, generate an operation program for solving the above problem, solve the problem according to the processing steps shown in the operation program, and , memorize and save the generated operating program,
The present invention is characterized in that, for problems that can be solved by a processing procedure similar to the processing procedure shown in the operating program, the problem is solved according to the stored operating program.

The invention of item 2 provides that in the invention of item 1, the self-state recognition unit, the external environment recognition unit, the problem-solving knowledge base, the meta-knowledge base, and the problem or command input means are located equally with respect to the inference engine. death.

The inference engine cyclically searches the self-state recognition unit, external environment recognition unit, problem-solving knowledge base, and meta-knowledge base each time it executes one execution function that makes up the operating program, and searches for new information. It is characterized by allowing the input of questions and information.

The invention of paragraph 3 is the invention of paragraph 1 or paragraph 2,
The present invention is characterized in that a natural language input means is provided as a means for inputting motion data, and motion commands are input in natural language.

[Examples] Examples of the present invention will be described below with reference to the drawings.

In this embodiment, an example will be described in which the inference control method is applied to the operation control of an autonomous mobile robot.

FIG. 1 is a block diagram showing a system configuration of an inference control method according to an embodiment of the present invention.

As shown in the figure, the inference control method of this embodiment includes an inference engine 1 as a control means, a self-state recognition unit 2 as an input unit, an external environment recognition unit 3, a natural language sentence analysis unit 6, and a problem solving unit. It has a problem-solving knowledge base 4 and a meta-knowledge base 5 as parts, as well as an output part 7.

The natural language sentence analysis unit 6 performs morphological analysis and syntactic analysis on the given natural language sentence, extracts the structure of dependencies with conceptual symbols as nodes, and converts the results into a frame structure. is sent to the inference engine l as natural language sentence interpretation data.

Note that the morphological analysis by the natural language analysis unit 6 is realized by using a conventional method. In addition, as a means of syntax analysis, context free grammar (context fr
analysis means based on EE grammar and augmented transition network.
Any method may be used as long as it extracts a dependency structure in which concept symbols are nodes and inter-concept dependency relationships are arcs, using t) or the like.

Here, a conceptual symbol is a symbol that expresses the meaning of a word without depending on a specific language.For example, "hon" in Japanese and "book" in English are both words that express the same concept. , this is expressed using a conceptual symbol, for example, CBOOK.

As a result, the entire system can be made compatible with various languages such as Japanese, English, and French by simply changing the morphological analysis section and the syntactic analysis section without changing the knowledge base.

The inference engine 1 extracts necessary information from input data, sets a problem-solving goal, and performs necessary processing.

The problem-solving knowledge base 4 classifies and stores knowledge for solving individual problems according to rules such as condition-result type, purpose-means type, and time/sequence type (hereinafter referred to as knowledge rules). . The classification and actual examples of this knowledge rule are explained in the third section.
As shown in the figure.

The meta-knowledge base 5 stores meta-knowledge for preventing the problem-solving control structure from becoming dependent on the rule type of the problem. Specifically, it is knowledge about how to interpret and execute each problem-solving knowledge and how to process external world data, self-data, and Japanese sentence interpretation data. Because the meta-knowledge is placed outside the control structure, the control structure does not depend on individual problems, and interruptions such as information from external world recognition and commands from the outside are allowed even in the middle of processing a certain problem. becomes. Even if a completely new type of problem-solving knowledge becomes necessary due to the realization of meta-knowledge, knowledge rules that indicate how to process that knowledge can be registered as meta-knowledge and a new problem-solving knowledge base can be added. , it becomes possible to expand the system without changing the control structure.

Figure 4 shows the relationship between problem rule types and meta-knowledge.

The self-state recognition unit 2 has slots in the form of a frame that indicate the internal state of the robot, and monitors which slot the robot's own state fits into. An example of a throat and its contents are shown in FIG.

The external environment recognition unit 3 has a plurality of coordinate systems, a label is given to each coordinate in each coordinate system, and the external environment recognition unit 3 is a 60-bottom system that stores the label as a value unique to the object on the coordinate. simply performs environmental recognition by searching the surrounding area and, if it finds the label of the object it is aiming for, reading its coordinates.

The output unit 7 outputs the progress and results of problem solving by the inference engine 1 through graphics display and inference process display.

Next, a problem solving procedure using the inference control method of this embodiment will be explained with reference to FIG. The problem solving according to this embodiment is performed based on a cyclic processing process as shown in the figure.

First, when a command is input as a natural language sentence, the natural language sentence analysis section 6 analyzes the command and sends it to the inference engine 1 as natural language sentence interpretation data.

Hereinafter, the processing contents of the inference engine 1 will be explained for each of the processes 1 to 5 shown in the figure.

In process 1, [If natural n-word sentence interpretation data is newly generated, a behavioral goal (goal) is generated from the interpretation data, and the result is stored in a storage area called a goal recognition slot; Clear the area where natural language sentence interpretation data was stored. ” will be processed.

In process 2, if a goal exists in the goal recognition slot y), a knowledge rule with that goal as the objective part is searched from the problem-solving knowledge base 4, and the content and type of the knowledge rule are determined. It is stored in a storage area called a rule slot, and the goal recognition slot is also cleared. In Process 2, which performs the process with the content, ``If a knowledge rule is stored in the rule slot, the metaknowledge base 5 is searched for metaknowledge for rule interpretation according to the rule type, and the metaknowledge is extracted.'' It is stored in a storage area called a slot.J is processed.

In process 3, if the meta-knowledge for rule interpretation is stored in the meta-knowledge throw/ In one execution of this process,
In many cases, it is not possible to generate a solution tree.A knowledge rule represents a subgoal to achieve a certain goal, and if the subgoal is an execution function that represents an action that can be executed immediately. In some cases, it may be necessary to proceed with problem solving using yet another knowledge rule. Therefore, in the latter case, a different knowledge rule will be used, and meta-knowledge will also be required depending on the rule type. Therefore, by repeating the above process cyclically until all the terminals of all sub-goals become executable functions and combining the respective results, one solution tree that analytically shows the solution to the problem is generated. This means ``If there is a solution tree (1) in the solution tree subslot, and if there is also a solution tree (2) in the storage area called the solution tree slot, then the appropriateness of the solution tree (2) is The solution tree (1) inserted at the position is the new solution tree,
If there is no solution tree, it is necessary to make the solution tree (1) a new solution tree, store the new solution tree in the solution tree slot, and clear the solution tree subslot. This is realized by processing the contents. At this time, instead of repeating only the above process continuously, the inference engine
This should be done only once per cycle. Therefore, self-state recognition and external environment recognition based on the inputs of the self-state recognition unit 2 and external environment recognition unit 3 are possible even during the generation of the solution tree, and the recognition results are required when generating the solution tree. It is also possible to support knowledge rules such as this, and it is also possible to add other command sentences in between.

In Process 3, further, ``If there is a solution tree in the solution tree slot, and there is no unprocessed subgoal in the solution tree, that is, all the ends of the solution tree are expanded to executable functions. If so, it is necessary to create an operation program, store it in a storage area called an operation program slot, and clear the solution tree slot.'' When creating an operating program, it is possible to proceed one step per cycle of the inference engine 1, as in the case of creating a solution tree. However, self-state recognition and external environment recognition are not particularly necessary for creating an operating program, and there is not much benefit even if interrupts are allowed during this time.In this embodiment, a function is provided to create an operating program all at once. This is achieved by calling it and processing it.

In process 4, ``If there is an operation program in the operation program slot 2, if the execution is not completed, execute l steps at a time, and if it is completed, it is necessary to clear the operation program slot.'' When executing an action program slot, the action contents are retrieved and executed sequentially from the beginning of the action program, and the point of interest is moved forward. However, in the case of a jump command, the jump destination is Shift your focus to the label, and if it's a label, skip it.

In process 5, ``If there is no value in each recognition slot and there is stacked (saved) recognition information, the latest saved information is returned to the original storage area. If there is a need to report on goal achievement, do so here.

A specific example of problem-solving processing when a command sentence is input in Japanese to an autonomous mobile robot equipped with the inference control method of this embodiment will be shown below.

For example, assume that the following command sentence is input.

“Go to the vending machine.

The natural language sentence analysis unit 6 that receives this sentence performs morphological analysis on this sentence. The results are as follows.

vending machine/in front of/in front of/to/go/ After completing the morphological analysis, syntactic analysis is performed considering the relationship between each morpheme, and the following dependency structure is created.

＼ CPFROMT: Noun (semantic information: location) ＼ The structure of modification is expressed by conceptual symbols,
Regarding the above Japanese sentence, “CEEND)l” is changed to “vending machine”, “CPFROMT” is changed to “before”, and “CG
ON corresponds to "go".

This modification then reads the structure and describes its semantic content in a frame structure. By writing in a frame structure, the form of the analysis result is less influenced by the style of the natural language sentence. The result described in the frame structure is as follows.

(doushi (CGO (semantic information: action))
eirei-flag t basyo-syuuten CPFRONT
0kakareru nil)
Conflict ・ 11 (1) (meis
hi (C: PFRONT (semantic information: location)) kitai CBEND M 1ka
kareru nil)...
(2) (+eishi (CBENDN (semantic information 2 entity)) kakareru nil)
*. @(3) The instruction sentence thus converted into a frame structure is passed to the next inference engine 1.

In the inference engine 1, first, in process 1, necessary information is extracted from the frame structure as a result of interpreting the command sentence and the sentence, and it is put into the form of a goal that the robot can understand.

From the frame structure above, (G GO:p dest C
A goal of the form PFRONT:kitai CBEND is generated.

When a goal is generated, the inference engine 1 searches the problem-solving knowledge base 4 for knowledge rules to achieve the goal while going through processes 1 to 5, and uses the meta-knowledge in the meta-knowledge base 5. By sequentially solving the subgoals, a solution tree is generated in the best possible way, and an operating program is generated from it.

First, a solution tree is generated. To do this, a knowledge rule with an objective that matches the goal is searched from the problem-solving knowledge base 4 that stores goal-means type knowledge rules (assuming that only one knowledge rule exists). In this example, , the following knowledge rules are obtained.

Rule l: (Rule-I T-ioD (CGo :p dest CPFRONT:kita
i CBEND M) ((CKNOWN :d obj CLOCATE:k
itai CBEND N)) ((CMOVE :p destCPFRONT:ki
tai CBEND M))) In addition, knowledge rules are expressed with conceptual symbols, such as “CKNOW”.
"N" corresponds to the Japanese word "fixed," and "CMOVE" corresponds to the Japanese word "move."

Rule 1 states, ``In order to achieve the goal of ``go to the vending machine,'' under the condition that the subgoal of ``I know the location of the vending machine'' is achieved,
All you have to do is achieve the subgoal of "move to the vending machine." ”. (conditional execution action command type) knowledge rule (see Figure 4).

Since this knowledge rule includes a subgoal, another knowledge rule is required to achieve the subgoal.

Rule 2: (Rule-2cho-01K (CKNOWN:d obj CLOCATE:ki
tai G BEND M) (OBJECT-PO3ITION CBEND M)
((CLOOK :d obj CLOGATE:ki
tai CBENII N))) Rule 2 states, ``In order to achieve the goal of ``I know the location of the vending machine'', the execution function 0BJE(:T-POSITION
(a function that obtains the position information of an object from the robot's own memory), and if the result is nil, the subgoal of ``visualize (in front of) the vending machine'' may be achieved. ”. (Unconditional execution confirmation command type)
This is the knowledge rule.

Rule 3: (Rule-37-010 (CLOOK :d obj CLOCATE:kit
ai CBENII N) ((SEARCH POSITION CBEND N)
)) Rule 3 is “Visually observe (in front of) the vending machine.”
In order to achieve this goal, the execution function 5EARG
)l POSITION (a function that obtains the position information of an object from the outside world using the robot's visual unit) is a knowledge rule (unconditional execution command type) that expresses the following.

Rule (Rule-4 T-01L (C MOVE :p dest C PFRONT:
kitai C BENII M) (SEARCH
OBJECT C BEND M :len O)((
CMOVE 1 :p org (get 'rob
position): p dest C:
BEND N: near t : speed 'normal))) Rule 4 states that ``In order to achieve the subgoal of ``move to the front of the vending machine'', the execution function SEAR(J
OBJECT (Using the robot's vision unit,
The execution function CMOVE 1 until the function that determines whether an object exists within the specified distance range becomes non-nil.
(The function that activates the robot's walking unit and moves in the specified direction) can be activated repeatedly. ”.

(unconditional execution loop imperative type) knowledge rule.

Developing the goal according to these knowledge rules results in the sixth
As shown in the figure, tree-structured data representing the hierarchical relationship between goals, subgoals, and execution functions can be obtained. Here, Figure 6(a) shows the goal, and Figure 6(b) shows the rule 1.
Figure 6 (C) shows the state where Rule 2 is applied, Figure 6 (d) shows the state where Rule 3 is applied, and Figure 6 (e) shows the state where Rule 4 is applied. Each is shown below.

However, in reality, it is necessary to add control information such as branching and repetition. What kind of control information must be added depends on the type of rule. The solution tree is tree-structured data that marks the position of the control information where the jump command is added.

The solution tree in this example is shown in FIG.

In the figure, “rn” means “RETURN-TO-NEW”.
T″ (jump command), rb” is “RETURN-”
To-BEFORE" (jump instruction). Once the solution tree is obtained, a motion program is generated based on this. In addition to the execution function at the end of the solution tree, a jump instruction is added at the position where the jump instruction is added. adds the appropriate jump instruction and a label indicating the destination of the jump instruction.

Here, the specifications of the operating program in this embodiment will be explained.

The operating program that the system generates when solving a problem is a list of indefinite length that includes control information and execution functions.

The control information here refers to jump instructions within the operating program, labels indicating jump destinations, etc. A list of control information in this embodiment is shown in FIG. 9 (nun shown in the figure).

now-A, nun-8 indicates a positive integer).

The execution function is one that represents a primitive motion for the robot, or a combination thereof. A list of the execution functions set in this embodiment is shown in FIG.

Note that there are some functions that are handled in slightly different ways, even though they are included in the abbreviations of execution functions. This is called a metafunction, and in the embodiment, three types shown in FIG. 811 are set.

Based on the above specifications, an operating program generated for the problem solving rule shown in FIG. 7 is shown below.

(OBJECT-POSITION CBEND
M) (RETURN-TO-NEXT 001.) (S
EARCHPOSITION CBfJID M)(S
EARCH0BJEC:T CBEND M:Ien
0)(RETURN-D0-NEXT 001)(
CMOVE i :p org (GET 'RCI
BOT'PCISHION): p dest CBE
ND M :near t :5peed 'NORMAL) (1'?ETURN-D0-BEFORE 001)
The flowchart corresponding to this operation program is the 8th
It will be as shown in the figure. In addition, function 1 in Fig. 8 is (OBJ
ECT-POSITION CBEND +4), function 2 is (SEARCHPO separate TON CBENI)
M), function 3 is (SEARCH0BJECT CBE
ND M :Ien O), function 4 is (C140VE
i:p org (GET 'ROBOT'PO
5ITION) :p dest CBEND M
:near t :5peed '%ORMAL).

Once the operating program is generated, the operating program is executed in process 4. The operating program is executed for each execution function. That is, each time an execution function is executed, each process shown in Figure 2 is executed, and each time a problem that should be solved with priority over the problem currently being executed (for example, a subgoal for a goal of a certain problem, Processing to respond to changes in the external environment, etc.) is allowed. If there is an input, it similarly generates and executes an operating program to solve the problem. If there is no input, the next execution function of the original operating program is executed.

Through the above processing, the robot completes the execution of the command "Go to the vending machine."

In this embodiment, the generated operating program is stored in the knowledge base and saved as new knowledge. and,
When solving a problem using the same processing procedure as "Go to the vending machine", that is, confirming the location of the object and moving it, the operation program generated in this example can be used. Search and read from the knowledge base and solve the problem accordingly.This eliminates the need to repeat the same inference and generate a solution tree each time when the same or similar processing is repeated. , you can solve problems more quickly.

[Effect of the invention] As explained above, the invention of item 1 generates a solution tree by referring to the problem-solving knowledge base and the meta-knowledge base when solving a problem, and generates a solution tree at the end of the solution tree. It generates an operating program that shows the processing steps to solve the above problem by arranging the execution functions to be executed in the order of execution, and is equipped with an inference engine that solves the problem according to the processing steps shown in the above operating program. It is possible to interrupt instructions between each execution function that makes up a program, and it is possible to achieve a high level of problem-solving ability that reflects the recognition results of external situations and the ability to interrupt, interrupt, and resume instructions at any time. There is an effect that it can be done.

Further, the present invention stores and saves the generated operation program in a knowledge base, and uses the stored operation program to solve problems that can be solved using processing steps similar to those shown in the operation program. By solving a problem according to the above, when resolving a problem experienced in the past, the problem can be solved quickly with simpler processing without repeating the same reasoning.

In the second aspect of the invention, the self-state recognition unit, the external environment recognition unit, the problem-solving knowledge base, the meta-knowledge base, and the problem or command input means are located equally with respect to the inference engine, and the inference engine has the following features: Every time one of the execution functions that make up the operating program is executed, the self-state recognition unit, external environment recognition unit, eight problem-solving knowledge bases, and meta-knowledge base are searched cyclically and new information or problems are input. By allowing other instructions to be easily interrupted while the operating program is being executed, the ability to solve problems that reflect the recognition results of external situations and the ability to interrupt, interrupt, and resume instructions at any time can be realized at a higher level. It has the effect of being able to.

The invention of item 3 has the effect that the natural language input/output source capability is dramatically improved because the data input means is provided with a natural language sentence analysis section.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the system configuration of a robot control method according to an embodiment of the present invention, Fig. 2 is a flowchart showing a problem solving procedure according to this embodiment, and Fig. 3 is a problem solving knowledge used in this embodiment. Figure 4 is a diagram showing the contents of the base, Figure 4 is a diagram showing the contents of the meta-knowledge used in this example, Figure 5 is a diagram showing the contents of the self-state recognition section used in this example, and Figure 6 is this example Figure 7 is a diagram showing an example of tree-structured data created when solving a problem, Figure 7 is a diagram showing a solution tree generated from the data in Figure 6, Figure 8 is from the solution tree in Figure 7. Figures 9 to 1 are diagrams showing flowcharts corresponding to generated operation programs.
FIG. 1 is a diagram showing the specifications of the operating program in this embodiment. 1: Inference engine 2: Self-state recognition unit 3: External environment recognition unit 4: Problem-solving knowledge base 5: Meta-knowledge base 6: Natural language sentence analysis unit 7: Output unit Applicant: Shiniskei Co., Ltd. Figure (e) Figure 4 ( Figure 10 Figure 11

Claims (3)

[Claims] (1) An inference engine as a control means, a self-state recognition unit that monitors its own state, an external environment recognition unit that monitors the state of the outside world, and a problem-solving knowledge base that stores knowledge rules for problem-solving. , a meta-knowledge base that stores information indicating how to perform problem solving and how to process data;
In an inference control method in artificial intelligence that solves problems according to information in the problem-solving knowledge base and meta-knowledge base, and recognizes external world conditions based on information from the self-state recognition unit and external environment recognition unit, the inference engine includes: In order to solve the above problem, a solution tree that analytically shows the solution to the problem is generated by referring to the above problem solving knowledge base and meta knowledge base, and the execution functions located at the ends of the above solution tree are arranged in the order of execution. Generate an operation program to solve the above problem, solve the problem according to the processing procedure shown in the operation program, memorize and save the generated operation program, and perform the same processing as the processing procedure shown in the operation program. An inference control method in artificial intelligence, characterized in that for problems that can be solved in steps, the problem is solved according to the stored operation program. (2) The self-state recognition unit, the external environment recognition unit, the problem-solving knowledge base, the meta-knowledge base, and the problem or command input means are located equally with respect to the inference engine, and the inference engine constitutes an operating program. Each time an execution function is executed, the self-state recognition unit, external environment recognition unit, problem-solving knowledge base, and meta-knowledge base are searched cyclically and new information and problems are allowed to be input. An inference control method in artificial intelligence according to claim 1. (3) The inference control method for artificial intelligence according to claim 1 or 2, characterized in that natural language input means is provided as means for inputting motion data, and motion commands in natural language are input.