JPH0289134A - Rule learning processing system - Google Patents

Abstract

PURPOSE:To set a system to deal with respective conditions which cannot be assumed by generalizing concrete names in inputs and an existed condition part with respect to the state inputs which do not satisfy any existed condition parts. CONSTITUTION:When a controller detects the conditions which have not previously been assumed, the concrete names appearing in the condition expression and the condition parts of a rule are replaced (generalization) by an abstracted concept, and the concrete names in the condition expression are substituted in the rule so as to lead a conclusion when the conditions are applied to the rule. Namely, generalization is performed in a new connection generation part 105 when one state in a state storage part 102 does not exist in a rule storage part 101, the conclusion part ('then' part) is converted based on a generalization method, corresponds to the condition part ('if' part), and it is added to the storage part 101 as the new rule. Thus, the system can deal with the condition which is not previously assumed to a certain degree.

Description

[Detailed Description of the Invention] [Industrial Field of Application J] The present invention relates to a rule learning processing method, and is particularly applicable to controlling equipment using rules, so that it can be applied to individual situations that the programmer could not have imagined. The present invention also relates to a rule learning processing method suitable for speeding up execution when similar conditions apply to the first order.

[Prior Art] Conventionally, automatic generation of operating procedures has only been performed within the range of situations that humans have assumed in advance. A representative example of automatic procedure generation is Miyakotori et al.'s FA transport vehicle group operation control simulator (Information Processing Society of Japan 34th National Conference Proceedings pp. 2173-2174); even in that case, the control rules are created by humans. is set assuming all possible cases,
It was necessary to reduce leakage by conducting numerous trials.

In addition, there are some systems in operation that handle abnormality processing, but they can only handle abnormalities that have been taught in advance by humans, and what's more, 80 to 90% of controller software programs.

There was a problem that the function was dominated by the automatic recovery function from an error. For this reason, even devices with simple functions require huge programs to operate automatically (F
ieldlng, P.J., HDicesare, F.
, ;Goldbogen, G., Hand Desro
chers+^,”Intelligentautom
error recovery in ma
nufacturingworkstations,'
Proc IEEE Ink, symp,
Intel.

Control (1987): pp, 280-27
5. ).

All of the above problems are caused by the need to prepare rules for each situation by anticipating all kinds of situations in advance.6 [Problems to be solved by the invention] Purpose of the present invention The object of the present invention is to provide a rule learning processing method in which a controller generalizes and applies predetermined conditions even to situations that the programmer could not have imagined in advance.

[Means for Solving the Problems] Four stages shown in Table 1 can be considered as environmental states to which the rule learning processing method of the present invention is applied.

That is, stage (0)! ! For normal operating conditions where

Procedures are generated according to rules set in advance by the system designer (programmer).

Step (1) For situations that can be assumed to be abnormal, take action according to rules set in advance by the system designer (programmer) for handling abnormalities.

Step (2) When the controller detects a situation that is not anticipated in advance, the situation expression and steps (0), (1
) Replace the concrete name that appears in the condition part of the rule with an abstracted concept (-filmization).

Now, if the situation applies to the rule, the specific name in the situation expression is substituted into the rule and a conclusion is drawn. . If the condition part of the risk avoidance rule is not met, this is registered as a new rule.

If the four-stage step (3) situation of applying the procedure generation rule or the result of step (2) meets the risk avoidance rule, it is determined that there is a dangerous system abnormality and the system is shut down.

This control is performed.

FIG. 1 is a diagram expressing the concept of control in the case of the present invention.
Arrows indicate what kind of processing should be used to deal with each of the states: , Gray, and Red.

Even if it does not fit the rules prepared in advance,
The learning action in step (2) above returns the robot to the normal motion trajectory.

[Operation] When generalizing rules, a hierarchical representation of the environment (stored in memory) is used.

1) The name of the object is replaced with a concept one level higher in the representation that describes the environment hierarchically (for example, as shown in FIG. 4(A)). 2) If matching is still not possible, replace the concept with a higher level concept.

This is done in two stages, for unexpected situations.

If the condition part (if part) can be compared and matched with the situation using this generalization, the name of the thing in the conclusion part (then part) is replaced with the name matched by the condition part.

At this stage, we have only replaced the names in the rules, so there is no guarantee that they will actually be usable in the environment. Next, we check for consistency with the model that describes the physical properties of the environment, and if they do not match, we modify new rules based on this environmental model.

Before executing the conclusion based on this new rule as an action, a check is further performed based on the risk avoidance rule.

If this is not the case and no error occurs during execution, the above rule is also registered as a new rule.

As a result, the effect of the clothing can be expected.

(1) It is possible to adapt to some extent to situations that the programmer did not envisage in advance.

(2) Since methods for dealing with new situations are automatically registered as rules, the same situation can be dealt with in the future with fewer inferences.

[Example 1] One procedure generation (workpiece conveyance) - Next, an example of the present invention will be described with reference to the drawings.

2 and 3 are configuration examples of a work transfer device to which the present invention is applied, respectively, and FIGS. 4(A) and 4(B) are diagrams representing the hierarchical relationships between the constituent elements and concepts of the environment, respectively.

Let us take the case of a workpiece conveyance system as an example. By teaching one simple example as a rule, it will be possible to automatically generate 9 procedures (1) even when transporting to other equipment for which the transport method has not been taught, and (2) even if more equipment is added and the layout changes.

The procedure generation involves the following process, and FIG. 5 shows a flowchart representing the process.

Process■ Search process for given rules■ Noun generalization and fitting process■ Feasibility check/consistency with environmental description content (
- Re-check before film formation) Process ■ Correction process based on failure reason ■ New rule registration ■, ■...■ in Figure 5 correspond to the relevant process ■, ■...■ are doing.

How to deal with the situation where the A99 procedure is not taught (1) First, in the layout shown in Figure 2, the *transport system includes the transport device P+Q+r+3 between the loader/unloader in and out for only the transport route 1n-A-out. 1n-p-^ using and manufacturing equipment A
It is a teaching in the form of -qr-s-out.

That is, the rule is if 1n-A-out then 1n-p-^-q-r-s-out. But now, if we want to carry it as 1n-B-out,
Conventional knowledge bases (expert systems), which have no applicable rules, would stop working at this point, but a single system can deal with this by using the learning function shown below.

Countermeasure (2) Search to see if there is a rule that applies if you generalize the noun in the if part (process ■). The second generalized concept is a "manufacturing device" as shown in FIG. 4(A). Therefore, the concrete object belonging to the same manufacturing equipment is A.
, B, so if we apply A, we can use the rule of countermeasure (1), so L ++ e n
Applying B instead of A in part, the tentative result is 1n-pB-q
- Obtain rs-out (process ■).

Countermeasure (3) Check the feasibility of the tentative results based on the constraints of the environmental model.

The system has an environment model that describes the size, connection relationships, and physical properties of the individual objects that make up the environment. In the case of this tentative result, the environment model returns an answer that the tentative result cannot be realized because it indicates that the transport device P goes to the manufacturing device B. This is because there are rules for realization conditions in the environment model, and the connection relationships of each device are checked. And as a reason for denial, p-B
It also returns a negative reason that the sequence −q violates the connection relation condition (process ■).

Countermeasure (4) Modify the then part of the temporary rule based on the reason for denial. Denial of the reason for denial is the corrective method.

In this case, the reason is a mismatch in the arrangement of devices.

The position of B is changed to achieve matching (process ①), and as a result, 1n-p-q-r-B-s-out is obtained.

Countermeasure (5) Register the newly obtained if 1n-B-out then 1n-p-q-r-B-s-out as a new rule (iIA procedure ■).

Also, when the same situation occurs, this rule is used, reducing the number of inferences and shortening the calculation time.

B. What to do when the environment changes Generally, environmental knowledge and procedure generation knowledge (rules) are often described separately.Even if the environment changes, environmental knowledge is connected to intuition, so it is easy to change. On the other hand, although it is ideal that procedure-generated knowledge is originally independent of environmental knowledge and unaffected by its changes, in reality it has to be a rule that depends on the environment. This is for the same reason that expert systems cannot be constructed without depending on the target to some extent. As a result, it was difficult to know where and how to change the rules, and there was a problem of lack of flexibility to change the environment.

This system solves this problem by automatically modifying the procedure generation rules based on the environmental model.

Let us consider a case where the device layout in FIG. 2 is changed as shown in FIG. 3 by adding device C and also changing the transport devices t, u, etc. Previously, rules were changed manually when the layout was changed. The automatic rule correction procedure of this system is as follows.

Step (1) Work is loaded from rin and loaded into 5 units [via C.

Assume that 1n-C-out is input as a question meaning "What route should I take to get to out?" Since C is a newly added device, naturally there is no procedure generation rule, and the rule search will fail (Fig. 5 Illustrated process
).

Step (2) The most recent one by generalizing the noun.

That is, 1nB-out is first generated, and this matches the previously registered rule (step 2 in FIG. 5).

Step (3) Here, the route generated based on 1n-B-out, 1n-p-q-r-B-s-ou%, is checked, and even though it is an already registered rule, it is set WI! It is pointed out that the connection relationships between the models are not consistent.

As a result, it becomes clear that it is necessary to modify the existing rules by changing the model (Step 2 in Figure 5).

Step (4) The reason for the denial is inconsistency with the device connection relationship,
Based on this, the then part of the existing rule is 1n-p-q-
r-B-3-out to 1n-p-q-tur
-B-s-out (Step 2 shown in Fig. 5).

Step (5) The then part corresponding to 1n-C-out becomes 1n-p-q-t-ur-B-s- by noun-membrane.
Based on out.

It is made as in-ρ-q-t-ur-C-s-out. However, this is denied by a consistency check using the device model (Step 2 shown in Figure 5).

Move R (6) The rule is revised based on the negative reason.

if 1n-C-out then 1n-p-q-t-C-ur-s-ou
t is obtained (Step 2 shown in Figure 5).

Step (7) Register as a new rule (Step 2 shown in Figure 5).

If concepts could be learned, procedural generative learning similar to that of humans would be possible, but this is not easily achieved with current symbol processing technology. On the contrary, it becomes necessary to prepare a huge amount of common sense in advance. Considering that the program for controlling the Ja device will be incorporated into Ja devices that will be moved in the future, we do not think it is appropriate to process a large amount of common sense on a large-scale computer.

[Example 2] Abnormality processing (workpiece conveyance device) - (Sensor substitution method used in one location can also be applied to other locations) Third
It is assumed that the system is normal if there is a sensor report stating that the workpiece passed along the path rp-q-t in the figure (abbreviated as p-q-L). Even if it is reported that ρ-[, the target point [ has been reached, so this is simply an abnormality in sensor q and not in the entire transport system. This is registered as a rule in advance. Specifically, the first order is as follows.

if p-q-t then system is normal
f p-t then system is normal and
5ensor q is abnormal-u-
As for r, the above rule is not registered.

Here, if there is a sensor report of tr, the rule does not have an interpretation for this case, so conventionally it would be an abnormality. In this system, we generalize this and apply it to the case of pt, and use it as a risk avoidance rule (in this case, there are no other abnormality reports), and use the rule to say, ``Sensor U is malfunctioning, but the system is functioning normally.'' It is registered as a rule that says ``Yes. That is, a first-order rule is generated.

if L-r then system is normal and
5. Ensor u is abnormal [Example 3] One step generation (manipulator) - (Generation of motion rules for handling parts for which motion rules are not described) For block A as shown in FIG. 6, in order for the hand of the manipulator to grasp Assume that the following behavior rule is set.

``If the target part is block A, grab it from the X or X direction.'' Let us now assume that block I3 shown in FIG. 7 is newly added to the environment model. From the hierarchical relationship of the environment model, block B, like block A, is represented by the concept of a dock, so the operation rules for handling block B are as follows:
Similarly to the above, it is created as follows based on the operation rule of block A.

"If the target part is block B, grab it from the X or Y direction." Next, this is checked using block B and the physical model of the hand in the manipulator 10.

Then, since the fingers of the hand do not open as much as the length of block B in the y direction, it becomes clear that it is impossible to grasp it from the y direction. Therefore, the rule is revised to ``If the target part is block B, start from the X direction.''

This is put into practice, and if it does not fall under the risk avoidance rules, it is registered as a new rule.

[Example 4] - Abnormality processing (manipulator) (Applying the countermeasure when a bolt is not found to the situation where a block is not found) One of the reasons why the manipulator 10 fails to grasp a part is that a person does not know the position information of the part. You may have entered it incorrectly. FIG. 8 shows a situation where the bolt 30 is supposed to be inserted into the indicated hole 20, but the bolt 30 is actually inserted next to the indicated hole 20. In order to deal with such a situation, let us assume that a rule has been set in which when a component cannot be grabbed at a designated position, the operator searches for the adjacent hole 21 while taking into account the spacing between the parts tables 40.

As shown in FIG. 1, it is assumed that one of the blocks 50 cannot be grabbed at the taught location 60. Traditionally, work would stop here. In the present invention.

Tracing the hierarchy of the environment model, he learns that there is a ``search method for missing bolt parts''.

Therefore, this is also applied when finding the block 50 to find the part, and if successful, this is newly registered as a rule regarding the search method for the block.

FIG. 9 shows a block diagram of an embodiment of the present invention.

In the figure, reference numeral 101 is a rule storage section, 102 is a state storage section, 103 is a rule operation section, 104 is a comparison and verification section, 105
106 represents a new combination generation unit, 106 represents an avoidance rule storage unit, and 107 represents a new rule registration processing unit.

In normal inference, the rule operation unit 103.

The comparison and matching unit 104 is used to compare and match the condition part (if part) of each rule in the rule storage unit 101 and each state in the state storage unit 102, and the conclusion of the rule that satisfies the three conditions is stored in the state storage unit. 102.

However, the rule storage section 101 may not contain any rules to which the state in the state storage section 102 applies. In such a case, according to the present invention, the new combination generation unit 105 performs generalization. The new combination generation unit 105 converts the conclusion part (then part) of the rule to correspond to one condition part based on the method used when converting it into a film. Then, this is set as a new rule in the rule storage unit 101.
In addition, it is checked against the state storage unit 102 to confirm that even if a new result is obtained, the result does not match the condition part of the avoidance rule in the avoidance rule storage unit IO6 (it is not dangerous). If the new rule clears the check by the avoidance rule.

The Qishin rule registration processing unit 107 registers it in the rule storage Al01.

[Effects of the Invention 1] As explained above, the present invention provides a system that can trace the hierarchy of an environment model, generate new rules, and automatically generate operation procedures to perform processing. It can be applied to etc.

[Brief explanation of drawings]

Fig. 1 is a diagram expressing the concept of control in the case of the present invention, Figs. 2 and 3 are examples of the configuration of a workpiece conveyance device to which the present invention is applied, and Figs. 4 (A) and (B) are the respective environment FIG. 5 is a flowchart representing the process of the present invention. FIGS. 6 to 8 are explanatory diagrams for explaining the embodiments of the present invention.
FIG. 9 shows a block diagram of the invention. In the figure, 10 is a manipulator, 20 is a hole, 30 is a bolt, 40 is a parts stand, 50 is a block, 60 is a place, 101
102 is a rule storage unit, 102 is a state storage unit, 103 is a rule operation unit 3104 is a comparison matching unit, 105 is a new combination generation unit,
106 is an avoidance rule storage unit, 107 is a new rule registration processing unit, in and out are loader and unloader, respectively.
A, B, and C are manufacturing equipment, respectively, ρ+ Q+ r+ 31
! l and u are transport devices, respectively.

Claims (1)

[Scope of Claims] A rule storage unit that stores a rule including a condition part having a plurality of conditions and a conclusion part inferred when the various conditions are satisfied, and a state inferred by this rule and corresponding to the above-mentioned conditions. a state storage section that stores the state of the controlled object to be controlled, and a rule operation section that compares and matches the state of this state storage section with the conditions of the condition section and stores the conclusion of the rule that satisfies the conditions in the state storage section. In a rule-based processing system that has a rule-based control logic and determines the processing content for a controlled object, for a state input that does not satisfy any of the existing condition parts, this input and the existing condition part Generalize the concrete name, compare and match again, and if it fits, substitute the concrete name that was in the input into the conclusion section of the rule to obtain a new conclusion, and define the conclusion and the actions to be performed based on the conclusion. A rule learning processing method characterized in that if a result does not meet the condition part of an avoidance rule, learning is performed to register the result as a new rule.