JPH05257695A - Automatic rule generating system - Google Patents

Abstract

PURPOSE:To shorten the verification time for validity of a rule, and to improve reliability by generating efficiently and automatically an optimal rule by learning. CONSTITUTION:Execution of a fuzzy inference is executed by fetching a rule 3 into a knowledge base 2, and executing it by an inference mechanism 1. In this case, by using a set of input/output data 7 consisting of optimal data in the past, a rule generating mechanism 6 generates automatically the rule. In this case, the rule generating mechanism 6 consists of a network weighting determining part 5 for detecting a relation between input and output variables in a network by learning using the network for showing a relation between the input and the output variables, and a rule extracting part 4 for extracting the rule from this network. That is, an arc having weight of a certain prescribed value or above is all selected, and the rule is generated from a set of the variable shown by the node being in both ends of those arcs and a label.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control device having a fuzzy inference mechanism in the fields of process control, equipment control and the like.

[0002]

2. Description of the Related Art A general knowledge acquisition method in a fuzzy inference mechanism is to draw out know-how consisting of "fuzzy" information from a skilled operator and the like, and based on this, inference rules and membership functions of input / output variables. It is a method of determining the shape of. However, it becomes difficult to express the rules without excess or deficiency as the number of rules increases. As described in Japanese Patent Application Laid-Open No. 2-89103, the conventional device is designed to facilitate adjustment of the membership function, but it is described how to efficiently describe or generate rules. Absent.

[0003]

The above-mentioned prior art does not consider the generation of rules, and it is necessary to perform a sufficient offline simulation while adjusting the rules and membership functions in the target system.

The present invention is to efficiently and automatically generate an optimum rule by learning.

[0005]

In order to achieve the above object, for each input variable and output variable, each label (the range of values that the variable can take is "large", "medium", "small"). A network consisting of nodes (contact points) corresponding to (classified using expressions such as) and weighted arcs (line segments) connecting them is provided. For this network, we dynamically change the weight of each arc in the network that was trained using the optimal input / output data set in the past, and find the relation between each variable by the magnitude of the weight.
It is designed to generate rules.

[0006]

The above-described network operates as follows when learning rules. That is, the input value to each node on the input side is determined according to the size of the input data of each input variable. Next, the input value is distributed to each arc connected from those nodes to the output side according to the size of the data of the output variable. By doing this, each arc is weighted and the same weight update (addition) is applied to all input / output data sets.
By performing, the weighting is performed in the network to represent the relationship between the input and output variables. Next, using this network, the rule extraction unit extracts rules as follows. That is, all arcs having weights equal to or greater than a certain fixed value are selected, and a rule is generated from a set of variables and labels represented by nodes at both ends of those arcs. For example, if the input side node has the label “medium” of the variable IV1
Then, if it on the output side is the label "small" of the variable OV2, the rule generated by this arc is "if I
If V1 = medium, then OV2 = small.

[0007]

EXAMPLES Examples of the present invention will be shown below.

In FIG. 1, fuzzy inference is performed by incorporating rule 3 into knowledge base 2 and executing this inference mechanism 1.
Is executed by executing. In the present invention,
The rule generation mechanism 6 automatically generates a rule using a set of input / output data 7 consisting of past optimum data. Here, the rule generation mechanism 6 uses a network representing the relationship between the input and output variables to find a relationship between the input and output variables in the network by learning, and a rule for extracting rules from this network. The extraction unit 4 is included.

Next, FIG. 2 shows input / output variables and networks. First, consider a system having n input variables and m output variables as shown in FIG.
Here, for the sake of simplicity, assuming that the range of possible values of individual variables is 0 to 100, as shown in FIG. 2- (2), "large", "medium", "small"
Label and the range of the label can be defined (membership function). By doing so, the goodness of fit for the input value of each label can be obtained as follows. That is, the input variable IV1 is shown in FIG.
If there is a membership function like (2), then IV1
= 10, the degree of "small" IV1 is 100, and when IV1 = 35, the degree of "small" IV1 is 2.
The degree of being 5 "medium" is 75.

In general, the fuzzy rule is OV1 = large when IV1 = medium.

Since the above expression is used, the number of rules that can be expressed is:

[0012]

[Equation 1]

(Lυk: number of labels in variable k) In the above-mentioned example, since lυk = 3, the number of expressible rules = 3n × 3m. As a result, FIG.
As shown in (3), the network corresponding to all the rules can be obtained by the nodes corresponding to the labels of the variables and the arcs connecting them (arcs from the input side to the output side).

Next, FIG. 3 shows a weight determining method in the network weight determining section. First, FIG.
-As in (1), it is assumed that there is a data file composed of a set of optimum input / output values accumulated in the past. here,
Input variable IVi (1 ≦ i ≦ n) and output variable OVj (1 ≦ j
Paying attention to ≦ m), we show how the weighting in the network changes by the first and second learning.
First, the first input data is IVi = 35, and the matching degree of each label with respect to this input value is “small” = 7.
5, "medium" = 25, "large" = 0 (Fig. 3-
(2)). Therefore, as shown in FIG. 3- (3), 75, 25, 0 are input to each node on the input side. Next, since the output value is OVj = 80, the conformity of each label to this output value is “small” = “medium” = 0 and “large” = 100. Therefore, 100% of the value from the input side is distributed to the node corresponding to "large" on the output side (Fig. 3- (3)). Here, the numbers in parentheses represent the weights up to the previous arc. The same is true in the second learning, and “small” = IVi = 20.
100, OVj = 60, “medium” = “large” =
Therefore, 100 is input from the node corresponding to “small” on the input side, and 50 is distributed to the nodes corresponding to “medium” and “large” on the output side. here,
The weights given by the first weighting are added, and as a result, after the second learning, the weight of each arc is determined as shown in FIG. 3- (5).

By repeating the above learning, the weight of each arc on the network is determined.

Next, FIG. 4 shows a rule extracting method in the rule extracting section. Regarding each arc in the above-mentioned weighted network, when attention is paid only to a certain value or more, it is assumed that the arc as shown in FIG. 4- (1) remains. Here, one arc corresponds to one rule, but the conditional part of two rules (if ...
It is possible to make one rule by combining (...) parts) or combining conclusion parts ("if ..." parts) like. Rules can be extracted from a certain arc (Fig. 4- (2)).

[0017]

According to the present invention, since rules are automatically generated, there is no need to create rules. Also,
Since the optimum data in the past and the data reflecting the know-how of the skilled operator are learned, it is possible to significantly shorten the rule validity verification time by the offline simulation. Furthermore, since the relationships between the input and output variables are extracted without contradiction, logical duplication and omission are eliminated, and the rule expression is simplified, so that speedup and reliability at the time of inference execution can be improved.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram.

FIG. 2 is a diagram showing a membership function and a network.

FIG. 3 is a diagram showing learning by a network weighting determination unit.

FIG. 4 is a diagram showing rule extraction by a rule extraction unit.

[Explanation of symbols]

1 ... Inference mechanism, 2 ... Knowledge base, 3 ... Rule, 4 ... Rule extraction unit, 5 ... Network weighting determination unit, 6 ... Rule generation mechanism, 7 ... Input / output data.

Claims (1)

[Claims] 1. A control device having a fuzzy inference mechanism including a rule generation mechanism for generating rules, a knowledge base for managing the rules, and an inference mechanism for fetching the rules from the knowledge base and executing inference. A rule automatic feature characterized by the fact that the rule generation mechanism is provided with a rule extraction unit that automatically extracts inference rules that allow the target system to perform the desired control by using the optimum input / output data set obtained by Generation method.