JPH06161764A - Multi-input fuzzy inferring method - Google Patents

Abstract

PURPOSE:To compress an information throughput and obtain an inference result at a high speed by specifying the maximum number of ignition rules for many inputted observation data and calculating a determinate operation value, obtained from a composite membership function of fuzzy inference while made nonfuzzy, from the values of a specific number of fuzzy numbers in the carrier set of a conclusion part. CONSTITUTION:For the fuzzy inference, two restriction conditions are set. Namely, the maximum number of ignition rules can be specified in advance as a condition 1. As a condition 2, the determinate operation value which should be obtained from composite MF as the fuzzy output of the fuzzy inference while made nonfuzzy can be calculated from the values of the specific number of nonfuzzy numbers in the carrier set of the conclusion part. Then tables 11 and 12 necessary for previously generated processing algorithm are provided so as to concretely provide the processing algorithm, and a program advances processing by referring to the tables.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a multi-input fuzzy inference processing algorithm, and more particularly to a multi-input fuzzy inference method for performing fuzzy inference on a large number of input observation data by processing a table pre-expanded in a storage element. It is a thing.

[0002]

2. Description of the Related Art The source of the basic concept of the present invention is a patent application filed on July 14, 1988 and August 23, 1988 (the former application is referred to as the first prior art and the latter application is referred to as the second prior art). )
Filed on January 16, 1990 (Japanese Patent Laid-Open No. 3-210633)
Japanese Laid-Open Patent Publication No. 3-2 (1st prior art)), these configurations are for pre-calculating an operation required for fuzzy inference and storing the operation result in a storage element. Discretization (digitization) for fuzzy computation
If it is stored in the address corresponding to each observed value of the input A and B input, when applying to the actual control, etc., by just deriving the observed value corresponding to the bit to the address pin, immediately The reasoning result can be output.

[0003]

However, in the storage element used in any of the above-mentioned prior arts, the sum of the number of bits assigned to the output and 2 (the sum of the number of bits assigned to each of the inputs A and B). A storage capacity equivalent to the value multiplied by is required, which becomes an exponentially huge storage capacity for multi-input observation data, and the amount of calculation for pre-storing in a storage element also increases dramatically. In reality, there arises a problem that it becomes difficult to realize this basic concept. That is, for example, when performing fuzzy inference for 6 inputs and 1 output,
If 8 bits are assigned to each input / output, it is actually 25.
This is because a storage capacity of 6 × 10 ¹² (256 terabytes) bytes is required.

In view of the problems of the prior art, the present invention has an object to provide a processing method capable of performing a fuzzy inference at a high speed with a small-scale storage element for a large number of input observation data.

[0005]

In order to achieve the above object, the present invention provides a processing algorithm for a high-speed processing method of fuzzy inference that imposes the following two constraint conditions upon fuzzy inference in the background. The feature is that the amount of information processing is compressed by and the required storage capacity is reduced.

Prior to the explanation of the constraints, the terms are first defined as follows to assist in understanding the present invention: The conditional part firing label means that the matching degree between the input observation data and the fuzzy label in the conditional part of the fuzzy production rule is not zero. 2. The rule firing degree means the degree of conformity with the condition part passed to the rule conclusion part. 3. The firing rule means that the degree of firing is not zero. 4. The membership function is simply referred to as MF.

Therefore, the above-mentioned constraint conditions are as follows: 1) constraint condition 1: the maximum value of the number of firing rules can be specified in advance: For example, as shown in FIG. 3 (A), an MF that does not overlap two adjacent labels, If defined for each input, there are two or one conditional fire label for one input. In other words, in a 2-input 1-output fuzzy inference generally used in fuzzy control, it means that there is a maximum of 2 × 2 = 4 firing rules. 2) Constraint 2: The definite operation value that should be defuzzified from the composite MF as the fuzzy output of fuzzy inference can be calculated from the value of a specific number of fuzzy numbers on the set of conclusion parts. : Simplified inference method (also called "height method") that uses a singleton (real number) instead of a fuzzy set for rule conclusion part labels, as shown in Fig. 3 (B)
And if a fuzzy set is adopted for the rule conclusion part,
In the algebraic product-addition centroid method, even if the MF of the conclusion label is replaced with a singleton, the value obtained by defuzzification is the same, so that it can be handled in the same way as the simplified inference method. Further, even in a general fuzzy inference method, a fuzzy set of conclusion part labels can be expressed as a set of a plurality of real numbers discretized on a base set.

Next, in order to solve the above problems, the configuration of an apparatus for implementing a fuzzy inference method for specifically realizing the processing algorithm will be described below. Device configuration 1. A software program used for an ordinary electronic computer is provided with a table necessary for a processing algorithm created in advance, and the program is configured to proceed with the processing by referring to the table.

Device configuration 2. A block in which observation data is input and processed is divided into A and B parallel blocks, and the processing is performed by a multi-CPU.

Device configuration 3. The storage elements are arranged in parallel and in multiple stages so that the observation data or the processing data is directly led to the address pin of the storage element in which the necessary result is stored in advance and the calculation result is obtained from the output section correspondingly.

Device configuration 4. The common circuit of the parallel portion divided into blocks is configured to operate in a time-division manner according to the count value of the counter.

[0012]

According to the present invention, however, by giving a constraint condition, the storage capacity of the storage element is compressed to a small scale for multi-input observation data, and the program maximizes the tables necessary for the processing algorithm. By referencing, it is possible to execute a software method that can obtain inference results at high speed. In addition, by dividing the memory element to which the observation data is input and processed and dividing it into blocks, and making the CPU perform parallel processing, the inference result can be obtained at a dramatically high speed, and the clock for the counter is eliminated to ensure high reliability. Can lead to inference results. In addition, the circuit scale can be reduced by sharing the respective parallel parts in which the memory elements are divided into blocks, or even if a malfunction of the clock occurs due to the use of the counter clock, a runaway does not occur. It works.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT An embodiment of the present invention will be exemplarily described in detail below with reference to the drawings. However, the dimensions, materials, shapes, and relative arrangements of the components described in this embodiment are not intended to limit the scope of the present invention thereto, unless otherwise specified, and are merely illustrative examples. It is nothing more than a thing.

First, the fuzzy inference processing algorithm of the present invention with the above constraint conditions will be described in detail below.
1 and 2 show two types of outputs, a block A that outputs the coordinates of the conclusion part Sington of the firing rule from the input observation data, and a block B that outputs the firing degree of each firing rule from the observation data of each input. , A block configuration diagram for implementing a processing method by multi-input fuzzy inference to lead to the defuzzification section 3 to obtain a definite output. For example, a membership function that does not overlap with seven labels 0 to 6 as shown in FIG. 3 (A) in each input is defined so as not to overlap every other adjacent label, and there are seven conclusion part labels. 3 inputs 1 using the simplified inference method represented by the singleton of
This is an example of performing fuzzy inference of output, which corresponds to the device configuration 3 and is achieved by FIG. 1 and FIG. 2 or FIG. 6 (A).

Next, the inside of each block of FIGS. 1 and 6B will be described below. Block A: The table 1 is used from the observation data of each input to obtain the number of the conditional firing label, and the table 2 is used from the combination of the numbers, so that the same number as the firing rule conclusion portion real value The coordinate value of is obtained and output. Also, after converting from the conditional part firing label number to the conclusion part label number using a specific table in the 3-input 1-output fuzzy inference by the address lookup method described in C language, the address lookup is also performed from the number. Another specific table of can be used to obtain the conclusion real value. In FIG. 1 and FIG. 2, since 3 bits are sufficient to express the number of condition part labels, 6 bits out of the 8 bits output of the storage element are used, so that 2 bits for one input can be expressed. The conditional part firing label is expressed.

Block B: The fuzzy logical product is extracted from the input observation data to obtain the degree of conformity of the condition part label using Table 3 and the degree of firing of each firing rule. As the method of the fuzzy logical product, there are a min operation, an algebraic product, etc., but in FIGS. 1 and 2, the table of the storage element in which the result is stored in advance is referred to or a normal calculation or operation circuit is used. It can be calculated by the min operation. In addition, 8 bits are used to express the conformity of each condition part label and the firing degree of each firing rule.
By using a word (2 bytes) storage element, it is possible to express the suitability of two conditional part firing labels for one input.

Defuzzification unit: The block A and the block B are integrated to output a defuzzified definite value. As a method of defuzzification, there are a barycentric method, a maximum height method, etc., but in FIG. 1 and FIG. 2, by referring to a table of storage elements in which results are stored in advance, or by a normal calculation or arithmetic circuit, The center of gravity method for calculating the moment (center of gravity) can be adopted, and a fixed output as shown in FIG. 5 (B) is obtained.

As described above, if two fuzzy labels do not overlap each other as shown in FIG. 3A with respect to the input of the observation data 1 to 3, the inside of the table 3 of FIG. The grades indicated by solid lines in (A) and (B), that is, the label suitability are stored for S1 in FIG. 1A and for output S2 in FIG. 1B in correspondence with the observation data 1 in FIG. ing. In FIG. 2, outputs T1 and T2 for observation data 2 and outputs U1 and U2 for observation data 3
Is the same as above. That is, assuming that a specific numerical value a is input to the observation data 1, the output A of the table 1 and the output S of the table 3 are as shown in FIG. 4 (C). Where A1 and A2 are membership labels NB and N
M ,. ． ． 0, 1, ... Corresponding to PB. ． ． 6 and S1 and S2 are "0," which is the grade of label conformity.
8-bit discretized value corresponding to 1 ", 0, 4,
4 ,. ． ． 255.

Regarding the inside of D in Table 2 in FIG. 1, for example, the conditional proposition If (observation data 1 = NM) & (observation data 2 = ZR) & (observation data 3) of fuzzy inference.
= PB) and the consequent part of then (output = PS), when rule j is fired, A1 = NM
(= 1), B1 = ZR (= 3), C1 = PB (= 7) are input to the table 2, and 170, which is the real value coordinate of the label PS, is stored in the address corresponding to the address bit. Is output from D1. On the other hand, depending on the min of the comparator 22 in FIG. 2, among S1, T1, and U1,
The smallest value is selected and output from V1 as the rule firing degree.

The defuzzification section 3 in FIG.
Eight real-valued moments (centers of gravity) indicated by D8 and V1 to V8 are taken and output as shown in FIG. 5 (C). or,
Among the eight real numbers, the ones having the same coordinates on the base set, that is, the eight firing rules having the same conclusion part label of the output are the sums as shown in FIG. 5 (A). It will be

The above fuzzy inference is based on the simplified method, but the fuzzy set of the conclusion part label is replaced with the min-algebraic product-addition-centroid method. It becomes the same as in FIG. 5 (C).
If the MAX type synthesis is performed as shown in FIG. 5A, the same conclusion label is collected and the center of gravity is calculated after the MAX operation, but there is a drawback in that the circuit actually increases.
From the control result, it is known that the addition type is one step better.

[0022]

As described above, according to the present invention, the maximum number of firing rules is specified with respect to a large number of input observation data, and the decision is made by defuzzifying from the synthetic membership function of fuzzy inference. The amount of information processing is compressed by imposing a precondition that the operation value can be calculated from the value of a certain number of fuzzy numbers on the set of conclusion parts, and is calculated and stored in advance in a storage element with a small storage capacity. As a result, the address lookup method that directly derives the inference result can be adopted. As a result, the inference can be executed at high speed without depending on the number of rules, and the inference method can be opened by adopting the simplification method. It has remarkable effects such as increasing free selectivity.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a block (A) and a defuzzification unit that implements a multi-input fuzzy inference method of the present invention.

FIG. 2 is a configuration diagram of a block (B).

FIG. 3 represents a membership function, and (A) shows a case where two adjacent labels do not overlap each other,
(B) shows cases where singletons (real numbers) that are not fuzzy sets are adopted in the rule conclusion part.

4A and 4B show a relationship between observation data and a membership value. FIG. 4A shows a membership grade (label compatibility) for the output S1, and FIG. 4B shows a membership grade for the output S2. , (C) show A output of Table 1 and S output of Table 3.

5A and 5B are diagrams for explaining a reasoning method of a conclusion part, FIG. 5A is a simplified reasoning method, FIG. 5B is a center of gravity method, and FIG. 5C is a fuzzy set of conclusion part labels.

FIG. 6 shows a configuration example of an apparatus for carrying out the present invention,
(A) shows a case of using a storage element in which necessary results are stored in advance, and (B) shows a block diagram in which FIGS. 1 and 2 are integrated.

[Explanation of symbols]

 1 block A 11 table 1 12 table 2 2 block B 21 table 3 22 comparator 3 non-fuzzy part 31 product-sum adder 32 adder 33 divider

Claims (1)

[Claims] 1. A fuzzy inference method in which a required operation is performed in advance for fuzzy inference, the operation result is stored in a storage element as a table, and a large amount of observation data is input from a sensor in the fuzzy inference method. , The maximum number of firing rules is specified in advance, and the defuzzified definite value from the composite membership function is calculated from the value of the defuzziness number in the conclusion part, which can be inferred by the address lookup method of the processing algorithm. A multi-input fuzzy inference method characterized in that.