KR101242509B1 - Design method of activation function for inference of fuzzy cognitive maps in mobile program and thereof system - Google Patents

Abstract

A method of designing an activation function for inferring fuzzy recognition degree suitable for use in a mobile environment, and a fuzzy recognition inference apparatus using the method are disclosed. The method of designing such an activation function reads the formula of the given activation function for inference of the variable fuzzy recognition degree in the mobile program, and if there is a matrix norm in calculating the lambda (λ) value contained in the activation function formula, After the operation is performed using a matrix norm with a small amount of computational processing to obtain the lambda value, the lambda value is substituted into the formula of the activation function to obtain the activation function. According to an embodiment of the present invention, since the computation time for obtaining the activation function is minimized or reduced while ensuring the inference stability of the fuzzy recognition degree, the inference using the fuzzy recognition degree is easily implemented in a mobile environment that is relatively inferior to a general computer use environment. .

Description

Design method of activation function for inferring fuzzy cognition suitable for use in mobile environment and apparatus for fuzzy cognitive maps using the method

The present invention relates to the field of fuzzy recognition inference, and more particularly, to a method of designing an activation function for inferring fuzzy recognition degree suitable for use in an environment of a mobile device such as a personal portable terminal, and a fuzzy recognition inference apparatus using the method.

Typically, inference of fuzzy cognitive maps is widely used for decision support.

Fuzzy Recognition is a knowledge expression model used in the fields of artificial intelligence and intelligent systems, and it graphically represents the causal knowledge of events. The causal relationship of a particular event to other events is expressed in a matrix using weights between [-1, 1], and the state of the event under consideration is expressed in a vector using state values between [0, 1]. . Therefore, it is used as a tool for decision making by predicting the result of the current situation in advance through inference that performs simple numerical operation between these matrices and vectors.

Activation function to limit the values of [0, 1] to values between [0, 1] when the matrix and vector operations used in inference occur when the state values of vector elements deviate from values between [0, 1]. Is used. Since the value of the state vector obtained through inference is determined by an activation function such as a sigmoid function, the determination of the activation function becomes very important. In the case of the sigmoid function, the value of λ, which is represented as a real value, is determined to correct the stability of inference.

Computation process to find the real value such as lambda value when determining the activation function is complicated and takes a relatively long time. In particular, when the activation function is determined through a relatively simple mobile program of a personal mobile terminal such as a mobile phone, it is very difficult to obtain an inference result of fuzzy recognition due to a long time operation.

An object of the present invention is to provide a method of designing an activation function for inferring fuzzy recognition degree suitable for use in a mobile environment, and a fuzzy recognition inference apparatus using the method.

Another technical problem to be solved by the present invention is to provide a method of designing an activation function that can minimize or reduce the computation time for obtaining an activation function while ensuring the inference stability of the fuzzy recognition degree, and a fuzzy recognition inference device accordingly.

Another technical problem to be solved by the present invention is to improve the reasoning stability of fuzzy recognition in a mobile environment, such as a smart phone that is relatively inferior to a general computer environment, while improving the ability to simply obtain the activation function in a shorter time A method and a fuzzy recognition accordingly provide an inference apparatus.

In order to achieve the above technical problem, a method of designing an activation function for inferring fuzzy recognition degree according to an aspect of an embodiment of the present invention,

Deriving a formula of a given activation function for inferring a variable fuzzy recognition degree on a mobile program;

Checking whether there is a matrix norm when obtaining a lambda (λ) value in the formula of the activation function;

Performing the operation using the matrix norm with less computational throughput than the Euclidean norm when the matrix norm is present to obtain the lambda (λ) value;

Substituting the lambda value into the formula of the activation function to obtain the activation function.

로서 나타날 수 있다. 여기서 λ는 0보다 큰 실수이다.In an embodiment of the present invention, the activation function may be a sigmoid function, in which case the formula of the sigmoid function is as follows.

It can appear as Where λ is a real number greater than zero.

의 모델로 나타날 수 있다. 여기서

는 퍼지 인식도의 가중치 행렬이다.In an embodiment of the present invention, the lambda (λ) value is expressed by the following equation to ensure global index stability.

Can be modeled on here

Is a weight matrix of fuzzy recognition degrees.

에서

로서 계산되는 프로비니우스 놈일 수 있다. 여기서

은 퍼지 인식도에서의 개념의 개수이고,

는 퍼지 인식도에서

번째 개념과

번째 개념 사이의 가중치이다.In an embodiment of the invention, the

in

Provinius norm calculated as here

Is the number of concepts in fuzzy perception,

In fuzzy awareness

The first concept

The weight between the first concept.

In an embodiment of the present invention, the mobile program may be mounted on a personal mobile terminal or a smart phone.

In an embodiment of the present disclosure, the activation function may be updated whenever the fuzzy recognition degree is changed.

In an embodiment of the present invention, the mobile program may be mounted in a portable multimedia player.

In an embodiment of the present invention, the inference of the fuzzy recognition degree may be used for decision support for the implementation of artificial intelligence.

In an embodiment of the present invention, when the lambda value is not obtained within a preset time when the calculation is performed using the matrix norm with less throughput than the Euclidean norm, the lambda (λ) is determined from the next calculation. The step of setting the value as 1 may be further provided.

In an embodiment of the present invention, when the lambda value is not obtained within a preset time when the calculation is performed using the matrix norm with less throughput than the Euclidean norm, the lambda (λ) is determined from the next calculation. The step of setting the value as 5 may be further provided.

와 같은 사인 형태의 함수가 될 수 있다. In an embodiment of the invention, the activation function is

Can be a sinusoidal function such as

According to another aspect of an embodiment of the present invention, the fuzzy recognition inference apparatus,

A memory for storing a formula of a given activation function for inferring varying fuzzy recognition degrees on a mobile program;

An input unit for receiving a concept value for inferring the fuzzy recognition degree;

When there is a matrix norm in obtaining the lambda value included in the formula of the activation function, the lambda value is obtained by performing a calculation using a matrix norm with less computational throughput than the Euclidean norm. and an inference calculating unit for substituting the value of lambda) into the formula of the activation function to obtain the activation function.

로서 나타날 수 있다. In an embodiment of the present invention, the fuzzy recognition inference device may be mounted on a smart phone. In an embodiment of the invention, the given activation function may be a sigmoid function, in which case the formula of the sigmoid function is

It can appear as

의 모델을 가질 수 있다. In an embodiment of the present invention, the lambda (λ) value is expressed by the following equation to ensure global index stability.

May have a model of

에서 상기

로서 계산되는 프로비니우스 놈일 수 있다. In an embodiment of the invention,

From above

Provinius norm calculated as

In an embodiment of the present invention, the activation function may be updated whenever the fuzzy recognition degree is changed, and the inference of the fuzzy recognition degree may be used for decision support for the implementation of artificial intelligence.

delete

delete

In an embodiment of the present invention, if the inference operation unit does not obtain the lambda value within a predetermined time when performing the operation using a matrix norm with less computational throughput than the Euclidean norm, the inference operation unit starts from the next calculation. You can set the lambda value as 1.

In an embodiment of the present invention, if the inference operation unit does not obtain the lambda value within a predetermined time when performing the operation using a matrix norm with less computational throughput than the Euclidean norm, the inference operation unit starts from the next calculation. You can set the lambda value as 5.

In an embodiment of the present disclosure, the apparatus may further include a communication unit connected to the inference operation unit to perform wireless communication with the outside, and a display unit displaying a result of the inference operation of the inference operation unit on a screen.

According to the exemplary embodiment of the present invention, since the computation time for obtaining the activation function is minimized or reduced while ensuring the inference stability of the fuzzy recognition degree, the inference using the fuzzy recognition degree can be easily performed in a mobile environment that is relatively inferior to a general computer use environment. Is implemented.

Thus, the user can easily and quickly obtain the result of inference while on the go while being portable, and software development in the mobile program is flexible and easier.

In addition, product competitiveness may be enhanced when a personal portable terminal capable of obtaining such inference results in a mobile environment appears.

1 is a view provided to explain the causal weight between concepts in the inference of fuzzy recognition degree to which the present invention is applied;
2 is an operational control flowchart of an activation function according to an embodiment of the present invention;
3 is a graph of an activation function applied to the present invention, and
4 is an exemplary block diagram of a fuzzy recognition inference apparatus for performing the operation control flow of FIG.

BRIEF DESCRIPTION OF THE DRAWINGS The above and other objects, features, and advantages of the present invention will become more apparent from the following description of preferred embodiments with reference to the attached drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments disclosed herein are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the invention to those skilled in the art, without intention other than to provide an understanding of the present invention.

In some drawings, the connection relationship of elements and lines is shown for an effective explanation of the technical contents, and other elements or circuit blocks may be further provided.

Each embodiment described and illustrated herein may also include complementary embodiments thereof, and the description of the derivation process of the activation function and the matrix and vector norms for inferring fuzzy recognition degree obscure the subject matter of the present invention. Note that it is omitted to avoid doing so.

1 is a view provided to explain the causal weight between concepts in the inference of the fuzzy recognition degree to which the present invention is applied.

(2)와

(4)사이의 인과 가중치

(3)는 구간 [-1, 1]의 값을 가지며, '

'은 개념

와

사이에 양의 인과 관계, '

'은 개념

(2)와

(4)사이에 음의 인과 관계, 그리고 '

'은 개념

와

사이사이에는 인과 관계가 없는 것을 의미한다. Referring to the drawings, concepts (or nodes)

(2) and

Causal weights between (4)

(3) has a value in the interval [-1, 1],

Silver concept

Wow

Positive causality between, '

Silver concept

(2) and

(4) a negative causal relationship between, and

Silver concept

Wow

There is no causal relationship between.

In other words, a row in an association matrix represents a causal effect of a causal node on another node, and a column represents a causal effect of a particular node from causative nodes.

In the matrix and vector calculation process used in the inference of the fuzzy recognition degree, a situation occurs that the value of the vector deviates between the values between [0, 1]. Therefore, the activation function is used to limit their values to values between [0, 1]. In general, the activation function has a lot of sigmoid function (sigmoid function).

The sigmoid function is expressed as the following equation (1).

(1)

(One)

Where λ is a variable that appears as a real value.

Since the value of the state vector obtained through inference at the fuzzy recognition degree according to Equation (2) is determined according to the activation function, the activation function is determined by the expert through experience and trial and error. In the case of the sigmoid function, the value of λ, which is represented as a real value, is determined. Often, the application of fuzzy perception uses λ = 1 or λ = 5.

는

번째 추론을 통해 얻은

번째 개념의 상태 값을 나타내고, 그 범위는

이다.

이며,

는

번째 개념과

번째 개개념 사이의 인과 가중치를 나타낸다. 그리고

은 퍼지 인식도의 개념의 개수이다. 또한, 식 (2)를 벡터 연산 형태로 표현하면 식 (3)과 같다. In equation (2)

The

From the second reasoning

State value of the first concept, and the range is

to be.

Is,

The

The first concept

Causal weights between the first concept. And

Is the number of concepts of fuzzy perception. In addition, when the expression (2) is expressed in the form of a vector operation, it is the same as the expression (3).

이고

이다.In other words,

ego

to be.

(2)

(2)

(3)

(3)

로 둘 때, 위의 식 (3)은 아래의 식 (4)와 같이 변형하여 표현할 수 있다.here

When it is set to, the above equation (3) can be expressed by modifying the following equation (4).

(4)

(4)

Some studies have suggested a method for automatically calculating the value of λ. Among them, a method for automatically determining the value of λ based on the characteristics of a matrix representing fuzzy recognition was proposed in 2010. This method, based on Lyapunov's theory of stability, ensures global exponential stability in the inference process. In other words,

(5)

(5)

to be.

In addition, although the design of the activation function is performed once when the initial fuzzy recognition degree is designed, it is necessary to automatically design the activation function accordingly when the fuzzy recognition degree is continuously updated.

Recently, as the demand for development of smartphone applications increases, a complex reasoning program for decision making that can be implemented in a mobile environment is required.

However, unlike the PC environment, the mobile environment in which various personal portable terminals such as smart phones and portable multimedia players are used has relatively low hardware performance and memory. Therefore, computational complexity should be as low as possible to provide effective results to users. In other words, it may be more effective to provide faster computation speeds, even if the results are rather poor. Therefore, it is necessary to facilitate the development of decision-related software using fuzzy recognition in a mobile environment by designing an activation function through fast computation.

의 모델을 사용하면,

의 연산 과정이 복잡하여 모바일 프로그램으로 구현하기에는 무리가 있다. 따라서 본 발명의 실시 예에서는 모바일 프로그램에서 단순 연산을 통해 활성화 함수를 설계할 수 있도록 한다. As a result, when implementing a decision support program using fuzzy recognition in a mobile environment, when the activation function of the sigmoid form is automatically calculated according to the constantly changing fuzzy recognition, as shown in Equation (5)

Using the model of,

The operation process of is complicated and it is impossible to implement it as a mobile program. Therefore, in the embodiment of the present invention, it is possible to design an activation function through a simple operation in a mobile program.

It is easy to implement into a mobile program when a simple operation is performed by reducing the computational complexity while obtaining an activation function that guarantees inference stability. Accordingly, in the embodiment of the present invention, as shown in FIG. 2, an activation function can be easily designed even in a mobile program by using a matrix norm having a simple operation among various matrix norms.

2 is an operational control flowchart of an activation function according to an exemplary embodiment of the present invention.

의 모델에서

의 유클리드 놈 계산이 요구되는 경우라면 단계 S23을 수행한다. 단계 S23은 행렬의 유클리드 놈 을 구하는 연산이 복잡하므로 이를 배제하고 대신에 연산이 간단한 프로비니우스 놈(Frobenius norm)을 사용하여 연산을 수행하는 단계이다. 즉, 하기의 케이스 1의 경우에 대신에 케이스 2의 경우로 연산을 수행하는 것이다. 즉, Referring to the drawing, when the function design step is entered in step S20, calculation of the activation function is started in step S21. In this case, as in the above formula (2) in step S22

In the model of

If Euclidean norm calculation is required, step S23 is performed. In step S23, since the operation of obtaining the Euclidean norm of the matrix is complicated, the operation is performed using the Frobenius norm, which is simple. That is, in case 1 below, calculation is performed in case 2 instead. In other words,

to be.

대신

로 대체하여 식을 전개한 후, 시그모이드 함수의 λ값의 조건을 계산하는 것이다. 이에 따르면, 추론의 안정성을 보장하면서 연산 속도가 더 빠른 결과를 얻는다.　　 As a result, the embodiment of the present invention proposes a new activation function design model that can be easily implemented in a mobile program because the operation is simple while ensuring the stability of inference. In other words, the process of deriving a condition that guarantees the stability of the already developed inference

instead

After substituting for and expanding the equation, the condition of the lambda value of the sigmoid function is calculated. This results in faster computation speed while ensuring the stability of inference.

대신

(Frobenius norm)으로 대체할 경우에 연산이 단순해지는 이유는 다음과 같다. As described above, in the sigmoid function used in the activation function for inferring fuzzy recognition, the weight matrix required for calculating the range of lambda values that guarantees the stability of inference is calculated.

instead

The reason why the operation is simplified when replacing with (Frobenius norm) is as follows.

항목은

의 연산 복잡도와 동일하다. 그러나 위 케이스 1의 연산에서는 거기에다 eigenvalue를 구하는 연산이 더 추가되는데, 이 과정의 연산이 매우 복잡하다. In the calculation in case 1 above

Item is

Is equal to the computational complexity of. However, in the case 1 operation above, there is an additional operation for obtaining an eigenvalue, which is very complicated.

의 모델이 연산되어지게 된다. 단계 S24는 시그모이드 함수의 연산이 완료되었는 지를 체크하는 단계로서, 상기 실수 값인 λ를 상기 활성화 함수의 공식에 대입하여 상기 활성화 함수를 얻으면 연산이 완료된다. As a result, by performing step S23,

The model of is computed. Step S24 is a step of checking whether the operation of the sigmoid function is completed. The operation is completed by substituting the real value lambda into the formula of the activation function to obtain the activation function.

The mobile program including the control flow as shown in FIG. 2 may be mounted on a personal portable terminal such as a portable multimedia player or a smart phone. In addition, the activation function may be updated whenever the fuzzy recognition degree is changed.

Inference of fuzzy recognition using FIG. 2 may be used for decision support for implementing AI. In FIG. 2, if the real value is not obtained within a preset time when performing the operation using the matrix norm which is simpler than the Euclidean norm, setting the lambda value as 1 or 5 from the next calculation time. May be further provided.

3 is a graph of an activation function applied to the present invention. Referring to the drawings, the graph SF of the sigmoid function is shown in the S form. It can be seen that the graph SF exists between the upper limit M2 given as 1 and the lower limit M1 given as 0. In the case of the embodiment of the present invention, the sigmoid function as shown in FIG. 3 is used, but without being limited thereto, the activation function may be a sinusoidal function.

4 is an exemplary block diagram of a fuzzy recognition inference apparatus for performing the operation control flow of FIG. 2.

Referring to the drawings, the fuzzy recognition inference apparatus includes a communication unit 10, a memory 20, an inference operation unit 30, a display unit 40, an input unit 50, and an output unit 60.

The memory 20 is connected to the inference calculator 30 through a signal line L2. The memory 20 may store a formula of a given activation function for inferring a fuzzy recognition degree that is variable on a mobile program. In addition, the memory 20 may store a program related to the operation of the inference calculator and a mobile program of the mobile device.

The input unit 50 is connected to the inference calculating unit 30 through the signal line L5 and receives a concept value for inferring the fuzzy recognition degree.

The inference calculator 30 is connected to the communication unit 10 through a signal line L1, is connected to the display unit 40 through a signal line L3, and the output unit 60 through a signal line L4. ).

The inference operator 30, which may be implemented as a microprocessor or a CPU, performs a calculation using a matrix norm which is simpler than an Euclidean norm when a matrix norm exists when obtaining a real value included in the formula of the activation function. The real value is obtained and the activation function is obtained by substituting the real value into the formula of the activation function.

In addition, the inference calculator 30 controls all operations for communication and multimedia operation of a portable communication terminal such as a smart phone according to a preset program.

의 모델로 주어지며, 상기 추론 연산부(30)는 연산동작을 수행 시, Similarly, the real value λ of the activation function is expressed by the following equation to ensure global index stability.

Given by the model, the inference calculation unit 30 when performing the operation,

에서 상기

를

로서 계산한다. remind

From above

To

.

The activation function may be updated whenever the fuzzy recognition degree is changed, and the inference of the fuzzy recognition degree may be used for decision support for the implementation of artificial intelligence.

On the other hand, if the inference operation unit 30 does not obtain the real value within a preset time when performing the operation using the matrix norm, which is simpler than the Euclidean norm, the inference operator 30 sets the lambda value to 1 or 5 from the next calculation. Can be set as

The result of the inference operation of the inference calculator may be displayed on the screen of the display unit 40 or output through the output unit 60.

As described above, according to the embodiment of the present invention, it is possible to implement a mobile program by reducing the computational complexity of the method of designing an activation function that guarantees global exponential stability of the inference process.

Therefore, when designing an activation function that guarantees stability in inference of fuzzy recognition in all fields for implementing decision support system using fuzzy recognition, it is possible to increase the computation speed by simplifying the calculation process while satisfying the existing conditions. .

By simplifying complex computations in fuzzy perception, enabling decision inference in mobile programs, various artificial intelligence technologies can be easily implemented by users anytime, anywhere.

In the above description, the embodiments of the present invention have been described with reference to the drawings, for example. However, it will be apparent to those skilled in the art that the present invention may be variously modified or changed within the scope of the technical idea of the present invention. . For example, in other cases, the internal circuit block configuration of the fuzzy recognition inference apparatus or the design method of the activation function may be variously changed or modified without departing from the technical spirit of the present invention.

Description of the Related Art [0002]
10:
20: Memory
30: inference calculator
40: display unit
50: input unit

Claims (24)

Deriving a formula of a given activation function for inferring varying fuzzy recognition degrees on a mobile program;
Checking whether there is a matrix norm when obtaining a lambda (λ) value in the formula of the activation function;
Performing the operation using the matrix norm with less computational throughput than the Euclidean norm when the matrix norm is present to obtain the lambda (λ) value;
And obtaining the activation function by substituting the lambda value into the formula of the activation function.
The method of claim 1, wherein the activation function is a sigmoid function.
The method of claim 2, wherein the formula of the sigmoid function is

A method of designing an activation function for inferring fuzzy recognition, characterized in that The method of claim 3, wherein the lambda (λ) value is as follows to ensure global index stability

A method of designing an activation function for inferring fuzzy recognition, characterized by a model of.
The method of claim 4, wherein

in

A method of designing an activation function for inferring fuzzy recognition, characterized in that it is a Provenius norm calculated as.
6. The method of claim 5, wherein the mobile program is mounted on a personal digital assistant or a smart phone.
6. The method of claim 5, wherein the activation function is updated each time the fuzzy recognition degree is changed.
6. The method of claim 5, wherein the mobile program is mounted on a portable multimedia player.
The method of claim 5, wherein the inference of the fuzzy recognition degree is used for decision support for the implementation of artificial intelligence.
2. The method of claim 1, wherein between the step of obtaining the lambda (λ) value and the step of obtaining the activation function, the lambda is set within a predetermined time when the operation is performed using a matrix norm with less computational throughput than the Euclidean norm ( and if the lambda value is not obtained, setting the lambda value as 1 from the next calculation.
2. The method of claim 1, wherein between the step of obtaining the lambda (λ) value and the step of obtaining the activation function, the lambda is set within a predetermined time when the operation is performed using a matrix norm with less computational throughput than the Euclidean norm ( and if the value of lambda) is not obtained, setting the lambda value as 5 from the next calculation.
The method of claim 1, wherein the activation function is a sinusoidal function.
A memory for storing a formula of a given activation function for inferring varying fuzzy recognition degrees on a mobile program;
An input unit for receiving a concept value for inferring the fuzzy recognition degree;
When there is a matrix norm in obtaining the lambda value included in the formula of the activation function, the lambda value is obtained by performing a calculation using a matrix norm with less computational throughput than the Euclidean norm. and an inference calculating unit for substituting a value of lambda) into a formula of the activation function to obtain the activation function.
The apparatus of claim 13, wherein the fuzzy recognition inference device is mounted on a smart phone.
15. The apparatus of claim 13, wherein the given activation function is a sigmoid function.
The method of claim 15, wherein the formula of the sigmoid function is as follows

Fuzzy recognition inference device, characterized in that appears as.
17. The method of claim 16, wherein the lambda value is equal to the following equation to ensure global index stability.

Fuzzy recognition inference device, characterized in that having a model of.
18. The method of claim 17,
remind

From above

Fuzzy Recognition inference device, characterized in that the Provenius norm calculated as.
19. The apparatus of claim 18, wherein the activation function is updated each time the fuzzy recognition degree is changed.
19. The apparatus of claim 18, wherein the inference of the fuzzy recognition degree is used for decision support for the implementation of artificial intelligence.
The lambda of claim 13, wherein the inference operation unit uses the matrix norm with less computational throughput than the Euclidean norm when the lambda value is not obtained within a predetermined time when the operation is performed. A fuzzy recognition reasoning apparatus characterized by setting the value of (λ) as 1.
The lambda of claim 13, wherein the inference operation unit uses the matrix norm with less computational throughput than the Euclidean norm when the lambda value is not obtained within a predetermined time when the operation is performed. A fuzzy recognition inference device characterized by setting the value of (λ) as 5.
15. The apparatus of claim 13, further comprising a communication unit connected to the inference operation unit to perform wireless communication with an outside, and a display unit displaying a result of the inference operation of the inference operation unit on a screen.
Leading a formula of a sigmoid activation function for inference of fuzzy recognition in a mobile program in a smartphone;
Performing a calculation using a Provenius norm with less computational throughput than the Euclidean norm when obtaining a lambda (λ) value in the formula of the sigmoid activation function;
Generating the activation function by substituting the lambda value into the formula of the sigmoid activation function.

Images