KR20200080164A - Memory management system and method considering application usage patterns analysis - Google Patents

Abstract

A memory management system comprises: a framework module for predicting an execution probability for each of at least one application by generating execution information on when the at least one application has been executed and training a long short-term memory (LSTM) network with the generated execution information; and a kernel module securing a memory by terminating applications in the order of the lower execution probability of the applications based on the predicted execution probability, thereby providing an effect of more efficiently using the memory.

Description

Memory management system and method considering application usage patterns analysis}

The present invention relates to a memory management system and method based on an application usage pattern analysis, and more specifically, a technique for more efficiently utilizing memory by analyzing user application usage information and execution probability.

With the development and widespread use of smartphones, various types of mobile applications are emerging. The number of mobile applications registered on the application download platforms of the representative mobile operating systems, Android and iOS, is currently about 2 million. Smartphone users usually install and use dozens to hundreds of applications, and smartphone applications are equipped with more features to improve user convenience. Accordingly, requirements such as the amount of main memory and computational processing performance are increasing.

Increasing main memory requirements can be addressed by hardware or software methods. The hardware method is a method of expanding the main memory of the smartphone, and the software method is a method of efficiently using the main memory through a memory management policy.

Android or iOS adopts Life Cycle Management, which releases some of the memory of the application according to the execution status of the application in order to quickly execute a user's application execution request with a limited system main memory capacity. Cache as many applications as possible in main memory so they can be loaded quickly when switching applications or running again. In a situation where the main memory is insufficient, the cached applications are started from applications that are considered to be infrequently used to prevent the user from feeling a decrease in system performance. In addition, a swap technique that uses a secondary storage device as a main memory space was considered to efficiently utilize the main memory of a smartphone.

In the case of a NAND flash memory-based storage device, ZRAM, which is used as a swap space by compressing a memory space due to a wear-out problem, or ZSWAP, which is used as a swap cache, is applied to a smartphone.

However, memory management techniques such as LRU applied to a smartphone have limitations in increasing execution time of frequently used applications because it cannot analyze a user's application usage pattern and use it to predict the next application to be used.

US Patent Publication No. 2016-0259557

Accordingly, the present invention has been proposed to solve the above-mentioned general problems, and it is an object of the present invention to analyze a user's application usage information and execution probability to utilize memory more efficiently.

In addition, the present invention aims to be able to utilize an efficient main memory and swap space by learning various usage patterns that appear for each application and the association between applications through a LSTM (Long Short-Term Memory) network.

The present invention also aims to increase the frequency with which applications with a high probability of execution are immediately restarted without being removed from memory.

In order to achieve the above object, the memory management system based on the application usage pattern analysis according to the technical idea of the present invention generates execution information on when at least one or more applications are executed, and LSTM ( Long Short-Term Memory) A framework module that predicts execution probability for each of the at least one application by learning a network, and secures memory by terminating applications in the order of low execution probability based on the predicted execution probability. Kernel module.

The framework module is an execution information generation unit that generates execution information on when at least one application has been executed, and inputs the generated execution information into a LSTM (Long Short-Term Memory) network to use the at least one application. It may include a network learning unit for learning the pattern, an execution probability prediction unit for predicting the execution probability for each of the at least one or more applications based on the learned usage pattern.

The kernel module is an execution probability storage unit for storing the predicted execution probability, and a memory for securing memory by terminating the application in the order of low application execution probability based on the stored execution probability when the preset usage limit value of the memory is exceeded. It may include a securing portion.

The execution probability stored in the execution probability storage unit may be updated whenever the execution probability is predicted by the execution probability prediction unit.

The memory securing unit may secure the memory by terminating the application until it is smaller than the usable reference value of the preset memory.

(시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서 LSTM에 데이터를 저장하는 메모리 셀),

(시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수),

(시간 t에서 LSTM의 출력 벡터)로 구성할 수 있다.The LSTM network is represented by Equations 1 to 6 below.

(Forget gate function that determines the amount of data to be discarded among the stored data at time t),

(At time t

Input gate function to generate data to be stored in),

(At time t

Input gate function to generate data to be stored in),

(Memory cell storing data in LSTM at time t),

(At time t

Input gate function to select the data to be stored in)

It can be configured as (the output vector of LSTM at time t).

<Equation 1>

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 2>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 3>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 4>

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀,

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수)(

: Memory cell that stores data in LSTM at time t,

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: At time t

Input gate function to generate data to be stored in,

: At time t

Input gate function to generate data to be stored in)

<Equation 5>

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to select data to be stored in,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 6>

: 시간 t에서 LSTM의 출력 벡터,

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀)(

: Output vector of LSTM at time t,

: At time t

Input gate function to select data to be stored in,

: Memory cell storing data in LSTM at time t)

The execution probability prediction unit may calculate execution probability values for each application through an execution probability value calculation formula defined by Equations 7 and 8 below.

<Equation 7>

의 j번째 인덱스, m: 집합 U의 원소의 수)(xt: input vector inputted to LSTM with application ui_t executed at time t, ui: application, ht: output vector outputted from LSTM by xt, i: application of execution at time t is the number of applications in set U Pointing index, j: vector

Jth index of m, number of elements in set U)

<Equation 8>

: 시간 t까지 실행된 애플리케이션 기록,

: 시간 t+1에서의 애플리케이션 ui의 실행확률 값, i: 시간 t에서 실행된 애플리케이션이 집합 U의 몇 번째 애플리케이션인지 가리키는 인덱스, j: 벡터

의 j번째 인덱스, m: 집합 U의 원소의 수)(

: Records the applications executed up to time t,

: The value of the execution probability of the application ui at time t+1, i: the index indicating the number of applications in the set U that are executed at time t, j: vector

Jth index of m, number of elements in set U)

In order to achieve the above object, a memory management method based on an application usage pattern analysis according to the technical idea of the present invention includes an execution information generation step of generating execution information on when at least one or more applications are executed in the execution information generation unit. , Network learning step of learning the usage pattern of the at least one application by inputting the generated execution information into the LSTM (Long Short-Term Memory) network in the network learning unit, based on the learned usage pattern in the execution probability prediction unit The execution probability prediction step of predicting the execution probability for each of the at least one application, the execution probability storage step of storing the predicted execution probability in the execution probability storage unit, and exceeding the available reference value of the memory set in the memory acquisition unit In this case, the method may include a memory securing step of securing the memory by terminating the applications in the order of low application execution probability based on the stored execution probability.

The execution probability stored in the execution probability storage step may be updated whenever the execution probability is predicted by the execution probability prediction step.

The memory securing step may secure the memory by terminating the application until it is smaller than the available reference value of the preset memory.

(시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서 LSTM에 데이터를 저장하는 메모리 셀),

(시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수),

(시간 t에서 LSTM의 출력 벡터)로 구성할 수 있다.The LSTM network is defined by Equations 1 to 6 below.

(Forget gate function that determines the amount of data to be discarded among the stored data at time t),

(At time t

Input gate function to generate data to be stored in),

(At time t

Input gate function to generate data to be stored in),

(Memory cell storing data in LSTM at time t),

(At time t

Input gate function to select the data to be stored in)

It can be configured as (the output vector of LSTM at time t).

<Equation 1>

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 2>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 3>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 4>

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀,

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수)(

: Memory cell that stores data in LSTM at time t,

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: At time t

Input gate function to generate data to be stored in,

: At time t

Input gate function to generate data to be stored in)

<Equation 5>

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to select data to be stored in,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 6>

: 시간 t에서 LSTM의 출력 벡터,

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀)(

: Output vector of LSTM at time t,

: At time t

Input gate function to select data to be stored in,

: Memory cell storing data in LSTM at time t)

In the execution probability prediction step, an execution probability value for each application may be calculated through an execution probability value calculation formula defined by Equations 7 and 8 below.

<Equation 7>

의 j번째 인덱스, m: 집합 U의 원소의 수)(xt: input vector inputted to LSTM with application ui_t executed at time t, ui: application, ht: output vector outputted from LSTM by xt, i: application of execution at time t is the number of applications in set U Pointing index, j: vector

Jth index of m, number of elements in set U)

<Equation 8>

: 시간 t까지 실행된 애플리케이션 기록,

: 시간 t+1에서의 애플리케이션 ui의 실행확률 값, i: 시간 t에서 실행된 애플리케이션이 집합 U의 몇 번째 애플리케이션인지 가리키는 인덱스, j: 벡터

의 j번째 인덱스, m: 집합 U의 원소의 수)(

: Records the applications executed up to time t,

: The value of the execution probability of the application ui at time t+1, i: the index indicating the number of applications in the set U that are executed at time t, j: vector

Jth index of m, number of elements in set U)

According to the memory management system and method based on the application usage pattern analysis according to the present invention,

First, it has an effect of using memory more efficiently by analyzing user application usage information and execution probability.

Second, it has the effect of utilizing the effective main memory and swap space by learning the association between various usage patterns and applications that appear for each application through the LSTM (Long Short-Term Memory) network.

Third, it has an effect of increasing the frequency that an application with a high probability of execution is immediately restarted without being removed from the memory.

1 is an embodiment of the present invention, a configuration diagram showing a memory management system based on application usage pattern analysis.
Figure 2 is an embodiment of the present invention, a flow chart showing a memory management method based on the application usage pattern analysis.
3 is a view showing the configuration of the system of the present invention applied to an Android operating system according to an embodiment of the present invention.

A memory management system and method based on application usage pattern analysis according to embodiments of the present invention will be described in detail with reference to the accompanying drawings. The present invention can be applied to various changes and may have various forms, and specific embodiments will be illustrated in the drawings and described in detail in the text. However, this is not intended to limit the present invention to a specific disclosure form, it should be understood to include all modifications, equivalents, or substitutes included in the spirit and scope of the present invention. In describing each drawing, similar reference numerals are used for similar components.

In addition, unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms, such as those defined in a commonly used dictionary, should be interpreted as having meanings consistent with meanings in the context of related technologies, and should not be interpreted as ideal or excessively formal meanings unless explicitly defined in the present application. Does not.

1 is a configuration diagram showing a memory management system based on an application usage pattern analysis as an embodiment of the present invention.

Referring to FIG. 1, a memory management system based on application usage pattern analysis according to an embodiment of the present invention may include a framework module 100 and a kernel module 200.

The framework module 100 generates execution information about when at least one or more applications have been executed, and trains a long short-term memory (LSTM) network with the generated execution information to execute each of the at least one application Probability can be predicted.

In more detail, the framework module 100 may include an execution information generation unit 110, a network learning unit 130, and an execution probability prediction unit 150.

The execution information generation unit 110 may generate execution information on when at least one application has been executed.

The network learning unit 130 may learn the usage pattern of the at least one application by inputting the generated execution information into an LSTM (Long Short-Term Memory) network.

The execution probability predicting unit 150 may predict the execution probability for each of the at least one application based on the learned usage pattern.

The kernel module 200 may secure the memory by terminating the applications in the order of low execution probability of the application based on the predicted execution probability.

In more detail, the kernel module 200 may include an execution probability storage unit 210 and a memory securing unit 230.

The execution probability storage unit 210 may store the predicted execution probability.

At this time, the execution probability stored in the execution probability storage unit 210 may be updated whenever the execution probability is predicted by the execution probability prediction unit 150.

The memory securing unit 230 may secure the memory by terminating the application in the order of low application execution probability based on the stored execution probability, when the preset usage limit of the memory is exceeded.

At this time, the memory securing unit 230 may secure the memory by terminating the application until it is smaller than the available reference value of the preset memory.

(시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서 LSTM에 데이터를 저장하는 메모리 셀),

(시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수),

(시간 t에서 LSTM의 출력 벡터)로 구성할 수 있다.Meanwhile, the LSTM network is defined by Equations 1 to 6 below.

(Forget gate function that determines the amount of data to be discarded among the stored data at time t),

(At time t

Input gate function to generate data to be stored in),

(At time t

Input gate function to generate data to be stored in),

(Memory cell storing data in LSTM at time t),

(At time t

Input gate function to select the data to be stored in)

It can be configured as (the output vector of LSTM at time t).

<Equation 1>

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 2>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 3>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 4>

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀,

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수)(

: Memory cell that stores data in LSTM at time t,

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: At time t

Input gate function to generate data to be stored in,

: At time t

Input gate function to generate data to be stored in)

<Equation 5>

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to select data to be stored in,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 6>

: 시간 t에서 LSTM의 출력 벡터,

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀)(

: Output vector of LSTM at time t,

: At time t

Input gate function to select data to be stored in,

: Memory cell storing data in LSTM at time t)

In addition, the execution probability prediction unit may calculate execution probability values for each application through an execution probability value calculation formula defined by Equations 7 and 8 below.

<Equation 7>

의 j번째 인덱스, m: 집합 U의 원소의 수)(xt: input vector inputted to LSTM with application ui_t executed at time t, ui: application, ht: output vector outputted from LSTM by xt, i: application of execution at time t is the number of applications in set U Pointing index, j: vector

Jth index of m, number of elements in set U)

<Equation 8>

: 시간 t까지 실행된 애플리케이션 기록,

: 시간 t+1에서의 애플리케이션 ui의 실행확률 값, i: 시간 t에서 실행된 애플리케이션이 집합 U의 몇 번째 애플리케이션인지 가리키는 인덱스, j: 벡터

의 j번째 인덱스, m: 집합 U의 원소의 수)(

: Records the applications executed up to time t,

: The value of the execution probability of the application ui at time t+1, i: the index indicating the number of applications in the set U that are executed at time t, j: vector

Jth index of m, number of elements in set U)

2 is a flowchart illustrating a memory management method based on application usage pattern analysis as an embodiment of the present invention.

Referring to FIG. 2, as an embodiment of the present invention, a memory management method based on application usage pattern analysis includes execution information generation step (S110), network learning step (S130), execution probability prediction step (S150), and execution probability storage step (S210), a memory securing step (S230).

In the execution information generation step, the execution information generation unit 110 may generate execution information on when at least one application has been executed (S110 ).

In the network learning step, the network learning unit 130 may input the generated execution information into an LSTM (Long Short-Term Memory) network to learn the usage pattern of the at least one application (S130).

In the execution probability prediction step, the execution probability for each of the at least one application may be predicted based on the learned usage pattern in the execution probability prediction unit 150 (S150 ).

The execution probability storage step may store the predicted execution probability in the execution probability storage unit 210 (S210).

At this time, the execution probability stored in the execution probability storage step S210 may be updated whenever the execution probability is predicted by the execution probability prediction step S150.

When the memory securing step exceeds the usable threshold value of the memory set in the memory securing unit 230, the application execution probability may be terminated in order of the application execution probability based on the stored execution probability to secure the memory (S230).

At this time, the memory securing step S230 may secure the memory by terminating the application until it is smaller than the available reference value of the preset memory.

(시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수),

(시간 t에서 LSTM에 데이터를 저장하는 메모리 셀),

(시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수),

(시간 t에서 LSTM의 출력 벡터)로 구성할 수 있다.Meanwhile, the LSTM network is defined by Equations 1 to 6 below.

(Forget gate function that determines the amount of data to be discarded among the stored data at time t),

(At time t

Input gate function to generate data to be stored in),

(At time t

Input gate function to generate data to be stored in),

(Memory cell storing data in LSTM at time t),

(At time t

Input gate function to select the data to be stored in)

It can be configured as (the output vector of LSTM at time t).

<Equation 1>

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 2>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 3>

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 4>

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀,

: 시간 t에서 저장된 데이터 중 버릴 데이터의 양을 결정하는 망각 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수,

: 시간 t에서

에 저장될 데이터를 생성하는 입력 게이트 함수)(

: Memory cell that stores data in LSTM at time t,

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: At time t

Input gate function to generate data to be stored in,

: At time t

Input gate function to generate data to be stored in)

<Equation 5>

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시그모이드 활성화 함수,

: 시간 t에서 LSTM에 입력되는 입력벡터,

:

함수에서 각

에 가중되는 벡터,

: 시간 t-1에서 LSTM의 출력 벡터,

:

함수에서 각

에 가중되는 벡터,

:

함수를 위한 바이어스 상수)(

: At time t

Input gate function to select data to be stored in,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)

<Equation 6>

: 시간 t에서 LSTM의 출력 벡터,

: 시간 t에서

에 저장될 데이터를 선택하는 입력 게이트 함수,

: 시간 t에서 LSTM에 데이터를 저장하는 메모리 셀)(

: Output vector of LSTM at time t,

: At time t

Input gate function to select data to be stored in,

: Memory cell storing data in LSTM at time t)

In addition, the execution probability prediction step (S150) may calculate execution probability values for each application through an execution probability value calculation formula defined by Equations 7 and 8 below.

<Equation 7>

의 j번째 인덱스, m: 집합 U의 원소의 수)(xt: input vector inputted to LSTM with application ui_t executed at time t, ui: application, ht: output vector outputted from LSTM by xt, i: application of execution at time t is the number of applications in set U Pointing index, j: vector

Jth index of m, number of elements in set U)

<Equation 8>

: 시간 t까지 실행된 애플리케이션 기록,

: 시간 t+1에서의 애플리케이션 ui의 실행확률 값, i: 시간 t에서 실행된 애플리케이션이 집합 U의 몇 번째 애플리케이션인지 가리키는 인덱스, j: 벡터

의 j번째 인덱스, m: 집합 U의 원소의 수)(

: Records the applications executed up to time t,

: The value of the execution probability of the application ui at time t+1, i: the index indicating the number of applications in the set U that are executed at time t, j: vector

Jth index of m, number of elements in set U)

Referring to the present invention in more detail, reference may be made to FIG. 3.

3 is a diagram illustrating a configuration diagram of the system of the present invention applied to an Android operating system according to an embodiment of the present invention.

As shown in FIG. 3, according to an embodiment of the present invention, the Android operating system to which the system of the present invention is applied includes the Android Framework 100 corresponding to the framework module and the Android Kernel 200 corresponding to the kernel module. can do.

Android Framework 100 may include Activity Manager Service and Context Management.

The Activity Manager Service may include an Activity/Process Context Generator 110 corresponding to an execution information generator. The Activity/Process Context Generator 110 of the Activity Manager Service can generate information about which application was executed and when it is delivered to Context Management.

The Activity/Process Context Generator 110 may be configured in a portion in charge of application execution and termination, such as Android's Activity Manager Service in a mobile operating system. This can generate situation information such as application execution, suspension, and termination.

In addition, the Activity/Process Context Generator 110 includes application usage information such as the name of the application package, a timestamp at the time the application is executed, and the ID of the process that is generated for each process created for the application execution, and to which each process belongs. Process creation information such as application name can be collected.

Meanwhile, in the Activity Manager Service including the Activity/Process Context Generator 110, when application execution and process creation occur, application usage information may be delivered to the Application Usage Trainer 130 and Application Usage Predictor 150 of Context Management. . In addition, when a process creation operation for executing an application occurs, corresponding situation information may be transmitted to the Application Usage Predictor 150.

Context Management may include an Application Usage Trainer 130 corresponding to the network learning unit, an Application Usage Predictor 150 corresponding to the execution probability prediction unit, and an LSTM Network.

The Application Usage Trainer 130 of Context Management can store the information received from the Activity/Process Context Generator 110 and train the LSTM Network.

The Application Usage Trainer 130 may receive information on which application is executed from the Activity/Process Context Generator 110 and store the IDs corresponding to the application in storage in chronological order. When training the LSTM network, the feed-forward network is generated by unfolding the LSTM network by the length of the data to be learned at a time, and then the BPTT algorithm that updates the network's weight parameters with the backpropagation algorithm can be used. LSTM Network can be learned to minimize Mean Squared Error (MSE), which is the difference between the predicted value and the actual value. The MSE to which the application execution probability is applied can be defined as in Equation 9 below.

<Equation 9>

: 시간 t의 애플리케이션 실행 기록 집합, i: 시간 t에서 실행된 애플리케이션이 집합 U의 몇 번째 애플리케이션인지 가리키는 인덱스, j: 벡터

의 j번째 인덱스, m: 집합 U의 원소의 수,

: 시간 t에서의 애플리케이션 ui의 실행확률 값,

: 시간 t에서 애플리케이션 ui가 실행되었는지 여부,

: 시간 t에서 실행된 애플리케이션

,

: 시간 t에서 애플리케이션 실행 기록의 가장 마지막 실행된 애플리케이션)(MSE: LSTM Network is the difference between the predicted value and the actual value,

: Application execution history set at time t, i: Index indicating the number of applications in application U executed at time t, j: Vector

Jth index of, m: number of elements in set U,

: Ui execution probability value at time t,

: Whether application ui was executed at time t,

: Application run at time t

,

: Last run application of the application run record at time t)

LSTM Network's LSTM can solve the gradient vanishing problem that occurs in the existing Recurrent Neural Network (RNN), and learn to relate the long-term and short-term data. The output vectors ht output from the LSTM at (Forget Gate) ft, (Input Gate) gt and it, memory cell ct, (Output Gate) ot, and time t constituting the LSTM are defined by Equations 1 to 6 Can. At this time, the parameters of the LSTM Network can be defined in Table 1 below.

<Table 1>

The relevant parameters for predicting the probability of application execution using LSTM Network can be defined in Table 2 below.

<Table 2>

를 통해 시간 t+1의 애플리케이션

의 실행 확률을 알아내기 위해, 실행 확률은

로 나타낼 수 있다. 가장 실행 가능성이 낮은 애플리케이션을 선택하기 위해, 시간 t에서 실행된 애플리케이션

를 LSTM 모델의 입력

로 변환하는 식과, 상기 수식 1 내지 수식 6에 정의된 LSTM 모델에 따른 연산을 수행하는 함수 lstm(

)를 상기 수식 7과 같이 정의할 수 있다.Record applications executed up to time t

Application of time t+1 through

To find the probability of execution, the probability of execution is

Can be represented as Application selected at time t to select the least viable application

LSTM model input

Function to perform the calculation according to the LSTM model defined in Equations 1 to 6 and the equation to convert to lstm(

) May be defined as Equation 7 above.

)의 출력을

에 대한 확률값으로 변환하기 위해, 상기 수식 8과 같이 Softmax 함수를 이용할 수 있다.Also, lstm(

) Output

In order to convert to a probability value for, a Softmax function can be used as in Equation 8 above.

The Application Usage Predictor 150 of Context Management predicts the probability of the application to be executed by the user next using the learned LSTM Network, and is managed by the Process Management 210 of the Android Kernel 200 for each prediction. The probability of execution can be recorded in the process information.

At this time, the Application Usage Predictor 150 of Context Management can use the received LSTM Network as well as receive additional information about which application has been executed from the Activity/Process Context Generator 110.

The Application Usage Predictor 150 may generate a list of application-process information called AppInfo as shown in Table 4 below after reading the application mapping table stored in the form of Table 3 below.

<Table 3>

(NAME: Full name of the application, ID: Unique ID of the application)

<Table 4>

(name: full name of the application, id: unique ID of the application, prob: probability of the application running, pids: list of pids of processes belonging to the application)

Mobile operating systems, such as Android, differ in the point of application use and process creation. Accordingly, the Application Usage Predictor 150 may receive information about when an application was executed and when and which application process was created from the Activity/Process Context Generator 110.

Algorithm 1 below may indicate a task to be performed when receiving information about which application is executed.

<Algorithm 1>

By calling LSTM Network as in Line 2-3, you can get the probability value that applications will be executed next, and save the value in the prob field while traversing the AppInfo list as in Line 4-5. If the process of the application has already been created, it is possible to record the probability value of the application in the process data structure of the kernel while traversing pids as shown in line 6-14 through File System Management. Since Android Kernel does not support floating point operation, probability value can be converted to Integer and recorded as in line 10.

Algorithm 2 below may indicate an operation to be performed when receiving information about which application process is generated.

<Algorithm 2>

Algorithm 2 may represent an operation performed when receiving information about which application process is created. When a process is created as in Line 2, the pid of the created process can be added to the list of pids of AppInfo of the application and the probability value can be recorded in the process data structure of the kernel.

The Android Kernel 200 may include File System Management, Process Management 210 corresponding to the execution probability storage unit, and Memory Management 230 corresponding to the memory securing unit. At this time, the Memory Management 230 may include a Memory Reclaimer.

File System Management may provide an information delivery interface for reflecting application execution probability information of the Application Usage Predictor 150 executed on the top of Android Kernel to each process information belonging to the application. The information delivery interface can be configured through procfs. procfs displays system information, such as process information, in a file structure on Unix-like operating systems, and can be used to change system parameters at runtime through the file input/output interface. In the present invention, the application usage predictor 150 may record an execution probability of an application through procfs in a process data structure managed by Process Management 210.

Process Management 210 can manage process information through the process data structure of Android Kernel 200.

Table 5

The form in which the execution probability value of the application to which the process belongs is recorded in Table 5 above, may be referred to as a form in which a prob field is added. Since the system process may cause a problem such as a system crash at the end, the execution probability value calculated through the Application Usage Predictor 150 can be applied only to processes belonging to an application directly executed by the user. To this end, prob is initialized to an arbitrary negative value to exclude system processes from main memory priority recall.

The memory reclaimer of the memory management 230 may search for the execution probability value recorded in the processes when the main memory shortage situation occurs, find the process having the lowest execution probability value, terminate it, and secure memory space. This has the advantage of increasing the frequency with which applications that are likely to run are not removed from memory and resume immediately.

The Memory Management 230 may call a Memory Reclaimer when a main memory shortage occurs. The following algorithm 3 may be referred to as an algorithm showing a memory recovery operation performed in the Memory Reclaimer.

<Algorithm 3>

Line 2-3 allows you to initialize a variable to point to the process where the probability value is minimal. In Line 4, it is possible to traverse the process list of Process Management and select the process with the lowest probability value among the processes subject to the main memory priority recovery as in lines 5-8.

As described above, the memory management system and method based on the application usage pattern analysis according to the embodiment of the present invention may be implemented in an actual Android smartphone to collect and analyze the execution time of the mobile application and characteristics of the application execution. As a target device to use for measuring its performance, a Google Nexus 6P smartphone equipped with a Qualcomm Snapdragon 810 processor can be used as shown in Table 6 below.

Table 6.

Cold Launch, Warm Launch, and Hot Launch exist in application execution. Hot Launch is a case in which all processes required to run an application have already been created and no new process needs to be created. In the case of Warm Launch, some processes for application execution are running, but one or more processes need to be newly created. In the case of Cold Launch, this is a case where all necessary processes are newly created because a process for running an application is not created.

As the performance measurement data, randomly selected user data was used from the application usage records of real users collected by LiveLab Research. The total number of application types in the application usage record was 35, as shown in Table 7 below.

Table 7

For cross-validation of learning results, 90% (4500) were used for learning and 10% (500) were used for testing out of a total of 5,000 usage records using 10-fold cross validation. Android provides application execution time information for Cold Launch and Warm Launch, but does not provide execution time for Hot Launch. Therefore, the average value of the cumulative execution time of applications during Cold and Warm Launch was compared using the existing technique and the technique of the present invention. In addition, the Activity Manager Service was modified to measure the number of times for each application execution type, and the average value was compared. The existing memory management system was designated as Android+ZRAM+LRU to indicate that the source code of the basic Android platform and its kernel was built, and the memory management system of the present invention was designated as Android+ZRAM+AMMS.

First, in order to check the performance of application execution prediction, 500 application test data were executed in order, and the number of occurrences of Hot Launch, Warm Launch, and Cold Launch were measured against the existing system and the proposed system. The results of calculating the average by accumulating the types of application execution that occur when executing 500 application usage records are shown in Table 8 below.

Table 8

The average number of Cold Launches, which took the longest to run Android+ZRAM+AMMS applications, was 76.4, about 10% less than 85 of Android+ZRAM+LRU. In addition, in the case of Warm Launch, which loads some application data into memory, the decrease was 10% from 14.4 to 12.8.

The system of the present invention accurately predicts the probability that the application will be executed than the existing system than the existing system, and it can be confirmed that the number of hot launches in which the application with a high probability of execution is directly executed in the main memory has increased.

Next, the application execution time was measured to determine the effect of the application execution prediction performance on the existing system and the system of the present invention on the application execution time. Application execution time refers to the time to draw a view on the screen after application loading, process creation, or activity creation. The execution time of the application was performed using the execution time measurement tool provided by the Android framework.

Since the execution time measurement tool provides execution time only for Cold Launch and Warm Launch, in the case of Hot Launch, where the entire application is cached in memory and executed immediately, the execution time was not measured. The results of calculating the average by accumulating execution times for Cold Launch and Warm Launch by running 500 application test data in order are shown in Table 9 below.

Table 9

The average cumulative time of Warm Launch was reduced by about 11% from 7095.8 ms of the existing system to 6302.8 ms of the present system. Cold Launch was reduced by about 18% from 106128.6ms of the existing system to 86810ms of the present system.

The average of the cumulative execution time of Warm Launch and Cold Launch is about 17% lower than that of the existing system. Since the execution time of Warm Launch and Cold Launch is longer than that of Hot Launch, the decrease of both Warm Launch and Cold Launch execution times indicates that the average execution time of the entire application decreases.

Since the present invention analyzes the probability of execution of an application and applies it to memory management, it is possible to averagely improve the execution speed of an application having a high probability of being executed by first retrieving the memory of the application having a low probability of execution in a memory shortage situation.

The hardware performance of smartphones has increased, and a large amount of main memory is required as users install and use applications that require high specifications on their smartphones. In order to solve this problem, a ZRAM swap technique that manages main memory more efficiently or a swap technique that utilizes auxiliary storage devices has been developed.

However, since ZRAM uses main memory, there is a limit to expand the size, and traditional swapping using auxiliary storage as a swap space takes a long time to respond and smartphones suffer from wear-out of Nand Flash memory. There are limitations that cannot be applied to. In addition, the effect of the existing LRU method memory management policy decreases as the number of applications used by the user increases.

To solve this problem, the memory management system of the present invention may collect and learn information that a user uses an application, and manage memory by predicting a probability that an application will be executed next. After analyzing the execution probability information, since the application with the lowest execution probability is removed from the memory space in a memory shortage situation, frequently used applications remain in memory as much as possible and have an effect of reducing the average execution time. As a result of performing a performance test based on the records used by the actual smart phone users, the present invention is about 10% of the number of times that applications that are more likely to be executed by users remain in the memory space for longer than the existing smart phone memory management technology and resume. It can be seen that it increased, and the average execution time of the application was reduced by about 17%.

Although the preferred embodiments of the present invention have been described above, the present invention can use various changes, modifications, and equivalents. It is clear that the present invention can be equally applied by appropriately modifying the above embodiments. Therefore, the above description is not intended to limit the scope of the present invention as defined by the following claims.

100: Framework module (Android Framework)
110: execution information generator (Activity/Process Context Generator)
130: Network Usage Department (Application Usage Trainer)
150: execution probability prediction unit (Application Usage Predictor)
200: kernel module (Android Kernel)
210: execution probability storage (Process Management)
230: Memory Management (Memory Management)

Claims (12)

A framework module that generates execution information on when at least one or more applications are executed, and predicts an execution probability for each of the at least one application by learning a long short-term memory (LSTM) network with the generated execution information ; And
A memory management system based on application usage pattern analysis, including; a kernel module that secures memory by terminating the application in the order of low execution probability of the application based on the predicted execution probability. The method of claim 1, wherein the framework module,
An execution information generation unit that generates execution information on when at least one application is executed;
A network learning unit that inputs the generated execution information into an LSTM (Long Short-Term Memory) network to learn a usage pattern of the at least one application; And
And an execution probability predicting unit for predicting an execution probability for each of the at least one application based on the learned usage pattern. The method of claim 1, wherein the kernel module,
An execution probability storage unit for storing the predicted execution probability; And
Memory management system based on the application usage pattern analysis, including; if the exceeding the available reference value of the preset memory, a memory securing unit for securing the memory by terminating the application in the order of low application execution probability based on the stored execution probability. According to claim 3, The execution probability stored in the execution probability storage unit,
A memory management system based on application usage pattern analysis that is updated whenever the execution probability is predicted by the execution probability prediction unit. The method of claim 3, wherein the memory securing unit,
A memory management system based on application usage pattern analysis that secures memory by terminating the application until it is less than the available reference value of the preset memory. The method of claim 2, wherein the LSTM network,
Defined by Equations 1 to 6 below

(Forget gate function that determines the amount of data to be discarded among the stored data at time t),

(At time t

Input gate function to generate data to be stored in),

(At time t

Input gate function to generate data to be stored in),

(Memory cell storing data in LSTM at time t),

(At time t

Input gate function to select the data to be stored in)

Memory management system based on application usage pattern analysis consisting of (output vector of LSTM at time t).
<Equation 1>

(

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 2>

(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 3>

(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 4>

(

: Memory cell that stores data in LSTM at time t,

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: At time t

Input gate function to generate data to be stored in,

: At time t

Input gate function to generate data to be stored in)
<Equation 5>

(

: At time t

Input gate function to select data to be stored in,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 6>

(

: Output vector of LSTM at time t,

: At time t

Input gate function to select data to be stored in,

: Memory cell storing data in LSTM at time t) The method of claim 6, wherein the execution probability prediction unit,
A memory management system based on an application usage pattern analysis that calculates an execution probability value for each application through an execution probability value calculation formula defined by Equations 7 and 8 below.
<Equation 7>

(xt: input vector inputted to LSTM with application ui_t executed at time t, ui: application, ht: output vector outputted from LSTM by xt, i: application of execution at time t is the number of applications in set U Pointing index, j: vector

Jth index of m, number of elements in set U)
<Equation 8>

(

: Records the applications executed up to time t,

: The value of the execution probability of the application ui at time t+1, i: the index indicating the number of applications in the set U that are executed at time t, j: vector

Jth index of m, number of elements in set U) An execution information generation step of generating execution information on when the at least one application is executed by the execution information generation unit;
A network learning step of learning the usage pattern of the at least one application by inputting the generated execution information into a long short-term memory (LSTM) network in a network learning unit;
An execution probability prediction step in which an execution probability prediction unit predicts an execution probability for each of the at least one application based on the learned usage pattern;
An execution probability storage step of storing the predicted execution probability in an execution probability storage unit; And
Based on the analysis of the application usage pattern, including: a memory securing step of securing the memory by terminating the application in the order of low application execution probability based on the stored execution probability, when the memory usage unit exceeds the preset usage standard of the memory; How to manage memory. The execution probability stored in the execution probability storage step is 10.
A memory management method based on an application usage pattern analysis that is updated whenever the execution probability is predicted by the execution probability prediction step. The method of claim 8, wherein the step of securing the memory,
A memory management method based on an application usage pattern analysis that secures a memory by terminating an application until it is smaller than the available reference value of the preset memory. The method of claim 8, wherein the LSTM network,
Application usage pattern analysis consisting of (Forget Gate) ft, (Input Gate) gt and it, memory cell ct, (Output Gate) ot, and output vector ht output from LSTM at time t defined by Equations 1 to 6 below Memory management method based on.
<Equation 1>

(

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 2>

(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 3>

(

: At time t

Input gate function to generate data to be stored in,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 4>

(

: Memory cell that stores data in LSTM at time t,

: Forgetting gate function which determines the amount of data to be discarded among the stored data at time t,

: At time t

Input gate function to generate data to be stored in,

: At time t

Input gate function to generate data to be stored in)
<Equation 5>

(

: At time t

Input gate function to select data to be stored in,

: Sigmoid activation function,

: Input vector input to LSTM at time t,

:

Each in function

Vector weighted to,

: LSTM output vector at time t-1,

:

Each in function

Vector weighted to,

:

Bias constant for function)
<Equation 6>

(

: Output vector of LSTM at time t,

: At time t

Input gate function to select data to be stored in,

: Memory cell storing data in LSTM at time t) The method of claim 11, wherein the predicting the probability of execution,
A memory management method based on an application usage pattern analysis that calculates an execution probability value for each application through an execution probability value calculation formula defined by Equations 7 and 8 below.
<Equation 7>

(xt: input vector inputted to LSTM with application ui_t executed at time t, ui: application, ht: output vector outputted from LSTM by xt, i: application of execution at time t is the number of applications in set U Pointing index, j: vector

Jth index of m, number of elements in set U)
<Equation 8>

(

: Records the applications executed up to time t,

: The value of the execution probability of the application ui at time t+1, i: the index indicating the number of applications in the set U that are executed at time t, j: vector

Jth index of m, number of elements in set U)

Images