WO2022037164A1

WO2022037164A1 - Labeled data storage server allocation method based on evolutionary optimization

Info

Publication number: WO2022037164A1
Application number: PCT/CN2021/096124
Authority: WO
Inventors: 郭肇禄; 陈远存; 谭力江; 张文生; 黄文俊; 蔡岳城
Original assignee: 广东奥博信息产业股份有限公司
Priority date: 2020-08-20
Filing date: 2021-05-26
Publication date: 2022-02-24
Also published as: CN112148213B; CN112148213A

Abstract

Disclosed is a labeled data storage server allocation method based on evolutionary optimization. In the present invention, an improved sine-cosine algorithm is used to solve the problem of labeled data storage server allocation. The improved sine-cosine algorithm comprises: firstly generating a random population, and setting the survival periods of individuals; next, generating guidance probabilities by means of comprehensively taking the priorities and survival periods of the individuals into consideration; then, choosing guidance individuals on the basis of the guidance probabilities, and executing sine-cosine operation operators based on the guidance individuals, choosing excellent individuals to enter a new-generation population, and then updating the survival periods of the individuals; and repeatedly executing this search process until a termination condition is satisfied. In the present invention, a search guidance direction is chosen by means of comprehensively taking the priorities and survival periods of individuals into consideration, thereby reducing the probability of falling into the local optimum, and the comprehensive storage efficiency of allocated labeled data storage servers can be improved.

Description

Annotation data storage server allocation method based on evolutionary optimization

technical field

The invention relates to the field of cloud computing, in particular to a method for allocating annotated data storage servers based on evolutionary optimization.

Background technique

Artificial intelligence is a hot technology in the modern information society. Artificial intelligence technology has been widely applied to all walks of life in society, such as: intelligent video surveillance, fingerprint recognition, handwriting recognition, voice recognition, etc. The support foundation of artificial intelligence is massive labeled data. Only with a large amount of labeled data can the accuracy and usability of artificial intelligence systems be improved.

Faced with massive labeled data, technicians often need to build cloud computing systems to store massive labeled data. In cloud computing systems that store annotation data, technicians often need to solve an annotation data storage server allocation problem: the cloud computing system is equipped with ND annotation data storage servers, each of which has an external storage capacity SV _dt , Internal storage volume MV _dt and energy consumption value per unit time PE _dt , where dimension subscript dt=1,2,...,ND; for a given labeling data storage task, it needs to be selected from ND labeling data storage servers To store the annotation data from the server, it is required to maximize the sum of the external storage capacity of the selected annotation data storage server, and the sum of the internal storage capacity of the selected annotation data storage server must be greater than the lower limit of the internal storage capacity LM, and also requires The sum of the energy consumption per unit time of the selected annotation data storage servers must be less than the energy consumption upper limit SE. The labeling data storage server allocation problem is an NP-complete problem. When the scale of the annotation data storage server is large, it is difficult for the traditional method to obtain an optimal annotation data storage server allocation scheme within an acceptable time. To this end, many researchers have proposed using evolutionary optimization algorithms that simulate the laws of nature to solve the problem.

The sine-cosine algorithm is an evolutionary optimization algorithm proposed in recent years. It has achieved certain results in optimizing many practical engineering problems. Methods[J].Acta Optics,2019,39(09):411-417.]. However, when the sine-cosine algorithm solves the problem of allocating annotated data storage servers, it is easy to fall into a local optimum, so that the overall storage efficiency of the allocated annotation data storage servers is not high.

SUMMARY OF THE INVENTION

The invention provides a method for allocating annotated data storage server based on evolution optimization. To a certain extent, it can overcome the problem that the traditional sine and cosine algorithm is prone to fall into local optimum when solving the problem of labeling data storage server allocation, and the invention can improve the comprehensive storage efficiency of the allocated labeling data storage server.

Technical scheme of the present invention: a method for allocating annotated data storage server based on evolution optimization, comprising the following steps:

Step 1, enter the number of labeled data storage servers ND;

Step 2, input the external storage volume SV _dt of the ND station labeling data storage server, the internal storage volume MV _dt and the energy consumption value per unit time PE _dt , where the dimension subscript dt=1,2,...,ND;

Step 3, input the lower limit of internal storage LM and the upper limit of energy consumption SE;

Step 4, set the population size NPS, the maximum number of iterations MIT and the maximum survival period MSL;

Step 5, set the current iteration number it=0;

Step 6: Randomly generate NPS individuals to form a population RP={RA ₁ ,RA ₂ ,...,RA _ki ,...,RA _NPS }, where RA _ki =[RA _ki,1 ,RA _ki,2 , ...,RA _ki,dt ,...,RA _ki,ND ] is the ki-th individual in the population, and the individual RA _ki stores the responsivity of the ND label data storage server; RA _ki,dt represents the individual RA Responsiveness of the dt-th labeled data storage server stored in _ki ; wherein, the individual subscripts ki=1,2,...,NPS, and the dimension subscripts dt=1,2,...,ND;

Step 7: Calculate the fitness value of the NPS individuals in the population; for the individual RA _ki , the calculation process of the fitness value PFit _ki is as follows: first, decode the responsiveness of the labeled data storage server of the ND station stored by the individual RA _ki into the state pool RT _ki , and then calculate the fitness value PFit _ki of individual RA _ki according to formula (1):

Among them, the state pool RT _ki stores the allocation status of the ND label data storage server, and RT _ki,dt represents the allocation status of the dt-th label data storage server stored in RT _ki ; aw1 is the memory penalty factor; aw2 is the energy consumption Penalty factor; wmv is the memory difference; wpe is the excess energy consumption; max is the function of taking the maximum value;

Step 8, set the survival period of each individual in the population SL _ki = 1, where the individual subscripts ki = 1, 2, ..., NPS;

Step 9: Find the individual with the smallest fitness value from the population and save it to the optimal individual GRA;

Step 10, calculate the integration coefficient kw according to formula (2)

Step 11: Calculate the bootstrap probability LP _ki of each individual in the population according to formula (3):

Among them, APFit _ki represents the priority of individual RA _ki , and the individual subscript ki=1,2,...,NPS;

Step 12, according to the leading probability LP _ki of each individual in the population, adopt the roulette selection method to select the leading individual EA _ki from the population;

Step 13, according to formula (4), perform the sine and cosine operation operator based on the guiding individual:

in

r2 and r4 are randomly generated, r2 is a random real number between [0,2×π], and π is a pi; r3 is a random real number between [0,2]; r4 is a random real number between [0,1] real number; sin is a sine function; cos is a cosine function; RU _ki is a new individual;

Step 14, calculate the fitness value UFit _ki of the newborn individual RU _ki according to formula (1);

Step 15, if the fitness value of the new individual RU _ki is less than the fitness value of the individual RA _ki , replace the individual RA _ki with the new individual RU _ki in the population, otherwise keep the individual RA _ki unchanged in the population;

Step 16, update the lifetime SL _ki according to formula (5):

Step 17, if SL _ki is greater than MSL, randomly generate an individual NA _ki , then replace the individual RA _ki with the individual NA _ki in the population and set SL _ki =1, otherwise keep the individual RA _ki unchanged in the population;

Step 18: Find the individual with the smallest fitness value from the population and save it to the optimal individual GRA;

Step 19, set the current iteration number it=it+1;

Step 20, if the current iteration it is greater than MIT, go to step 21, otherwise go to step 10;

In step 21, the optimal individual GRA is decoded into the allocation state of the labeled data storage server, that is, the allocation result of the labeled data storage server is obtained.

The beneficial effects of the invention are as follows: when the traditional sine and cosine algorithm optimizes the allocation problem of labeling data storage servers, both the sine search strategy and the cosine search strategy directly use the optimal individual in the population as the search guide direction, which makes the optimized search show better performance. Large greed often leads to a lack of diversity in the population, which is prone to the problem of falling into local optimum. In order to improve the deficiencies of the traditional method, the present invention adopts the improved sine and cosine algorithm to optimize the allocation problem of the labeling data storage server. In the improved sine cosine algorithm, the fitness value of the individual is set as the priority of selecting the search guidance direction, and the survival period of the individual in the population is also set as another criterion for selecting the search guidance direction. The improved sine and cosine algorithm combines the priority and survival period of the individual to select the search guidance direction of the individual, which maintains the diversity of the population, reduces the probability of falling into a local optimum, and can improve the integration of the assigned annotation data storage server. storage efficiency.

Description of drawings

Figure 1 is a flow chart of a method for assigning annotated data storage servers based on evolutionary optimization.

detailed description

The technical solutions of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

Example:

The present embodiment combines the accompanying drawings, and the specific implementation steps of the present invention are as follows:

Step 1, enter the number of labeled data storage servers ND=150;

The labeling data storage server refers to a server used to store labeling data; wherein, labeling data refers to data that has undergone data cleaning and has been labeled; the external storage means that the hard disk of the labeling data storage server is free Space size; the internal storage volume refers to the size of the memory of the labeling data storage server; the energy consumption value per unit time refers to the power consumption of the labeling data storage server working for 1 hour;

Step 3, input the lower limit of internal storage LM=2600 and the upper limit of energy consumption SE=68000;

Step 4, set the population size NPS=100, the maximum number of iterations MIT=3000 and the maximum survival period MSL=160;

Step 5, set the current iteration number it=0;

Step 6: Randomly generate NPS individuals to form a population RP={RA ₁ ,RA ₂ ,...,RA _ki ,...,RA _NPS }, where RA _ki =[RA _ki,1 ,RA _ki,2 , ...,RA _ki,dt ,...,RA _ki,ND ] is the ki-th individual in the population, and the individual RA _ki stores the responsivity of the ND label data storage server; RA _ki,dt represents the individual RA The responsivity of the dt-th labeled data storage server stored in _ki ; wherein, the individual subscripts ki=1,2,...,NPS, and the dimension subscripts dt=1,2,...,ND; the The responsivity of the annotation data storage server is a real number between [0, 1];

The specific process of decoding the responsiveness of the ND station annotation data storage server stored in the individual RA _ki into the state pool RT _ki is as follows: first, four roundings are performed on the responsiveness of the ND station annotation data storage server stored in the individual RA _ki . The five inputs are processed into integers, and ND integers with the value of 0 or 1 are obtained, and then the obtained ND integers with the value of 0 or 1 are stored in the state pool RT _ki in turn; the ND values saved in the state pool RT _ki are An integer of 0 or 1 is the allocation status of the ND annotation data storage server; the value of 1 in the state pool RT _ki indicates that the corresponding annotation data storage server is allocated to complete the given storage task; the value in the state pool RT _ki A value of 0 indicates that the corresponding annotation data storage server is not allocated to complete the given storage task;

Step 10, calculate the integration coefficient kw according to formula (2)

in

r2 is a random real number between [0,2×π], and π is a pi; r3 is a random real number between [0,2]; r4 is a random real number between [0,1]; sin is a sine function ; cos is the cosine function; RU _ki is the new individual;

Step 16, update the lifetime SL _ki according to formula (5):

Step 19, set the current iteration number it=it+1;

Claims

The method for allocating annotated data storage servers based on evolutionary optimization is characterized by comprising the following steps:

Step 1, enter the number of labeled data storage servers ND;

Step 2, input the external storage amount SV dt of the ND station labeling data storage server, the internal storage amount MV dt and the energy consumption value per unit time PE dt , wherein the dimension subscript dt=1,2,...,ND;

Step 3, input the lower limit of internal storage LM and the upper limit of energy consumption SE;

Step 4, set the population size NPS, the maximum number of iterations MIT and the maximum survival period MSL;

Step 5, set the current iteration number it=0;

Step 6: Randomly generate NPS individuals to form a population RP={RA 1 ,RA 2 ,...,RA ki ,...,RA NPS }, where RA ki =[RA ki,1 ,RA ki,2 , ...,RA ki,dt ,...,RA ki,ND ] is the ki-th individual in the population, and the individual RA ki stores the responsivity of the ND label data storage server; RA ki,dt represents the individual RA Responsiveness of the dt-th labeling data storage server stored in ki ; where the individual subscript ki=1,2,...,NPS;

Step 7: Calculate the fitness value of the NPS individuals in the population; for the individual RA ki , the calculation process of the fitness value PFit ki is as follows: first, decode the responsiveness of the labeled data storage server of the ND station stored by the individual RA ki into the state pool RT ki , and then calculate the fitness value PFit ki of individual RA ki according to formula (1):

Among them, the state pool RT ki stores the allocation status of the ND label data storage server, and RT ki,dt represents the allocation status of the dt-th label data storage server stored in RT ki ; aw1 is the memory penalty factor; aw2 is the energy consumption Penalty factor; wmv is the memory difference; wpe is the excess energy consumption; max is the function of taking the maximum value;

Step 8, set the survival period of each individual in the population SL ki = 1, where the individual subscripts ki = 1, 2, ..., NPS;

Step 9: Find the individual with the smallest fitness value from the population and save it to the optimal individual GRA;

Step 10, calculate the integration coefficient kw according to formula (2):

Step 11: Calculate the bootstrap probability LP ki of each individual in the population according to formula (3):

Among them, APFit ki represents the priority of individual RA ki , and the individual subscript ki=1,2,...,NPS;

Step 12, according to the leading probability LP ki of each individual in the population, adopt the roulette selection method to select the leading individual EA ki from the population;

Step 13, according to formula (4), perform the sine and cosine operation operator based on the guiding individual:

in
r2 is a random real number between [0,2×π], and π is a pi; r3 is a random real number between [0,2]; r4 is a random real number between [0,1]; sin is a sine function ; cos is the cosine function; RU ki is the new individual;

Step 14, calculate the fitness value UFit ki of the new individual RU ki ;

Step 15, if the fitness value of the new individual RU ki is less than the fitness value of the individual RA ki , replace the individual RA ki with the new individual RU ki in the population, otherwise keep the individual RA ki unchanged in the population;

Step 16, update the lifetime SL ki according to formula (5):

Step 17, if SL ki is greater than MSL, randomly generate an individual NA ki , then replace the individual RA ki with the individual NA ki in the population and set SL ki =1, otherwise keep the individual RA ki unchanged in the population;

Step 18: Find the individual with the smallest fitness value from the population and save it to the optimal individual GRA;

Step 19, set the current iteration number it=it+1;

Step 20, if the current iteration it is greater than MIT, go to step 21, otherwise go to step 10;

In step 21, the optimal individual GRA is decoded into the allocation state of the labeled data storage server, that is, the allocation result of the labeled data storage server is obtained.