CN110780938A - Computing task unloading method based on differential evolution in mobile cloud environment - Google Patents

Computing task unloading method based on differential evolution in mobile cloud environment Download PDF

Info

Publication number
CN110780938A
CN110780938A CN201910880150.9A CN201910880150A CN110780938A CN 110780938 A CN110780938 A CN 110780938A CN 201910880150 A CN201910880150 A CN 201910880150A CN 110780938 A CN110780938 A CN 110780938A
Authority
CN
China
Prior art keywords
population
individuals
individual
fitness
scaling factor
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910880150.9A
Other languages
Chinese (zh)
Other versions
CN110780938B (en
Inventor
毛莺池
王瑄
平萍
王龙宝
黄倩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN201910880150.9A priority Critical patent/CN110780938B/en
Publication of CN110780938A publication Critical patent/CN110780938A/en
Application granted granted Critical
Publication of CN110780938B publication Critical patent/CN110780938B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/44—Arrangements for executing specific programs
    • G06F9/445—Program loading or initiating
    • G06F9/44594—Unloading
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Biology (AREA)
  • Physiology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a computing task unloading method based on differential evolution in a mobile cloud environment, which comprises the following steps: converting the inferred calculation process into a task graph, and constructing a task unloading model; measuring the individual similarity of the population to obtain an initialized population with the maximum difference; weighting and fusing population evolution algebra and individual fitness to adjust scaling factors, and selecting a variation strategy according to the scaling factors; generating cross individuals by mixing the dimension components of the target individual and the variant individual, comparing the fitness of the cross individuals with the fitness of the target individual, and keeping the individuals with better fitness to enter the next generation; measuring the aggregation degree of population individuals according to the variance of population fitness, and randomly selecting part of individuals to carry out secondary variation; judging whether the iteration times are met, if so, outputting the codes of the optimal individuals in the population, and otherwise, continuing the iteration; and decoding the codes of the optimal individuals in the population into a task unloading scheme, and outputting the scheme. The algorithm of the invention has strong optimizing capability and can effectively shorten the task response time under the condition of meeting the cost constraint.

Description

Computing task unloading method based on differential evolution in mobile cloud environment
Technical Field
The invention belongs to the field of mobile cloud computing, and particularly relates to a computing task unloading method based on differential evolution in a mobile cloud environment.
Background
With the development of mobile devices and embedded devices, deep learning calculation is required to be applied to the devices. CNN is an important branch of deep learning, and has been widely used in the fields of speech recognition, document analysis, language detection, and image recognition. The CNN generally adopts a cloud computing solution as a computing-intensive network, but this limits an application program that needs real-time response, and may cause problems of privacy disclosure, increased energy consumption overhead, and the like. In recent years, mobile terminal devices with smaller storage space and limited computing capability cannot meet the storage and computing requirements of the CNN model, so that it is also difficult to directly perform inference and computation of the CNN model locally. The mobile cloud computing is a cloud computing technology based on terminal equipment such as a mobile phone, and breaks through resource limitation of a mobile terminal by using cloud storage and computing resources. The mobile cloud computing mainly enhances the data processing capacity of the mobile equipment and reduces the energy consumption of the mobile phone through task unloading. Task unloading refers to sending part or all of tasks on the mobile device to the cloud platform for processing, so that the purposes of reducing computing delay, saving energy consumption, protecting data privacy and the like are achieved. A collaborative inference mode based on a task unloading technology adopts an end cloud collaborative method, a computing task is dynamically deployed between a cloud and an edge, and the method becomes a new direction for CNN computing optimization research in a mobile cloud environment by combining the advantage of strong cloud computing capability and the advantage of low transmission delay of a mobile end. The core idea is to divide the neural network by taking a layer as granularity, wherein part of the layers are used for performing inference calculation at a mobile terminal, and the other part of the layers are unloaded to the cloud. The existing research based on collaborative inference mostly takes minimizing time delay or energy consumption as a single unloading target, neglects the influence of user Qos requirement on task unloading decision, and often cannot meet the requirement of the user.
Disclosure of Invention
The purpose of the invention is as follows: in order to solve the problems in the prior art and solve the problem that the unloading strategy taking the minimum time delay as a single target is difficult to meet the Qos requirement of a user, the invention provides a computing task unloading method based on differential evolution in a mobile cloud environment, which can improve the optimization searching capability of an algorithm, effectively shorten the task response time and adaptively make an unloading decision according to the set cost constraint of a cloud platform and different network speeds.
The technical scheme is as follows: in order to achieve the above object, the present invention provides a method for offloading computing tasks based on differential evolution in a mobile cloud environment, comprising the following steps:
(1) converting the inference calculation process into a task graph, establishing a response time and execution cost model, and establishing a task unloading model for minimizing time delay under cost constraint;
(2) measuring individual similarity of the population by using the weighted hamming distance to obtain an initialized population with the maximum difference;
(3) calculating the fitness of each individual, comprehensively considering the evolution algebra and the fitness of the individual, weighting and fusing two influence factors and dynamically adjusting a scaling factor;
(4) performing binary conversion on the original mutation operation, and selecting a mutation strategy according to a scaling factor;
(5) mixing the target individual and each dimension component of the variant individual to generate crossed individuals; comparing the fitness of the crossed individuals with that of the target individuals, and reserving the individuals with better fitness to enter the next generation by using a greedy strategy;
(6) measuring the aggregation degree of population individuals by using a population fitness variance according to a secondary variation mechanism, if the population fitness is smaller than a threshold value, selecting optimal individuals from a population, randomly selecting partial individuals, and randomly disturbing each dimension component of the individuals;
(7) judging whether the iteration times are met, if so, outputting the codes of the optimal individuals in the population, and otherwise, returning to the step (3) to continue the iteration;
(8) and (5) taking the encoding and decoding of the optimal individual in the population as a task optimal unloading scheme, and outputting the scheme.
Further, the step (2) of initializing the population by weighting the hamming distance specifically comprises the following steps:
defining the weighted hamming distance function as:
Figure BDA0002205627150000021
wherein HD ijRepresenting the weighted hamming distance between individuals i and j,
Figure BDA0002205627150000022
representing the j-th dimension of the i-th individual, a weight being set for each binary bit, a kWeights representing the k-th dimension components of the individual, a k＝2 k(ii) a Randomly generating individuals with 3 times of population scale, selecting one as an initial individual to be added into a population set, calculating the sum of the weighted hamming distances of the rest individuals and all individuals in the population set, adding the individual with the maximum sum value into the set, and continuously iterating until the number of the individuals in the set reaches the population scale to realize the maximum difference of the initial individuals.
Further, the scaling factor in step (3) is adaptively adjusted in the following specific steps:
considering that the variation of the scaling factor along with the population evolution algebra conforms to the Logistic model, the obtained dynamic adjustment formula is as follows:
Figure BDA0002205627150000023
wherein F (t) represents a scaling factor of population evolution to generation t, F minAnd F maxMinimum and maximum values of the scaling factor are represented, respectively, α being the decay rate;
the adjustment of the scaling factor can be refined to an individual level through the individual fitness; the scaling factor formula is adjusted according to the individual fitness as follows:
Figure BDA0002205627150000031
wherein F i(t) represents the scaling factor at the t-th generation of the individual i in the population, f i(t) is the fitness value of the individual i, f b(t) and f w(t) respectively representing the optimal fitness and the worst fitness of individuals in the population of the t generation;
and (3) synthesizing population evolution algebra and considering two factors of individual fitness to obtain a scaling factor weighted self-adaptive adjustment formula:
Figure BDA0002205627150000032
wherein
Figure BDA0002205627150000033
Represents the scaling factor of the individual i in the population after the dynamic adjustment in the t generation, and mu is a weight factor.
Further, the specific steps of binary conversion of the variant policy in step (4) are as follows:
for binary variables, the difference vector can be obtained through XOR operation; the multiplication and addition may be replaced by an and operation and or operation, respectively; the DE/rand/1 mutation strategy formula after binary conversion is as follows:
Figure BDA0002205627150000034
wherein X i(t) denotes the ith individual in the tth generation population, r 1，r 2And r 3Is the randomly selected three individual numbers in the t generation population, H i(t) is the post-mutation vector; the + represents an or operation and the + represents an or operation, the representative and operation is the sum of the values,
Figure BDA0002205627150000036
represents an exclusive or operation; w is a random binary string with each bit W jDetermined by the following equation:
Figure BDA0002205627150000037
further, the specific steps of the secondary variation in the step (6) are as follows:
defining a population fitness variance σ 2Comprises the following steps:
Figure BDA0002205627150000038
wherein f is avg(t) is the average fitness of the population of the t generation; when sigma is 2When the average value of the N & MF variables is smaller than a given threshold epsilon, selecting N & MF individuals and optimal individuals randomly, wherein MF is a random variation proportion, and randomly disturbing each dimension component of the individuals; the random variation formula is:
Figure BDA0002205627150000039
wherein Representing the j dimension component of the ith individual in the t generation population; if the fitness of the individual after random variation is increased, the variation individual is reserved.
Partial tasks are unloaded into the cloud to run in a terminal cloud cooperation mode, and the task response time can be effectively shortened. The problem of making CNN inference task unloading strategy in the mobile cloud environment is a 0-1 integer programming problem with constraint conditions, and the problem is NP-hard problem. When the scale of the problem is large, the solution using conventional mathematical methods is inefficient. The invention provides a computing task unloading method based on differential evolution in a mobile cloud environment, aiming at the defect that the existing computing task unloading method is poor in effect in the mobile cloud environment. The differential evolution algorithm is a global search optimization algorithm for disturbing individuals through differential vectors among population individuals, and is simple in principle, few in control parameters, high in reliability, strong in robustness and good in optimization performance. In the differential evolution algorithm, initial individuals are usually generated by a random function, and due to the randomness generated by the individuals, similar individuals in an initial population are excessive, so that the search range of the algorithm is limited, and the algorithm is easy to fall into local optimum. The magnitude of the scaling factor has an effect on the degree of scaling of the differential variables, with larger scaling factors facilitating the search for potential solutions over a large range, and smaller scaling factors speeding up the convergence of the algorithm. When the algorithm falls into the local optimum, because the individuals in the population tend to be the same, variant individuals are difficult to generate, and the algorithm is difficult to jump out of the local optimum solution.
Has the advantages that: compared with the prior art, the invention has the following advantages:
aiming at the computing task unloading method based on differential evolution in the mobile cloud environment, the invention establishes a computing task unloading scheme, designs a binary differential evolution algorithm of weighted adaptive variation and random quadratic variation, solves the problem of the optimal unloading method and effectively shortens the task response time.
Drawings
FIG. 1 is a computing task offload application scenario in a particular embodiment;
FIG. 2 is a flow chart of the present invention;
FIG. 3 is a diagram illustrating an example of a method for offloading computing tasks based on differential evolution in a mobile cloud environment in an exemplary embodiment;
fig. 4 is a diagram illustrating an example of individual codes in a computing task offloading method based on differential evolution in a mobile cloud environment in an embodiment.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
The invention discloses a computing task unloading method based on differential evolution in a mobile cloud environment. A calculation task unloading scheme is established, a binary differential evolution algorithm of weighted self-adaptive variation and random quadratic variation is designed, the problem of the optimal unloading method is solved, and the task response time is effectively shortened.
FIG. 1 is a computing task offload application scenario of the present invention.
When a user submits computing tasks such as CNN image recognition, a computing task unloading strategy can be formulated in an end cloud cooperation mode, a neural network is segmented by taking a layer as granularity, the computing tasks are unloaded to run in the local and cloud, a part of layers carry out inference calculation at a mobile end, and the other part of layers are unloaded to the cloud. The computing tasks are executed through cooperation of the end clouds, and the total time of task response can be effectively achieved.
FIG. 2 is a flowchart of a computing task offloading method based on differential evolution in a mobile cloud environment.
Step A: and converting the inference calculation process into a task graph, establishing a response time and execution cost model, and constructing a task unloading model for minimizing time delay under cost constraint. Each subtask in the task graph has two choices in scheduling, and the subtask is operated at a mobile terminal or unloaded to a micro cloud for operation. The response time of the CNN inference task includes run time and transfer time. And obtaining the running time of the subtask through a running time prediction module. If the execution positions of the subtasks i and j are the same, the transmission time is 0; if the execution positions are different, the transmission time is as follows:
t ij＝d ij/v
wherein d is ijRepresenting the amount of data transfer between subtasks i and j, and v representing the transfer speed between the mobile device and the cloud. In addition, the time required for the input data to be transmitted from the mobile terminal to the cloud is d oV, time d required for output data volume to be transmitted from cloud to mobile terminal n/v。
Set predecessor task set of subtask i as
Figure BDA0002205627150000051
m iIs the number of predecessor tasks. The subtask i can only start running after all predecessor tasks are completed, so the start time st of the subtask i iComprises the following steps:
st i＝max{ft k+t ki(x i-x k) 2，k∈J i}
wherein t is kiRepresenting the transmission time, x, between subtasks k and i iIndicating the offloading decision of subtask i. Get the completion time ft of subtask i iComprises the following steps:
Figure BDA0002205627150000052
wherein
Figure BDA0002205627150000053
And
Figure BDA0002205627150000054
the execution time of the subtask i on the mobile terminal and the cloud is respectively.
To calculate the completion time of the entire task graph, the calculation may be started from the start subtask until the completion time ft of the last subtask is obtained n. The completion time for obtaining the whole task graph is as follows:
T(X)＝ft n+x nd n/v
the user unloads the task into the micro cloud to operate, and certain cost needs to be paid. If the computing cost of the cloud unit time is price, the total cost of the task running in the cloud can be represented as:
a task unloading scheme X is formulated, so that on the premise of meeting cost constraints, the task response time is shortened as much as possible, and an objective function is as follows:
minT(X)
C(X)≤C max
wherein C is maxFor the maximum acceptable cost.
And B: and measuring the individual similarity of the population by using the weighted hamming distance to obtain the initialized population with the maximum difference. Each individual is binary coded and each individual in the population represents a task offloading scheme. 0 represents that the subtask runs in the mobile side, and 1 represents that the subtask runs in the cloud side.
Fig. 3 illustrates an example of task offloading for end cloud collaboration.
FIG. 3 is a task diagram after conversion of the inference calculation process of the inclusion v1 modelThere are 9 subtasks. The black circles represent that the subtasks run on the mobile side, and the white circles represent that the subtasks run on the cloud. As can be seen in FIG. 3, subtask 1 runs on the cloud, x 11 is ═ 1; subtask 2 runs on the cloud, x 21 is ═ 1; subtask 3 runs on the mobile side, x 30; the subtask 4 runs on the mobile side, x 40; subtask 5 runs on the mobile side, x 50; subtask 6 runs on the cloud, x 61 is ═ 1; the subtask 7 runs on the mobile side, x 70; the subtask 8 runs on the mobile side, x 80; the subtask 9 runs on the mobile side, x 9＝0。
Fig. 4 is an individual code corresponding to the task offloading method.
Because the initial individuals of the standard differential evolution algorithm are generated by the random function, the similar individuals in the initial population are possibly too many, the search range of the algorithm is limited, and the algorithm is easy to fall into local optimum. In generating the initial population, it is preferable to have the individuals distributed as uniformly as possible in the solution space, defining a weighted hamming distance function as:
Figure BDA0002205627150000061
wherein HD ijRepresenting the weighted hamming distance between individuals i and j,
Figure BDA0002205627150000062
representing the j-th dimension of the i-th individual, a weight being set for each binary bit, a kWeights representing the k-th dimension components of the individual, a k＝2 k. Randomly generating individuals with 3 times of population scale, selecting one as an initial individual to be added into a population set, calculating the sum of the weighted hamming distances of the rest individuals and all individuals in the population set, adding the individual with the maximum sum value into the set, and continuously iterating until the number of the individuals in the set reaches the population scale to realize the maximum difference of the initial individuals.
And C: calculating the fitness of each individual, comprehensively considering the evolution algebra and the fitness of the individual, weighting and fusing two influence factors and dynamically adjusting the scaling factor. Considering that the variation of the scaling factor along with the population evolution algebra conforms to the Logistic model, the obtained dynamic adjustment formula is as follows:
Figure BDA0002205627150000071
wherein F (t) represents a scaling factor of population evolution to generation t, F minAnd F maxRepresenting the minimum and maximum values of the scaling factor, respectively, α being the decay rate.
The adjustment of the scaling factor can be refined to individual level by individual fitness. The scaling factor formula is adjusted according to the individual fitness as follows:
Figure BDA0002205627150000072
wherein F i(t) represents the scaling factor at the t-th generation of the individual i in the population, f i(t) is the fitness value of the individual i, f b(t) and f w(t) respectively representing the best fitness and the worst fitness of individuals in the population of the t generation.
And (3) synthesizing population evolution algebra and considering two factors of individual fitness to obtain a scaling factor weighted self-adaptive adjustment formula:
Figure BDA0002205627150000073
wherein
Figure BDA0002205627150000074
Represents the scaling factor of the individual i in the population after the dynamic adjustment in the t generation, and mu is a weight factor.
Step D: and carrying out binary conversion on the original mutation operation, and selecting a mutation strategy according to the scaling factor. The method specifically comprises the following steps:
and D1, performing binary conversion on the mutation strategy. For binary variables, the difference vector can be obtained through XOR operation; the multiplication and addition may be replaced by an and operation and an or operation, respectively. The DE/rand/1 mutation strategy formula after binary conversion is as follows:
Figure BDA0002205627150000075
wherein X i(t) denotes the ith individual in the tth generation population, r 1，r 2And r 3Is the randomly selected three individual numbers in the t generation population, H iAnd (t) is a post-mutation vector. The + represents an or operation and the + represents an or operation,
Figure BDA0002205627150000076
the representative and operation is the sum of the values,
Figure BDA0002205627150000077
representing an exclusive or operation. W is a random binary string with each bit W jDetermined by the following equation:
Figure BDA0002205627150000078
step D2: the mutation strategy is selected according to the scaling factor. The final differential variation formula is as follows:
Figure BDA0002205627150000081
wherein X best(t) represents the best individual in the current population.
Step E: mixing the target individual and each dimension component of the variant individual to generate crossed individuals; and comparing the fitness of the crossed individuals with that of the target individuals, and reserving the individuals with better fitness to enter the next generation by using a greedy strategy. The method specifically comprises the following steps:
step E1: and mixing the target individual and the variable individual with each dimension component to generate cross individuals. The formula for the crossover operation is:
Figure BDA0002205627150000083
wherein CR is a cross factor and has a value range of [0,1 ]]。j randIs [1, D ]]Random integers within the interval are used to ensure that at least one-dimensional component of the crossover entities is from the variant entity.
Step E2: and comparing the fitness of the crossed individuals with that of the target individuals, and reserving the individuals with better fitness to enter the next generation by using a greedy strategy. For the minimization problem, the formula for the selection operation is:
Figure BDA0002205627150000084
wherein f (U) i(t)) and f (X) i(t)) respectively represent the fitness of the crossover individual and the target individual.
Step F: and measuring the aggregation degree of the population individuals by using the variance of the population fitness according to a secondary variation mechanism, if the population fitness is smaller than a threshold value, selecting the optimal individuals from the population, randomly selecting part of individuals, and randomly disturbing each dimension component of the individuals. Defining a population fitness variance σ 2Comprises the following steps:
Figure BDA0002205627150000085
wherein f is avgAnd (t) is the average fitness of the population of the t generation. When sigma is 2And when the average value is less than a given threshold epsilon, randomly selecting N & MF individuals (MF is a random variation proportion) and the optimal individual, and randomly disturbing each dimension component of the individuals. The random variation formula is:
Figure BDA0002205627150000091
wherein Representing the j dimension component of the ith individual in the t generation population. If the fitness of the individual after random variation is increased, the variation individual is reserved.
Step G: judging whether the iteration times are met, if so, outputting the codes of the optimal individuals in the population, and otherwise, returning to the step (3) to continue the iteration;
step H: and (5) taking the encoding and decoding of the optimal individual in the population as a task optimal unloading scheme, and outputting the scheme. Each individual is binary coded and each individual in the population represents a task offloading scheme. 0 represents that the subtask runs in the mobile side, and 1 represents that the subtask runs in the cloud side.

Claims (5)

1. A computing task unloading method based on differential evolution in a mobile cloud environment is characterized by comprising the following steps:
(1) converting the inference calculation process into a task graph, establishing a response time and execution cost model, and establishing a task unloading model for minimizing time delay under cost constraint;
(2) measuring individual similarity of the population by using the weighted hamming distance to obtain an initialized population with the maximum difference;
(3) calculating the fitness of each individual, comprehensively considering the evolution algebra and the fitness of the individual, weighting and fusing two influence factors and dynamically adjusting a scaling factor;
(4) performing binary conversion on the original mutation operation, and selecting a mutation strategy according to a scaling factor;
(5) mixing the target individual and each dimension component of the variant individual to generate crossed individuals; comparing the fitness of the crossed individuals with that of the target individuals, and reserving the individuals with better fitness to enter the next generation by using a greedy strategy;
(6) measuring the aggregation degree of population individuals by using a population fitness variance according to a secondary variation mechanism, if the population fitness is smaller than a threshold value, selecting optimal individuals from a population, randomly selecting partial individuals, and randomly disturbing each dimension component of the individuals;
(7) judging whether the iteration times are met, if so, outputting the codes of the optimal individuals in the population, and otherwise, returning to the step (3) to continue the iteration;
(8) and (5) taking the encoding and decoding of the optimal individual in the population as a task optimal unloading scheme, and outputting the scheme.
2. The method for offloading computing tasks based on differential evolution in a mobile cloud environment according to claim 1, wherein the step (2) of initializing the population by weighting hamming distance specifically comprises the following steps:
defining the weighted hamming distance function as:
Figure FDA0002205627140000011
wherein HD ijRepresenting the weighted hamming distance between individuals i and j,
Figure FDA0002205627140000012
representing the j-th dimension of the i-th individual, a weight being set for each binary bit, a kWeights representing the k-th dimension components of the individual, a k=2 k(ii) a Randomly generating individuals with 3 times of population scale, selecting one as an initial individual to be added into a population set, calculating the sum of the weighted hamming distances of the rest individuals and all individuals in the population set, adding the individual with the maximum sum value into the set, and continuously iterating until the number of the individuals in the set reaches the population scale to realize the maximum difference of the initial individuals.
3. The differential evolution-based computing task offloading method in a mobile cloud environment according to claim 1, wherein the scaling factor in step (3) is adaptively adjusted according to the following specific steps:
considering that the variation of the scaling factor along with the population evolution algebra conforms to the Logistic model, the obtained dynamic adjustment formula is as follows:
Figure FDA0002205627140000021
wherein F (t) represents a scaling factor of population evolution to generation t, F minAnd F maxMinimum and maximum values of the scaling factor are represented, respectively, α being the decay rate;
the adjustment of the scaling factor can be refined to an individual level through the individual fitness; the scaling factor formula is adjusted according to the individual fitness as follows:
wherein F i(t) represents the scaling factor at the t-th generation of the individual i in the population, f i(t) is the fitness value of the individual i, f b(t) and f w(t) respectively representing the optimal fitness and the worst fitness of individuals in the population of the t generation;
and (3) synthesizing population evolution algebra and considering two factors of individual fitness to obtain a scaling factor weighted self-adaptive adjustment formula:
Figure FDA0002205627140000027
wherein
Figure FDA0002205627140000028
Represents the scaling factor of the individual i in the population after the dynamic adjustment in the t generation, and mu is a weight factor.
4. The differential evolution-based computing task offloading method in a mobile cloud environment according to claim 1, wherein the specific steps of binary conversion of the variance policy in step (4) are as follows:
for binary variables, the difference vector can be obtained through XOR operation; the multiplication and addition may be replaced by an and operation and or operation, respectively; the DE/rand/l mutation strategy formula after binary conversion is as follows:
Figure FDA0002205627140000023
wherein X i(t) denotes the ith individual in the tth generation population, r 1，r 2And r 3Is the randomly selected three individual numbers in the t generation population, H i(t) isA post-mutation vector; the + represents an or operation and the + represents an or operation,
Figure FDA0002205627140000024
the representative and operation is the sum of the values,
Figure FDA0002205627140000025
represents an exclusive or operation; w is a random binary string with each bit W jDetermined by the following equation:
Figure FDA0002205627140000026
5. the differential evolution-based computing task offloading method in a mobile cloud environment according to claim 1, wherein the specific steps of the quadratic variation in the step (6) are as follows:
defining a population fitness variance σ 2Comprises the following steps:
Figure FDA0002205627140000031
wherein f is avg(t) is the average fitness of the population of the t generation; when sigma is 2When the average value of the N & MF variables is smaller than a given threshold epsilon, selecting N & MF individuals and optimal individuals randomly, wherein MF is a random variation proportion, and randomly disturbing each dimension component of the individuals; the random variation formula is:
Figure FDA0002205627140000032
wherein
Figure FDA0002205627140000033
Representing the j dimension component of the ith individual in the t generation population; if the fitness of the individual after random variation is increased, the variation individual is reserved.
CN201910880150.9A 2019-09-18 2019-09-18 Computing task unloading method based on differential evolution in mobile cloud environment Active CN110780938B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910880150.9A CN110780938B (en) 2019-09-18 2019-09-18 Computing task unloading method based on differential evolution in mobile cloud environment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910880150.9A CN110780938B (en) 2019-09-18 2019-09-18 Computing task unloading method based on differential evolution in mobile cloud environment

Publications (2)

Publication Number Publication Date
CN110780938A true CN110780938A (en) 2020-02-11
CN110780938B CN110780938B (en) 2021-02-09

Family

ID=69383799

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910880150.9A Active CN110780938B (en) 2019-09-18 2019-09-18 Computing task unloading method based on differential evolution in mobile cloud environment

Country Status (1)

Country Link
CN (1) CN110780938B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112083967A (en) * 2020-08-18 2020-12-15 深圳供电局有限公司 Unloading method of cloud edge computing task, computer equipment and storage medium
CN112200495A (en) * 2020-11-12 2021-01-08 同济大学 Power system dynamic scheduling method based on improved differential evolution algorithm
CN113347277A (en) * 2021-07-15 2021-09-03 湘潭大学 Unloading distribution method based on task segmentation in edge calculation
CN113610116A (en) * 2021-07-14 2021-11-05 上海工程技术大学 Fault diagnosis method for adaptive differential evolution algorithm optimized support vector machine
CN114051266A (en) * 2021-11-08 2022-02-15 首都师范大学 Wireless body area network task unloading method based on mobile cloud-edge computing
CN114143814A (en) * 2021-12-13 2022-03-04 华北电力大学（保定） Multitask unloading method and system based on heterogeneous edge cloud architecture

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104484756A (en) * 2014-12-15 2015-04-01 国家电网公司 Method for improving enterprise energy efficiency of wind-and-light power storage and generation microgrid
CN108564592A (en) * 2018-03-05 2018-09-21 华侨大学 Based on a variety of image partition methods for being clustered to differential evolution algorithm of dynamic
CN110087257A (en) * 2019-04-24 2019-08-02 重庆邮电大学 A kind of task discharge mechanism and method for supporting mobile edge calculations

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104484756A (en) * 2014-12-15 2015-04-01 国家电网公司 Method for improving enterprise energy efficiency of wind-and-light power storage and generation microgrid
CN108564592A (en) * 2018-03-05 2018-09-21 华侨大学 Based on a variety of image partition methods for being clustered to differential evolution algorithm of dynamic
CN110087257A (en) * 2019-04-24 2019-08-02 重庆邮电大学 A kind of task discharge mechanism and method for supporting mobile edge calculations

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
IVANOE DE FALCO 等: "《Mapping of time-consuming multitask applications on a cloud system by multiobjective Differential Evolution》", 《PARALLEL COMPUTING》 *
吴亮红 等: "《自适应二次变异差分进化算法》", 《控制与决策》 *
林涛 等: "《基于改进差分进化算法的云计算任务调度策略》", 《传感器与微系统》 *
陈华 等: "《基于logistic模型的自适应差分进化算法》", 《控制与决策》 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112083967A (en) * 2020-08-18 2020-12-15 深圳供电局有限公司 Unloading method of cloud edge computing task, computer equipment and storage medium
CN112083967B (en) * 2020-08-18 2023-10-20 深圳供电局有限公司 Cloud edge computing task unloading method, computer equipment and storage medium
CN112200495A (en) * 2020-11-12 2021-01-08 同济大学 Power system dynamic scheduling method based on improved differential evolution algorithm
CN112200495B (en) * 2020-11-12 2023-05-02 同济大学 Dynamic scheduling method of power system based on improved differential evolution algorithm
CN113610116A (en) * 2021-07-14 2021-11-05 上海工程技术大学 Fault diagnosis method for adaptive differential evolution algorithm optimized support vector machine
CN113610116B (en) * 2021-07-14 2023-07-04 上海工程技术大学 Fault diagnosis method for optimizing support vector machine by self-adaptive differential evolution algorithm
CN113347277A (en) * 2021-07-15 2021-09-03 湘潭大学 Unloading distribution method based on task segmentation in edge calculation
CN114051266A (en) * 2021-11-08 2022-02-15 首都师范大学 Wireless body area network task unloading method based on mobile cloud-edge computing
CN114051266B (en) * 2021-11-08 2024-01-12 首都师范大学 Wireless body area network task unloading method based on mobile cloud-edge calculation
CN114143814A (en) * 2021-12-13 2022-03-04 华北电力大学（保定） Multitask unloading method and system based on heterogeneous edge cloud architecture
CN114143814B (en) * 2021-12-13 2024-01-23 华北电力大学（保定） Multi-task unloading method and system based on heterogeneous edge cloud architecture

Also Published As

Publication number Publication date
CN110780938B (en) 2021-02-09

Similar Documents

Publication Publication Date Title
CN110780938B (en) Computing task unloading method based on differential evolution in mobile cloud environment
CN113242568B (en) Task unloading and resource allocation method in uncertain network environment
US20230116117A1 (en) Federated learning method and apparatus, and chip
US20240135191A1 (en) Method, apparatus, and system for generating neural network model, device, medium, and program product
CN110531996B (en) Particle swarm optimization-based computing task unloading method in multi-micro cloud environment
Zhang et al. DRL-driven dynamic resource allocation for task-oriented semantic communication
CN113537365B (en) Information entropy dynamic weighting-based multi-task learning self-adaptive balancing method
CN109787696B (en) Cognitive radio resource allocation method based on case reasoning and cooperative Q learning
TW202141363A (en) Adaptive quantization for execution of machine learning models
CN113573363A (en) MEC calculation unloading and resource allocation method based on deep reinforcement learning
CN113051130A (en) Mobile cloud load prediction method and system of LSTM network combined with attention mechanism
CN113610227A (en) Efficient deep convolutional neural network pruning method
CN113946423A (en) Multi-task edge computing scheduling optimization method based on graph attention network
CN116009990B (en) Cloud edge collaborative element reinforcement learning computing unloading method based on wide attention mechanism
CN117202264A (en) 5G network slice oriented computing and unloading method in MEC environment
CN115499511B (en) Micro-service active scaling method based on space-time diagram neural network load prediction
CN116976461A (en) Federal learning method, apparatus, device and medium
CN115914230A (en) Adaptive mobile edge computing unloading and resource allocation method
CN113157453B (en) Task complexity-based high-energy-efficiency target detection task dynamic scheduling method
Chen et al. DWFed: A statistical-heterogeneity-based dynamic weighted model aggregation algorithm for federated learning
Huang et al. Latency guaranteed edge inference via dynamic compression ratio selection
CN113747500A (en) High-energy-efficiency low-delay task unloading method based on generation countermeasure network in mobile edge computing environment
CN114648021A (en) Question-answering model training method, question-answering method and device, equipment and storage medium
CN113570036A (en) Hardware accelerator architecture supporting dynamic neural network sparse model
Zhen et al. A Secure and Effective Energy-Aware Fixed-Point Quantization Scheme for Asynchronous Federated Learning.

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant