A cloud task model approach for high-precision artificial neural network optimization algorithm

Abstract

The present invention proposes a cloud task model method of a high-precision artificial neural network optimization algorithm, including: step 1: initialization; step 2: after the initialization is completed, start the optimal solution search process based on SCVNNA; step 3: the global exploration and local development behavior of the individual will depend on the value of the current difference adjustment operator z, if z is greater than a random number between 0 and 1, then the difference adjustment population will be executed to update their positions; step 4: once the maximum number of iterations is reached, the search process stops, the position of the individual is converted into a decision variable, and used as the optimal cloud computing task scheduling scheme. The present invention has significant advantages, including improving the accuracy and computational efficiency of task scheduling, and significantly reducing the idleness and waste of resources. At the same time, the scheme improves the robustness and optimization effect of the system by combining global exploration with local optimization, and is suitable for efficient scheduling and management of large-scale cloud computing tasks.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud task optimization scheduling, and in particular to a cloud task model method of a high-precision artificial neural network optimization algorithm. Background Art

[0002] As the number of cloud users continues to increase, it is imperative to provide users with high-quality services. Therefore, how to implement a reasonable and efficient task scheduling strategy under constraints is a huge challenge facing current cloud computing technology. Each task may have different requirements in terms of resource bandwidth, storage capacity, resource cost, and response time, making the optimization problem particularly complex. In addition, the heterogeneity and dynamics in the cloud computing environment further exacerbate the complexity of these problems. Effective task scheduling is a necessary condition for achieving optimal user service quality. Summary of the invention

[0003] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a cloud task model method of a high-precision artificial neural network optimization algorithm, including:

[0004] Step 1: Initialization, including establishing the mapping between cloud computing tasks and neural network individuals, and initializing individuals and weights; at the same time, initializing the dimension of the search space, the upper and lower bounds of the population, and the maximum number of iterations;

[0005] Step 2: After initialization, the optimal solution search process based on SCVNNA begins; at this stage, the fitness value of the individual is calculated based on its position information, and the individual with the minimum fitness value will be regarded as the current optimal solution and recorded. At the same time, sparse connection is performed to adjust the weight;

[0006] Step 3: The global exploration and local exploitation behaviors of individuals will depend on the value of the current difference adjustment operator z. If z is greater than a random number between 0 and 1, the difference adjustment population will be executed to update their positions. Otherwise, the chaotic transfer function operator will be executed to update the positions.

[0007] Step 4: Once the maximum number of iterations is reached, the search process stops and the individual positions are converted into decision variables and used as the optimal cloud computing task scheduling solution.

[0008] In a preferred embodiment of the present invention, the step 1 comprises:

[0009] Compute Node and a set of computational tasks ,in, , the final scheduling result can be expressed by the matrix It is expressed as follows:

[0010] ;

[0011] in, is a decision variable, if When, it means The task is in Otherwise, ,During the scheduling process, tasks will not be assigned repeatedly, that is, each task can only be assigned to one virtual machine for execution; at the same time, each virtual machine will be assigned at least one task to ensure that all virtual machines participate in task execution;

[0012] The cost is used as the objective function to represent the resource cost. is the price adjustment factor, and the expression of the cost function is:

[0013] ,

[0014] ,

[0015] ,

[0016] in, is the time cost function, is the load capacity function, is the price cost function;

[0017] is the decision variable, Indicates the total number of virtual machines in the collection. Represents the total number of task sets, is the price adjustment factor, which is used to adjust the proportional relationship between resource bandwidth and economic cost;

[0018] It is The load capacity of each task, It is the smallest load capacity among all tasks. Indicates The minimum load capacity among the resource nodes is used as the normalization benchmark;

[0019] Indicates The cost vector value in each task, Indicates The cost vector value in each virtual machine; Indicates The time cost value in each task, Indicates The time cost value in a virtual machine, Indicates time The minimum execution time of all tasks in the

[0020] Since the results represented by the three matrices of resource nodes and task nodes have different standards, they need to be normalized, so the above three objective functions are expressed as follows:

[0021] ,

[0022] ,

[0023] ,

[0024] in, is the normalized time cost function, which is used to evaluate the time cost of task execution. is the normalized load capacity function, which is used to evaluate the load cost of the system. is the resource cost function, which is used to evaluate the economic cost of resource use;

[0025] Indicates that among all tasks and virtual machines The maximum value of

[0026] Indicates that among all tasks and virtual machines The maximum value of

[0027] Is taken in all tasks and virtual machines The maximum value of

[0028] Different cloud computing systems or computing resource providers may have different requirements for task execution performance. , and are the weights of time cost, load capacity and cost function respectively, the smallest It is the optimal solution for cloud computing task scheduling:

[0029] ,

[0030] represents the optimal objective function value of cloud computing task scheduling, It means to find the weighted minimum value, which is minimized by weighted combination of multiple cost functions.

[0031] In a preferred embodiment of the present invention, the step 1 further comprises:

[0032] Initial population , indicating that the algorithm is At the iteration, the set of individuals in the population is is the population size, and the update is through the weight matrix , for any number of iterations , It is Individual vectors, It is weight vectors, each have Dimension, where ,therefore:

[0033] ,

[0034] in, Indicates The individual in At the iteration Solutions in dimensional space;

[0035] After creating the initial weights, the weight matrix in Eq. Select the solution corresponding to the target The weight of , therefore, the sum of the weight matrices satisfies the following conditions:

[0036] ,

[0037] in, It is The individual in The weight values ​​on the dimensions are taken as weight vectors An element in

[0038] ,

[0039] in, express is a random number uniformly distributed between 0 and 1. is the population size.

[0040] In a preferred embodiment of the present invention, the step 2 comprises:

[0041] ,

[0042] , ,

[0043] in Including 10% , is the number of iterations When , the sparse weight matrix, in which some weight values ​​are set to zero, is equivalent to them not participating in the crossover process, which improves the efficiency of individuals in the crossover process and their ability to jump out of the local optimum, increases the diversity of the solution space, and thus ensures the superiority of the entire group. Therefore, the introduction of the sparse connection strategy can jump out of the local optimal solution;

[0044] exist The formula for generating a new population in iterations can be expressed as:

[0045] ,

[0046] in Indicates the population size; Indicates the current iteration number; Indicates that the solution is At the iteration, The weighted solution of each individual;

[0047] Update the weight vector according to the following formula:

[0048] ,

[0049] in, Indicated in At the iteration, The weight vector of each individual, Indicated in At the iteration, The weight vector of each individual, For the The target weight vector obtained at the iteration is, Represents a random number.

[0050] In a preferred embodiment of the present invention, step 3 comprises:

[0051] By integrating the Tent chaos operator with the sine-cosine strategy, a new Tent sine-cosine chaos operator, namely the Tent difference adjustment operator, is proposed to generate chaotic candidate solutions. The calculation formula is as follows:

[0052] ,

[0053] Indicated in The value after iterations is Indicates the number of mappings; Indicates The value after iterations; express A random number between ; Indicates the current iteration number, Indicates the maximum number of iterations; and To control the constant;

[0054] The difference-adjusted population is used to replace the biased population, and more information is introduced to adjust the randomness. The calculated new individual position is affected by the target position and the existing individual position, and the individual position is adjusted more specifically, which helps to improve the search efficiency and convergence ability of the algorithm. Therefore, the difference-adjusted population is defined as follows:

[0055] ,

[0056] in, It is The individual in At the iteration, in dimension By adjusting the scaling factor and mapping constants You can reach different locations around the current optimal solution, mapping constants is a random number between (0,2), It is The target solution at the iteration is Indicates The individual in At the iteration The solution in dimensional space, Indicates the total number of individuals in a population; by reducing The value of to achieve the difference adjustment, so the scaling factor Define the following formula:

[0057] ,

[0058] in, express A random number between is a new nonlinear convergence factor, and its formula is as follows:

[0059] ,

[0060] in, Indicates the current iteration number, represents the maximum number of iterations, To control the constant.

[0061] In a preferred embodiment of the present invention, step 4 comprises:

[0062] The Circle mapping operator is introduced into the transfer function operator, and a chaotic transfer function operator method is used. The operator transfers the new mode solution of the current position to the target solution. The new position in the closer search space is defined as follows:

[0063] ,

[0064] in, Indicates The individual in The solution at the iteration

[0065] Indicates The individual in At the iteration Solutions in dimensional space;

[0066] Indicates the current iteration number;

[0067] is a random number between 0 and 1, which follows a uniform distribution;

[0068] Represents the Circle mapping operator;

[0069] It is The target solution at the iteration;

[0070] represents the total number of individuals in the population, The calculation formula is:

[0071] ,

[0072] in, Represents the Circle mapping operator;

[0073] It is The parameter value at the iteration;

[0074] and is a mapping constant;

[0075] is the modulo operator.

[0076] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0077] 1. Improve computational efficiency

[0078] By deploying the artificial neural network optimization algorithm in the cloud task model, the present invention can utilize the high parallelism and scalability of cloud computing resources to significantly improve computing efficiency. Distributed computing in the cloud can effectively decompose complex optimization tasks and process multiple individuals or different neural network levels in parallel, thereby accelerating convergence and shortening computing time.

[0079] 2. Enhanced optimization accuracy

[0080] This method achieves high-precision control of the neural network optimization process through sparse connection strategy, dynamic weight matrix update and other technologies. The adaptive adjustment of the weight matrix combined with task relevance analysis makes the optimization process more refined, reduces the possibility of overfitting and underfitting, and thus improves the generalization ability of the neural network.

[0081] 3. Adaptive task allocation and load balancing

[0082] The cloud task model can adaptively allocate computing resources according to the complexity and resource requirements of different tasks to ensure balanced distribution of task loads. This resource allocation strategy can not only improve the execution efficiency of tasks, but also optimize the utilization of cloud computing resources and avoid resource waste.

[0083] 4. High robustness and fault tolerance

[0084] Since the optimization algorithm runs in the cloud, the system has higher robustness. Even in the case of a single point failure or network interruption, the task can continue to run on other nodes, ensuring the continuity of the entire optimization process. In addition, the multi-copy backup mechanism in the cloud computing environment improves data security and fault tolerance.

[0085] 5. Support multi-task collaborative optimization

[0086] The present invention can process multiple different optimization tasks simultaneously in a cloud environment and share some computing resources or intermediate results through a collaborative optimization mechanism. Multi-task collaborative optimization can effectively improve the utilization efficiency of computing resources. At the same time, when dealing with complex multi-task problems, they can learn from each other and optimize to enhance the overall optimization effect.

[0087] 6. Dynamic Adjustment and Scalability

[0088] The cloud task model method supports dynamic adjustment of optimization strategies and resource allocation according to task complexity, and adapts to neural network optimization tasks of different scales and complexities. By dynamically adjusting the number of cloud computing nodes and task allocation strategies, the present invention can flexibly expand and improve adaptability when facing large-scale neural network optimization.

[0089] 7. Reduce costs and energy consumption

[0090] Due to the on-demand use of cloud computing resources, users can dynamically apply for and release resources according to task requirements. This flexibility not only reduces computing costs, but also reduces the waste of computing resources, thereby reducing energy consumption and improving overall green computing capabilities.

[0091] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0093] Figure 1 It is a schematic diagram of the process of the present invention.

[0094] Figure 2 is the best total cost obtained by increasing the number of iterations over 100 tasks.

[0095] Figure 3 is the best time spent on 100 tasks by increasing the number of iterations.

[0096] Figure 4 is the best price achieved by increasing the number of iterations over 100 tasks.

[0097] Figure 5 is the optimal load cost obtained by increasing the number of iterations among 100 tasks.

[0098] Figure 6 is the best total cost achieved by each method in a small-scale test.

[0099] Figure 7 is the best time achieved by each method in a small-scale test.

[0100] Figure 8 is the best price achieved for each method in a small-scale test.

[0101] Fig. 9 is the optimal load cost achieved by each method in a small-scale test.

[0102] Fig.10 is the best total cost obtained by increasing the number of iterations over 1000 tasks.

[0103] Fig.11 is the best time spent by increasing the number of iterations among 1000 tasks.

[0104] Fig.12is the best price achieved by increasing the number of iterations over 1000 tasks.

[0105] Fig.13 is the optimal load cost obtained by increasing the number of iterations in 1000 tasks.

[0106] Fig.14 is the best total cost achieved by each method in large-scale testing.

[0107] Fig.15 is the best time cost achieved by each method in large-scale testing.

[0108] Fig.16 is the best price achieved for each method in large-scale testing.

[0109] Fig.17 is the best load cost achieved by each method in large-scale testing. DETAILED DESCRIPTION

[0110] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0111] 1. Cloud computing task scheduling optimization problem

[0112] In cloud computing, task scheduling strategies directly affect the utilization efficiency of system resources. Therefore, how to effectively allocate user-submitted jobs to computing resources (such as virtual machines) becomes the core issue of cloud task scheduling. This patent application focuses on the performance of different scheduling methods, assuming that all tasks are logically independent of each other. Therefore, in a cloud environment, task scheduling can be summarized into three steps:

[0113] 1. Map tasks to resources using specific strategies based on information about tasks and available computing resources;

[0114] 2. The scheduler generates an optimized execution plan to meet the demand;

[0115] 3. The execution plan is passed to the underlying task processing layer for execution, and the execution results are fed back to the user.

[0116] 1.1. Modeling of cloud computing tasks and computing resources

[0117] In order to describe the optimization process of the scheduling model in the cloud computing environment in detail, the following method is adopted: and a set of computational tasks ,in, , the final scheduling result can be expressed by the matrix It is expressed as follows:

[0118] (1)

[0119] in, is a decision variable, if When, it means The task is in Otherwise, ,During the scheduling process, tasks will not be assigned repeatedly, that is, each task can only be assigned to one virtual machine for execution; at the same time, each virtual machine will be assigned at least one task to ensure that all virtual machines participate in task execution;

[0120] In the context of task scheduling, in order to evaluate the overall processing capacity and resource utilization of the cloud computing system, we use three attributes to describe each resource node: time cost (i.e., the time cost of processing capacity), load capacity (i.e., the response to processing load), and cost (i.e., the cost of processing resource bandwidth). Time cost is expressed in terms of the computing power of the CPU, and load capacity is based on the node memory size. The bandwidth of a resource node is affected by the CPU power and memory size, that is, the larger the CPU power and memory size, the higher the bandwidth of the node resources. In other words, the memory resources of a node (measured in megabytes) depend on the values ​​of time cost, load capacity, and cost.

[0121] In order to describe the general processing power and resource consumption of cloud computing systems in task scheduling scenarios, three attributes (time cost, load capacity, and cost price) are used to represent each resource node. The underlying computing system is modeled as three vectors. That is, the time cost vector, the load capacity vector, and the cost price vector. At the same time, each task corresponds to three vectors, namely the time cost vector, the load capacity vector, and the cost price vector.

[0122] Multi-objective scheduling optimization model

[0123] In order to effectively process a given set of tasks, the underlying computing system is expected to be able to process the given set of tasks efficiently and energy-efficiently, while excelling in terms of performance and resource consumption. Specifically, the system should efficiently utilize CPU power and memory while minimizing the overall resource usage cost. Obviously, the experiment aims to solve the multi-objective task scheduling optimization problem in cloud computing systems, where time cost, load capacity and resource bandwidth are set as three optimization objectives. Similar to the model with constraints on CPU and memory, the time cost function is expressed by formula (2), the load capacity function is expressed by equation (3), and the resource bandwidth can be used to calculate indicators such as energy consumption and economic cost, so the resource cost can be further expressed by these indicators. This patent application uses cost as the objective function to express resource cost. is the price adjustment factor. The cost function is expressed by equation (4).

[0124] (2)

[0125] (3)

[0126] (4)

[0127] in, is the time cost function, is the load capacity function, is the price cost function;

[0128] is the decision variable, Indicates the total number of virtual machines in the collection. Represents the total number of task sets, is the price adjustment factor, which is used to adjust the proportional relationship between resource bandwidth and economic cost;

[0129] It is The load capacity of each task, It is the smallest load capacity among all tasks. Indicates The minimum load capacity among the resource nodes is used as the normalization benchmark;

[0130] Indicates The cost vector value in each task, Indicates The cost vector value in each virtual machine; Indicates The time cost value in each task, Indicates The time cost value in a virtual machine, Indicates time The minimum execution time of all tasks within.

[0131] In the time cost function In, the time cost is obtained by summing up the difference between the execution time of each task and the execution time of the minimum task, which depends on the processing power of the CPU. Since we are more interested in the task processing power from the system perspective rather than the service perspective, the entire execution time is used instead of the completion time. At the same time, the minimum execution time of tasks in the system is considered, which more comprehensively evaluates the overall processing power of the system and the overall processing efficiency at different time scales. Assuming that the system is efficient, after a virtual machine completes the task assigned to it, it will enter a sleep state.

[0132] Load Capacity Function Considering the load performance between resources and tasks in a specific state, the load capacity of the system is evaluated by calculating the ratio of the load value of each virtual machine and task to the minimum load value. The smaller the value of the function, the better the load performance of the system, and vice versa. In a computing system, the time of task execution and resource utilization have an important impact on the price cost. Therefore, in order to consider the tasks assigned to each virtual machine, the cost function Price adjustment factor introduced This means that lightweight tasks have a lower unit time cost than heavyweight tasks.

[0133] Obviously, our goal is to minimize the values ​​of the above three functions, which is a multi-objective optimization problem. Since each function pursues different goals, they may conflict with each other. For example, using a more powerful CPU can increase the task processing speed, but the cost will also increase. For another example, a virtual machine with a larger memory can load more tasks, but if the CPU computing power is insufficient, the overall task execution time may be extended.

[0134] 1.2. Scheduling Optimization Model

[0135] Since the results represented by the three matrices of resource nodes and task nodes have different standards, they need to be normalized, so the above three objective functions are expressed as follows:

[0136] (5)

[0137] (6)

[0138] (7)

[0139] in, is the normalized time cost function, which is used to evaluate the time cost of task execution. is the normalized load capacity function, which is used to evaluate the load cost of the system. is the resource cost function, which is used to evaluate the economic cost of resource use;

[0140] Indicates that among all tasks and virtual machines The maximum value of

[0141] Indicates that among all tasks and virtual machines The maximum value of

[0142] Is taken in all tasks and virtual machines The maximum value of

[0143] Different cloud computing systems or computing resource providers have different requirements for task execution performance. , and are the weights of the time cost, load capacity and cost function respectively. Since the time cost, load capacity and cost are considered equally important in the task scheduling of this patent application, we set , the smallest It is the optimal solution for cloud computing task scheduling.

[0144] (8)

[0145] represents the optimal objective function value of cloud computing task scheduling, It means finding the weighted minimum value, which is minimized by weighted combination of multiple cost functions (time cost, load cost, resource cost).

[0146] 2. Cloud computing task scheduling based on neural network algorithm with sparse connection strategy and difference adjustment operator

[0147] Artificial neural network (ANN) is a mathematical model based on the principle of neural network in biology, which aims to simulate the mechanism of human brain nervous system to process complex information. ANN is mainly used for prediction, by receiving input data and output data and inferring the relationship between them. The input data of ANN is usually obtained through experiments, calculations, etc., and the error between the predicted solution and the target solution is reduced by continuously iteratively adjusting the weight value, such as Figure 1As shown. However, sometimes we don’t know what the target solution is. To solve this problem, the author of the NNA evolutionary algorithm takes the current optimal solution as the target solution and continuously adjusts the weight value of each neuron to find a better target solution. Compared with other advanced algorithms, the NNA algorithm has the advantages of simplicity and fewer parameters. However, in order to further improve the task scheduling process, which is prone to falling into local optimal solutions and slow convergence, a cloud computing task scheduling based on a neural network algorithm with sparse connection strategy and difference adjustment operator is proposed.

[0148] 2.1 Initial stage of using sparse connection strategy

[0149] In the NNA algorithm, the initial population , indicating that the algorithm is At the iteration, the set of individuals in the population is is the population size, and the update is through the weight matrix , for any number of iterations , It is Individual vectors, It is weight vectors, each , have Dimension, where ,therefore:

[0150] (9)

[0151] in, Indicates The individual in At the iteration Solutions in dimensional space;

[0152] In order to ensure that the generated new mode does not have large deviations during the creation and movement process, the weights of the mode solution need to be restricted to allow the NNA solution to control the movement with slight deviations. After creating the initial weights, the weight matrix is ​​calculated in the manner of equation (10): Select the solution corresponding to the target The weight of . Therefore, the sum of the weight matrices satisfies the following conditions:

[0153] (10)

[0154] in, It is The individual in The weight values ​​on the dimensions are taken as weight vectors An element in

[0155] (11)

[0156] in, express is a random number uniformly distributed between 0 and 1. is the population size.

[0157] In addition, in order to overcome the problem of falling into local optimality in the initialization phase, the sparse connection idea in the neural network is applied to the algorithm construction phase. Sparse connection in deep learning is a method of forming a sparse connection structure by setting some connections in the neural network to 0. This structure imitates the sparse coding principle in the biological neural system, that is, only some neurons are activated during information transmission. The goal of the sparse connection strategy is to reduce the number of model parameters and prevent falling into local optimality while maintaining model performance. When initializing the weights of SCVNNA, 10% of the weights are randomly set to 0 to prevent the algorithm from falling into local optimality during the construction phase. The specific details are reflected in the following equations.

[0158] (12)

[0159] , (13)

[0160] in Including 10% , is the number of iterations When , the sparse weight matrix is ​​obtained. In this matrix, some weight values ​​are set to zero, which is equivalent to that they do not participate in the crossover process, which improves the efficiency of individuals in the crossover process and their ability to jump out of the local optimum, increases the diversity of the solution space, and thus ensures the superiority of the entire group. Therefore, the introduction of the sparse connection strategy can jump out of the local optimal solution.

[0161] exist The formula for generating a new population in iterations can be expressed as formula (14):

[0162] (14)

[0163] in Indicates the population size; Indicates the current iteration number; Indicates that the solution is At the iteration, The weighted solution of each individual.

[0164] 2.2 Update the weight matrix

[0165] Update the weight vector according to the following formula :

[0166] (15)

[0167] in, Indicated in At the iteration, The weight vector of each individual, Indicated in At the iteration, The weight vector of each individual, For the The best target weight vector obtained at the iteration, Represents a random number, which is used to introduce randomness and increase the exploration ability of the algorithm.

[0168] 2.3 Difference Adjustment Operator

[0169] The ergodic and random characteristics of chaotic variables can be used to optimize the algorithm and improve the diversity of the population. By integrating the Tent chaos operator with the sine-cosine strategy, a new Tent chaos operator, the Tent sine-cosine chaos operator (difference adjustment operator), is proposed to generate chaotic candidate solutions. The calculation formula is as follows:

[0170] (16)

[0171] Indicated in The value after iterations is Indicates the number of mappings; Indicates The value after iterations; express A random number between ; Indicates the current iteration number, Indicates the maximum number of iterations; and To control the constants, they are set to 20 and 2 respectively.

[0172] In the bias operator, the bias population is randomly generated according to the upper and lower bounds, which leads to excessive dispersion of individual positions and makes it difficult to effectively guide individuals toward the target position, thus reducing the search efficiency. Therefore, the difference-adjusted population is used to replace the bias population, introducing more information to adjust the randomness. The calculated new individual positions are affected by the target position and the existing individual positions, and the individual positions are adjusted more specifically, which helps to improve the search efficiency and convergence ability of the algorithm. Therefore, the difference-adjusted population is defined as follows:

[0173] (17)

[0174] in, It is The individual in At the iteration, in dimension By adjusting the scaling factor and mapping constants You can reach different locations around the current optimal solution, mapping constants is a random number between (0,2), It is The target solution at the iteration is Indicates The individual in At the iteration The solution in dimensional space, Indicates the total number of individuals in a population; by reducing The value of to achieve the difference adjustment, so the scaling factor Define the following formula:

[0175] (18)

[0176] in, express A random number between is a new nonlinear convergence factor, and its formula is as follows:

[0177] (19)

[0178] in, Indicates the current iteration number, represents the maximum number of iterations, It is a control constant and its value is 20.

[0179] 2.4 Transfer Operator

[0180] In order to improve the quality of the solution and bring it closer to the target solution, a transfer function operator is introduced. This operator transfers the new mode solution at the current position to a position closer to the target solution. A new position in the closer search space. Therefore, a method for transfer function operator (TF) is proposed to define the process as follows.

[0181] (20)

[0182] in, Indicates The individual in The solution at the iteration

[0183] Indicates The individual in At the iteration Solutions in dimensional space;

[0184] Indicates the current iteration number;

[0185] is a random number between 0 and 1, which follows a uniform distribution;

[0186] Represents the Circle mapping operator;

[0187] It is The target solution at the iteration;

[0188] represents the total number of individuals in the population, The calculation formula is:

[0189] (twenty one)

[0190] in, Represents the Circle mapping operator;

[0191] It is The parameter value at the iteration;

[0192] and is a mapping constant;

[0193] is the modulo operator.

[0194] However, compared with other advanced algorithms, the NNA algorithm has the advantages of simplicity and fewer parameters. In order to further improve the problem of the task scheduling process being prone to falling into local optimal solutions and slow convergence, we proposed two optimization strategies.

[0195] 2.5 Scheduling Process

[0196] Based on the above design, the SCVNNA cloud task scheduling method is summarized. Its main implementation can be divided into the following four main steps: Figure 1 shown.

[0197] Step 1: Initialization. This mainly includes establishing the mapping between cloud computing tasks and neural network individuals, and initializing individuals and weights. At the same time, initialize parameters such as the dimension of the search space, the upper and lower bounds of the population, and the maximum number of iterations.

[0198] Step 2: Once the initialization is completed, the optimal solution search process based on SCVNNA begins. In this stage, the fitness value of the individual is calculated based on its position information. The individual with the minimum fitness value will be regarded as the current optimal solution and recorded. At the same time, sparse connection is performed to adjust the weights and execute formulas (12) and (13).

[0199] Step 3: The global exploration and local development behaviors of individuals will depend on the value of the current difference adjustment operator z. If z is greater than a random number between 0 and 1, the difference adjustment population will be executed to update their positions, that is, according to formulas (16), (17), (18) and (19). Otherwise, the transfer operator is executed to update the position.

[0200] Step 4: Once the maximum number of iterations is reached, the search process stops and the location of the prey is converted into a decision variable , and as the best cloud computing task scheduling solution.

[0201] To date, almost all metaheuristic algorithms have been used for cloud computing task scheduling, but the NNA algorithm, as a relatively new method, has not been widely studied. We provide an effective task scheduling solution based on the NNA algorithm and improve its convergence and search capabilities through two optimization strategies. Experimental evaluation shows that our method performs well in cloud task scheduling and is better than current metaheuristic algorithms.

[0202] 4. Experimental Evaluation

[0203] In this section, we evaluate the performance of the proposed SCVNNA and compare it with some current methods based on a set of simulation-based experiments.

[0204] 4.1 Experimental Setup

[0205] like Figure 2 to Figure 17 As shown, in order to evaluate their performance in simulated cloud computing task scheduling, the SCVNNA algorithm is compared with the ant optimization algorithm (ACO), particle swarm optimization algorithm (PSO), whale optimization algorithm (WOA), genetic algorithm (GA) and NNA, and all algorithms are implemented using MATLABR 2022b.

[0206] The performance is analyzed using time, load, price cost, and total cost as metrics, which are detailed in the cost model in Section 3. The experiments will be divided into two groups based on the number of tasks scheduled in small-scale and large-scale computing scenarios. In addition, the population size will be set to 50 and the number of VMs to 40. We choose to set the number of iterations to 100 because a larger number of iterations brings obvious advantages, which is similar to some recent research results.

[0207] 4.2 Experimental Results

[0208] To fully evaluate the performance of each algorithm, for each experiment we will increase the number of iterations to report the normalized values ​​obtained and report the actual overhead by varying the number of tasks.

[0209] 4.2.1 Comparison with Small-Scale Tasks

[0210] To comprehensively evaluate the performance of each algorithm, normalized values ​​were calculated by increasing the number of iterations and varying the actual cost of the number of tasks to better understand how each algorithm performs in different situations.

[0211] When the number of small-scale tasks increases from 100 to 1000, Figure 2 From the comparison results in , we can see that with the increase in the number of iterations, the total cost of all algorithms shows a downward trend, indicating that they are all effective in the cloud computing environment. In particular, the total cost of the SCVNNA algorithm is always lower than that of the WOA, PSO, GA, ACO, and NNA algorithms, showing its superiority in search ability. In addition, with the increase in the number of iterations, SCVNNA finds the optimal solution faster than NNA, indicating that the introduced sparse connection strategy can effectively overcome the problem of NNA falling into the local optimal solution during the initialization phase. Fig.12 and Fig.13 The results of normalized load overhead and price cost shown in show that the costs of the six algorithms gradually decrease with the increase of the number of iterations, indicating that they can effectively reduce the cost to search for the optimal solution. SCVNNA performs better than other algorithms in reducing price overhead and load overhead, reducing them by more than 10% and more than 20%, respectively, indicating that the introduced difference adjustment operator significantly improves its search efficiency and convergence ability. However, according to Figure 3 The time cost results of the SCVNNA algorithm are higher than those of other algorithms, suggesting that it may have certain limitations in utilizing CPU performance. Although SCVNNA may not perform well in terms of local optimal solutions (time cost), it performs well in terms of global optimal solutions (total cost), demonstrating its superior ability in solving complex optimization problems.

[0212] As the number of tasks increases, Figure 10~Figure 13 The performance comparison between different algorithms is shown. Fig.10 From the normalized total cost data in , we can see that SCVNNA outperforms other algorithms, showing that it can effectively reduce the system resource consumption for small-scale tasks in cloud computing. Figure 11~Figure 13It shows that as the number of tasks increases, each algorithm shows an increasing trend in task execution time, system load, and cost. This is because as the number of tasks increases, the system load will also increase accordingly, resulting in an increase in time and cost. In terms of time cost, the consumption of the six algorithms is almost the same, but in terms of load and price cost, the SCVNNA algorithm performs better, especially when the number of tasks increases. This shows that the SCVNNA algorithm enhances population diversity and improves system memory utilization in a cloud computing environment by adjusting operators by differences.

[0213] 4.2.2 Comparison of Large-Scale Task Scheduling

[0214] In large-scale tests, the number of tasks is gradually increased from 1000 to 10000. Figures 14 to 17 The comparison of all six algorithms by increasing the number of iterations in 1000 tasks is shown, and it can be seen that the total cost, time, price, and load overhead of each algorithm gradually decreases, which once again proves their advantages in cloud task scheduling. In particular, SCVNNA performs best in total cost, time, price, and load overhead compared to the other five algorithms, further verifying its ability to more effectively utilize system resources in handling large-scale task execution in cloud computing, while reducing resource consumption of large-scale tasks.

[0215] From the actual cost of system time, load and price, it can be seen that the advantage of SCVNNA becomes more significant when the number of tasks increases, thanks to the sparse connection strategy that improves its search efficiency. This further confirms that the performance of SCVNNA in cloud computing task scheduling is significantly better than ACO, PSO, GA, WOA and NNA algorithms.

[0216] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A cloud task model method for a high-precision artificial neural network optimization algorithm, characterized in that: include: Step 1: Initialization, including establishing the mapping between cloud computing tasks and neural network individuals, and initializing individuals and weights; at the same time, initializing the dimension of the search space, the upper and lower bounds of the population, and the maximum number of iterations; Step 2: After initialization, the optimal solution search process based on SCVNNA begins; In this stage, the fitness value of the individual is calculated based on its position information. The individual with the minimum fitness value will be regarded as the current optimal solution and recorded. At the same time, sparse connection is performed to adjust the weight; Step 3: The global exploration and local development behaviors of individuals will depend on the value of the current difference adjustment operator z. If z is greater than a random number between 0 and 1, the difference adjustment population will be executed to update their positions. Otherwise, the chaotic transfer function operator will be executed to update the positions. Step 4: Once the maximum number of iterations is reached, the search process stops and the individual positions are converted into decision variables and used as the optimal cloud computing task scheduling solution; The step 2 comprises: ， ， ， in Including 10% , is the number of iterations When , the sparse weight matrix, in which some weight values ​​are set to zero, is equivalent to them not participating in the crossover process, which improves the efficiency of individuals in the crossover process and their ability to jump out of the local optimum, increases the diversity of the solution space, and thus ensures the superiority of the entire group. Therefore, the introduction of the sparse connection strategy can jump out of the local optimal solution; exist The formula for generating a new population in iterations can be expressed as: ， in Indicates the population size; Indicates the current iteration number; Indicates that the solution is At the iteration, The weighted solution of each individual; Update the weight vector according to the following formula: ， in, Indicated in At the iteration, The weight vector of each individual, Indicated in At the iteration, The weight vector of each individual, For the The target weight vector obtained at the iteration is, Represents a random number; The step 3 comprises: By integrating the Tent chaos operator with the sine-cosine strategy, a new Tent sine-cosine chaos operator, namely the Tent difference adjustment operator, is proposed to generate chaotic candidate solutions. The calculation formula is as follows: ， Indicated in The value after iterations is Indicates the number of mappings; Indicates The value after iterations; express A random number between ; Indicates the current iteration number, Indicates the maximum number of iterations; and To control the constant; The difference-adjusted population is used to replace the biased population, and more information is introduced to adjust the randomness. The calculated new individual position is affected by the target position and the existing individual position, and the individual position is adjusted more specifically, which helps to improve the search efficiency and convergence ability of the algorithm. Therefore, the difference-adjusted population is defined as follows: ， in, It is The individual in At the iteration, in dimension By adjusting the scaling factor and mapping constants You can reach different locations around the current optimal solution, mapping constants is a random number between (0,2), It is The target solution at the iteration is Indicates The individual in At the iteration The solution in dimensional space, Indicates the total number of individuals in a population; by reducing The value of is used to achieve the difference adjustment, so the scaling factor Define the following formula: ， in, express A random number between is a new nonlinear convergence factor, and its formula is as follows: ， in, Indicates the current iteration number, represents the maximum number of iterations, To control the constant.

2. The cloud task model method of a high-precision artificial neural network optimization algorithm according to claim 1 is characterized in that: The step 1 comprises: Compute Node and a set of computational tasks ,in, , the final scheduling result can be expressed by the matrix It is expressed as follows: ； in, is a decision variable, if When, it means The task is in Otherwise, ,During the scheduling process, tasks will not be assigned repeatedly, that is, each task can only be assigned to one virtual machine for execution; at the same time, each virtual machine will be assigned at least one task to ensure that all virtual machines participate in task execution; The cost is used as the objective function to represent the resource cost. is the price adjustment factor, and the expression of the cost function is: ， ， ， in, is the time cost function, is the load capacity function, is the price cost function; is the decision variable, Indicates the total number of virtual machines in the collection. Represents the total number of task sets, is the price adjustment factor, which is used to adjust the proportional relationship between resource bandwidth and economic cost; It is The load capacity of each task, It is the smallest load capacity among all tasks. Indicates The minimum load capacity among the resource nodes is used as the normalization benchmark; Indicates The cost vector value in each task, Indicates The cost vector value in each virtual machine; Indicates The time cost value in each task, Indicates The time cost value in a virtual machine, Indicates time The minimum execution time of all tasks in the Since the results represented by the three matrices of resource nodes and task nodes have different standards, they need to be normalized, so the above three objective functions are expressed as follows: ， ， ， in, is the normalized time cost function, which is used to evaluate the time cost of task execution. is the normalized load capacity function, which is used to evaluate the load cost of the system. is the resource cost function, which is used to evaluate the economic cost of resource use; Indicates that among all tasks and virtual machines The maximum value of Indicates that among all tasks and virtual machines The maximum value of Is taken in all tasks and virtual machines The maximum value of Different cloud computing systems or computing resource providers may have different requirements for task execution performance. , and are the weights of time cost, load capacity and cost function respectively, the smallest It is the optimal solution for cloud computing task scheduling: ， represents the optimal objective function value of cloud computing task scheduling, It means to find the weighted minimum value, which is minimized by weighted combination of multiple cost functions.

3. The cloud task model method of a high-precision artificial neural network optimization algorithm according to claim 1 is characterized in that: The step 1 comprises: Initial population , indicating that the algorithm is At the iteration, the set of individuals in the population is is the population size, and the update is through the weight matrix , for any number of iterations , It is Individual vectors, It is weight vectors, each have Dimension, where ,therefore: ， in, Indicates The individual in At the iteration Solutions in dimensional space; After creating the initial weights, the weight matrix in Eq. Select the solution corresponding to the target The weight of , therefore, the sum of the weight matrices satisfies the following conditions: ， in, It is The individual in The weight values ​​on the dimensions are taken as weight vectors An element in ， in, express is a random number uniformly distributed between 0 and 1. is the population size.

4. A cloud task model method of a high-precision artificial neural network optimization algorithm according to any one of claims 1 to 3, characterized in that: The step 3 comprises: The Circle mapping operator is introduced into the transfer function operator, and a chaotic transfer function operator method is used. The operator transfers the new mode solution of the current position to the target solution. The new position in the closer search space is defined as follows: ， in, Indicates The individual in The solution at the iteration Indicates The individual in At the iteration Solutions in dimensional space; Indicates the current iteration number; is a random number between 0 and 1, which follows a uniform distribution; Represents the Circle mapping operator; It is The target solution at the iteration; represents the total number of individuals in the population, The calculation formula is: ， in, Represents the Circle mapping operator; It is The parameter value at the iteration; and is a mapping constant; is the modulo operator.

Images