Policy optimization method, device, equipment and storage medium

Abstract

Embodiments of the present disclosure provide a strategy optimization method and device, equipment and a storage medium. The method comprises: obtaining initial population information for task allocation in a computing power network, the initial population information comprising a plurality of allocation strategies and target function values corresponding to the plurality of allocation strategies, the target function values corresponding to the plurality of allocation strategies being determined by a multi-objective function; based on the multi-objective function and the initial population information, performing multiple population evolutions using a diffusion graph-based evolution algorithm and a mutual information grouping-based evolution algorithm to obtain target population information; and selecting a final strategy from the target population information. Thus, in the process of optimizing and solving the task allocation strategy of the computing power network, the multi-objective optimization requirement of the computing power network is met, the solving efficiency is improved and the diversity of the solution is maintained by using the diffusion graph-based evolution algorithm and the mutual information grouping-based evolution algorithm, and the quality of the task allocation strategy provided for the computing power network is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of computer, and particularly, to a strategy optimization method and device, equipment and storage medium. BACKGROUND

[0002] An algorithm network is a collaborative network formed by integrating distributed algorithm resources, and appropriate algorithm resources need to be allocated to multiple tasks in the algorithm network.

[0003] In related technologies, the task allocation strategy of the algorithm network mainly depends on heuristic algorithms, linear programming algorithms or deep learning algorithms for optimization and solution. The heuristic algorithm is often aimed at a single optimization target, and it is difficult to take into account the multi-objective optimization needs of the algorithm network; the linear programming algorithm has a large computational overhead for solving large-scale task scheduling problems, and it is difficult to meet the real-time task scheduling requirements of the algorithm network; the deep learning algorithm needs a large amount of historical data for model training, and the model generalization ability is insufficient.

[0004] It can be seen that the optimization and solution process of the task allocation strategy of the algorithm network has the shortcomings of single optimization target, low optimization efficiency and insufficient solution diversity. SUMMARY

[0005] Embodiments of the present disclosure provide a strategy optimization method, device, equipment and storage medium to solve the problems of single optimization target, low optimization efficiency and insufficient solution diversity in the optimization and solution process of the task allocation strategy of the algorithm network.

[0006] In a first aspect, embodiments of the present disclosure provide a strategy optimization method, comprising: obtaining initial population information for task allocation in an algorithm network, the initial population information comprising a plurality of allocation strategies and target function values corresponding to the plurality of allocation strategies respectively, the target function values corresponding to the plurality of allocation strategies respectively being determined by a set multi-objective function, each of the allocation strategies comprising algorithm node identifiers allocated to a plurality of tasks respectively; based on the multi-objective function and the initial population information, performing multiple population evolutions using a diffusion graph-based evolution algorithm and a mutual information grouping-based evolution algorithm to obtain target population information; and selecting a final strategy for node allocation of the plurality of tasks in the target population information.

[0007] In a second aspect, the embodiments of the present disclosure provide a strategy optimization apparatus, comprising: an acquisition unit configured to acquire initial population information for task allocation in a computing power network, the initial population information comprising a plurality of allocation strategies and target function values corresponding to the plurality of allocation strategies respectively, the target function values corresponding to the plurality of allocation strategies respectively being determined by a set multi-objective function, each of the allocation strategies comprising a plurality of computing power node identifiers to which a plurality of tasks are respectively allocated; an evolution unit configured to perform multiple population evolutions based on the multi-objective function and the initial population information by using a diffusion graph-based evolution algorithm and a mutual information grouping-based evolution algorithm to obtain target population information; and a strategy selection unit configured to select a final strategy for node allocation of the plurality of tasks in the target population information.

[0008] In a third aspect, the embodiments of the present disclosure provide an electronic device, comprising: at least one processor and a memory; the memory stores computer-executable instructions; and the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the strategy optimization method according to the first aspect.

[0009] In a fourth aspect, the embodiments of the present disclosure provide a computer-readable storage medium, the computer-readable storage medium storing computer-executable instructions, and when a processor executes the computer-executable instructions, the strategy optimization method according to the first aspect is implemented.

[0010] In a fifth aspect, the embodiments of the present disclosure provide a computer program product, the computer program product containing computer-executable instructions, and when a processor executes the computer-executable instructions, the strategy optimization method according to the first aspect is implemented.

[0011] The strategy optimization method, apparatus, device and storage medium provided by the embodiments of the present disclosure can improve the efficiency of the optimization solving process and the diversity of the solution in the process of optimizing the task allocation strategy of the computing power network, and finally provide a high-quality task allocation strategy for the computing power network. BRIEF DESCRIPTION OF DRAWINGS

[0012] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the strategy optimization method provided in the embodiments of this disclosure;

[0014] Figure 2 A schematic diagram illustrating the process of population evolution using a diffusion graph-based evolutionary algorithm provided in this embodiment of the disclosure;

[0015] Figure 3 A schematic diagram illustrating the process of population evolution using an evolutionary algorithm based on mutual information grouping, provided in an embodiment of this disclosure;

[0016] Figure 4 A flowchart illustrating the strategy optimization method provided in this embodiment of the disclosure;

[0017] Figure 5 A structural block diagram of the strategy optimization device provided in the embodiments of this disclosure;

[0018] Figure 6 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0020] In the process of optimizing the task allocation strategy of computing power network, it is difficult to meet the multi-objective optimization requirements of computing power network, and there are also problems such as low optimization efficiency and insufficient diversity of solutions.

[0021] Therefore, there is an urgent need for an optimization method that can meet the multi-objective optimization requirements of computing networks, maintain solution diversity, and efficiently search for high-quality task allocation schemes under limited computing budgets, so as to improve the overall operating efficiency and energy efficiency of computing networks.

[0022] This disclosure provides a strategy optimization method, apparatus, device, and storage medium. It utilizes population evolution to optimize task allocation strategies for computing networks. During the optimization process, a multi-objective function is used as the optimization objective, satisfying the multi-objective optimization requirements of computing networks. Two optimization algorithms are employed: a diffusion graph-based evolutionary algorithm and a mutual information grouping-based evolutionary algorithm. The diffusion graph-based algorithm improves the search efficiency for high-quality individuals during population evolution, thus increasing the convergence speed of the task allocation strategy optimization process. The mutual information grouping-based algorithm maintains the diversity of individual evolution during population evolution, thus preserving the diversity of solutions in the task allocation strategy optimization. Therefore, this disclosure provides a method that satisfies the multi-objective optimization requirements of computing networks while maintaining solution diversity and efficiently searching for high-quality task allocation schemes within a limited computational budget.

[0023] For example, the application scenarios of this disclosure embodiment are as follows:

[0024] In the application scenarios of this disclosure, a computing power network and a monitoring terminal for the computing power network may be included. The computing power network includes multiple computing power nodes, each of which can be a terminal or a server, used to provide computing resources (i.e., computing power resources). The monitoring terminal of the computing power network is communicatively connected to the multiple computing power nodes in the computing power network and is used to monitor the multiple computing power nodes in real time. The strategy optimization method provided in this disclosure embodiment can be executed on the monitoring terminal of the computing power network to generate a task allocation strategy, and the task allocation strategy can be provided to the computing power network or the scheduling terminal of the computing power network so that the computing power network allocates computing tasks to the corresponding computing power nodes for execution according to the task allocation strategy.

[0025] The monitoring end of the algorithm network can be a terminal or a server. Terminals include mobile phones and computers, while servers include cloud servers and local servers.

[0026] Figure 1 This is a schematic flowchart of the strategy optimization method provided in an embodiment of this disclosure. Figure 1 As shown, the strategy optimization method includes the following steps S101~S103:

[0027] S101, for task allocation in the computing power network, obtain initial population information. The initial population information includes multiple allocation strategies and the objective function values ​​corresponding to the multiple allocation strategies. The objective function values ​​corresponding to the multiple allocation strategies are determined by the set multi-objective function. Each allocation strategy includes the identifier of the computing power node to which multiple tasks are allocated.

[0028] Since the objective function values ​​corresponding to the multiple allocation strategies are determined by the set multi-objective function, each allocation strategy can correspond to multiple objective function values.

[0029] Each allocation strategy includes the identifiers of computing nodes to be allocated to multiple tasks (i.e., multiple tasks to be allocated in the computing power network). Each allocation strategy uses integer encoding, and one allocation strategy can represent a complete task allocation scheme.

[0030] Among them, multiple tasks can be tasks that arrive in the computing power network within the current task cycle.

[0031] In this embodiment, for task allocation in the computing power network, multiple allocation strategies can be initialized, such as random initialization of multiple allocation strategies. Through a multi-objective function, the objective function values ​​corresponding to each of the multiple allocation strategies are determined, and these multiple allocation strategies and their corresponding objective function values ​​are used as initial population information. In the initial population information, multiple allocation strategies are equivalent to multiple chromosomes, i.e., one chromosome represents one allocation strategy, and one allocation strategy contains the identifiers of the computing power nodes allocated to multiple tasks. The identifiers of the computing power nodes allocated to multiple tasks are equivalent to multiple genes on the chromosome, and the length of the chromosome (i.e., the number of genes) is equal to the number of tasks.

[0032] S102, based on a multi-objective function and initial population information, uses a diffusion graph-based evolutionary algorithm and a mutual information grouping-based evolutionary algorithm to perform multiple population evolutions to obtain the target population information.

[0033] In any one of the multiple population evolutions, either a diffusion graph-based evolutionary algorithm or a mutual information grouping-based evolutionary algorithm is used. In other words, the diffusion graph-based evolutionary algorithm and the mutual information grouping-based evolutionary algorithm appear in different evolutionary stages.

[0034] In the evolutionary iterations using a diffusion graph-based evolutionary algorithm, the algorithm guides the search direction of allocation strategies during population evolution based on the similarity between multiple allocation strategies in the current population information, significantly improving the convergence speed of individuals towards high-quality solution regions. Therefore, using a diffusion graph-based evolutionary algorithm can effectively shorten the algorithm's convergence time, reduce the number of calculations of the objective function, and thus improve the optimization efficiency of task allocation strategies, meeting the real-time scheduling requirements of the algorithm network.

[0035] In the evolutionary iterations using the mutual information-based grouping evolutionary algorithm, the mutual information-based grouping evolutionary algorithm utilizes the mutual information between the variable values ​​of multiple policy variables in the allocation strategy and multiple objective functions to group and evolve multiple policy variables, thereby maintaining the diversity of solutions. This provides the scheduler of the computing network with diversified task allocation strategies that take into account multiple performance indicators, and also avoids blind search in the population evolution process, improves the efficiency of solution space exploration, enhances the optimization efficiency of task allocation strategies, and meets the real-time scheduling requirements of the algorithm network.

[0036] In this embodiment, the number of evolutionary iterations applicable to the diffusion graph-based evolutionary algorithm and the mutual information grouping-based evolutionary algorithm can be predefined. For the number of evolutionary iterations applicable to the diffusion graph-based algorithm, the population evolution is performed using the diffusion graph-based algorithm based on the multi-objective function and the population information related to that evolutionary iteration (if that evolutionary iteration is the first time, then the population information related to that iteration is the initial population information; otherwise, the population information related to that iteration is the population information after the previous evolution) to obtain the evolved population information. For the number of evolutionary iterations applicable to the mutual information grouping-based algorithm, the evolved population information is obtained using the mutual information grouping-based evolutionary algorithm based on the multi-objective function and the population information related to that evolutionary iteration (refer to the above explanation).

[0037] S103, in the target population information, assign and select the final strategy for nodes of multiple tasks.

[0038] There can be multiple final strategies.

[0039] In this embodiment, a final strategy can be selected from multiple allocation strategies based on the objective function values ​​corresponding to those strategies in the target population information. After determining the final strategy, it can be sent to the scheduler of the computing power network. The scheduler can then use the final strategy as the computing power network scheduling scheme and complete the computing power network scheduling according to the scheme.

[0040] In this embodiment of the disclosure, a multi-objective function is used to meet the multi-objective optimization requirements of the computing power network. An evolutionary algorithm based on diffusion graphs and an evolutionary algorithm based on mutual information grouping are used as optimization algorithms to improve the convergence speed of the task allocation strategy optimization solution process and maintain the diversity of solutions. This improves the optimization efficiency and quality of the task allocation scheme of the computing power network, and provides a diversified task allocation strategy that takes into account multiple performance indicators for the scheduling of the computing power network.

[0041] In some embodiments, before obtaining initial population information for task allocation in a computing power network, the strategy optimization method further includes: obtaining real-time state information corresponding to multiple computing power nodes and attribute information corresponding to multiple tasks in the computing power network; and determining the objective function values ​​corresponding to multiple allocation strategies based on the real-time state information corresponding to multiple computing power nodes and the attribute information corresponding to multiple tasks through a multi-objective function. Thus, by utilizing the real-time state information of the computing power nodes and the attribute information corresponding to multiple tasks, real-time and accurate input information is provided for calculating the objective value of the multi-objective function, improving the accuracy of the objective function values ​​corresponding to the multiple allocation strategies.

[0042] The real-time status information for each of the multiple computing nodes may include one or more of the following: node computing power, node power parameters, and node link bandwidth. Furthermore, more node status parameters, such as power consumption, central processing unit (CPU) utilization, and network latency, can be added to the real-time status information based on the target optimization requirements of the computing network.

[0043] The attribute information corresponding to each of the multiple tasks may include one or more of the following: task computation volume, task data volume, and task arrival time. Task arrival time refers to the time when the task arrives at the computing power network. Furthermore, based on the target optimization requirements of the computing power network, more task attribute parameters can be added to the attribute information corresponding to each of the multiple tasks, such as task deadlines and task priorities.

[0044] In this embodiment, the monitoring terminal of the computing power network can obtain the real-time status information of multiple computing power nodes and the attribute information of multiple tasks in the computing power network when it detects that multiple tasks have arrived in the computing power network, or when it detects that the current time has entered the task allocation cycle, or when it receives the task allocation strategy generation instruction.

[0045] The real-time status information corresponding to multiple computing nodes and the attribute information corresponding to multiple tasks can constitute the input dataset. The real-time status information corresponding to multiple computing nodes represents the "supply capacity" of the computing network, and the attribute information corresponding to multiple tasks represents the "computing demand" of the computing network that requires computing resource scheduling. The input dataset formed by the combination of the two can be used as the input to the policy optimization algorithm provided in this embodiment, and in particular, it can be used as the input to a multi-objective function.

[0046] Optionally, after obtaining the real-time status information corresponding to multiple computing power nodes and the attribute information corresponding to multiple tasks, outlier processing is performed on the real-time status information corresponding to multiple computing power nodes and the attribute information corresponding to multiple tasks to avoid negative interference from outliers to subsequent target value calculation, population evolution and other processes.

[0047] Outliers refer to data points that significantly deviate from the normal range or statistical distribution in the real-time status information corresponding to multiple computing nodes and the attribute information corresponding to multiple tasks. These data points may be caused by sensor malfunctions, network latency fluctuations, recording errors, etc. For example, the utilization rate of the central processing unit (CPU) of the computing node suddenly exceeds 100%, the power parameter value is negative, or the power parameter value exceeds the rated power of the computing node.

[0048] In some embodiments, the multi-objective function may include a first objective function and a second objective function. The first objective function minimizes the average task completion time, and the second objective function minimizes the total task energy consumption. Thus, by using the first and second objective functions, the dual-objective optimization requirements of the computing network for minimizing both the average task completion time and the total task energy consumption are satisfied. This allows subsequent population optimizations to simultaneously achieve better allocation strategies in both the average task completion time and the total task energy consumption, thereby improving the overall operating efficiency and energy efficiency of the computing network.

[0049] When the multi-objective function may include a first objective function and a second objective function, the objective function values ​​corresponding to the multiple allocation strategies include the average task completion time and the total task energy consumption corresponding to the multiple allocation strategies.

[0050] In one possible implementation, the average task completion time corresponding to multiple allocation strategies is calculated through the following process: For each allocation strategy, based on the real-time status information of multiple computing nodes and the attribute information of multiple tasks, the task transmission time and task execution time when multiple tasks are allocated according to the allocation strategy are calculated using a first objective function. The average completion time of the tasks allocated according to the allocation strategy is then calculated by adding the task transmission time and task execution time of each task according to the allocation strategy. This yields the average task completion time corresponding to the allocation strategy. Thus, the accuracy of the average task completion time is improved by taking into account both the task transmission time and task execution time of each task when allocated according to the allocation strategy.

[0051] Optionally, the first objective function and the second objective function are expressed as follows:

[0052]

[0053] in, This indicates the average time to complete the task. M represents the total energy consumption of the task, and M represents the total number of tasks. This represents the amount of data for the j-th task. This indicates the available bandwidth allocated to the i-th node for the j-th task, through... Divide by Obtain the transmission time of the j-th task assigned to the i-th node; This represents the computational cost of the j-th task. This represents the computation speed at which the j-th task is assigned to the i-th node, through... Divide by The execution time (also known as task computation time) of the j-th task assigned to the i-th node is obtained. This represents the power of the execution process when the j-th task is assigned to the i-th node. , and / or The sum can be a preset value, such as an empirical value pre-set based on the node signal of the computing node and the task size of the computing task, or it can be obtained by the monitoring end of the computing network through regression analysis of the operating data of multiple computing nodes.

[0054] In some embodiments, during multiple population evolutions, a diffusion graph-based evolutionary algorithm is used for the first K evolutions, and a mutual information grouping-based evolutionary algorithm is used for the (K+1)th to Lth evolutions, where L is the total number of evolutions and K is a set threshold. Thus, in the early stages of population evolution, the diffusion graph-based algorithm improves the convergence speed of individuals towards high-quality solution regions, while in the later stages, the mutual information grouping-based algorithm maintains solution diversity, achieving a balance between the optimization efficiency and policy quality of the task allocation strategy.

[0055] Figure 2 This is a schematic diagram illustrating the process of population evolution using a diffusion graph-based evolutionary algorithm, as provided in an embodiment of this disclosure. Figure 2 As shown, a population evolution process using a diffusion graph-based evolutionary algorithm includes the following steps S201~S208:

[0056] S201. From the current population information, select multiple original strategies and the objective function values ​​corresponding to the multiple original strategies respectively. If the current evolution is the first evolution, the current population information is the initial population information. If the current evolution is not the first evolution, the current population information is the intermediate population information after the previous evolution.

[0057] The original strategy refers to the allocation strategy that participates in the current sub-evolution.

[0058] In this embodiment, the current population information includes multiple allocation strategies and the objective function values ​​corresponding to each of the multiple allocation strategies. Based on the objective function values ​​corresponding to each of the multiple allocation strategies, multiple original strategies can be selected from the multiple allocation strategies through an environment selection operation, and the objective function values ​​corresponding to each of the multiple original strategies can be obtained from the objective function values ​​corresponding to each of the multiple allocation strategies.

[0059] In one possible implementation, based on the objective function values ​​corresponding to multiple allocation strategies, multiple original strategies are selected from the multiple allocation strategies through an environment selection operation. This includes: performing non-dominated sorting on the multiple allocation strategies according to their respective objective function values ​​to obtain multiple non-dominated layers and the correspondence between the multiple allocation strategies and the multiple non-dominated layers; filling the multiple non-dominated layers with multiple allocation strategies layer by layer according to the correspondence between the multiple allocation strategies and the multiple non-dominated layers, until the s-th non-dominated layer stops filling, where the number of allocation strategies in the multiple non-dominated layers is greater than or equal to a set strategy number threshold, where s is greater than or equal to 1; calculating the difference between the strategy number threshold and the number of allocation strategies in the first s-1 non-dominated layers to obtain the remaining number of strategies to be selected; selecting the remaining number of strategies to be selected from the unfilled multiple allocation strategies according to their objective function values ​​and filling them into the s-th non-dominated layer; and determining all allocation strategies in the first s-th non-dominated layers as multiple original strategies. In this way, from the multiple allocation strategies contained in the current population information, the superior allocation strategy is selected to participate in the population evolution, so as to evolve into a better allocation strategy.

[0060] In the process of selecting the remaining number of allocation strategies to fill the s-th non-dominated layer based on the objective function values ​​corresponding to the unfilled allocation strategies, the crowding degree of each of the unfilled allocation strategies can be evaluated according to the objective function values ​​corresponding to the unfilled allocation strategies, and the target crowding degree corresponding to each of the unfilled allocation strategies can be obtained. Then, the remaining number of allocation strategies to fill the s-th non-dominated layer can be selected from the unfilled allocation strategies in descending order of the target crowding degree corresponding to each of the unfilled allocation strategies.

[0061] Based on the objective function values ​​corresponding to the unfilled allocation strategies, the crowding of each unfilled allocation strategy is evaluated. This may include: sorting the unfilled allocation strategies according to the magnitude of their objective function values ​​to obtain a strategy sequence; setting the target crowding of the first allocation strategy to infinity, where the first allocation strategy is the target crowding of the first and last allocation strategies in the strategy sequence; and determining the target crowding of the second allocation strategy based on the objective function values ​​corresponding to its adjacent allocation strategies, the maximum value among the objective function values ​​corresponding to the unfilled allocation strategies, and the minimum value among the objective function values ​​corresponding to the unfilled allocation strategies, where the second allocation strategy is any allocation strategy in the strategy sequence other than the first allocation strategy.

[0062] It should be noted that, due to the use of multiple objective functions, one allocation strategy can correspond to multiple objective function values. When sorting the objective function values ​​corresponding to the multiple unfilled allocation strategies, multiple strategy sequences can be obtained under each objective function for the objective function values ​​corresponding to the multiple unfilled allocation strategies. The multiple strategy sequences correspond one-to-one with the multiple objective functions.

[0063] For example, taking the second allocation strategy as the i-th allocation strategy in the strategy sequence, the formula for calculating the target crowding degree of the second allocation strategy is:

[0064]

[0065] in, Y represents the target crowding degree of the i-th allocation strategy, and Y represents the number of objective functions. The maximum value among the objective function values ​​of multiple unfilled allocation strategies under the y-th objective function. Let be the minimum of the objective function values ​​among the multiple unfilled allocation strategies under the y-th objective function. Let the objective function value of the (i+1)th allocation policy in the policy sequence corresponding to the y-th objective function be determined under the y-th objective function. Let be the objective function value of the (i-1)th allocation policy in the policy sequence corresponding to the y-th objective function under the y-th objective function.

[0066] S202, based on the similarity between multiple original policies, the multiple original policies are projected from the policy space to the diffusion graph space to obtain the multidimensional feature key values ​​corresponding to the multiple original policies respectively. The multidimensional feature key values ​​corresponding to the multiple original policies respectively form the diffusion graphs corresponding to the multiple original policies respectively.

[0067] In this embodiment, the policy space is a data space formed by multiple original policies. Taking M tasks as an example, each original policy includes the identifiers of the computing power nodes allocated to the M tasks, so each original policy is an M-dimensional vector, and the policy space is composed of multiple M-dimensional vectors. The diffusion graph is a multi-dimensional feature key value, and the diffusion graph space is a data space composed of multiple multi-dimensional feature key values. Based on the similarity between multiple original policies, the multiple original policies are projected from the policy space onto the diffusion graph space, so that the multiple original policies can be represented as corresponding multi-dimensional feature key values. The multi-dimensional feature key values ​​reflect the similarity between the original policies, thereby using the similarity between multiple original policies to guide the evolutionary direction of individuals in the population evolution process, that is, to guide the search direction of the allocation policy, and improve the convergence speed of population evolution.

[0068] In one possible implementation, S202 includes: determining a first similarity matrix and a second similarity matrix based on multiple original strategies, wherein the first similarity matrix includes the similarity between every two original strategies, and the second similarity matrix includes the total similarity corresponding to each of the multiple original strategies; determining a probability matrix based on the first and second similarity matrices, wherein the element value located in the i-th row and j-th column of the probability matrix represents the probability of the i-th original strategy transitioning to the j-th original strategy; and performing eigenvalue decomposition on the probability matrix to obtain the multidimensional feature key values ​​corresponding to each of the multiple original strategies.

[0069] In this implementation, the Gaussian kernel function is used to calculate the similarity between every two original policies, resulting in a first similarity matrix. In the first similarity matrix, the element in the i-th row and j-th column represents the similarity between the i-th and j-th original policies. The element values ​​of each row in the first similarity matrix are summed to obtain the degree value of each row. A degree matrix, i.e., the second similarity matrix, is constructed based on the degree values ​​of each row. In the second similarity matrix, the i-th diagonal element (i.e., the element in the i-th row and i-th column) represents the total similarity between the i-th original policy and the multiple original policies. Inverting the second similarity matrix yields its inverse. Multiplying the first similarity matrix by this inverse constructs the Markov transition matrix, also known as the probability matrix. The element in the i-th row and j-th column of the Markov transition matrix represents the probability of transitioning from the i-th original policy to the j-th original policy during the diffusion process. Eigenvalue decomposition of the Markov transition matrix yields multiple eigenvectors corresponding to each original policy. Combining these eigenvectors according to a predetermined dimension generates multiple eigenvalues ​​for each original policy in the diffusion graph space. These eigenvectors and eigenvalues ​​constitute the multidimensional feature keys for each original policy. Thus, through this matrix calculation process, multiple original policies are projected from the policy space onto the diffusion graph space based on their similarity.

[0070] For example, the formula for calculating the value of the element located in the i-th row and j-th column in the first similarity matrix is ​​expressed as:

[0071]

[0072] in, This represents the value of the element in the i-th row and j-th column of the first similarity matrix, where c is a preset parameter representing the kernel width parameter of the Gaussian kernel function; This represents the i-th original policy among multiple original policies. Let represent the j-th original policy among multiple original policies, and exp() represent an exponential function with the natural constant e as the base.

[0073] For example, the formula for calculating the value of the element in the i-th row of the second similarity matrix is ​​expressed as follows:

[0074]

[0075] in, This represents the value of the element in the i-th row of the second similarity matrix.

[0076] For example, the formula for calculating the probability matrix is ​​expressed as follows:

[0077]

[0078] Where P represents the probability matrix.

[0079] For example, the eigenvalue decomposition formula of the probability matrix is ​​expressed as:

[0080]

[0081] in, Represents the k-th eigenvalue. Let represent the k-th eigenvector. The k-th eigenvalue is formed by the k-th eigenvalue and the k-th eigenvector in the multidimensional eigenvalue. The value of k is an integer in the range 1 to d, where d is greater than 1.

[0082] Therefore, the multidimensional feature key values ​​corresponding to the original strategy can be represented as:

[0083]

[0084] S203, in the diffusion graph space, perform crossover and mutation operations on the multidimensional feature key values ​​corresponding to multiple original strategies to obtain the multidimensional feature key values ​​corresponding to multiple new strategies.

[0085] In this embodiment, in the diffusion graph space, cross-mutation operations are performed on the multidimensional feature key values ​​corresponding to multiple original strategies. For example, the values ​​of two multidimensional feature key values ​​are swapped, or the values ​​of a single multidimensional feature key value are changed, to generate multiple new multidimensional feature key values. These multiple new multidimensional feature key values ​​are the multidimensional feature key values ​​corresponding to multiple new strategies.

[0086] S204, based on the mapping relationship between the policy space and the diffusion graph space, transforms the multidimensional feature key values ​​corresponding to multiple new policies to the policy space to obtain multiple new policies.

[0087] In this embodiment, multiple original policies can be mapped from the policy space to the diffusion graph space. Naturally, the multidimensional feature keys corresponding to multiple new policies can also be transformed from the diffusion graph space to the policy space. A mapping relationship between the policy space and the diffusion graph space can be established based on the multiple original policies and their corresponding multidimensional feature keys. For example, a formula or model representing the mapping relationship between the policy space and the diffusion graph space can be used. Then, through this mapping relationship, the multidimensional feature keys corresponding to the multiple new policies can be transformed to the policy space, resulting in multiple new policies.

[0088] In one possible implementation, S204 includes: training a data mapping model using the multidimensional feature key values ​​corresponding to multiple original policies and the multiple original policies themselves as training data; and using the data mapping model to transform the multidimensional feature key values ​​corresponding to multiple new policies from the diffusion graph space to the policy space, thereby obtaining multiple new policies. Thus, by training the data mapping model, the model learns the data relationship between the policy space and the diffusion graph space, improving the accuracy of transforming the multidimensional feature key values ​​corresponding to multiple new policies from the diffusion graph space to the policy space.

[0089] The data mapping model can be a neural network model. During the training of the data mapping model, the multidimensional feature key values ​​corresponding to multiple original strategies can be used as multiple training samples, and the multiple original strategies can be used as sample labels corresponding to multiple training samples. The specific training process can refer to the training of neural network models, and the model parameters can be adjusted through feedforward and backpropagation.

[0090] S205, based on the multidimensional feature key values ​​corresponding to the multiple original strategies and the multidimensional feature key values ​​corresponding to the multiple new strategies, target value prediction is performed on the multiple original strategies and the multiple new strategies respectively, to obtain the predicted target values ​​corresponding to the multiple original strategies and the predicted target values ​​corresponding to the multiple new strategies respectively.

[0091] In this embodiment, calculating the objective function value using a multi-objective function may be time-consuming and computationally expensive. Therefore, the objective value can be predicted for each of the original and new strategies based on the multi-dimensional feature key values ​​corresponding to the original strategies and the new strategies, respectively. This yields the predicted objective values ​​for each of the original and new strategies. By replacing the objective function calculation with the objective value prediction, the computational load in the task allocation strategy optimization process is reduced, thereby improving the efficiency of task allocation strategy optimization.

[0092] In one possible implementation, S205 includes: training a first prediction model using multidimensional feature key values ​​and objective function values ​​corresponding to multiple original strategies as training data. The first prediction model is used to predict the target value of the strategy to be predicted based on the multidimensional feature key values ​​of the strategy to be predicted; inputting the multidimensional feature key values ​​corresponding to multiple original strategies into the first prediction model, and using the first prediction model to predict the target value of the multiple original strategies to obtain the predicted target value corresponding to each of the multiple original strategies; inputting the multidimensional feature key values ​​corresponding to multiple new strategies into the first prediction model, and using the first prediction model to predict the target value of the multiple new strategies to obtain the predicted target value corresponding to each of the multiple new strategies. Thus, by training the first prediction model, it learns the ability to predict target values. During population evolution, the first prediction model predicts the target value of the original strategy and the target value of the new strategy using the first prediction model, without needing to calculate the target value through a multi-objective function, effectively reducing the computational load during population evolution and improving the efficiency of population evolution.

[0093] The first prediction model can be a neural network model, such as a backpropagation neural network model. During the training of the first prediction model using multidimensional feature key values ​​and objective function values ​​corresponding to multiple original strategies as training data, the multidimensional feature key values ​​corresponding to multiple original strategies are input into the first prediction model as multiple training samples, and the objective function values ​​corresponding to multiple original strategies are used as sample labels for the multiple training samples.

[0094] S206, select multiple evolutionary strategies from multiple original strategies and multiple new strategies based on the predicted target values ​​corresponding to multiple original strategies and multiple new strategies respectively.

[0095] In this embodiment, environmental selection is performed on multiple original strategies and multiple new strategies based on the predicted target values ​​corresponding to multiple original strategies and multiple new strategies, as well as a preset selection quantity. The best multiple strategies (which can be either original strategies or new strategies) are selected as evolutionary strategies. For example, u evolutionary strategies are selected, where u is greater than 1.

[0096] S207 determines the objective function values ​​corresponding to multiple evolutionary strategies through a multi-objective function.

[0097] S208, add multiple evolutionary strategies and their corresponding objective function values ​​to the current population information to obtain the population information after the current evolution.

[0098] In this embodiment, after selecting multiple evolutionary strategies, in order to ensure the accuracy of the target values ​​corresponding to the strategies in the population information, the target function values ​​corresponding to the multiple evolutionary strategies are determined by a multi-objective function. Then, the multiple evolutionary strategies and the target function values ​​corresponding to the multiple evolutionary strategies are added to the current population information to obtain the population information after the current evolution, thereby realizing the update of the population information.

[0099] When the evolutionary algorithm based on the diffusion graph is used in the first to Kth evolutions, the population information after the current evolution can be used as the current population information in the next evolution.

[0100] In this embodiment of the disclosure, during the optimization of the task allocation strategy of the computing power network, an evolutionary algorithm based on diffusion graphs is used to perform the above-mentioned population evolution operation. This enables the similarity between multiple original strategies to guide the search direction of the allocation strategy during the population evolution process, significantly improving the convergence speed of individual populations towards high-quality solution regions, effectively shortening the algorithm convergence time, reducing the number of calculations of the objective function, and thus improving the optimization efficiency of the task allocation strategy to meet the real-time scheduling requirements of the algorithm network.

[0101] Figure 3 This is a schematic diagram illustrating the process of population evolution using an evolutionary algorithm based on mutual information grouping, provided in an embodiment of this disclosure. Figure 3 As shown, a population evolution process using an evolutionary algorithm based on mutual information grouping includes the following steps S301~S309:

[0102] S301, Select multiple original strategies and their corresponding objective function values ​​from the current population information. The current population information is the intermediate population information after the last evolution.

[0103] The original strategy refers to the allocation strategy that participates in the current sub-evolution.

[0104] In this embodiment, the current population information includes multiple allocation strategies and the objective function values ​​corresponding to each of the multiple allocation strategies. Based on the objective function values ​​corresponding to each of the multiple allocation strategies, multiple original strategies can be selected from the multiple allocation strategies through an environment selection operation, and the objective function values ​​corresponding to each of the multiple original strategies can be obtained from the objective function values ​​corresponding to each of the multiple allocation strategies.

[0105] The selection of multiple original strategies from multiple allocation strategies through the environment selection operation can be referred to the implementation method provided in S201, which will not be elaborated here.

[0106] S302, normalize and calculate the mutual information of the multiple computing node identifiers and the objective function values ​​corresponding to the multiple original strategies respectively, to obtain the mutual information between the multiple strategy variables and the multiple objective functions. The multiple computing node identifiers included in the multiple original strategies are the variable values ​​of the multiple strategy variables in the multiple original strategies respectively.

[0107] Among them, multiple policy variables correspond one-to-one with multiple tasks to be allocated in the computing power network. Each allocation policy contains multiple policy variables and their corresponding variable values, which are the computing power node identifiers.

[0108] In this embodiment, numerical normalization is used to normalize the multiple computing node identifiers and the objective function values ​​corresponding to the multiple original strategies, respectively, to obtain normalized multiple computing node identifiers and normalized multiple objective function values. Mutual information calculation is performed on the normalized multiple computing node identifiers and normalized multiple objective function values ​​to obtain the mutual information between each pair of multiple computing node identifiers and multiple objective functions.

[0109] In one possible implementation, the computing node identifiers included in the multiple original strategies are discretized into B equal-length intervals to obtain multiple normalized computing node identifiers; similarly, the objective function values ​​corresponding to the multiple original strategies are discretized into B equal-length intervals to obtain multiple normalized objective function values. B is a preset integer. This achieves consistent normalization processing for multiple original strategies and multiple objective function values, facilitating improved accuracy in mutual information calculation.

[0110] Optionally, the identifiers of the computing power nodes included in the j-th original strategy are discretized into B equal-length intervals, using the following formula:

[0111]

[0112] Where c is a parameter that is not equal to 0. This represents the computing node identifier corresponding to the i-th strategy variable in multiple original strategies. This represents the computing node identifier corresponding to the i-th strategy variable in the j-th original strategy. Represents the normalized result , This represents the minimum value of the i-th strategy variable among the corresponding computing node identifiers in multiple original strategies. This represents the maximum value of the computing node identifier corresponding to the i-th strategy variable in multiple original strategies.

[0113] The formula for discretizing the objective function values ​​corresponding to multiple original strategies into a second number of equal-length intervals to obtain multiple normalized objective function values ​​is similar to the formula above.

[0114] Optionally, mutual information is calculated for the normalized identifiers of multiple computing nodes and the normalized objective function values, using the following formula:

[0115]

[0116] in, P(a) , P(b) and P(a,b) The calculation formula is as follows:

[0117]

[0118]

[0119]

[0120] in, This represents the mutual information between the i-th policy variable and the m-th objective function. This represents the i-th policy variable after normalization. Let m be the m-th objective function after normalization, N represent the number of original policies, and a and b represent... The normalized policy value and sum The interval indices corresponding to the normalized objective function value in multiple equal-length intervals. P(a) express Marginal empirical probability, P(b) express Marginal empirical probability, P(a,b) express and The joint empirical probability.

[0121] S303: Based on the mutual information between multiple policy variables and multiple objective functions, the multiple policy variables are grouped to obtain multiple sub-policy groups.

[0122] In this embodiment, after obtaining the mutual information between multiple policy variables and multiple objective functions pairwise, for each policy variable, the mutual information between the policy variable and the multiple objective functions can be superimposed to obtain the sum of the mutual information of the policy variable on the multiple objective functions, that is, the sum of the mutual information corresponding to the policy variable. Based on the sum of the mutual information corresponding to each policy variable, the multiple policy variables are sorted and grouped to obtain multiple sub-policy groups.

[0123] A sub-strategy group may include one or more sub-strategies, and a sub-strategy may include the value of a strategy variable or the values ​​corresponding to multiple strategy variables. For example, in one sub-strategy group, each sub-strategy includes the value of the first strategy variable, and in another sub-strategy group, each sub-strategy includes the values ​​of the second and third strategy variables.

[0124] S304 generates multiple new sub-policies by performing intra-group crossover mutation on multiple sub-policy groups.

[0125] In this embodiment, within each sub-strategy group: based on the sub-strategies contained in the sub-strategy group, intra-group crossover mutation is performed to obtain multiple new sub-strategies generated within the sub-strategy group.

[0126] S305, predict the target value of the sub-strategy in multiple sub-strategy groups to obtain the predicted target value of the sub-strategy in multiple sub-strategy groups.

[0127] In this embodiment, calculating the objective function value using a multi-objective function may be time-consuming and computationally expensive. Therefore, objective value prediction is performed on the sub-policies in multiple sub-policy groups. This objective value prediction replaces the objective function calculation, reducing the computational load in the task allocation strategy optimization process and improving the efficiency of task allocation strategy optimization.

[0128] In one possible implementation, multiple second prediction models are trained using multiple sub-policy groups and the objective function values ​​corresponding to the original policies to which each sub-policy group belongs, as training data. These second prediction models are used to predict the target values ​​of the sub-policy groups, with each sub-policy group corresponding one-to-one with a different second prediction model. The target values ​​of the sub-policies within the multiple sub-policy groups can be predicted using these multiple second prediction models, resulting in predicted target values ​​for the sub-policies within each sub-policy group. Therefore, by training a corresponding second prediction model for each sub-policy group, the accuracy of the second prediction models in predicting the target values ​​of new sub-policies within the sub-policy groups is improved. This replaces the calculation of target values ​​using multiple objective functions, effectively reducing the computational load in the population evolution process and improving the efficiency of population evolution.

[0129] The second prediction model can be a neural network model, such as a backpropagation neural network model. During the training of multiple second prediction models, using the objective function values ​​corresponding to the original policies of the multiple sub-policy groups and their respective sub-policy groups as training data, for each second prediction model, multiple sub-policies within the corresponding sub-policy group are used as multiple training samples, and the objective function values ​​corresponding to the original policies of the sub-policies of the second prediction model are used as the sample labels for each training sample, thus achieving supervised training of the second prediction model.

[0130] S306, Based on the predicted target value corresponding to the sub-strategy in multiple sub-strategy groups, select sub-strategies from multiple sub-strategy groups to obtain multiple target sub-strategies.

[0131] In this embodiment, for each sub-strategy group: the environment selection can be performed on the sub-strategy group based on the predicted target value corresponding to the sub-strategy in the multiple sub-strategy groups and the preset selection number, and the best multiple sub-strategies can be selected as the target sub-strategy.

[0132] S307, multiple objective sub-policies are concatenated to obtain multiple evolutionary policies.

[0133] In this embodiment, multiple policy variables correspond one-to-one with multiple tasks. According to the correspondence between multiple policy variables and multiple tasks, the variable values ​​of policy variables in multiple target sub-policies can be concatenated to obtain multiple evolutionary policies.

[0134] For example, there are four tasks, each corresponding to a policy variable, resulting in four policy variables: Policy Variable 1, Policy Variable 2, Policy Variable 3, and Policy Variable 4. After grouping, these four policy variables are divided into two groups: Policy Variable 1 and Policy Variable 3 are in one sub-policy group, and Policy Variable 2 and Policy Variable 4 are in another sub-policy group. In the sub-policy group containing Policy Variable 1 and Policy Variable 3, the variable values ​​corresponding to Policy Variable 1 in various original policies are paired with the variable values ​​corresponding to Policy Variable 3 in various original policies to form multiple sub-policies. By performing intra-group cross-mutation on these sub-policies, multiple new sub-policies are obtained. Similarly, in the sub-policy group containing Policy Variable 2 and Policy Variable 4, the variable values ​​corresponding to Policy Variable 2 in various original policies are paired with the variable values ​​corresponding to Policy Variable 4 in various original policies to form multiple sub-policies. By performing intra-group cross-mutation on these sub-policies, multiple new sub-policies are obtained. A new sub-strategy is generated; from the sub-strategy group containing strategy variable 1 and strategy variable 3, a sub-strategy is selected to obtain the target sub-strategy from that sub-strategy group; from the sub-strategy group containing strategy variable 2 and strategy variable 4, a sub-strategy is selected to obtain the target sub-strategy from that sub-strategy group; the target sub-strategy from the sub-strategy group containing strategy variable 1 and strategy variable 3, and the target sub-strategy from the sub-strategy group containing strategy variable 2 and strategy variable 4 are concatenated in the order of strategy variable 1, strategy variable 2, strategy variable 3, and strategy variable 4 to obtain the complete allocation strategy.

[0135] S308 determines the objective function values ​​corresponding to multiple evolutionary strategies through a multi-objective function.

[0136] S309, add multiple evolutionary strategies and their corresponding objective function values ​​to the current population information to obtain the population information after the current evolution.

[0137] In this embodiment, after selecting multiple evolutionary strategies, in order to ensure the accuracy of the target values ​​corresponding to the strategies in the population information, the target function values ​​corresponding to the multiple evolutionary strategies are determined by a multi-objective function. Then, the multiple evolutionary strategies and the target function values ​​corresponding to the multiple evolutionary strategies are added to the current population information to obtain the population information after the current evolution, thereby realizing the update of the population information.

[0138] If the current evolution is the last evolution (i.e., the Lth evolution), the population information after the current evolution is the target population information; otherwise, the population information after the current evolution can be used as the current population information in the next evolution.

[0139] In this embodiment of the disclosure, during the optimization of the task allocation strategy of the computing power network, an evolutionary algorithm based on mutual information grouping is used to perform the above-mentioned population evolution operation. This realizes the calculation of mutual information between the variable values ​​corresponding to the policy variables in the allocation strategy and the objective value of the allocation strategy under the objective function. The decision variables in the allocation strategy are grouped and evolved, thereby maintaining the diversity of solutions. This provides the scheduler of the computing power network with diversified task allocation strategies that take into account multiple performance indicators. It also avoids blind search in the population evolution process, improves the efficiency of solution space exploration, enhances the optimization efficiency of task allocation strategy, and meets the real-time scheduling requirements of the algorithm network.

[0140] As an example, Figure 4 A flowchart illustrating the strategy optimization method provided in this embodiment of the disclosure. Figure 4 As shown, the archive is equivalent to the population information (such as the current population information) in the above embodiments, and the training samples are equivalent to multiple allocation strategies and their corresponding objective function values ​​selected from the population information. Multiple population evolutions are performed through multiple iterations. In each iteration, i.e., during each population evolution process, it is determined whether the current iteration number is less than 'a' (e.g., K in the aforementioned embodiments). If it is less than 'a', a diffusion graph-based evolutionary strategy (i.e., the diffusion graph-based evolutionary algorithm in the aforementioned embodiments) is used for the current evolution; otherwise, a mutual information-based evolutionary operation (i.e., the mutual information-based evolutionary algorithm in the aforementioned embodiments) is used for the current evolution. In the current evolution, the true evaluations (i.e., objective function values) of the selected individuals (i.e., allocation strategies) can be obtained. Based on the selected individuals and their true evaluations, the archive and training samples are updated for the next iteration, i.e., the next population evolution. After the iteration ends, the final archive (i.e., the target population information) can be output (i.e., obtained).

[0141] likeFigure 4 As shown, in the process of the current evolution using a diffusion graph-based evolution strategy, the evolution strategy in the original space (i.e., the strategy space in the aforementioned embodiment) can be mapped to multiple subspaces, which are the diffusion graph spaces in the aforementioned embodiment. Then, crossover and mutation operations are performed to obtain multiple new individuals. Environmental selection is performed on the multiple new individuals, and the selected individuals are mapped to the original space.

[0142] like Figure 4 As shown, in the evolutionary operation based on mutual information, during the current evolution, the evolutionary strategy in the original space can be divided into multiple subgroups based on mutual information, namely the multiple sub-strategy groups in the aforementioned embodiment; crossover and mutation operations are performed based on multiple subgroups to obtain multiple new individuals; environmental selection is performed on multiple new individuals; and the selected individuals are spliced ​​together and mapped to the original space.

[0143] Corresponding to the strategy optimization method in the above embodiments, Figure 5 This is a structural block diagram of a strategy optimization apparatus provided in an embodiment of this disclosure. For ease of explanation, only the parts relevant to the embodiments of this disclosure are shown. (Refer to...) Figure 5 The strategy optimization device 500 includes:

[0144] The acquisition unit 501 is used to acquire initial population information for task allocation in the computing power network. The initial population information includes multiple allocation strategies and the objective function values ​​corresponding to the multiple allocation strategies. The objective function values ​​corresponding to the multiple allocation strategies are determined by a set multi-objective function. Each allocation strategy includes the identifier of the computing power node to which multiple tasks are allocated.

[0145] Evolutionary unit 502 is used to perform multiple population evolutions based on a multi-objective function and initial population information, using a diffusion graph-based evolutionary algorithm and a mutual information grouping-based evolutionary algorithm, to obtain the target population information.

[0146] The strategy selection unit 503 is used to assign and select the final strategy for nodes of multiple tasks in the target population information.

[0147] In some embodiments, during multiple population evolutions, in the first to Kth evolutions, a diffusion graph-based evolutionary algorithm is used for multiple population evolutions, and in the K+1 to Lth evolutions, a mutual information grouping-based evolutionary algorithm is used for multiple population evolutions, where L is the total number of evolutions and K is a set threshold.

[0148] In some embodiments, the evolutionary unit 502 is specifically used for: selecting multiple original strategies and their corresponding objective function values ​​from the current population information; when the current evolution is the first evolution, the current population information is the initial population information; when the current evolution is not the first evolution, the current population information is the intermediate population information after the previous evolution; based on the similarity between the multiple original strategies, projecting the multiple original strategies from the policy space onto the diffusion graph space to obtain the multidimensional feature key values ​​corresponding to the multiple original strategies, and forming diffusion graphs corresponding to the multiple original strategies based on the multidimensional feature key values ​​corresponding to the multiple original strategies; performing crossover and mutation operations on the multidimensional feature key values ​​corresponding to the multiple original strategies in the diffusion graph space to obtain the multidimensional feature key values ​​corresponding to the multiple new strategies; based on the policy space and... The mapping relationship in the diffusion graph space is used to transform the multidimensional feature key values ​​corresponding to multiple new strategies into the policy space, resulting in multiple new strategies. Based on the multidimensional feature key values ​​corresponding to the original strategies and the new strategies, target value prediction is performed on both strategies, resulting in predicted target values ​​for both strategies. Based on these predicted target values, multiple evolutionary strategies are selected. The objective function values ​​for each evolutionary strategy are determined using a multi-objective function. Finally, the evolutionary strategies and their corresponding objective function values ​​are added to the current population information to obtain the population information after the current evolution.

[0149] In some embodiments, the evolution unit 502 is specifically used to: determine a first similarity matrix and a second similarity matrix based on multiple original strategies, wherein the first similarity matrix includes the similarity between every two original strategies and the second similarity matrix includes the total similarity corresponding to each of the multiple original strategies; determine a probability matrix based on the first similarity matrix and the second similarity matrix, wherein the element value located in the i-th row and j-th column of the probability matrix represents the probability of the i-th original strategy transitioning to the j-th original strategy; and perform eigenvalue decomposition on the probability matrix to obtain the multidimensional feature key values ​​corresponding to each of the multiple original strategies.

[0150] In some embodiments, the evolution unit 502 is specifically used to: train a data mapping model using the multidimensional feature key values ​​corresponding to the multiple original policies and the multiple original policies as training data; and through the data mapping model, transform the multidimensional feature key values ​​corresponding to the multiple new policies from the diffusion graph space to the policy space to obtain multiple new policies.

[0151] In some embodiments, the evolution unit 502 is specifically used to: train a first prediction model using multidimensional feature key values ​​corresponding to multiple original strategies and objective function values ​​corresponding to multiple original strategies as training data; the first prediction model is used to predict the target value of the strategy to be predicted based on the multidimensional feature key values ​​of the strategy to be predicted; input the multidimensional feature key values ​​corresponding to multiple original strategies into the first prediction model, and predict the target value of multiple original strategies through the first prediction model to obtain the predicted target value corresponding to multiple original strategies; input the multidimensional feature key values ​​corresponding to multiple new strategies into the first prediction model, and predict the target value of multiple new strategies through the first prediction model to obtain the predicted target value corresponding to multiple new strategies.

[0152] In some embodiments, the evolution unit 502 is specifically used to: select multiple original strategies and their corresponding objective function values ​​from the current population information, wherein the current population information is the intermediate population information after the last evolution; normalize and calculate the mutual information of the multiple computing node identifiers included in the multiple original strategies and the objective function values ​​corresponding to the multiple original strategies, thereby obtaining the mutual information between multiple strategy variables and multiple objective functions pairwise, wherein the multiple computing node identifiers included in the multiple original strategies are the variable values ​​of the multiple strategy variables in the multiple original strategies; and group the multiple strategy variables according to the mutual information between the multiple strategy variables and multiple objective functions pairwise to obtain multiple sub-strategies. The process involves: generating multiple new sub-policies by performing intra-group crossover mutations on each sub-policy group; predicting the target values ​​of the sub-policies within each sub-policy group; selecting sub-policies from the sub-policy groups based on the predicted target values; concatenating these target sub-policies to obtain multiple evolutionary policies; determining the objective function values ​​for each evolutionary policy using a multi-objective function; and adding the evolutionary policies and their corresponding objective function values ​​to the current population information to obtain the population information after the current evolution.

[0153] The strategy optimization device provided in this embodiment can be used to execute the technical solutions of the above-described strategy optimization method embodiments. Its implementation principle and technical effects are similar, and will not be repeated here.

[0154] refer to Figure 6The diagram illustrates a structural schematic of an electronic device 600 suitable for implementing embodiments of the present disclosure. The electronic device 600 can be a terminal device or a server. The terminal device can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, personal digital assistants (PDAs), portable Android devices (PADs), portable media players (PMPs), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0155] like Figure 6 As shown, electronic device 600 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage device 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of electronic device 600. The processing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0156] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 An electronic device 600 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0157] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, it performs the functions defined in the methods of embodiments of this disclosure.

[0158] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0159] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0160] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods shown in the above embodiments.

[0161] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0162] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0163] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The names of the units are not necessarily limiting in certain circumstances; for example, an acquisition unit can also be described as "a unit that acquires the page image and page description text of a webpage to be detected".

[0164] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0165] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0166] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0167] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0168] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A strategy optimization method, characterized in that, include: For task allocation in a computing power network, initial population information is obtained. The initial population information includes multiple allocation strategies and objective function values ​​corresponding to the multiple allocation strategies. The objective function values ​​corresponding to the multiple allocation strategies are determined by a set multi-objective function. Each allocation strategy includes the identifier of the computing power node to which multiple tasks are allocated. Based on the multi-objective function and the initial population information, multiple population evolutions are performed using a diffusion graph-based evolutionary algorithm and a mutual information grouping-based evolutionary algorithm to obtain the target population information. In the multiple population evolutions, the diffusion graph-based evolutionary algorithm is used in the first to Kth evolutions, and the mutual information grouping-based evolutionary algorithm is used in the (K+1)th to Lth evolutions, where L is the total number of evolutions and K is a set threshold. In the target population information, a final strategy is selected for the nodes of the multiple tasks; A single population evolution process using the aforementioned diffusion graph-based evolutionary algorithm includes: From the current population information, select multiple original strategies and the objective function values ​​corresponding to the multiple original strategies respectively. If the current evolution is the first evolution, the current population information is the initial population information. If the current evolution is not the first evolution, the current population information is the intermediate population information after the previous evolution. Based on the similarity between the multiple original strategies, the multiple original strategies are projected from the policy space onto the diffusion graph space to obtain the multidimensional feature key values ​​corresponding to the multiple original strategies respectively. The multidimensional feature key values ​​corresponding to the multiple original strategies respectively form the diffusion graphs corresponding to the multiple original strategies respectively. In the diffusion graph space, crossover and mutation operations are performed on the multidimensional feature key values ​​corresponding to the multiple original strategies to obtain the multidimensional feature key values ​​corresponding to the multiple new strategies. Based on the mapping relationship between the policy space and the diffusion graph space, the multidimensional feature key values ​​corresponding to the multiple new policies are transformed to the policy space to obtain the multiple new policies; Based on the multidimensional feature key values ​​corresponding to the multiple original strategies and the multiple new strategies respectively, target value prediction is performed on the multiple original strategies and the multiple new strategies respectively to obtain the predicted target values ​​corresponding to the multiple original strategies and the predicted target values ​​corresponding to the multiple new strategies respectively. Based on the predicted target values ​​corresponding to the multiple original strategies and the predicted target values ​​corresponding to the multiple new strategies, multiple evolutionary strategies are selected from the multiple original strategies and the multiple new strategies; The objective function values ​​corresponding to the multiple evolutionary strategies are determined using the multi-objective function. The multiple evolutionary strategies and their corresponding objective function values ​​are added to the current population information to obtain the population information after the current evolution.

2. The strategy optimization method according to claim 1, characterized in that, The step of projecting the multiple original policies from the policy space onto the diffusion graph space based on the similarity between the multiple original policies to obtain the multidimensional feature key values ​​corresponding to the multiple original policies respectively includes: Based on the plurality of original strategies, a first similarity matrix and a second similarity matrix are determined. The first similarity matrix includes the similarity between every two original strategies, and the second similarity matrix includes the total similarity corresponding to each of the plurality of original strategies. Based on the first similarity matrix and the second similarity matrix, a probability matrix is ​​determined. The element value in the i-th row and j-th column of the probability matrix represents the probability that the i-th original strategy will be transferred to the j-th original strategy among the multiple original strategies. The probability matrix is ​​subjected to eigenvalue decomposition to obtain the multidimensional feature key values ​​corresponding to the multiple original strategies.

3. The strategy optimization method according to claim 1, characterized in that, The process of transforming the multidimensional feature key values ​​corresponding to the plurality of new strategies to the policy space based on the mapping relationship between the policy space and the diffusion graph space, thereby obtaining the plurality of new strategies, includes: The data mapping model is trained using the multidimensional feature key values ​​corresponding to the multiple original strategies and the multiple original strategies as training data. The data mapping model is used to transform the multidimensional feature key values ​​corresponding to the multiple new strategies from the diffusion graph space to the strategy space, thereby obtaining the multiple new strategies.

4. The strategy optimization method according to claim 1, characterized in that, The step of predicting target values ​​for the multiple original strategies and the multiple new strategies based on their respective multidimensional feature key values ​​and the multidimensional feature key values, to obtain the predicted target values ​​for the multiple original strategies and the multiple new strategies, includes: The first prediction model is trained using the multidimensional feature key values ​​corresponding to the multiple original strategies and the objective function values ​​corresponding to the multiple original strategies as training data. The first prediction model is used to predict the target value of the strategy to be predicted based on the multidimensional feature key values ​​of the strategy to be predicted. The multidimensional feature key values ​​corresponding to the multiple original strategies are input into the first prediction model, and the target value of the multiple original strategies is predicted by the first prediction model to obtain the predicted target value corresponding to the multiple original strategies. The multidimensional feature key values ​​corresponding to the multiple new strategies are input into the first prediction model, and the target values ​​of the multiple new strategies are predicted by the first prediction model to obtain the predicted target values ​​corresponding to the multiple new strategies.

5. The strategy optimization method according to claim 1, characterized in that, A single population evolution process using the aforementioned mutual information-based grouping evolution algorithm includes: From the current population information, select multiple original strategies and the objective function values ​​corresponding to the multiple original strategies respectively. The current population information is the intermediate population information after the last evolution. The multiple computing node identifiers included in the multiple original strategies and the objective function values ​​corresponding to the multiple original strategies are normalized and their mutual information is calculated to obtain the mutual information between multiple strategy variables and multiple objective functions. The multiple computing node identifiers included in the multiple original strategies are the variable values ​​of the multiple strategy variables in the multiple original strategies. Based on the mutual information between the multiple policy variables and the multiple objective functions, the multiple policy variables are grouped to obtain multiple sub-policy groups; By performing intra-group crossover mutation on the multiple sub-strategy groups, multiple new sub-strategies are generated in the multiple sub-strategy groups respectively. Target value prediction is performed on the sub-strategies in the plurality of sub-strategy groups to obtain the predicted target values ​​corresponding to the sub-strategies in the plurality of sub-strategy groups; Based on the predicted target value corresponding to the sub-strategy in the plurality of sub-strategy groups, sub-strategy selection is performed in the plurality of sub-strategy groups to obtain a plurality of target sub-strategies; The multiple target sub-strategies are concatenated to obtain multiple evolutionary strategies; The objective function values ​​corresponding to the multiple evolutionary strategies are determined using the multi-objective function. The multiple evolutionary strategies and their corresponding objective function values ​​are added to the current population information to obtain the population information after the current evolution.

6. A strategy optimization device, characterized in that, include: The acquisition unit is used to acquire initial population information for task allocation in the computing power network. The initial population information includes multiple allocation strategies and objective function values ​​corresponding to the multiple allocation strategies. The objective function values ​​corresponding to the multiple allocation strategies are determined by a set multi-objective function. Each allocation strategy includes the identifier of the computing power node to which multiple tasks are allocated. An evolutionary unit is used to perform multiple population evolutions based on the multi-objective function and the initial population information, using an evolutionary algorithm based on a diffusion graph and an evolutionary algorithm based on mutual information grouping, to obtain target population information. In the multiple population evolutions, the evolutionary algorithm based on the diffusion graph is used in the first to Kth evolutions, and the evolutionary algorithm based on mutual information grouping is used in the (K+1)th to Lth evolutions, where L is the total number of evolutions and K is a set threshold. The strategy selection unit is used to assign and select the final strategy for the nodes of the multiple tasks from the target population information. The evolutionary unit, in a single population evolution process using the diffusion graph-based evolutionary algorithm, includes: From the current population information, select multiple original strategies and the objective function values ​​corresponding to the multiple original strategies respectively. If the current evolution is the first evolution, the current population information is the initial population information. If the current evolution is not the first evolution, the current population information is the intermediate population information after the previous evolution. Based on the similarity between the multiple original strategies, the multiple original strategies are projected from the policy space onto the diffusion graph space to obtain the multidimensional feature key values ​​corresponding to the multiple original strategies respectively. The multidimensional feature key values ​​corresponding to the multiple original strategies respectively form the diffusion graphs corresponding to the multiple original strategies respectively. In the diffusion graph space, crossover and mutation operations are performed on the multidimensional feature key values ​​corresponding to the multiple original strategies to obtain the multidimensional feature key values ​​corresponding to the multiple new strategies. Based on the mapping relationship between the policy space and the diffusion graph space, the multidimensional feature key values ​​corresponding to the multiple new policies are transformed to the policy space to obtain the multiple new policies; Based on the multidimensional feature key values ​​corresponding to the multiple original strategies and the multiple new strategies respectively, target value prediction is performed on the multiple original strategies and the multiple new strategies respectively to obtain the predicted target values ​​corresponding to the multiple original strategies and the predicted target values ​​corresponding to the multiple new strategies respectively. Based on the predicted target values ​​corresponding to the multiple original strategies and the predicted target values ​​corresponding to the multiple new strategies, multiple evolutionary strategies are selected from the multiple original strategies and the multiple new strategies; The objective function values ​​corresponding to the multiple evolutionary strategies are determined using the multi-objective function. The multiple evolutionary strategies and their corresponding objective function values ​​are added to the current population information to obtain the population information after the current evolution.

7. The strategy optimization apparatus according to claim 6, characterized in that, When the evolutionary unit projects the multiple original policies from the policy space to the diffusion graph space based on the similarity between the multiple original policies to obtain the multidimensional feature key values ​​corresponding to the multiple original policies, it includes: Based on the plurality of original strategies, a first similarity matrix and a second similarity matrix are determined. The first similarity matrix includes the similarity between every two original strategies, and the second similarity matrix includes the total similarity corresponding to each of the plurality of original strategies. Based on the first similarity matrix and the second similarity matrix, a probability matrix is ​​determined. The element value in the i-th row and j-th column of the probability matrix represents the probability that the i-th original strategy will be transferred to the j-th original strategy among the multiple original strategies. The probability matrix is ​​subjected to eigenvalue decomposition to obtain the multidimensional feature key values ​​corresponding to the multiple original strategies.

8. The strategy optimization apparatus according to claim 6, characterized in that, When the evolutionary unit transforms the multidimensional feature key values ​​corresponding to the multiple new strategies to the policy space based on the mapping relationship between the policy space and the diffusion graph space, to obtain the multiple new strategies, the process includes: The data mapping model is trained using the multidimensional feature key values ​​corresponding to the multiple original strategies and the multiple original strategies as training data. The data mapping model is used to transform the multidimensional feature key values ​​corresponding to the multiple new strategies from the diffusion graph space to the strategy space, thereby obtaining the multiple new strategies.

9. An electronic device, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the strategy optimization method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by the processor, implement the strategy optimization method as described in any one of claims 1 to 5.

Images