Multi-target edge computing task scheduling method and device, equipment and medium

Abstract

The application provides a multi-target edge computing task scheduling method, device, equipment and medium, and relates to the fields of edge computing task scheduling and communication network resource allocation. The method comprises the following steps: obtaining task attribute information, node state information and network state information of each to-be-scheduled task in a current scheduling period, so as to construct a corresponding multi-target optimization model; generating an initial candidate scheduling scheme set according to task allocation constraints of each to-be-scheduled task and each candidate execution node; adopting a hybrid expert agent model to perform performance prediction on each candidate scheduling scheme in the initial candidate scheduling scheme set, so as to obtain a predicted performance result; performing pre-screening on the initial candidate scheduling scheme set according to the predicted performance result, so as to obtain an optimal candidate scheme subset; performing real performance evaluation and screening on the optimal candidate scheme subset according to the multi-target optimization model, so as to generate a target scheduling decision of the current scheduling period; and the technical effect of improving the multi-target collaborative scheduling capability is achieved.

Description

Technical Field

[0001] This application relates to the field of edge computing task scheduling and communication network resource allocation, and in particular to a method, apparatus, device and medium for multi-objective edge computing task scheduling. Background Technology

[0002] In edge computing scenarios, tasks are typically scheduled through a collaborative offloading mechanism involving local, edge node, and cloud environments to meet the computing and latency requirements of different business operations.

[0003] However, existing scheduling schemes mostly rely on rule-based strategies, single agent prediction, or traditional evolutionary search, which makes it difficult to simultaneously consider latency, energy consumption, and link occupancy when node states, link conditions, and task characteristics change dynamically. At the same time, the performance evaluation accuracy for heterogeneous tasks is limited, and the computational overhead of evaluating a large number of candidate schemes one by one is high, affecting the real-time performance and stability of scheduling.

[0004] Therefore, how to improve the multi-objective collaborative scheduling capability of heterogeneous tasks in a dynamic edge computing environment has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a multi-objective edge computing task scheduling method, apparatus, device, and medium to solve the aforementioned technical problems. This method addresses the needs of task offloading and resource collaborative allocation in dynamic edge computing environments. It comprehensively considers the tasks to be scheduled, candidate execution nodes, and the network association status between tasks and nodes. Around multiple optimization objectives such as service latency, system energy consumption, and link resource usage, it jointly models, evaluates, and optimizes task allocation decisions within the current scheduling cycle, thereby improving the multi-objective collaborative scheduling capability in heterogeneous task scenarios.

[0006] Firstly, this application provides a multi-objective edge computing task scheduling method, including:

[0007] Obtain task attribute information for each task to be scheduled within the current scheduling period, node status information for each candidate execution node, and network status information between the task to be scheduled and the candidate execution nodes;

[0008] Based on task attribute information, node status information, and network status information, a corresponding multi-objective optimization model is constructed. The multi-objective optimization model includes an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, as well as task latency constraints and node queuing stability constraints.

[0009] Based on the task allocation constraints of each task to be scheduled and each candidate execution node, an initial set of candidate scheduling schemes is generated, which contains multiple candidate scheduling schemes.

[0010] A pre-built hybrid expert agent model is used to predict the performance of each candidate scheduling scheme in the initial candidate scheduling scheme set, and the predicted performance results are obtained.

[0011] Based on the predicted performance results, the initial set of candidate scheduling schemes is pre-screened to obtain a subset of preferred candidate schemes;

[0012] The optimal candidate scheme subset is evaluated and screened based on the multi-objective optimization model to generate the target scheduling decision for the current scheduling cycle.

[0013] Secondly, this application provides a multi-target edge computing task scheduling device, comprising:

[0014] The acquisition module is used to acquire the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution node within the current scheduling period.

[0015] The construction module is used to build a corresponding multi-objective optimization model based on task attribute information, node status information and network status information. The multi-objective optimization model includes an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, as well as task latency constraints and node queuing stability constraints.

[0016] The first generation module is used to generate an initial candidate scheduling scheme set based on the task allocation constraints of each task to be scheduled and each candidate execution node. The initial candidate scheduling scheme set contains multiple candidate scheduling schemes.

[0017] The prediction module is used to perform performance prediction on each candidate scheduling scheme in the initial candidate scheduling scheme set using a pre-built hybrid expert agent model, and obtain the predicted performance results.

[0018] The filtering module is used to pre-filter the initial candidate scheduling scheme set based on the predicted performance results to obtain a subset of preferred candidate schemes;

[0019] The second generation module is used to evaluate and screen the subset of preferred candidate schemes based on the multi-objective optimization model, and generate the target scheduling decision for the current scheduling cycle.

[0020] Thirdly, this application provides an electronic device, including: a memory and a processor;

[0021] The memory stores the instructions that the computer executes;

[0022] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0023] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0024] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0025] This application provides a multi-objective edge computing task scheduling method, apparatus, device, and medium. By acquiring the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution nodes within the current scheduling period, a multi-objective optimization model is constructed, including the objective function of average service latency, the objective function of total system energy consumption, the objective function of backhaul link occupancy cost, and corresponding constraints. An initial set of candidate scheduling schemes is generated by combining task allocation constraints. The performance of the candidate scheduling schemes is predicted and pre-screened using a pre-constructed hybrid expert agent model. Then, the optimal candidate scheme subset is evaluated and screened based on the real performance of the multi-objective optimization model. This approach can take into account the joint optimization of latency, energy consumption, and link occupancy in a dynamic edge computing environment, and reduce the computational overhead caused by the precise evaluation of a massive number of candidate schemes one by one, thereby improving the overall multi-objective collaborative scheduling capability of heterogeneous tasks. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0027] Figure 1 This application provides a schematic diagram of an application data processing system architecture.

[0028] Figure 2 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 1 ;

[0029] Figure 3 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 2 ;

[0030] Figure 4 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 3 ;

[0031] Figure 5 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 4 ;

[0032] Figure 6 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 5 ;

[0033] Figure 7 This is a schematic diagram of the structure of the multi-target edge computing task scheduling device provided in the embodiments of this application;

[0034] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0035] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0036] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0037] Edge computing task scheduling technology is widely used in business scenarios where local terminals, edge nodes and cloud nodes work together. It is especially suitable for scenarios that are sensitive to latency and computing power, such as augmented reality / virtual reality (AR / VR), real-time games, high-definition audio and video processing and smart wearable interaction.

[0038] In this type of scenario, user terminals continuously generate tasks with different data volumes, computing requirements, and latency requirements. These tasks can be executed locally on the terminal, offloaded to the edge node on the access side, or further transmitted to the cloud node for execution. The system typically relies on the wireless link from the terminal to the edge and the backhaul link from the edge to the cloud to complete the collaborative processing.

[0039] Existing edge computing scheduling schemes typically employ rule-driven offloading, single agent model prediction, or traditional evolutionary search methods to allocate tasks. The basic approach generally involves selecting a local, edge, or cloud execution location for a task based on link bandwidth, node load, task size, or local latency estimation. Some schemes also use agent models to predict scheme performance or use evolutionary search to find a better solution among multiple candidate schemes.

[0040] However, most of the aforementioned technical solutions struggle to simultaneously balance average service latency, total system energy consumption, and backhaul link occupancy in dynamic environments. Especially during periods of sudden task concurrency spikes, fluctuating node computing power, or link congestion, scheduling based solely on a single metric or coarse-grained predictions can easily lead to situations where some tasks experience reduced latency but increased backhaul link occupancy, or where reduced node energy consumption results in queuing and backlog at edge nodes.

[0041] Meanwhile, heterogeneous tasks differ significantly in computational and transmission dependencies, making it difficult to accurately assess the true performance of candidate solutions when using a unified prediction method. On the other hand, performing real calculations on a large number of candidate solutions one by one would result in high computational overhead, making it difficult for the scheduling process to meet real-time business needs, thereby affecting service stability and user experience.

[0042] In view of this, how to improve the multi-objective collaborative scheduling capability of heterogeneous tasks in a dynamic edge computing environment has become an urgent technical problem to be solved.

[0043] To address the aforementioned issues, this application provides a multi-objective edge computing task scheduling method, apparatus, device, and medium. Under a collaborative architecture comprised of user terminals, edge nodes, and cloud nodes, it acquires task attribute information for each task to be scheduled within the current scheduling cycle, node status information for each candidate execution node, and network status information between the task to be scheduled and the candidate execution nodes. Based on this, it constructs a multi-objective optimization model including an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, and corresponding constraints. Combined with task allocation constraints, it generates an initial set of candidate scheduling schemes. A pre-built hybrid expert agent model is used to predict and pre-screen the performance of the candidate scheduling schemes. Finally, based on the multi-objective optimization model, a subset of the preferred candidate schemes is evaluated and screened for real performance. This approach can simultaneously optimize latency, energy consumption, and link occupancy in a dynamic edge computing environment, reducing the computational overhead of accurately evaluating a massive number of candidate schemes one by one, thereby improving the overall multi-objective collaborative scheduling capability for heterogeneous tasks.

[0044] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0045] Figure 1 This is a schematic diagram of an application data processing system architecture provided in an embodiment of this application. The application data processing system is a computer device. Figure 1 As shown, the above architecture includes at least one of a data acquisition device 101, a processing device 102, and a display device 103.

[0046] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the architecture of the application data processing system. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. Figure 1 The components shown can be implemented in hardware, software, or a combination of both.

[0047] In the specific implementation process, the data acquisition device 101 may include an input / output interface or a communication interface, and the data acquisition device 101 can be connected to the processing device through the input / output interface or the communication interface.

[0048] The processing device 102 can first obtain three types of status information of tasks, nodes and networks within the current scheduling period, construct a multi-objective optimization model containing time delay, energy consumption, link cost objectives and corresponding constraints, generate an initial candidate scheduling scheme set that satisfies the allocation constraints, complete performance prediction and pre-screening through a hybrid expert agent model, and then generate the target scheduling decision for the current scheduling period through real evaluation and screening by the multi-objective model.

[0049] The display device 103 can also be a touch screen or the screen of a terminal device, used to receive user commands while displaying the above-mentioned content, so as to realize interaction with the user.

[0050] It should be understood that the aforementioned processing device can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0051] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0052] Figure 2 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the multi-target edge computing task scheduling method provided in this embodiment includes:

[0053] S201. Obtain the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution node within the current scheduling period.

[0054] For example, task attribute information is used to characterize the characteristics of each task to be scheduled, serving as the task-side input for subsequent construction of multi-objective optimization models and evaluation of candidate solutions. Specifically, it includes the amount of task data, the amount of task computation, and the maximum tolerable latency. Node status information is used to characterize the running status of each candidate execution node, serving as the node-side input for scheduling decisions and feasibility assessments. Specifically, it includes the node's available computing power, the number of tasks to be processed, and the computational energy consumption coefficient. Network status information is used to characterize the status of the communication link between the task to be scheduled and the candidate execution node, serving as the network-side input for calculating transmission latency, link occupancy, and backhaul cost. Specifically, it includes the link transmission rate, the link bandwidth between nodes, and the total bandwidth of the backhaul link.

[0055] In some implementations, auxiliary status parameters such as resource utilization and link congestion status may also be optionally obtained, but these are not necessary limitations of this step.

[0056] In this application, task attribute information, node status information and network status information can be collected and organized within the current scheduling cycle, and corresponding status information can be generated for use in subsequent steps.

[0057] Based on the above analysis, this step unifies the relevant states of tasks, nodes, and links as inputs for subsequent scheduling modeling and scheme evaluation, making scheduling decisions in dynamic environments comparable.

[0058] It should be understood that the above examples are merely illustrative and not limiting, and do not affect the scope of protection of this application.

[0059] S202. Based on the task attribute information, node status information, and network status information, construct the corresponding multi-objective optimization model.

[0060] In this embodiment, the multi-objective optimization model includes an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, as well as task latency constraints and node queuing stability constraints.

[0061] For example, a multi-objective optimization model is used to transform the task scheduling problem into an optimization problem that solves multiple objectives simultaneously under constraints. The average service latency objective function characterizes the overall service latency level of the tasks to be scheduled; the total system energy consumption objective function characterizes the total energy consumption of all tasks during allocation and execution within the current scheduling cycle; the backhaul link occupancy cost objective function characterizes the degree of link resource occupancy by tasks assigned to the corresponding execution nodes; the task latency constraint limits the actual processing latency of a single task to no more than the corresponding maximum tolerable latency; and the node queuing stability constraint limits the task arrival rate on candidate execution nodes to no more than their service rate, thereby determining the range of feasible solutions.

[0062] Optionally, the allocation relationship between tasks and candidate execution nodes can be defined to represent the task scheduling result in the candidate scheduling scheme.

[0063] For any candidate scheduling scheme, the service latency of a task on the corresponding candidate execution node can be decomposed into the sum of transmission latency, queuing latency, and execution latency. Transmission latency can be determined based on the task data volume and the corresponding link transmission rate; execution latency can be determined based on the task's computational requirements and the node's available computing power; and queuing latency can be estimated based on the node's queue status. The average service latency objective function can be expressed as the statistical result of the service latency of all tasks to be scheduled. The total system energy consumption objective function consists of the execution energy consumption of each task on the corresponding execution node and the necessary transmission energy consumption. Computational energy consumption can be determined by the task's computational requirements and the node's computational energy consumption coefficient; and transmission energy consumption can be estimated based on the task data volume and link transmission status. The backhaul link occupancy cost objective function can be defined based on the degree of occupancy of backhaul link resources by the tasks assigned to relevant nodes.

[0064] In terms of constraint construction, the task latency constraint requires that for each task, its actual service latency is not greater than the maximum tolerable latency of that task; the node queuing stability constraint requires that for each candidate execution node, the arrival rate of the tasks assigned to that node is less than its service rate.

[0065] When the above constraints are not met, the corresponding candidate solution is not feasible. Based on the above analysis, it can be seen that by incorporating task latency, node energy consumption, and backhaul link resources into a unified model, and by limiting the feasible domain through latency constraints and queuing stability constraints, the subsequent screening process can simultaneously exclude timeout solutions and congestion solutions, so that the scheduling result no longer depends on a single indicator.

[0066] It should be understood that the above examples are merely illustrative and not limiting, and do not affect the scope of protection of this application.

[0067] S203. Generate an initial set of candidate scheduling schemes based on the task allocation constraints of each task to be scheduled and each candidate execution node.

[0068] In this embodiment, the initial candidate scheduling scheme set contains multiple candidate scheduling schemes.

[0069] For example, task allocation constraints are used to define the allocation relationship between the task to be scheduled and the candidate execution nodes. Candidate scheduling schemes represent a specific task allocation method. The initial set of candidate scheduling schemes carries multiple candidate scheduling schemes and serves as input for subsequent performance prediction, pre-screening, and actual performance evaluation. The basic form of task allocation constraints is that each task to be scheduled satisfies a preset allocation relationship with the candidate execution nodes within the current scheduling cycle.

[0070] The system can construct an initial set of candidate scheduling schemes based on the task allocation constraints between each task to be scheduled and each candidate execution node. Each candidate scheduling scheme in the initial set corresponds to a different task allocation method. After each candidate scheduling scheme is generated, it can be stored in the form of a task-to-node mapping and associated with the status information collected in that period.

[0071] This step forms a candidate solution space that satisfies the task allocation constraints, and this set is sent to the next step for performance prediction.

[0072] Based on the above analysis, it can be seen that the initial set of candidate scheduling schemes is not the final decision that is directly output, but a structured expansion of feasible allocation relationships, thereby supporting subsequent scheduling search.

[0073] It should be understood that the above examples are merely illustrative and not limiting, and do not affect the scope of protection of this application.

[0074] S204. Using a pre-built hybrid expert agent model, the performance of each candidate scheduling scheme in the initial candidate scheduling scheme set is predicted to obtain the predicted performance results.

[0075] For example, a hybrid expert agent model is used to predict the low-cost performance of candidate scheduling schemes, mainly outputting two predicted values: average service latency and total system energy consumption. The calculation logic for backhaul link occupancy cost is simple and does not require agent model prediction; it can be directly calculated based on the task data volume and backhaul link bandwidth parameters. These three factors together constitute the predicted performance results of the candidate scheduling schemes. This model consists of multiple expert prediction networks and a gating mechanism. The gating mechanism dynamically selects or weights multiple expert networks based on the input state. Each expert network processes the performance mapping relationship under different input conditions, and finally, the predicted performance results of each candidate scheduling scheme are output through weighted fusion.

[0076] It is worth noting that, compared to using a single agent model for performance prediction, this application uses a gating mechanism to dynamically divide tasks of different types, enabling different expert networks to learn the performance mapping relationship corresponding to different task features, thereby reducing prediction interference between heterogeneous tasks.

[0077] In one possible implementation, the hybrid expert agent model is constructed through the following steps:

[0078] Obtain real performance evaluation records of historical scheduling tasks; determine the local state features and real performance labels of the corresponding tasks based on the real performance evaluation records; construct the corresponding training sample set based on the local state features and real performance labels; obtain the initial hybrid expert network; iteratively train the initial hybrid expert network with prediction error loss and expert load balancing loss as joint optimization objectives until the loss converges, and obtain the hybrid expert agent model.

[0079] In this embodiment, the training of the hybrid expert agent model takes prediction error and expert load balancing as joint optimization objectives, and the corresponding formula for the total loss function is as follows:

[0080]

[0081] in, The total loss function for model training. This is the prediction error loss term, used to constrain the deviation between the predicted values ​​and the actual performance values ​​of the surrogate model. These are the weighting coefficients for the two types of loss terms, used to balance prediction accuracy and expert load balancing. This is the expert load balancing loss term, used to prevent gating networks from concentrating tasks on a small number of expert networks for an extended period.

[0082] For example, after obtaining the actual performance evaluation records of historical scheduling tasks, the task attribute information, node state information, and network state information in the records are aligned. Local state features related to candidate scheduling schemes are extracted, and corresponding actual performance labels are generated based on the actual execution results. These labels are then paired one-to-one to form a training sample set. Subsequently, the training sample set is input into the initial hybrid expert network, and predicted values ​​are obtained through forward propagation. The difference between the predicted values ​​and the actual performance labels is calculated as the prediction error loss. At the same time, the allocation ratio of samples received by each expert network is statistically analyzed to form the expert load balancing loss. The joint optimization objective is composed of the prediction error loss and the expert load balancing loss. During training, the expert network parameters and the gated network parameters are updated through backpropagation until the joint loss converges, thereby obtaining a stable hybrid expert agent model.

[0083] The aforementioned hybrid expert agent model can receive local input features corresponding to candidate scheduling schemes, output prediction results of performance indicators, and improve its fitting ability to heterogeneous task states through the division of labor learning among multiple expert networks. Since the training samples are directly derived from the real performance evaluation records of historical scheduling tasks, the correspondence between the model output and the real scheduling results is more explicit; due to the introduction of expert load balancing loss, the sample distribution among multiple expert networks during training is more even, and the model parameter updates are more stable.

[0084] In summary, the hybrid expert agent model can reduce the probability of expert network load imbalance while ensuring prediction accuracy, enabling it to more accurately represent the real performance results of different schemes in the subsequent candidate scheduling scheme selection, and providing a reliable performance prediction basis for multi-objective edge computing task scheduling.

[0085] S205. Based on the predicted performance results, the initial candidate scheduling scheme set is pre-screened to obtain a subset of preferred candidate schemes.

[0086] For example, pre-screening is used to quickly filter the initial set of candidate scheduling schemes based on predicted performance results and constraints. A preferred subset of candidate schemes is used to carry the better candidate scheduling schemes retained after the quick screening, and its size is smaller than the initial set of candidate scheduling schemes. The predicted performance results, as the basis for pre-screening, include at least the predicted average service latency, the predicted total system energy consumption, and the predicted backhaul link occupancy cost for each candidate scheduling scheme.

[0087] The system can first perform feasibility pre-filtering on the initial candidate scheduling scheme set based on constraints, eliminating candidate scheduling schemes that clearly do not meet the constraints; then, it can perform optimization screening on the remaining candidate scheduling schemes based on the predicted performance results to obtain a subset of optimal candidate schemes. Optimization screening can be based on a comprehensive comparison of multiple objectives.

[0088] In one possible embodiment, the size of the preferred subset of candidate schemes can be set based on the computing resources of the current scheduling device or the time budget of the scheduling cycle.

[0089] Through this step, the initial set of candidate scheduling schemes is compressed into a smaller subset of preferred candidate schemes that are closer to the final feasible optimal solution, and the real performance evaluation in the next step is triggered.

[0090] Based on the above analysis, the pre-screening step eliminates obviously infeasible and obviously inferior solutions based on the output of the proxy model, so that the real performance evaluation is focused only on a limited number of high-value candidate solutions, thereby enabling the entire scheduling process to adapt to the requirements of real-time business for periodic response.

[0091] It should be understood that the above examples are merely illustrative and not limiting, and do not affect the scope of protection of this application.

[0092] S206. Based on the multi-objective optimization model, the subset of preferred candidate schemes is evaluated and screened for real performance, and the target scheduling decision for the current scheduling cycle is generated.

[0093] For example, real performance evaluation is used to perform actual calculations on candidate scheduling schemes in the subset of preferred candidate schemes based on a multi-objective optimization model, in order to verify the prediction results and output reliable performance indicators; the target scheduling decision of the current scheduling cycle is used to represent the final output task allocation result and guide the actual execution of each task to be scheduled on the corresponding execution node. The screening process is based on the real performance evaluation results, and at the same time, it combines task latency constraints and node queuing stability constraints to make a final feasibility and quality judgment on the candidate schemes.

[0094] The system reads candidate scheduling schemes from the subset of preferred candidate schemes one by one, and performs real performance calculations based on a multi-objective optimization model to obtain the real performance indicators corresponding to each candidate scheduling scheme. After completing the calculation of the real performance indicators, the system checks the task latency constraints and node queuing stability constraints again, eliminates candidate scheduling schemes that do not meet the constraints, and selects the target scheduling decision for the current scheduling cycle from the remaining candidate schemes.

[0095] Once the target scheduling decision is determined, the corresponding task assignment or scheduling control operations can be executed based on the target scheduling decision. This step constitutes a closed loop from prediction and screening to final implementation. The aforementioned pre-screening only narrows the search scope under low-cost conditions, while this step verifies each retained solution through real model calculations to ensure that the final output target scheduling decision meets the requirements of multi-objective joint optimization and the corresponding constraints.

[0096] Based on the above analysis, it can be seen that this application reduces the computational power consumption during the cycle caused by performing real calculations on all candidate schemes in a dynamic edge computing environment through a two-stage solution process of prediction, pre-screening, and real performance evaluation. At the same time, the reliability and feasibility of the scheduling results are guaranteed by the final real verification.

[0097] It should be understood that the above examples are merely illustrative and not limiting, and do not affect the scope of protection of this application.

[0098] This application provides a multi-objective edge computing task scheduling method. By acquiring the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution nodes within the current scheduling period, a multi-objective optimization model is constructed, which includes an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, and corresponding constraints. An initial set of candidate scheduling schemes is generated by combining task allocation constraints. A pre-constructed hybrid expert agent model is used to predict and pre-screen the performance of the candidate scheduling schemes. Then, the optimal candidate scheme subset is evaluated and screened based on the multi-objective optimization model. This method can take into account the joint optimization of latency, energy consumption, and link occupancy in a dynamic edge computing environment, and reduce the computational overhead caused by the precise evaluation of a massive number of candidate schemes one by one. In this way, the multi-objective collaborative scheduling capability of heterogeneous tasks is improved as a whole.

[0099] Figure 3 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 2 ,like Figure 3 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the specific construction process of the multi-objective optimization model, including:

[0100] S301. Obtain the corresponding first parameter information based on the task attribute information.

[0101] In this embodiment, the first parameter information includes the amount of task input data, the amount of computation required, and the maximum tolerable latency parameter.

[0102] For example, this step extracts the core basic parameters required for modeling from the task dimension. The task input data volume is used to characterize the data scale transmitted across nodes in the task. The computational demand is used to characterize the amount of computing resources required for the task to complete the computation. The maximum tolerable latency parameter is the longest end-to-end processing time that the task can accept, which serves as the judgment benchmark and modeling input boundary for subsequent latency constraints.

[0103] Specifically, this step extracts the core basic parameters required for modeling from the task dimension, providing task-side input for the subsequent construction of objective functions and constraints. The attribute information of the i-th task to be scheduled within the current scheduling cycle can be uniformly represented in the form of triples, as shown in the following formula:

[0104]

[0105] in, For the first time in the current scheduling period One task to be scheduled For the first The amount of input data for each task to be scheduled is determined by the task generator based on the actual size of the data to be processed or uploaded. For the first The computational requirements of each task to be scheduled are estimated based on the task type, complexity parameters, and historical execution records. For the first The maximum tolerable latency for each pending task is pre-configured based on the task's real-time requirements and quality of service standards. The number of the task to be scheduled, with a value range of 1. , where N is the total number of tasks to be scheduled in the current scheduling period.

[0106] S302. Obtain the corresponding second parameter information based on the node status information.

[0107] In this embodiment, the second parameter information includes the node's available computing power, the number of tasks to be processed, and the computing energy consumption coefficient.

[0108] For example, this step extracts the core state parameters required for modeling from the node dimension. The available computing power of a node is used to characterize the effective computing power that the node can currently schedule. The number of tasks to be processed is used to reflect the current task backlog and load congestion of the node. The calculated energy consumption coefficient is used to characterize the energy consumption level corresponding to the unit computing power output of the node, providing data support for energy consumption target calculation and queuing stability determination.

[0109] Specifically, this step extracts the core state parameters required for modeling from the node dimension, providing node-side data support for energy consumption calculation, latency calculation, and stability determination. The available computing power of each candidate execution node is calculated by combining the total computing power of the node with the current resource utilization rate, as shown in the following formula:

[0110]

[0111] in, For the first The available computing power value of each candidate execution node represents the effective computing power that the node can currently be scheduled for. For the first The rated total computing power of each candidate execution node can be obtained directly from the node's operating system resource interface. For the first The resource utilization rate of each candidate execution node is obtained by the node status acquisition module, which acquires the current computing and storage resource usage. This is the candidate execution node's ID, with a value range of [value range missing]. M represents the total number of candidate execution nodes. At that time, it represents the local node of the user terminal. Represents edge computing nodes. Represents a cloud computing node.

[0112] The number of pending tasks q obtained in this step is synchronized. j The current inventory of tasks waiting to be processed by the node can be read directly from the node's task queue; calculate the energy consumption coefficient k. j The hardware power consumption characteristics of the node processor are pre-configured, with values ​​ranging from

[10] . -28 10 -26 ].

[0113] S303. Obtain the corresponding third parameter information based on the network status information.

[0114] In this embodiment, the third parameter information includes the link transmission rate, the inter-node link bandwidth, and the total backhaul link bandwidth.

[0115] For example, this step extracts the core link parameters required for modeling from the network dimension. The link transmission rate is used to calculate the time taken for tasks to be transmitted between nodes, the link bandwidth between nodes is used to limit the upper limit of the transmission capacity for task migration and offloading, and the total bandwidth of the backhaul link is used to characterize the overall carrying capacity of the link from the edge node to the cloud backbone, providing a basis for the quantitative calculation of link occupancy costs.

[0116] Specifically, this step extracts the core link parameters required for modeling from the network dimension, providing network-side input for transmission delay calculation, migration cost accounting, and link occupancy quantification. This can be represented as a task. Corresponding candidate execution node The link transmission rate can be obtained through the gateway, routing device, or operating system network monitoring interface. It can be represented as the original execution node. With the target candidate execution node The link bandwidth between them is used to calculate the transmission time during task migration. It can be expressed as the maximum total bandwidth of the backhaul links available in the edge computing environment, representing the overall carrying capacity of the link from the edge node to the cloud backbone.

[0117] By integrating the three types of information mentioned above—task attributes, node status, and network link—we can obtain the task. Relative to candidate execution nodes The joint state vector serves as the unified input basis for multi-objective modeling and surrogate prediction, and the corresponding formula is as follows:

[0118]

[0119] in, For the task Relative to candidate execution nodes The joint state vector fully covers all the basic state information required for task scheduling decisions.

[0120] S304. Based on the first parameter information, the second parameter information, and the third parameter information, construct the corresponding average service delay objective function, the total system energy consumption objective function, the backhaul link occupancy cost objective function, as well as the task delay constraint and the node queuing stability constraint.

[0121] For example, when constructing the average service latency objective function, the task transmission latency is calculated based on the task input data volume and link transmission rate, and the task computation latency is calculated based on the computational demand and available computing power of the nodes. The node queuing latency and task migration latency are then combined to form the expression for calculating the average service latency. When constructing the system total energy consumption objective function, the computational energy consumption coefficient of each node and the task computational demand are combined to summarize the computational and transmission energy consumption of all nodes, forming the expression for calculating the system total energy consumption. When constructing the backhaul link occupancy cost objective function, the resource occupancy ratio of the backhaul link is quantified by combining the task data volume offloaded to the cloud and the total bandwidth of the backhaul link, forming the expression for calculating the link occupancy cost. The task latency constraint is set to ensure that the total processing latency of a single task does not exceed the corresponding maximum tolerable latency parameter. The node queuing stability constraint is set based on the ratio of the node's pending task arrival rate to the node's service rate, ensuring that the node task queue does not experience infinite backlog.

[0122] Specifically, this step, based on three types of fundamental parameters, sequentially constructs computational expressions for three types of optimization objectives and two types of constraint decision rules, providing the core computational logic for the multi-objective optimization model. Before construction, binary decision variables are first defined to uniformly represent the allocation relationship between tasks and nodes:

[0123]

[0124] By all The resulting binary decision matrix is ​​denoted as Each set of binary decision matrices corresponds to a candidate scheduling scheme. To ensure the validity of the scheduling scheme, a task integrity constraint is set to ensure that each task is assigned to only one execution node within a scheduling cycle:

[0125]

[0126] in, For the task Assigned to nodes The binary decision variable is used to mark the node affiliation of a task, and X is the binary decision matrix corresponding to the candidate scheduling scheme, which contains the allocation results of all tasks in the current scheduling period.

[0127] Furthermore, the objective function for average service latency is constructed as follows:

[0128] Task When assigned to candidate execution node j, the total service latency consists of four parts: transmission latency, queuing latency, execution latency, and migration latency, as shown in the following formula:

[0129]

[0130] in, For the task Assigned to nodes Total end-to-end service latency at that time For the task Transmission from terminal to node Transmission delay, For the task At the node Queuing delay in the task queue For the task At the node The computation execution latency on the above, For the task Migration latency that occurs when a cross-node switch occurs.

[0131] The specific calculation methods for each component's delay are as follows:

[0132] Transmission delay: Calculated as the ratio of the amount of input data to the corresponding link transmission rate.

[0133]

[0134] Execution latency: Calculated by the ratio of the task's computational requirements to the actual available computing power allocated to the task by the node.

[0135]

[0136] Among them, the actual available computing power of a node allocated to a single task This is obtained by evenly distributing the node's available computing power among all assigned tasks on that node:

[0137]

[0138] in, Candidate execution nodes Assigned to task The actual available computing power.

[0139] Migration latency: Occurs only when the execution node of the task is inconsistent between the previous scheduling cycle and the current scheduling cycle, and consists of the time spent on migration data transmission and the additional time spent on state switching.

[0140]

[0141] in, This is an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. It is used to determine whether a node switch has occurred in the task. For the task The execution node number corresponding to the previous scheduling cycle is obtained from the historical scheduling record. For the task The amount of status data that needs to be synchronously transmitted during migration is determined by the size of the data to be transferred from the original execution node. The additional state transition time incurred for switching task nodes includes the time overhead of pausing, saving and resuming task execution.

[0142] Queuing delay: Calculated based on a queuing theory model, it is obtained as the reciprocal of the difference between the node service rate and the task arrival rate.

[0143]

[0144] Among them, the node task arrival rate The number of tasks assigned to this node within a unit scheduling cycle is calculated as follows:

[0145]

[0146] Average service rate of nodes It is calculated from the ratio of the available computing power of a node to the average computing demand of the assigned tasks:

[0147]

[0148] Average computational requirements for node-assigned tasks for:

[0149]

[0150] in, Candidate execution nodes The task arrival rate represents the number of tasks arriving at that node per unit of time. Candidate execution nodes The average service rate represents the number of tasks that the node can process per unit of time. The duration of a single scheduling cycle. Candidate execution nodes The average computational requirement for all tasks currently allocated.

[0151] Based on the total service latency of a single task, the average service latency of all tasks within the current scheduling period is taken to obtain the objective function for average service latency:

[0152]

[0153] in, The objective function for average service latency characterizes the overall real-time performance of the scheduling scheme, and the optimization direction is to minimize this value.

[0154] Furthermore, the objective function for the total energy consumption of the system is constructed as follows:

[0155] The energy consumption during task execution includes two parts: computing energy consumption and transmission energy consumption. Computing energy consumption is determined by the node's computing power, energy consumption coefficient, and task computation amount, while transmission energy consumption is determined by the terminal's transmission power, data volume, and transmission rate.

[0156] The formulas for calculating energy consumption and transmitting energy consumption are as follows:

[0157]

[0158] in, For the task At the node The computational energy consumed during execution Candidate execution nodes The computational energy consumption coefficient is determined by the hardware characteristics of the node processor. For the task Upload to node The resulting transmission energy consumption, Perform tasks for user terminals The transmission power during data upload is determined by the terminal hardware configuration.

[0159] The computational and transmission energy consumption of all tasks within the current scheduling period are summarized to obtain the system's total energy consumption objective function:

[0160]

[0161] in, Let be the objective function for the total energy consumption of the system, which characterizes the overall energy consumption level of the scheduling scheme, and the optimization direction is to minimize this value.

[0162] Furthermore, the objective function for backhaul link occupancy cost is constructed as follows:

[0163] By statistically analyzing the proportion of all task data volume offloaded to the cloud to the total backhaul bandwidth, the resource consumption of the backhaul link is quantified, and the corresponding formula is as follows:

[0164]

[0165] in, Let be the objective function for backhaul link occupancy cost, representing the proportion of resources occupied by the scheduling scheme on the backbone backhaul links. The optimization direction is to minimize this value. This is a set of cloud node IDs, corresponding to all cloud candidate execution nodes that can participate in task execution.

[0166] Furthermore, the constraints are constructed:

[0167] Task latency constraints: Ensure that the total processing latency of a single task does not exceed its maximum tolerable latency, thus meeting the task's real-time requirements.

[0168]

[0169] Node queuing stability constraints: Ensure that the task arrival rate of a node is lower than its service rate, thus preventing the node task queue from accumulating indefinitely.

[0170]

[0171] S305. Based on the objective functions of average service delay, total system energy consumption, backhaul link occupancy cost, task delay constraints, and node queuing stability constraints, construct a multi-objective optimization model.

[0172] For example, this step integrates the three types of objective functions and two types of constraints into a unified optimization framework, forming a complete multi-objective optimization model. This model uses a binary decision matrix X as the optimization variable, minimizes the average service latency, total system energy consumption, and backhaul link occupancy cost as joint optimization objectives, and uses task latency and node queuing stability as mandatory boundary constraints. All candidate scheduling schemes undergo unified standard performance evaluation based on this model, ensuring the consistency and computability of scheduling decisions in dynamic edge environments.

[0173] The multi-objective edge computing task scheduling method provided in this application extracts core parameters of the task to be scheduled, candidate execution nodes, and network dimensions step by step, providing comprehensive basic data support for multi-objective modeling. Based on three types of parameters, it constructs objective functions for latency, energy consumption, and link occupancy, along with constraints for latency and queue stability, forming multi-dimensional scheduling optimization judgment rules. Integrating the objectives and constraints into a unified multi-objective optimization model ensures consistent evaluation criteria for candidate solutions, guaranteeing the consistency and computability of scheduling decisions in dynamic edge environments.

[0174] Figure 4 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 3 ,like Figure 4 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the process for obtaining the predicted performance results, including:

[0175] S401. For a single candidate scheduling scheme, determine the candidate execution node allocated to each task to be scheduled in the candidate scheduling scheme.

[0176] For example, this step matches the target execution node corresponding to each task to be scheduled in the current scheduling cycle based on the task-node allocation mapping relationship built into the candidate scheduling scheme, and clarifies the deployment affiliation of all tasks under a single scheduling scheme, providing accurate positioning basis for subsequent load statistics and feature extraction on the node side.

[0177] S402. Based on the allocation results of the candidate scheduling scheme, calculate the total concurrent task load of the corresponding candidate execution node under the candidate scheduling scheme.

[0178] For example, this step iterates through all the scheduled tasks assigned to a single candidate execution node under the current scheduling scheme, adds up the computational requirements corresponding to each task, and combines them with the computational load base of the node's original pending tasks to calculate the total concurrent task computational load of the node after the scheme takes effect, accurately reflecting the actual computational pressure level of the node after task allocation.

[0179] Specifically, this step calculates the total computational requirements of other tasks allocated to the same node as the target task under a single candidate scheduling scheme, reflecting the intensity of intra-node resource competition faced by the task during execution. The corresponding formula is as follows:

[0180]

[0181] in, On candidate execution node j, with task The total computational demand of other tasks allocated simultaneously, i.e., the total computational load of concurrent tasks, where k is the number of the task to be scheduled, used to traverse the tasks. All other tasks besides those.

[0182] S403. Based on the task attribute information corresponding to each task to be scheduled, the node status information of the candidate execution nodes, and the total load of concurrent tasks, generate a local input feature vector for each task to be scheduled.

[0183] For example, this step extracts the task input data volume, computational requirements and other attribute parameters of the corresponding task to be scheduled, as well as the available computing power and energy consumption coefficient of the allocated nodes and other status parameters. These are then superimposed with the aforementioned total computational load parameters of concurrent tasks, and after feature dimension alignment and normalization, they are spliced ​​together to form a standardized input feature vector that represents the local operating environment of a single task.

[0184] Specifically, this step integrates four types of features: task attributes, node status, historical scheduling, and concurrent load. After dimensional alignment and normalization, a standardized local input feature vector is formed, which serves as the input to the hybrid expert agent model. The corresponding formula is as follows:

[0185]

[0186] in, For the task The corresponding local input feature vector contains all the local state features required for the surrogate model to predict.

[0187] S404. Input the local input feature vector of each task to be scheduled into the hybrid expert agent model. After calculating and fusing the predictions by multiple expert prediction networks in the hybrid expert agent model according to their weights, the prediction performance results corresponding to each candidate scheduling scheme are obtained.

[0188] For example, multiple expert prediction networks within the model perform specific predictions for latency, total system energy consumption, and backhaul link occupancy cost. The backhaul link occupancy cost does not require prediction by a proxy model and can be directly calculated based on the task data volume and backhaul link bandwidth parameters. The system combines the latency prediction, energy consumption prediction, and the directly calculated backhaul link occupancy cost to form the predicted performance result of the candidate scheduling scheme.

[0189] Specifically, this step completes performance prediction through a process of "gated distribution - expert prediction - weighted fusion - global aggregation". The specific calculation process is as follows:

[0190] Gated network expert weight calculation: Adaptively calculate the activation weights of each expert network based on the input feature vector to achieve division of labor prediction for heterogeneous tasks.

[0191]

[0192] in, Let be the gating weight vector corresponding to task i, containing the activation weight values ​​of each expert network. The parameter matrix of the gated network is learned through the model training process. This is a noise term used to increase the exploratory nature of expert selection and prevent the model from getting trapped in local optima. To reduce the computational cost of proxy prediction by retaining only the weights of the K highest-scoring experts, E represents the total number of expert networks. To normalize the retained expert scores, output the gating weights used for weighted fusion.

[0193] Local prediction by expert network: The activated expert network outputs the latency and energy consumption prediction results for the corresponding task based on the input features.

[0194]

[0195] in, Let be the local performance prediction result of the e-th expert network for task i. Let be the nonlinear mapping function of the e-th expert network, which is learned through training. To handle latency for task i predicted by the e-th expert network, Process the energy consumption for task i predicted by the e-th expert network.

[0196] Weighted fusion yields single-task prediction results: Under the influence of gating weights, the prediction results of each expert are weighted and fused to obtain the final prediction performance for a single task.

[0197]

[0198] in, Let be the final prediction performance vector for task i. Let be the gating weight value corresponding to task i on the e-th expert network.

[0199] Global aggregation yields the overall prediction results for the solution: the predicted values ​​of all tasks within the current candidate solution are summarized to obtain the prediction performance indicators for the entire solution. Among these, link occupancy cost does not require proxy prediction and can be directly calculated from the task data volume and bandwidth parameters.

[0200] The predicted average service latency and the predicted total system energy consumption are as follows:

[0201]

[0202] The three types of prediction indicators together constitute the prediction performance results of candidate scheduling scheme X, providing an efficient basis for performance judgment in the subsequent pre-screening stage.

[0203] The multi-target edge computing task scheduling method provided in this application provides an accurate positioning basis for subsequent load statistics and feature extraction by clearly defining the allocation correspondence between the tasks to be scheduled and the candidate execution nodes. It calculates the total concurrent task computing load of each node under the corresponding scheme, accurately reflecting the actual computing pressure of the nodes after task allocation. It integrates task attributes, node status, and concurrent load to generate standardized feature vectors, providing unified and standardized input data for the proxy model. Through multi-group expert prediction network division of labor calculation and weighted fusion, it quickly outputs the performance prediction results of candidate schemes, providing an efficient judgment basis for subsequent pre-screening.

[0204] Figure 5 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 4 ,like Figure 5 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the specific process for obtaining a subset of preferred candidate solutions, including:

[0205] S501. Based on the node queuing stability constraint, perform feasibility screening on each candidate scheduling scheme in the initial candidate scheduling scheme set, eliminate candidate scheduling schemes that do not meet the node queuing stability constraint, and obtain the corresponding feasible candidate scheme set.

[0206] For example, each scheme in the initial candidate scheduling scheme set corresponds to a complete mapping between scheduled tasks and candidate execution nodes. This step, for a single candidate scheduling scheme, calculates the total computational demand of tasks allocated to each candidate execution node, combines this with task arrival interval characteristics to obtain the node task arrival rate, and then calculates the number of tasks a node can process per unit time (i.e., the node service rate) based on the node's available computing power. This verifies whether the node queuing stability constraint holds. If all nodes in a scheme satisfy the constraint, it is included in the feasible candidate scheme set; if any node does not satisfy the constraint, it is directly determined as infeasible and eliminated, thus excluding invalid schemes that would lead to an infinite backlog of node task queues from the source.

[0207] Specifically, this step reuses the node queuing stability constraint formula in the multi-objective optimization model to perform low-cost hard constraint verification on each candidate scheduling scheme:

[0208]

[0209] For a single candidate scheduling scheme, the task arrival rate and service rate of each node are quickly calculated. If any node in the scheme does not meet the above constraints, it is determined to be infeasible and directly eliminated. Schemes that meet the constraints of all nodes are retained and form a set of feasible candidate schemes. Invalid schemes that would lead to an infinite backlog of node queues are eliminated from the source.

[0210] S502. Based on the predicted performance results, sort each candidate scheduling scheme in the feasible candidate scheme set in a non-dominated order, and select a preset number of candidate scheduling schemes with the highest ranking to form a subset of preferred candidate schemes.

[0211] For example, this step uses three predicted performance metrics output by the hybrid expert agent model—average service latency, total system energy consumption, and backhaul link occupancy cost—as evaluation dimensions. Based on the Pareto dominance rule, feasible candidate solutions are hierarchically divided: if solution A is not inferior to solution B in all objectives, and is superior in at least one objective, then solution A dominates solution B. Solutions without a dominant relationship are grouped into the same non-dominated level. Solutions are selected sequentially from the highest level, stopping when the cumulative number reaches a preset value, forming a subset of preferred candidate solutions. The preset number can be pre-set according to the computational resource quota and real-time requirements of the scheduling cycle, or dynamically adjusted according to the real-time system load status. Within the feasible solution range, solutions with better comprehensive performance across multiple objectives are selected, effectively reducing the processing scale of subsequent real-world performance evaluation.

[0212] Specifically, this step uses the predicted performance results as the evaluation dimension to construct a three-dimensional pre-screening performance vector, and performs ranking and screening based on the Pareto dominance rule. The formula corresponding to the pre-screening performance vector is as follows:

[0213]

[0214] in, Let q be the pre-screening performance vector corresponding to the q-th candidate scheduling scheme. Let be the binary decision matrix corresponding to the q-th candidate scheduling scheme.

[0215] Based on this vector, non-dominated sorting and crowding distance calculation are performed. A preset number of schemes with better multi-objective comprehensive performance are selected according to the level to form a subset of preferred candidate schemes, which effectively reduces the processing scale of subsequent real performance evaluation.

[0216] The multi-objective edge computing task scheduling method provided in this application performs feasibility screening through node queuing stability constraints, which can eliminate invalid schemes that may cause node queue backlog in advance, reducing the amount of invalid computation in subsequent evaluation stages from the source. By performing non-dominated sorting screening based on multi-dimensional predicted performance results, it can accurately retain schemes with better overall multi-objective performance, effectively reducing the processing scale of subsequent real performance evaluation and improving overall screening efficiency and accuracy.

[0217] Figure 6 A flowchart illustrating the multi-target edge computing task scheduling method provided in this application embodiment. Figure 5 ,like Figure 6 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the specific process for generating the target scheduling decision for the current scheduling period, including:

[0218] S601. Based on the multi-objective optimization model, calculate the actual performance index of each candidate scheduling scheme in the subset of preferred candidate schemes.

[0219] For example, this step invokes the real evaluation calculation logic corresponding to the multi-objective optimization model. For each candidate scheduling scheme, it accurately calculates the end-to-end actual completion latency of the task when executed on the corresponding candidate execution node, the actual total system energy consumption generated by node computation and transmission, and the actual backhaul link occupancy overhead generated by data backhaul. The three types of calculation results are integrated into a standardized real performance index vector. This calculation strictly follows the objective function expression of the multi-objective optimization model, and the output result is an accurate numerical value, which is different from the fast approximate prediction of the hybrid expert agent model. It is used for the final scheme selection and population iteration.

[0220] Specifically, this step calls the objective function of the multi-objective optimization model to perform accurate real performance calculations on the preferred candidate solutions, obtaining the real performance vector, as shown in the following formula:

[0221]

[0222] The value of this vector is obtained by rigorous calculation of the objective function and is an accurate true performance value, which is different from the approximate prediction value output by the hybrid expert agent model. It is used for the final scheme selection and population iteration.

[0223] S602. Based on the task delay constraints, eliminate candidate scheduling schemes that do not meet the task delay constraints to obtain a set of qualified candidate schemes.

[0224] For example, this step verifies the actual completion latency of each scheduled task in the candidate scheduling scheme that has completed the actual performance calculation. It compares the actual latency of each task with the maximum tolerable latency parameter of the corresponding task. If the actual latency of any task exceeds the constraint threshold, the scheme is determined to be unsatisfactory and is eliminated. Schemes that satisfy the latency constraints for all tasks are retained and form a set of qualified candidate schemes, ensuring that all schemes entering subsequent iterations are feasible in terms of latency.

[0225] Specifically, this step reuses the task delay constraint formula from the multi-objective optimization model to verify the feasibility of candidate solutions for completing real performance calculations:

[0226]

[0227] The actual completion time of each task in the scheme is verified one by one. If the actual time of any task exceeds the constraint threshold, the scheme is determined to be unsatisfactory and is eliminated. Schemes that satisfy the time constraints for all tasks are retained and form a set of qualified candidate schemes.

[0228] S603, Obtain the current generation population scheme.

[0229] In this embodiment, the current generation population scheme is the set of candidate scheduling schemes corresponding to the current iteration round. The current generation population scheme of the first iteration is composed of the initial set of candidate scheduling schemes, and the current generation population scheme of non-first iterations is the next generation population scheme generated in the previous iteration.

[0230] For example, the current generation population scheme is the set of parent candidates used to generate offspring schemes during the evolutionary iteration process, carrying the high-quality scheduling schemes retained from the previous iteration. In the first iteration, the initially generated complete set of candidate scheduling schemes is directly reused as the first generation parent, without the need for additional generation; in subsequent iterations, the next generation population scheme output from the previous environmental selection is directly used as the parent, forming intergenerational scheme transfer and evolutionary continuation.

[0231] Specifically, the initial population for the first iteration is constructed using a random generation method that satisfies the task integrity constraint, as shown in the following formula:

[0232]

[0233] in, This refers to the initial population for the first iteration, i.e., the current generation population scheme for the first round. X represents the initial population size, i.e., the number of candidate scheduling schemes it contains. k Let be the binary decision matrix corresponding to the k-th candidate scheduling scheme.

[0234] For the current generation population scheme that is not in the first iteration, the next generation population scheme generated in the previous iteration is directly adopted to achieve the transfer and evolution of schemes between generations.

[0235] S604. Merge the set of qualified candidate schemes with the current generation population scheme, and perform non-dominated sorting and crowding distance calculation based on the actual performance indicators to select the next generation population scheme.

[0236] For example, this step merges the newly generated qualified candidate schemes with the parent population schemes into a unified candidate pool. Pareto non-dominated sorting is performed based on three optimization dimensions of real performance indicators to classify the schemes into superior and inferior levels. Within each level, the crowding distance of each scheme is calculated to measure the sparsity of the scheme distribution in the target space. During selection, schemes are included sequentially from high to low non-dominated levels, with priority given to schemes with larger crowding distances within the same level, until a preset population size is reached, forming a next-generation population scheme that balances elite preservation and population diversity.

[0237] Specifically, this step merges the newly generated qualified candidate schemes with the parent population into a unified candidate pool, performs an environment selection operation based on actual performance indicators, and generates the next generation population. The corresponding formula is as follows:

[0238]

[0239] in, For the next generation population scheme in the (t+1)th iteration, Let t be the current generation population scheme in the t-th iteration. Let be the set of qualified candidate solutions obtained after the t-th iteration through actual evaluation and constraint removal. To perform the environment selection operation based on real performance indicators and non-dominated sorting and crowding distance calculation, schemes are selected from high to low non-dominated levels, and schemes with higher crowding are retained at the same level until the preset population size is reached.

[0240] S605. Repeatedly perform the iterative operations of candidate scheme expansion, pre-screening and real performance evaluation on the next generation population scheme until the preset termination condition is met, and generate the target scheduling decision for the current scheduling cycle.

[0241] For example, this step uses the generated next-generation population scheme as the new parent generation, and iteratively executes the complete process of candidate scheme crossover and mutation expansion, surrogate model pre-screening, and multi-objective model real evaluation to continuously optimize the population quality. When the preset maximum number of iterations is reached, the optimal performance of the population does not improve significantly for several consecutive rounds, or the objective function converges, the iteration process is terminated, and the candidate scheduling scheme with the best overall performance is selected from the final population as the target scheduling decision output for the current scheduling cycle.

[0242] Specifically, this step uses the generated next-generation population scheme as the new parent generation, and iteratively executes the complete process of candidate scheme expansion, proxy pre-screening, real performance evaluation, and environment selection to continuously optimize the population quality. The iteration process terminates when the preset maximum number of iterations is reached, the optimal performance of the population shows no significant improvement for multiple consecutive rounds, or the objective function converges, and the final set of optimal candidate schemes is output.

[0243]

[0244] in, This is the final set of preferred candidate solutions after the iteration terminates. Let L be the binary decision matrix corresponding to the Lth final preferred candidate scheduling scheme, where L is the number of candidate scheduling schemes that are ultimately retained.

[0245] In one possible implementation, candidate scheme expansion includes:

[0246] Two parent candidate scheduling schemes are selected from the current generation population scheme, and a crossover operation is performed in the task allocation dimension to generate an initial offspring candidate scheduling scheme. The resource utilization rate of the candidate execution nodes corresponding to each task to be scheduled in the initial offspring candidate scheduling scheme is detected. When the resource utilization rate exceeds the preset load threshold, and / or the number of tasks to be processed exceeds the preset queue threshold, each candidate execution node is sorted according to the available computing power value. The target candidate execution node is selected from the preset number of candidate execution nodes at the top of the sort, and the corresponding task to be scheduled is reallocated to the target candidate execution node to generate the final offspring candidate scheduling scheme.

[0247] For example, the current generation population scheme can consist of multiple candidate scheduling schemes, each scheme corresponding to the mapping relationship between tasks and edge nodes on the access side, cloud nodes, or local nodes on the terminal. When selecting two parent generation candidate scheduling schemes, they can be determined from the population based on the non-dominated sorting results and congestion distance, in order to preserve the diversity of different allocation patterns. When the crossover operation is performed on the task allocation dimension, some task allocation fragments in the parent generation scheme can be exchanged, while maintaining the unique allocation constraint of tasks, so that the same task to be scheduled still corresponds to a unique candidate execution node after crossover. The initial child generation candidate scheduling scheme after crossover then enters the resource detection module to statistically analyze the utilization rate of the node corresponding to each task to be scheduled. The statistical quantity can be calculated by the node's currently occupied computing power, the computing power of tasks waiting to be processed, and the total available computing power of the node.

[0248] When the resource utilization of a candidate execution node exceeds a preset load threshold, and / or the number of tasks to be processed exceeds a preset queue threshold, the system will sort all candidate execution nodes by their available computing power and select a target candidate execution node from the top-ranked nodes (a preset number). This preset number can be associated with the total number of nodes in the current scheduling cycle to limit the search scope for reallocation. The reallocation process for tasks to be scheduled can be completed based on task computation requirements, node available computing power, and network transmission status, migrating tasks originally assigned to high-load nodes to the target candidate execution node, forming the final offspring candidate scheduling scheme.

[0249] The candidate scheme expansion process generates offspring candidate schemes through crossover and mutation operations. The formula for the expanded offspring set is as follows:

[0250]

[0251] in, Let be the set of offspring candidate scheduling schemes obtained through crossover and mutation expansion in the t-th iteration. The population expansion factor is used to control the generation size of the offspring scheme. This represents the number of candidate scheduling schemes for the current generation of the population.

[0252] In practical applications, this sorting and redistribution logic can also be combined with other load characterization parameters, which are not limited in this application.

[0253] The above scheme introduces load detection and computing power ranking mechanisms after cross-generation, enabling local corrections to be completed during the formation of offspring schemes, thereby maintaining a good load distribution for subsequent input selection. Since reallocation is triggered only when resource utilization exceeds a threshold, and the reallocation targets are limited to nodes with high available computing power, the offspring schemes maintain strong feasibility and improve the balance and iterative solution stability of candidate scheduling schemes.

[0254] The multi-target edge computing task scheduling method provided in this application first quantifies empirical judgments into a judgment matrix through pairwise comparisons. Then, it uses a consistency check to eliminate judgment results with significant logical contradictions, thereby ensuring that weight calculations are based on reliable data. Subsequently, multiple alternative measures are compared under the same evaluation scale through a total ranking weight, thus avoiding bias caused by a single indicator dominating decision-making. This makes the final selected target scale prevention and removal measures more in line with the comprehensive management needs under the current operating conditions, thereby improving the objectivity and consistency of scale prevention and removal measure selection, reducing subjective bias caused by manual experience selection, and enabling different alternative measures to be ranked comparablely under a unified weight system. Due to the introduction of the consistency check mechanism, it can also effectively avoid erroneous decisions caused by the distortion of the judgment matrix, thereby improving the matching degree between the scale prevention and removal scheme and the on-site scaling risk, reducing reagent waste, ineffective construction, and equipment operation fluctuations, and enhancing the stability and economy of scaling management in oilfield production processes.

[0255] Figure 7 This is a schematic diagram of the structure of a multi-target edge computing task scheduling device provided in an embodiment of this application. The device in this embodiment can be in software and / or hardware form. For example... Figure 7 As shown in the embodiment of this application, the multi-target edge computing task scheduling device 700 includes: an acquisition module 701, a construction module 702, a first generation module 703, a prediction module 704, a filtering module 705, and a second generation module 706.

[0256] The acquisition module 701 is used to acquire the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution node within the current scheduling period.

[0257] The construction module 702 is used to construct a corresponding multi-objective optimization model based on task attribute information, node status information and network status information. The multi-objective optimization model includes an average service delay objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, as well as task delay constraints and node queuing stability constraints.

[0258] The first generation module 703 is used to generate an initial candidate scheduling scheme set based on the task allocation constraints of each task to be scheduled and each candidate execution node, wherein the initial candidate scheduling scheme set contains multiple candidate scheduling schemes.

[0259] The prediction module 704 is used to perform performance prediction on each candidate scheduling scheme in the initial candidate scheduling scheme set using a pre-built hybrid expert agent model, and obtain the predicted performance results.

[0260] The filtering module 705 is used to pre-filter the initial candidate scheduling scheme set based on the predicted performance results to obtain a subset of preferred candidate schemes;

[0261] The second generation module 706 is used to evaluate and screen the subset of preferred candidate schemes based on the multi-objective optimization model, and generate the target scheduling decision for the current scheduling period.

[0262] In one possible implementation, the building module 702 is further configured to:

[0263] Based on the task attribute information, obtain the corresponding first parameter information, which includes the task input data volume, computational requirements, and maximum tolerable latency parameter;

[0264] Based on the node status information, obtain the corresponding second parameter information, which includes the node's available computing power, the number of tasks to be processed, and the computing energy consumption coefficient parameters.

[0265] Based on the network status information, obtain the corresponding third parameter information, which includes the link transmission rate, inter-node link bandwidth, and total backhaul link bandwidth.

[0266] Based on the first parameter information, the second parameter information, and the third parameter information, respectively construct the corresponding objective functions for average service latency, total system energy consumption, and backhaul link occupancy cost, as well as the task latency constraints and node queuing stability constraints.

[0267] Based on the objective functions of average service delay, total system energy consumption, backhaul link occupancy cost, task delay constraints, and node queuing stability constraints, a multi-objective optimization model is constructed.

[0268] In one possible implementation, the prediction module 704 is further configured to:

[0269] For a single candidate scheduling scheme, determine the candidate execution node assigned to each task to be scheduled within the candidate scheduling scheme;

[0270] Based on the allocation results of the candidate scheduling schemes, calculate the total concurrent task computation load of the corresponding candidate execution nodes under the candidate scheduling schemes;

[0271] Based on the task attribute information corresponding to each task to be scheduled, the node status information of the candidate execution nodes, and the total load of concurrent tasks, a local input feature vector is generated for each task to be scheduled.

[0272] The local input feature vector of each task to be scheduled is input into the hybrid expert agent model. After multiple expert prediction networks in the hybrid expert agent model calculate and fuse them according to weights, the prediction performance result corresponding to each candidate scheduling scheme is obtained.

[0273] In one possible implementation, the filtering module 705 is further configured to:

[0274] Based on the node queuing stability constraint, each candidate scheduling scheme in the initial candidate scheduling scheme set is screened for feasibility, and candidate scheduling schemes that do not meet the node queuing stability constraint are eliminated to obtain the corresponding set of feasible candidate schemes.

[0275] Based on the predicted performance results, each candidate scheduling scheme in the feasible candidate scheme set is sorted in a non-dominated manner, and a preset number of candidate scheduling schemes with the highest ranking are selected to form a subset of preferred candidate schemes.

[0276] In one possible implementation, the second generation module 706 is further configured to:

[0277] Based on the multi-objective optimization model, calculate the actual performance index of each candidate scheduling scheme in the subset of preferred candidate schemes;

[0278] Based on the task delay constraints, candidate scheduling schemes that do not meet the task delay constraints are eliminated, resulting in a set of qualified candidate schemes.

[0279] Obtain the current generation population scheme, where the current generation population scheme is the set of candidate scheduling schemes corresponding to the current iteration round. The current generation population scheme of the first iteration is composed of the initial set of candidate scheduling schemes, and the current generation population scheme of non-first iterations is the next generation population scheme generated in the previous iteration.

[0280] The set of qualified candidate schemes is merged with the current generation population scheme. Based on the actual performance indicators, non-dominated sorting and crowding distance calculation are performed to select the next generation population scheme.

[0281] Repeatedly perform the iterative operations of candidate scheme expansion, pre-screening and real performance evaluation on the next generation population scheme until the preset termination condition is met, and generate the target scheduling decision for the current scheduling cycle.

[0282] In one possible implementation, the second generation module 706 is further configured to expand candidate solutions, including:

[0283] Select two parent candidate scheduling schemes from the current generation population scheme, perform a crossover operation in the task allocation dimension, and generate an initial offspring candidate scheduling scheme.

[0284] Detect the resource utilization rate of the candidate execution nodes corresponding to each task to be scheduled in the initial offspring candidate scheduling scheme;

[0285] When the resource utilization rate exceeds the preset load threshold, and / or the number of tasks to be processed exceeds the preset queue threshold, each candidate execution node is sorted according to its available computing power value. The target candidate execution node is selected from the preset number of candidate execution nodes at the top of the sort, and the corresponding tasks to be scheduled are reallocated to the target candidate execution node to generate the final offspring candidate scheduling scheme.

[0286] In one possible implementation, the prediction module 704 is also used for a hybrid expert agent model, which is constructed through the following steps:

[0287] Obtain real performance evaluation records of historical scheduled tasks;

[0288] Based on real performance evaluation records, determine the local state characteristics and real performance labels of the corresponding tasks;

[0289] Construct a corresponding training sample set based on local state features and true performance labels;

[0290] Obtain the initial hybrid expert network;

[0291] Using prediction error loss and expert load balancing loss as joint optimization objectives, the initial hybrid expert network is iteratively trained until the loss converges, resulting in a hybrid expert agent model.

[0292] The multi-target edge computing task scheduling device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0293] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 800 provided in this embodiment includes at least one processor 801 and a memory 802. Optionally, the device 800 further includes a communication component 803. The processor 801, memory 802, and communication component 803 are connected via a bus.

[0294] In a specific implementation, at least one processor 801 executes computer execution instructions stored in memory 802, causing at least one processor 801 to perform the above-described method.

[0295] The specific implementation process of processor 801 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0296] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0297] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0298] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0299] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0300] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0301] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0302] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0303] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0304] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0305] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0306] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0307] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0308] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A multi-target edge computing task scheduling method, characterized in that, include: Obtain the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution node within the current scheduling period; Based on the task attribute information, the node status information, and the network status information, a corresponding multi-objective optimization model is constructed. The multi-objective optimization model includes an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, as well as task latency constraints and node queuing stability constraints. Based on the task allocation constraints of each scheduled task and each candidate execution node, an initial candidate scheduling scheme set is generated, wherein the initial candidate scheduling scheme set contains multiple candidate scheduling schemes; A pre-built hybrid expert agent model is used to predict the performance of each candidate scheduling scheme in the initial candidate scheduling scheme set, and the predicted performance results are obtained. Based on the predicted performance results, the initial candidate scheduling scheme set is pre-screened to obtain a subset of preferred candidate schemes; The optimal candidate scheme subset is evaluated and screened based on the multi-objective optimization model to generate the target scheduling decision for the current scheduling period.

2. The method according to claim 1, characterized in that, The step involves constructing a corresponding multi-objective optimization model based on the task attribute information, the node state information, and the network state information, and by introducing task latency constraints and node queuing stability constraints. This model includes: Based on the task attribute information, obtain the corresponding first parameter information, wherein the first parameter information includes the amount of task input data, the amount of computation required, and the maximum tolerable latency parameter; Based on the node status information, the corresponding second parameter information is obtained, wherein the second parameter information includes the node's available computing power, the number of tasks to be processed, and the computing energy consumption coefficient parameters; Based on the network status information, the corresponding third parameter information is obtained, wherein the third parameter information includes the link transmission rate, the inter-node link bandwidth, and the total backhaul link bandwidth. Based on the first parameter information, the second parameter information, and the third parameter information, the corresponding average service latency objective function, the total system energy consumption objective function, the backhaul link occupancy cost objective function, the task latency constraint, and the node queuing stability constraint are constructed respectively. The multi-objective optimization model is constructed based on the average service latency objective function, the total system energy consumption objective function, the backhaul link occupancy cost objective function, the task latency constraint, and the node queuing stability constraint.

3. The method according to claim 1, characterized in that, The pre-built hybrid expert agent model is used to predict the performance of each candidate scheduling scheme in the initial candidate scheduling scheme set, and the predicted performance results are obtained, including: For a single candidate scheduling scheme, determine the candidate execution node assigned to each scheduled task in the candidate scheduling scheme; Based on the allocation results of the candidate scheduling scheme, calculate the total concurrent task computation load of the corresponding candidate execution node under the candidate scheduling scheme; Based on the task attribute information corresponding to each scheduled task, the node status information of the candidate execution node, and the total load of the concurrent tasks, a local input feature vector is generated for each scheduled task. The local input feature vector of each of the tasks to be scheduled is input into the hybrid expert agent model. After calculation by multiple expert prediction networks in the hybrid expert agent model and fusion according to weights, the prediction performance result corresponding to each candidate scheduling scheme is obtained.

4. The method according to claim 2, characterized in that, The step of pre-screening the initial candidate scheduling scheme set based on the predicted performance results to obtain a preferred candidate scheme subset includes: Based on the node queuing stability constraint, each candidate scheduling scheme in the initial candidate scheduling scheme set is screened for feasibility, and candidate scheduling schemes that do not meet the node queuing stability constraint are eliminated to obtain the corresponding feasible candidate scheme set. Based on the predicted performance results, each candidate scheduling scheme in the feasible candidate scheme set is sorted in a non-dominated manner, and a preset number of the candidate scheduling schemes with the highest ranking are selected to form the preferred candidate scheme subset.

5. The method according to claim 4, characterized in that, The step of evaluating and filtering the subset of preferred candidate schemes based on the multi-objective optimization model to generate the target scheduling decision for the current scheduling period includes: Based on the multi-objective optimization model, calculate the actual performance index of each candidate scheduling scheme in the subset of preferred candidate schemes; Based on the task delay constraints, candidate scheduling schemes that do not meet the task delay constraints are eliminated to obtain a set of qualified candidate schemes. Obtain the current generation population scheme, wherein the current generation population scheme is the set of candidate scheduling schemes corresponding to the current iteration round, the current generation population scheme of the first iteration is composed of the initial candidate scheduling scheme set, and the current generation population scheme of non-first iterations is the next generation population scheme generated in the previous iteration. The qualified candidate scheme set is merged with the current generation population scheme, and non-dominated sorting and crowding distance calculation are performed according to the actual performance index to select the next generation population scheme. The iterative operations of candidate scheme expansion, pre-screening, and real performance evaluation are repeatedly performed on the next-generation population scheme until the preset termination condition is met, thereby generating the target scheduling decision for the current scheduling cycle.

6. The method according to claim 5, characterized in that, The expansion of the candidate solutions includes: Two parent generation candidate scheduling schemes are selected from the current generation population scheme, and a crossover operation is performed in the task allocation dimension to generate an initial offspring generation candidate scheduling scheme. Detect the resource utilization rate of the candidate execution node corresponding to each scheduled task in the initial offspring candidate scheduling scheme; When the resource utilization rate exceeds a preset load threshold, and / or when the number of tasks to be processed exceeds a preset queue threshold, each candidate execution node is sorted according to its available computing power value, a target candidate execution node is selected from the preset number of candidate execution nodes at the top of the sort, and the corresponding task to be scheduled is reallocated to the target candidate execution node to generate the final offspring candidate scheduling scheme.

7. The method according to any one of claims 1-6, characterized in that, The hybrid expert agent model is constructed through the following steps: Obtain real performance evaluation records of historical scheduled tasks; Based on the actual performance evaluation records, determine the local state characteristics and actual performance labels of the corresponding tasks; Based on the local state features and the true performance labels, construct the corresponding training sample set; Obtain the initial hybrid expert network; Using prediction error loss and expert load balancing loss as joint optimization objectives, the initial hybrid expert network is iteratively trained until the loss converges, thus obtaining the hybrid expert agent model.

8. A multi-target edge computing task scheduling device, characterized in that, include: The acquisition module is used to acquire the task attribute information of each task to be scheduled, the node status information of each candidate execution node, and the network status information between the task to be scheduled and the candidate execution node within the current scheduling period. The construction module is used to construct a corresponding multi-objective optimization model based on the task attribute information, the node status information and the network status information. The multi-objective optimization model includes an average service latency objective function, a total system energy consumption objective function, a backhaul link occupancy cost objective function, as well as task latency constraints and node queuing stability constraints. The first generation module is used to generate an initial candidate scheduling scheme set based on the task allocation constraints of each scheduled task and each candidate execution node, wherein the initial candidate scheduling scheme set contains multiple candidate scheduling schemes. The prediction module is used to perform performance prediction on each of the candidate scheduling schemes in the initial candidate scheduling scheme set using a pre-built hybrid expert agent model, and obtain the predicted performance results. The filtering module is used to pre-filter the initial candidate scheduling scheme set based on the predicted performance results to obtain a subset of preferred candidate schemes; The second generation module is used to perform real performance evaluation and screening of the subset of preferred candidate schemes based on the multi-objective optimization model, and generate the target scheduling decision for the current scheduling period.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images