Cloud edge multi-agent-based resource scheduling result generation method and device, equipment and storage medium

Abstract

The application discloses a cloud-edge multi-agent-based resource scheduling result generation method and device, equipment and a storage medium, and relates to the technical field of resource scheduling. The method is applied to a multi-agent cluster system, and the system comprises a mapping module, an edge node agent and a cloud agent. The method comprises the following steps: performing feature extraction and classification on user resource use data by using the mapping module; and using a reinforcement learning strategy to optimize resource request information according to user classification labels and federated learning shared model parameters by using the edge node agent. The cloud agent obtains a resource scheduling result by combining a global resource state, a credit score and a double-weight mechanism of cooperation contribution, thereby realizing efficient scheduling of resources, improving the flexibility and efficiency of resource scheduling, and reducing the risk of task overload.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a method, apparatus, device, and storage medium for generating resource scheduling results based on cloud-edge multi-agent systems. Background Technology

[0002] With the rapid development of IoT, 5G communication, big data, and artificial intelligence technologies, edge computing, as an important extension of cloud computing, is gradually forming a new intelligent scheduling architecture centered on cloud-edge collaboration. By deploying resources such as computing, storage, and power scheduling closer to the user end, it can not only effectively shorten response time and reduce network transmission load, but also better meet the needs of high real-time application scenarios such as industrial control, smart transportation, and intelligent security.

[0003] Current research has attempted to address resource allocation issues in edge computing using reinforcement learning and intelligent scheduling algorithms. These methods typically involve independent training modes for cloud agents and edge agents, aiming to optimize the global objective while considering local resource constraints. Furthermore, some schemes introduce multi-dimensional dynamic scoring models to determine the value of user task scheduling, and distributed edge shared recognition to enhance the accuracy of user classification profiles. However, the processing of user classification profiles is relatively coarse, failing to fully utilize the advantages of distributed edge shared recognition, resulting in inaccurate resource scheduling. Additionally, edge node agents, when making task scheduling decisions, primarily rely on the current instantaneous state, rarely considering the temporal fluctuations in task load, which increases the risk of task overload. Therefore, there is an urgent need for improved multi-agent-based resource scheduling methods to enhance the flexibility and efficiency of resource scheduling. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, device, and storage medium for generating resource scheduling results based on cloud-edge multi-agent systems, aiming to solve the technical problem of how to improve the efficiency and flexibility of resource scheduling.

[0005] To achieve the above objectives, this application proposes a method for generating resource scheduling results based on cloud-edge multi-agent systems. The method is applied to a multi-agent cluster system, which includes a mapping module, edge node agents, and cloud-based agents. The method includes:

[0006] Obtain user resource usage data;

[0007] Feature extraction is performed on the user resource usage data to obtain feature vectors, which include feature vectors of the mean, variance, maximum value, and minimum value of CPU, GPU, memory, and storage.

[0008] The feature vectors are classified using a cloud-edge collaborative clustering algorithm to obtain user classification labels;

[0009] Input the user classification tags into a preset classification model to obtain resource request information;

[0010] The resource request information is optimized according to a reinforcement learning strategy to obtain the target resource request information;

[0011] Based on the target resource request information, combined with the global resource status and a dual-weighting mechanism of credit score and collaborative contribution, the resource scheduling result is obtained.

[0012] In one embodiment, the step of classifying the feature vector using a cloud-edge collaborative clustering algorithm to obtain user classification labels includes:

[0013] Initialize cluster centers to obtain a preset number of initial cluster center points;

[0014] The feature vectors are classified locally to obtain local cluster centers;

[0015] The local cluster center is encrypted and uploaded to the cloud-based intelligent agent, so that the cloud-based intelligent agent can perform global aggregation calculation on the local cluster center to obtain the global cluster center;

[0016] The average cluster center is calculated based on the global cluster center.

[0017] The average cluster centers are subjected to global balanced classification processing according to a preset load level standard to generate calibrated classification labels.

[0018] The calibrated classification labels are added to the feature vector to obtain the enhanced feature vector;

[0019] The initial cluster centers are updated based on the enhanced feature vector to obtain the updated cluster centers;

[0020] Based on the updated cluster centers, user category labels are obtained.

[0021] In one embodiment, before the step of obtaining the resource request strategy based on the user classification tags input into a preset classification model, the following steps are included:

[0022] Acquire historical resource scheduling data and construct an initial classification model. The historical resource scheduling data includes users' historical resource usage records, execution status of scheduling tasks, resource allocation results, and corresponding energy consumption data.

[0023] The historical resource scheduling data is preprocessed to obtain the processed data;

[0024] User behavior characteristics are extracted from the processed data.

[0025] The user behavior features are input into the initial classification model for training and validation to obtain the preset classification model.

[0026] In one embodiment, the step of inputting the user classification tags into a preset classification model to obtain resource request information includes:

[0027] Based on the user classification tags input into the preset classification model, the resource demand type is determined. The resource demand type includes high computing demand, low latency sensitivity, high storage demand, and high bandwidth demand.

[0028] Based on the resource demand type, a corresponding resource request baseline value is obtained, which is optimized based on historical scheduling data and user behavior patterns.

[0029] Select the resource request baseline value corresponding to the resource demand type as the basic resource request quantity;

[0030] Based on the aforementioned basic resource request volume, a random perturbation mechanism is introduced for adjustment to generate a differentiated resource request strategy;

[0031] Resource request information is obtained based on the differentiated resource request strategy.

[0032] In one embodiment, the step of obtaining the resource scheduling result based on the target resource request information combined with the global resource status and a dual-weighting mechanism of credit score and cooperation contribution includes:

[0033] Obtain the global resource status and coordination contribution, wherein the coordination contribution is calculated based on the results of the preset classification model;

[0034] The coordination contribution is normalized to obtain the processed coordination contribution.

[0035] The load rate is calculated based on the target resource request information and the global resource status.

[0036] The overload determination result is obtained by comparing the load rate with the preset safety threshold;

[0037] A credit score is obtained based on the overload determination result;

[0038] The resource request priority is calculated based on the credit score and the processed coordination contribution.

[0039] The target resource request information is sorted according to the resource request priority to determine the resource allocation order;

[0040] The resource scheduling result is obtained by combining the resource allocation order with the target resource request information and the global resource status.

[0041] In one embodiment, the step of obtaining a credit score based on the overload determination result includes:

[0042] Obtain an initial credit score;

[0043] When the overload determination result indicates that an overload has occurred, the initial credit score is multiplied by a preset penalty coefficient to obtain the first credit score.

[0044] When the overload determination result is that no overload situation has occurred, the initial credit score is multiplied by the preset reward coefficient to obtain the second credit score.

[0045] In one embodiment, after the step of obtaining the resource scheduling result based on the target resource request information combined with the global resource status and a dual-weighting mechanism of credit score and cooperation contribution, the method includes:

[0046] Get your current credit score;

[0047] The target cluster center points in the cloud-edge collaborative clustering algorithm are updated according to a first preset time to obtain the updated cluster center points;

[0048] The current credit score is ranked and adjusted according to a second preset time to obtain the exploration rate;

[0049] The target resource request information is optimized based on the updated cluster center points and the exploration rate to obtain optimized target resource request information;

[0050] Based on the optimized target resource request information, the step of obtaining the resource scheduling result by combining the target resource request information with the global resource status and the dual weighting mechanism of credit score and cooperation contribution is executed.

[0051] Furthermore, to achieve the above objectives, this application also proposes a resource scheduling result generation device based on cloud-edge multi-agent systems, the resource scheduling result generation device based on cloud-edge multi-agent systems comprising:

[0052] The acquisition module is used to acquire user resource usage data;

[0053] The feature extraction module is used to extract features from the user resource usage data to obtain feature vectors, which include feature vectors of the mean, variance, maximum value, and minimum value of CPU, GPU, memory, and storage.

[0054] The classification module is also used to classify the feature vectors using a cloud-edge collaborative clustering algorithm to obtain user classification labels;

[0055] The processing module is also used to input the user classification tags into a preset classification model to obtain resource request information;

[0056] The optimization module is used to optimize the resource request information according to the reinforcement learning strategy to obtain the target resource request information;

[0057] The results module is used to obtain resource scheduling results based on the target resource request information combined with the global resource status and a dual-weighting mechanism of credit score and collaborative contribution.

[0058] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the resource scheduling result generation method based on cloud-edge multi-agent as described above.

[0059] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the resource scheduling result generation method based on cloud-edge multi-agent as described above.

[0060] This application extracts and classifies features from user resource usage data through a mapping module. Edge node agents optimize resource request information using reinforcement learning strategies based on user classification labels and federated learning shared model parameters. The cloud agent combines global resource status, credit score, and collaborative contribution as dual weights to obtain resource scheduling results, achieving efficient resource scheduling, improving the flexibility and efficiency of resource scheduling, and reducing the risk of task overload. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart illustrating the first embodiment of the cloud-edge multi-agent resource scheduling result generation method of this application;

[0063] Figure 2 This is a flowchart illustrating the second embodiment of the cloud-edge multi-agent resource scheduling result generation method of this application;

[0064] Figure 3This is a flowchart illustrating the third embodiment of the cloud-edge multi-agent resource scheduling result generation method of this application;

[0065] Figure 4 This is a schematic diagram of the module structure of the resource scheduling result generation device based on cloud-edge multi-agents in the first embodiment of the resource scheduling result generation method based on cloud-edge multi-agents of this application.

[0066] Figure 5 This is a schematic diagram of the hardware operating environment involved in the cloud-edge multi-agent resource scheduling result generation method in the embodiments of this application.

[0067] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0068] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0069] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0070] With the rapid development of information technology, technologies such as the Internet of Things, 5G communication, big data, and artificial intelligence are changing our lives and work at an unprecedented pace. Against this backdrop, edge computing, as an important extension of cloud computing, is gradually becoming one of the key technologies to meet the demands of high real-time application scenarios. By deploying resources such as computing, storage, and power scheduling closer to the user end, edge computing can not only effectively shorten response latency but also significantly reduce network transmission load, which is particularly important for fields such as industrial control, intelligent transportation, and intelligent security. However, traditional centralized scheduling mechanisms in cloud centers are inadequate for large-scale systems due to problems such as hierarchical fragmentation, coarse value assessment, and lack of resource prediction. Although some existing research attempts to solve the resource allocation problem in edge computing using reinforcement learning and intelligent scheduling algorithms, these methods often lack a dynamic collaborative mechanism between cloud-based and edge-based agents, neglect the recovery of historical scheduling benefits when determining the value of user task scheduling, and fail to fully consider the temporal fluctuations of task load.

[0071] Therefore, this application proposes a resource scheduling result generation method based on cloud-edge multi-agent to solve the above problems. The main solution of this application embodiment is as follows: acquire user resource usage data; extract features from the user resource usage data to obtain feature vectors, which include feature vectors of the mean, variance, maximum, and minimum values ​​of CPU, GPU, memory, and storage; classify the feature vectors using a cloud-edge collaborative clustering algorithm to obtain user classification labels; input the user classification labels into a preset classification model to obtain resource request information; optimize the resource request information according to a reinforcement learning strategy to obtain target resource request information; and obtain resource scheduling results based on the target resource request information combined with the global resource status and a dual-weight mechanism of credit score and collaborative contribution.

[0072] Based on the above, this application also provides a method for generating resource scheduling results based on cloud-edge multi-agent systems, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the resource scheduling result generation method based on cloud-edge multi-agent systems according to this application. The method is a multi-agent cluster system, which includes a mapping module, edge node agents, and cloud agents.

[0073] In this embodiment, the method for generating resource scheduling results based on cloud-edge multi-agent systems includes steps S10 to S60:

[0074] Step S10: Obtain user resource usage data.

[0075] It's important to note that in acquiring user resource usage data, the system needs to collect detailed information from various edge node agents, including but not limited to real-time usage of resources such as CPU, GPU, memory, storage, and bandwidth. These edge node agents refer to computing resource points deployed close to the data source or user end. These nodes are typically located at the network edge, capable of directly processing locally generated data and providing low-latency and high-efficiency service responses. Edge node agents can be physical devices such as routers, switches, and dedicated edge servers, or they can be sensors or other IoT devices with computing capabilities. This data not only reflects the user's current operational needs but also implies potential future service request patterns.

[0076] Step S20: Extract features from user resource usage data to obtain feature vectors.

[0077] It's important to note that the feature vector includes the mean, variance, maximum, and minimum values ​​for CPU, GPU, memory, and storage. The mean reflects the average usage level of users on that resource dimension, helping the system understand users' regular needs. The variance reveals the fluctuations in user resource usage; a higher variance may indicate unstable user needs. The maximum and minimum values ​​identify the highest and lowest usage needs of users on that resource dimension, respectively, helping the system assess the range of users' potential resource requirements. By combining these statistical features into a feature vector, the system can comprehensively characterize users' usage habits and needs across different resource dimensions.

[0078] Specifically, the mapping module is responsible for extracting and processing historical resource usage data from edge node agents, performing statistical feature calculations in 6-hour time windows. This process involves calculating the mean, variance, maximum, and minimum values ​​of various indicators (such as CPU utilization, GPU computing power utilization, memory and storage I / O, etc.) to form feature vectors that reflect user behavior patterns. Notably, to more accurately capture load characteristics, the system also analyzes overall energy consumption, typically obtained through real-time sampling and accumulation using power sensors integrated into the edge node agents. Overall energy consumption reflects the energy consumption of users during resource usage, which is crucial for achieving green computing and power scheduling. By comprehensively considering resource usage and energy consumption, the system can optimize energy efficiency and reduce operating costs while meeting user needs. Furthermore, direct hardware detection methods ensure the authenticity and reliability of the data, avoiding errors introduced by model predictions.

[0079] Step S30: The feature vectors are classified using a cloud-edge collaborative clustering algorithm to obtain user classification labels.

[0080] It's important to note that cloud-edge collaborative clustering is a classic clustering method widely used in data mining and machine learning. Its core idea is to divide the data into K clusters, ensuring that data points within each cluster are as similar as possible, while data points between clusters are as different as possible. In this way, users with similar resource usage characteristics can be grouped into the same category, allowing for the development of targeted resource request strategies for each category.

[0081] Further, step S30 includes: initializing cluster centers to obtain a preset number of initial cluster center points. Specifically, in the initial stage of the clustering process, the system initializes the cluster centers of the cloud-edge collaborative clustering algorithm according to a preset number of categories K. The selection of these initial center points has a certain impact on the convergence speed and quality of the clustering results. The initialization method may include randomly selecting the current feature vector as the initial center, or adopting preset rules. For example, based on historical experience, some typical load characteristic values ​​(such as 20%, 40%, and 70% load rates) can be selected as preliminary cluster center points. In addition, a set of representative initial cluster centers can be determined based on previously collected historical scheduling data through empirical settings or simple statistical analysis, laying the foundation for subsequent iterative optimization. Subsequently, the feature vectors are locally classified to obtain local cluster centers. Specifically, each feature vector is assigned to the nearest category according to its Euclidean distance from the initial cluster center point. After the assignment is completed, the mean of all feature vectors in each category is recalculated to obtain new "target cluster center points". This process iterates until the change in cluster centers is less than the preset convergence threshold. At this point, the local clustering results and stable local cluster centers of each edge node agent are obtained.

[0082] Next, the local cluster centers are encrypted and uploaded to the cloud-based intelligent agent. The cloud-based agent then performs a global aggregation calculation on the local cluster centers to obtain the global cluster centers. Based on these global cluster centers, an average cluster center is calculated. Specifically, after receiving the encrypted local cluster centers, the cloud-based intelligent agent decrypts them and performs a global aggregation calculation. This aggregation process is not simply a summation; it uses a weighted average based on the contribution or weight of each edge node intelligent agent to comprehensively reflect the clustering information of all edge node intelligent agents, generating a globally representative set of cluster centers—the global cluster centers. After obtaining the global cluster centers, the average cluster centers are calculated. The average cluster centers refer to the central points obtained through this global aggregation that comprehensively reflect the cluster distribution of the entire system. Then, a global balanced classification process is performed on the average cluster centers according to a preset load level standard to generate calibrated classification labels. This preset load level standard is predefined; for example, load is divided into four levels: low load (0%-30%), medium load (30%-60%), high load (60%-90%), and abnormal load exceeding 90%. Based on these fixed standards, user classification information uploaded from the edge or the aggregated cluster centers are calibrated to ensure that the final user classification labels are consistent with the global resource management and scheduling strategy (i.e., "user profile"), resulting in calibrated classification labels. Next, the calibrated classification labels are added to the feature vector to obtain an enhanced feature vector. Specifically, the calibrated classification labels are appended as new feature dimensions or "labels" to the original user feature vector. For example, the original feature vector might only contain CPU utilization and memory usage, while this step adds feature columns such as "CPU load level," "GPU load level," "bandwidth load level," "storage load level," and "memory load level." The initial cluster centers are then updated based on the enhanced feature vectors to obtain updated cluster centers. Finally, the updated cluster centers are used for partitioning to obtain user classification labels. Specifically, this is a key step in a closed-loop iterative optimization. After obtaining enhanced feature vectors containing load level information, these new feature vectors are used to retrain or iterate the cloud-edge collaborative clustering algorithm. In the next clustering cycle, these enhanced feature vectors will be used to calculate new cluster centers. Furthermore, the system periodically (e.g., every 6 hours) updates the cluster centers in the K-means algorithm, then partitions the updated cluster centers, and uses a cascading lightweight transmission of cluster centers to implement encrypted communication between the cloud and edge agents to keep the partitioned data confidential, thus obtaining the final user classification labels.

[0083] This approach allows the system to not only identify user groups with similar resource usage patterns but also dynamically adjust classification criteria to adapt to changing needs. This lays the foundation for efficient and flexible resource scheduling, ensuring resources are allocated appropriately across different demands and improving overall service quality and user experience.

[0084] Step S40: Input the user category tags into the preset category model to obtain resource request information.

[0085] It should be noted that before inputting user classification labels into the preset classification model, parameter sharing and encryption mechanisms are implemented before cluster center aggregation. First, each edge node agent assigns its corresponding cluster center... After encryption, the data is uploaded to the cloud-based intelligent agent. The encryption method is divided into an initial phase and an update phase. In the initial phase, the edge node intelligent agent uploads the number of tags. The cloud-based intelligent agent sends out a set of random numbers. Edge node intelligent agents upload to clustering center The cloud-based intelligent agent performs corresponding processing to obtain the true cluster centers. Then, it performs global cluster center aggregation calculations on the true cluster centers. During the update phase, the cloud-based intelligent agent no longer distributes random number sets; instead, edge nodes upload a set of [historical cluster centers + latest cluster centers]. The cloud-based intelligent agent then performs corresponding processing to obtain the true cluster centers. The specific formula is:

[0086]

[0087] in, For nodes The weight, This represents the total number of nodes. After completing the parameter sharing and global cluster center aggregation, deep learning techniques (such as Convolutional Neural Networks, CNNs) are used for identification and feature extraction to obtain a series of fused feature vectors. Next, the fused feature vectors are input into a pre-defined classification model to generate specific resource request information. This pre-defined classification model is typically a machine learning model trained on a large amount of historical data, capable of predicting a user's current and future resource needs based on the input feature vectors.

[0088] Further, before step S40: acquire historical resource scheduling data and construct an initial classification model. Historical resource scheduling data includes users' historical resource usage records, the execution status of scheduling tasks, resource allocation results, and corresponding energy consumption data. Then, preprocess the historical resource scheduling data to obtain processed data. Specifically, since the acquired data is raw and disorganized, a series of preprocessing operations are required. This includes data cleaning (removing erroneous or incomplete records), format standardization (ensuring compatibility between data from different sources), and feature engineering (extracting features useful for subsequent analysis). For example, convert time-series data into fixed time windows (e.g., every 6 hours), and statistically calculate resource usage within each time period to obtain features such as mean, variance, maximum, and minimum values. Then, extract user behavior features based on the processed data. Finally, input the user behavior features into the initial classification model for training and validation to obtain the preset classification model. Feature extraction is not limited to calculating basic statistics but may also involve more complex features, such as periodic pattern recognition and anomaly detection. For example, by analyzing users' resource usage at different times of the day, patterns with distinct work habits can be identified; or potential anomalies can be identified by detecting sudden increases or decreases in resource usage. Once user behavioral characteristics are obtained, they can be used as input to begin training an initial classification model. This process typically involves the selection and application of machine learning algorithms, such as decision trees, support vector machines, and neural networks. During training, the model learns how to accurately predict future behavioral patterns based on known historical data. Simultaneously, to ensure the model's effectiveness and reliability, rigorous validation is required. This can be achieved through techniques such as cross-validation, ensuring that the model performs well not only on the training set but also exhibits stable performance on unseen data.

[0089] Step S50: Optimize the resource request information according to the reinforcement learning strategy to obtain the target resource request information.

[0090] It's important to note that reinforcement learning learns optimal decision-making strategies through the interaction between the agent and its environment to maximize long-term cumulative rewards. Specifically, the edge node agent selects actions (i.e., resource request quantity and priority) based on its current state (such as resource needs, credit score, and collaborative contribution), and adjusts its strategy based on reward signals from the environment (such as energy savings and scheduling delays). When optimizing resource request information, the agent aims to find a balance: satisfying its own resource needs while avoiding system overload or resource waste due to excessive requests. To this end, the agent employs reinforcement learning algorithms such as Q-learning, learning the optimal resource request strategy through continuous trial and error. After each scheduling, the agent calculates the reward value based on actual energy consumption, latency, and other indicators, and updates the Q-value table, thereby gradually optimizing the resource request strategy. The specific update formula is as follows:

[0091]

[0092]

[0093]

[0094]

[0095] in, As the weighting coefficient, it needs to satisfy... , For your own credit score, The average credit score of nodes of the same type. Indicates the current state. This indicates the action to be taken in the current state. This indicates the instant reward received. The learning rate is set to 0.1 by default. This is the discount factor, with a default value of 0.9. Indicates the updated status. This indicates the action to be taken corresponding to the updated state. Specifically, it refers to the action taken by the agent in the current state. Select action The decision is initially scored and rewarded, and then updated to obtain the updated state. It depends on the new state. Regarding one's previous state in the current situation The decision to choose an action is evaluated post-hocly. This post-hoc score is then compared to the predicted initial score. Based on the error between the two scores, a certain percentage (learning rate) is applied. The original Q-value is updated to more closely approximate the post-event rating. Current resource utilization is monitored in real-time by the system to obtain the utilization rates of various resources (such as CPU, GPU, memory, and storage), which can be achieved through system performance counters or resource management APIs. Historical request success rate is calculated by analyzing historical data to determine the proportion of user resource requests that were successfully fulfilled over a past period. The number of CPU cores requested is determined based on the user's application requirements and system resource status; the number of GPUs requested is determined for tasks requiring graphics processing or parallel computing; and the amount of memory requested is determined based on the memory requirements of the user's task.

[0096] By continuously iterating and updating the Q value, the agent gradually optimizes its resource request strategy and obtains the target resource request information.

[0097] Step S60: Based on the target resource request information and combined with the global resource status and the dual-weighting mechanism of credit score and collaboration contribution, the resource scheduling result is obtained.

[0098] It's important to note that, firstly, the cloud-based agent needs to comprehensively assess the overall resource status of the system. This includes, but is not limited to, the current usage of computing resources (such as CPU and GPU availability), storage resources, network bandwidth, and predicted future load trends. Through comprehensive analysis of this information, the cloud-based agent can identify bottlenecks in the system and adjust resource allocation strategies accordingly to avoid overload or resource waste. A dual-weighting mechanism of credit score and collaborative contribution is also introduced. Each edge node agent has a credit score based on its historical performance. This score reflects not only the node's historical behavior (e.g., whether it adheres to resource request agreements) but also its contribution to the overall system's collaboration. Specifically, collaborative contribution measures the improvement in classification accuracy provided by the node when participating in federated learning. After clarifying the global resource status and updating the credit scores and collaborative contributions of each node, the cloud-based agent will make resource scheduling decisions based on a complex priority weight formula. This formula dynamically adjusts the weights based on factors such as the node's current overload status and historical average latency, ensuring that latency optimization is prioritized when the overall system is overloaded, and that more collaboration is incentivized when there are significant differences in collaborative contributions between nodes. Furthermore, if a node is detected to be occupying too many resources for an extended period, its scheduling priority will be appropriately reduced to prevent resource monopoly. Finally, after comprehensive consideration of the above steps, the cloud-based agent will generate a specific resource allocation result and distribute it to the corresponding edge node agents for execution.

[0099] Furthermore, following step S60, the system also includes: obtaining the current credit score. First, the system needs to obtain the current credit score of each edge node agent. The credit score reflects the node's historical performance, including whether it adheres to resource request agreements and its contribution to the overall system's collaboration. The initial value is uniformly set to 1.0 (range [0.5, 5.0]), and dynamically adjusted based on the node's performance. For example, if a node continuously... If a node does not cause system overload, its credit score will be reset to the initial value of 1.0; conversely, if a node causes system overload, its credit score will be adjusted according to the penalty coefficient. Reduce. Then, update the target cluster centers in the cloud-edge collaborative clustering algorithm according to the first preset time interval to obtain the updated cluster centers. Specifically, the objective function in the cloud-edge collaborative clustering algorithm is used to update the cluster centers. This process is executed according to the first preset time interval, such as once every 6 hours. First, the edge node intelligent agent uploads the local cluster centers to the cloud intelligent agent. This data is encrypted to protect privacy. After receiving the data from each node, the cloud intelligent agent calculates the global cluster centers and sends them to each edge node intelligent agent to update the preset classification model. This periodic update mechanism helps improve the global classification accuracy and ensures the accurate identification and classification of user behavior features. Then, rank and adjust the current credit score according to the second preset time interval to obtain the exploration rate. Specifically, for the current credit score, the system ranks and adjusts it according to the second preset time interval (e.g., 1 hour) to obtain the exploration rate. The dynamic adjustment formula for the exploration rate is:

[0100]

[0101] in, The base exploration rate, with a default value of 0.2, This represents the percentile ranking of the current credit score across all nodes. The number of nodes. The exploration rate determines the balance between exploration and exploitation for the agent. Specifically, if a node's credit score ranks high among its peers (e.g., top 10%), its exploration rate... The target resource request rate will be lowered to a minimum of 0.05, meaning that the node is more inclined to utilize known best strategies rather than explore new ones. Conversely, nodes with lower credit scores will have higher exploration rates, encouraging them to try different resource request volumes to discover new, efficient strategies. Finally, the target resource request information is optimized based on the updated cluster centers and exploration rates, resulting in optimized target resource request information. Step S60 is then executed based on this optimized information. This process forms a closed-loop optimization, ensuring not only the effective utilization of resources but also promoting the healthy and stable operation of the entire system, improving service quality, and meeting diverse user needs.

[0102] This embodiment extracts and classifies user resource usage data through a mapping module. Edge node agents optimize resource request information using reinforcement learning strategies based on user classification labels and federated learning shared model parameters. The cloud agent combines global resource status, credit score, and collaborative contribution as dual weights to obtain resource scheduling results, achieving efficient resource scheduling, improving its flexibility and efficiency, and reducing the risk of task overload.

[0103] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The method for generating resource scheduling results based on cloud-edge multi-agent systems, step S30, further includes steps S201 to S205:

[0104] Step S201: Determine the resource demand type by inputting the user classification tags into the preset classification model.

[0105] It should be noted that user classification tags are obtained through feature extraction and cluster analysis of users' historical resource usage data. These tags accurately reflect users' resource usage behavior patterns. Inputting user classification tags into a preset classification model can further determine the types of users' resource needs, thereby providing a basis for generating resource request strategies.

[0106] The resource requirements mentioned above include high computing demand, low latency sensitivity, high storage demand, and high bandwidth demand. High computing demand users typically require significant CPU and GPU resources to handle complex computational tasks, such as deep learning training and large-scale data analysis. Low latency sensitivity users have extremely high responsiveness requirements and need to quickly obtain computational results, such as in real-time video processing and online games; therefore, they are more sensitive to network bandwidth and latency. High storage demand users require substantial storage resources to store data, such as for data backup and large-scale file storage. High bandwidth demand users require high bandwidth to transmit large amounts of data, such as for high-definition video streaming and large-scale data transfer. By determining the user's resource demand type, the system can formulate differentiated resource request strategies for different types of users. For example, for high computing demand users, the system can prioritize allocating more CPU and GPU resources; for low latency sensitive users, the system can optimize network configuration to reduce latency; for high storage demand users, the system can allocate more storage space; and for high bandwidth demand users, the system can prioritize ensuring network bandwidth. This differentiation strategy can improve the accuracy and efficiency of resource allocation, better meet users' personalized needs, and optimize overall resource utilization efficiency.

[0107] Step S202: Obtain the corresponding resource request baseline value based on the resource demand type.

[0108] It's important to note that once the user's resource requirement type is determined, the next step is to optimize the resource request baseline based on historical scheduling data and user behavior patterns. For each resource requirement type, the system has a pre-defined baseline as the basic resource request volume. These baselines are not fixed; they are the result of optimization and adjustment based on historical scheduling data and user behavior patterns. For example, for users with high computing demands, the system may refer to the best resource configurations used in previous similar task executions, including the required number of CPU cores and memory size; while for users with high bandwidth demands, the focus will be on network bandwidth requirements and the optimal time window selection.

[0109] Step S203: Select the resource request baseline value corresponding to the resource demand type as the basic resource request quantity.

[0110] It's important to note that selecting the baseline resource request value corresponding to the resource demand type as the basic resource request volume is a crucial step in the entire process. This step ensures that each user's resource request meets their basic needs without causing unnecessary waste or overload. For example, if a user is marked as having high storage requirements, the system will provide them with sufficient storage space to support their data operation needs while avoiding over-allocation that leads to resource idleness.

[0111] Step S204: Adjust the resource request strategy by introducing a random disturbance mechanism based on the basic resource request volume.

[0112] It should be noted that, to further improve the effectiveness and adaptability of the resource request strategy, the system introduces a random perturbation mechanism. This method allows for the addition of a certain amount of random variation to the basic resource request volume, as shown in the following formula:

[0113]

[0114] in, Indicates the amount of basic resource requests. Indicates the exploration rate. This represents a randomly generated perturbation value. This mechanism aims to balance "exploration" (trying new possibilities) with "utilization" (relying on known best practices), helping the agent discover potential better solutions while maintaining flexibility to handle unforeseen circumstances. Combining the basic resource request volume with the random perturbation mechanism, the system can generate personalized, differentiated resource request strategies for each user. These strategies not only consider the user's immediate needs but also incorporate predictions of future trends, making resource allocation more scientific and rational. For example, during peak periods, the system may slightly increase the resource quotas for some high-load users to prevent service interruptions; while during off-peak periods, resource allocation can be appropriately reduced to save costs.

[0115] Step S205: Obtain resource request information according to the differentiated resource request strategy.

[0116] It should be noted that, based on the generated differentiated resource request strategy, the system forms specific resource request information and sends it to the cloud-based intelligent agent for processing. Upon receiving these requests, the cloud-based intelligent agent uses a dual-weighting mechanism, combining global resource status, credit score, and collaborative contribution, to make the final resource scheduling decision. This process ensures that resources can achieve maximum utilization and service quality while meeting user needs.

[0117] In summary, the process from user classification tags to resource request information formation not only effectively solves the problems of hierarchical fragmentation and crude value assessment inherent in traditional centralized scheduling schemes, but also adapts to increasingly complex application scenarios through continuous self-learning and optimization, providing users with a more personalized and efficient service experience. Furthermore, this approach provides strong technical support for future large-scale IoT deployments, promoting the development of intelligent management.

[0118] This embodiment determines the type of resource demand by user classification tags, optimizes the baseline value of resource requests based on historical data, and generates differentiated resource request strategies by combining the basic resource request volume with a random perturbation mechanism, thereby improving the accuracy and efficiency of resource allocation, reducing energy consumption, and enhancing system response speed and service quality.

[0119] Based on the first embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 The step S60, which generates resource scheduling results based on cloud-edge multi-agent systems, further includes steps S301 to S308:

[0120] Step S301: Obtain the global resource status and coordination contribution.

[0121] It's important to note that the global resource status includes the current usage and predicted trends of computing resources (such as CPU and GPU), storage resources, and network bandwidth. By monitoring these metrics in real time, the cloud-based agent can gain a comprehensive understanding of the system's operational status and identify potential bottlenecks or overload risks. The collaborative contribution reflects the degree to which each edge node agent contributes to improving the overall system's classification accuracy within the federated learning framework. The collaborative contribution is calculated based on the results of the preset classification model. Specifically, each node uploads its corresponding cluster centers to the cloud-based agent, which then calculates the global cluster centers based on this data and updates the preset classification model accordingly. During this process, the node's collaborative contribution is measured by the improvement in classification accuracy after its participation. For example, if a node's data significantly improves the global classification accuracy, that node will receive a higher collaborative contribution score.

[0122] Step S302: Normalize the coordination contribution to obtain the processed coordination contribution.

[0123] It should be noted that the collaborative contribution reflects the improvement in global classification accuracy by edge node agents within the federated learning framework. However, since different nodes may have different computing power, data volume, and participation levels, directly comparing their raw contributions may lead to unfair results. Therefore, normalization mapping of all node contributions to a uniform standard range (e.g., between 0 and 1) can more fairly reflect the actual contributions of each node and provide a reliable basis for subsequent resource scheduling decisions. Specifically, the raw collaborative contribution values ​​of all participating nodes are collected. Then, the maximum (Max) and minimum (Min) values ​​of these values ​​are determined. Next, the contribution of each node is linearly transformed using the following formula:

[0124]

[0125] This method allows the contributions of all nodes to be compared on the same scale, avoiding bias caused by differences in node size. Furthermore, this normalization method preserves the relative relationships between the original contributions, ensuring that high-contribution nodes maintain their dominant position after adjustment.

[0126] Step S303: Calculate the load rate based on the target resource request information and the global resource status.

[0127] It's important to note that the target resource request information includes the specific requirements of the edge node agent for resources such as CPU, GPU, memory, storage, and bandwidth. This information reflects the user's expected usage over a future period. Meanwhile, the global resource status provides a detailed view of the currently available resources across the entire system, including the actual capacity and allocated amount of each resource pool. By combining the target resource request quantity with the global resource status, the load rate for each resource dimension can be calculated. The specific load rate can be calculated using the following formula:

[0128]

[0129] Step S304: Based on the comparison between the load rate and the preset safety threshold, the overload determination result is obtained.

[0130] It's important to note that safety thresholds are typically set as empirical values ​​to ensure the system can provide stable service quality even under high load. For example, for CPU resources, the safety threshold might be set to 80%, meaning that when the CPU load reaches or exceeds 80%, the system considers the resource to be at risk of overload. Specifically, for each resource dimension, if its load rate is below the safety threshold, it indicates that the resource is currently in a safe state and can continue to accept new resource requests; conversely, if the load rate equals or exceeds the safety threshold, an overload alarm is triggered. At this point, the system needs to take immediate action to prevent service quality degradation or system crashes due to excessive resource consumption. For example, this can be achieved by dynamically adjusting resource allocation strategies, prioritizing the needs of critical tasks, or temporarily rejecting non-urgent task requests until the load returns to normal levels.

[0131] Step S305: Obtain the credit score based on the overload determination result.

[0132] It's important to note that after the system determines an overload by comparing the load rate with a preset safety threshold, it updates the credit score of each node accordingly. If a node is determined not to have caused an overload, its credit score will either increase or remain unchanged. This reflects the node's positive contribution to the overall health of the system. Conversely, if a node's resource requests lead to an overload, its credit score will be penalized to varying degrees depending on the specific level of overload.

[0133] Specifically, step S305 includes: obtaining an initial credit score. Specifically, each edge node agent has an initial credit score, typically set to 1.0 and within the range of [0.5, 5.0]. This initial credit score reflects that the node is considered harmless to the system at the initial stage, i.e., it is assumed that they will not cause system overload or performance degradation. This setting helps provide a level playing field for all nodes and also facilitates dynamic adjustments based on actual performance later on.

[0134] Furthermore, when the overload determination result indicates an overload situation, the initial credit score is multiplied by a preset penalty coefficient to obtain the first credit score. Specifically, when the overload determination result shows that a node's resource requests have caused system overload, the system will adjust the node's initial credit score according to the preset penalty coefficient. The penalty coefficient is a custom value, usually set between 0.1 and 0.2, and the specific value can be determined according to actual conditions and needs. For example, if a node's initial credit score is 1.0, and the penalty coefficient is set to 0.2, then the node's new credit score (first credit score) will become... This mechanism effectively lowers the priority of nodes that frequently cause system overload, prompting these nodes to reassess and optimize their resource request strategies to avoid future overload situations.

[0135] Furthermore, when the overload assessment result indicates that no overload has occurred, the initial credit score is multiplied by a preset reward coefficient to obtain a second credit score. If the overload assessment result indicates that none of a node's resource requests have caused system overload, then that node's credit score will be increased. This is usually adjusted by applying a preset reward coefficient. The reward coefficient is also a custom value, with a recommended range between 0.05 and 0.1. For example, if a node's initial credit score is 1.0 and the reward coefficient is set to 1.1, then the new credit score (second credit score) will be... This approach not only encourages nodes to maintain good behavioral patterns, but also promotes the stability and efficiency of the entire system.

[0136] Step S306: Calculate the resource request priority based on the credit score and the processed coordination contribution.

[0137] It should be noted that the credit score reflects a node's historical performance and its contribution to system stability, while the coordination contribution measures the specific contribution of a node to improving global classification accuracy within the federated learning framework. After the system obtains the latest credit score and normalized coordination contribution for each edge node agent, these two metrics jointly determine the priority of the node when requesting resources. The specific formula is as follows:

[0138]

[0139] in, Indicates credit score, This is the weighting coefficient for the credit score. The weighting coefficient is the inverse of the historical average latency. The weighting coefficients for the normalized coordination contribution, and satisfying the following conditions: Historical average latency refers to the average time required for a node to complete a task over a past period. It is calculated using a sliding window with a window size of one hour. To further optimize resource scheduling, the system can dynamically adjust weighting coefficients based on actual operating conditions. For example, in certain scenarios, if the system is primarily affected by latency, weighting coefficients can be appropriately increased. The value of is adjusted to give more weight to the historical average latency of nodes; conversely, if the main problem facing the system is excessive resource consumption, the value can be increased. The weighting of credit scores emphasizes their importance.

[0140] Step S307: Sort the target resource request information according to the resource request priority to determine the resource allocation order.

[0141] It's important to note that the system sorts all target resource requests in descending order based on the resource request priority of each node (combined with credit score and coordination contribution). This means the highest priority requests are prioritized, with lower priority requests following in descending order. For example, in a scenario with multiple types of resource requirements (such as CPU, GPU, memory, etc.), priority sorting ensures that nodes that contribute most to system stability and performance improvement receive their requests first. After sorting, the system processes resource requests one by one in this order until all available resources are allocated appropriately or all requests are processed. In resource-constrained situations, low-priority requests may not immediately receive all required resources. In such cases, the system can adopt flexible strategies, such as partially fulfilling the request or temporarily queuing it. In this way, the system can utilize existing resources to serve as many users as possible while ensuring that critical tasks are not affected.

[0142] Step S308: Based on the resource allocation order and the target resource request information and global resource status, obtain the resource scheduling result.

[0143] It's important to note that after prioritizing resource requests, the system dynamically matches them based on the current global resource status (such as available computing resources, storage capacity, and network bandwidth for cloud-based and edge-node agents). For each high-priority resource request, the system checks if there are sufficient idle resources available for allocation. If resources are sufficient, the request is fulfilled directly; if resources are insufficient, partial allocation is attempted or the request is placed in a waiting queue, with adjustments made based on subsequent resource releases. Simultaneously, the scheduling process considers factors such as task type, latency sensitivity, and energy consumption constraints to further optimize resource combination and configuration. The final resource scheduling result not only reflects the principles of fairness and efficiency but also dynamically adapts to changes in system load, improving overall service quality and resource utilization.

[0144] This embodiment calculates resource request priority by using credit score and coordination contribution, and performs load analysis and overload determination by combining global resource status, thereby achieving dynamic, fair and efficient resource scheduling, improving system stability and resource utilization, and enhancing overall service quality and scheduling flexibility.

[0145] Based on the first embodiment of this application, this application also provides a resource scheduling result generation device based on cloud-edge multi-agent systems. Please refer to... Figure 4 The device includes:

[0146] Module 10 is used to obtain user resource usage data.

[0147] The feature extraction module 20 is used to extract features from user resource usage data to obtain feature vectors. The feature vectors include the feature vectors of the mean, variance, maximum value, and minimum value of CPU, GPU, memory, and storage.

[0148] The classification module 30 is also used to classify the feature vectors using a cloud-edge collaborative clustering algorithm to obtain user classification labels.

[0149] The processing module 40 is also used to obtain resource request information by inputting user classification tags into a preset classification model.

[0150] The optimization module 50 is used to optimize the resource request information according to the reinforcement learning strategy to obtain the target resource request information.

[0151] The results module 60 is used to obtain resource scheduling results based on the target resource request information combined with the global resource status and a dual-weighting mechanism of credit score and collaborative contribution.

[0152] The cloud-edge multi-agent resource scheduling result generation device provided in this application, employing the cloud-edge multi-agent resource scheduling result generation method in the above embodiments, can solve the technical problem of how to improve the efficiency and flexibility of resource scheduling. Compared with the prior art, the beneficial effects of the cloud-edge multi-agent resource scheduling result generation device provided in this application are the same as those of the cloud-edge multi-agent resource scheduling result generation method provided in the above embodiments, and other technical features in the cloud-edge multi-agent resource scheduling result generation device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0153] In one embodiment, the classification module 30 is further configured to initialize cluster centers to obtain a preset number of initial cluster center points; perform local classification on the feature vectors to obtain local cluster centers; encrypt the local cluster centers and upload them to the cloud-based intelligent agent, so that the cloud-based intelligent agent can perform global aggregation calculation on the local cluster centers to obtain global cluster centers; calculate the average cluster centers based on the global cluster centers; perform global balanced classification processing on the average cluster centers according to a preset load level standard to generate calibrated classification labels; add the calibrated classification labels to the feature vectors to obtain enhanced feature vectors; update the initial cluster center points based on the enhanced feature vectors to obtain updated cluster centers; and divide the data based on the updated cluster centers to obtain user classification labels.

[0154] In one embodiment, the processing module 40 is further configured to acquire historical resource scheduling data and construct an initial classification model. The historical resource scheduling data includes the user's historical resource usage records, the execution status of scheduling tasks, resource allocation results, and corresponding energy consumption data. The historical resource scheduling data is preprocessed to obtain processed data. User behavior features are extracted based on the processed data. The user behavior features are input into the initial classification model for training and verification to obtain a preset classification model.

[0155] In one embodiment, the processing module 40 is further configured to: determine the resource demand type based on the user classification tags input into a preset classification model; the resource demand type includes high computing demand, low latency sensitivity, high storage demand, and high bandwidth demand; obtain the corresponding resource request baseline value based on the resource demand type; optimize the resource request baseline value based on historical scheduling data and user behavior patterns; select the resource request baseline value corresponding to the resource demand type as the basic resource request quantity; adjust the basic resource request quantity by introducing a random perturbation mechanism to generate a differentiated resource request strategy; and obtain resource request information based on the differentiated resource request strategy.

[0156] In one embodiment, the result module 60 is further configured to obtain the global resource status and coordination contribution, wherein the coordination contribution is calculated based on the results of a preset classification model; the coordination contribution is normalized to obtain a processed coordination contribution; the load rate is calculated based on the target resource request information and the global resource status; an overload determination result is obtained by comparing the load rate with a preset safety threshold; a credit score is obtained based on the overload determination result; a resource request priority is calculated based on the credit score and the processed coordination contribution; the target resource request information is sorted according to the resource request priority to determine the resource allocation order; and a resource scheduling result is obtained by combining the target resource request information and the global resource status according to the resource allocation order.

[0157] In one embodiment, the result module 60 is further configured to obtain an initial credit score; when the overload determination result indicates that an overload has occurred, the initial credit score is multiplied by a preset penalty coefficient to obtain a first credit score; when the overload determination result indicates that no overload has occurred, the initial credit score is multiplied by a preset reward coefficient to obtain a second credit score.

[0158] In one embodiment, the result module 60 is further configured to: obtain the current credit score; update the target cluster center points in the cloud-edge collaborative clustering algorithm according to a first preset time to obtain updated cluster center points; rank and adjust the current credit score according to a second preset time to obtain an exploration rate; optimize the target resource request information based on the updated cluster center points and exploration rate to obtain optimized target resource request information; and execute the step of obtaining resource scheduling results based on the optimized target resource request information combined with the global resource status and the dual weighting mechanism of credit score and collaborative contribution.

[0159] This application provides a cloud-edge multi-agent resource scheduling result generation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the cloud-edge multi-agent resource scheduling result generation method in the above embodiment 1.

[0160] The following is for reference. Figure 5This document illustrates a schematic diagram of a cloud-edge multi-agent resource scheduling result generation device suitable for implementing embodiments of this application. The cloud-edge multi-agent resource scheduling result generation device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 5 The resource scheduling result generation device based on cloud-edge multi-agent shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0161] like Figure 5 As shown, the cloud-edge multi-agent resource scheduling result generation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the cloud-edge multi-agent resource scheduling result generation device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the cloud-edge multi-agent resource scheduling result generation device to wirelessly or wiredly communicate with other devices to exchange data. Although various cloud-edge multi-agent resource scheduling result generation devices are shown in the figures, it should be understood that it is not required to implement or possess all of them. More or fewer may be implemented alternatively.

[0162] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0163] The cloud-edge multi-agent resource scheduling result generation device provided in this application, employing the cloud-edge multi-agent resource scheduling result generation method described in the above embodiments, can solve the technical problem of how to improve the efficiency and flexibility of resource scheduling. Compared with the prior art, the beneficial effects of the cloud-edge multi-agent resource scheduling result generation device provided in this application are the same as those of the cloud-edge multi-agent resource scheduling result generation method provided in the above embodiments, and other technical features in this cloud-edge multi-agent resource scheduling result generation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0164] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0165] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0166] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the cloud-edge multi-agent resource scheduling result generation method in the above embodiments.

[0167] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible storage medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable storage medium may be transmitted using any suitable storage medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0168] The aforementioned computer-readable storage medium may be included in a cloud-edge multi-agent resource scheduling result generation device; or it may exist independently and not be assembled into a cloud-edge multi-agent resource scheduling result generation device.

[0169] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a cloud-edge multi-agent resource scheduling result generation device, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0170] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or can be implemented using a combination of dedicated hardware and computer instructions.

[0171] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0172] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described cloud-edge multi-agent resource scheduling result generation method, thereby solving the technical problem of how to improve the efficiency and flexibility of resource scheduling. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the cloud-edge multi-agent resource scheduling result generation method provided in the above embodiments, and will not be repeated here.

[0173] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the resource scheduling result generation method based on cloud-edge multi-agent as described above.

[0174] The computer program product provided in this application can solve the technical problem of how to improve the efficiency and flexibility of resource scheduling. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the cloud-edge multi-agent resource scheduling result generation method provided in the above embodiments, and will not be repeated here.

[0175] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for generating resource scheduling results based on cloud-edge multi-agent systems, characterized in that, The method includes: Obtain user resource usage data; Feature extraction is performed on the user resource usage data to obtain feature vectors, which include feature vectors of the mean, variance, maximum value, and minimum value of CPU, GPU, memory, and storage. The feature vectors are classified using a cloud-edge collaborative clustering algorithm to obtain user classification labels; Input the user classification tags into a preset classification model to obtain resource request information; The resource request information is optimized according to a reinforcement learning strategy to obtain the target resource request information; Based on the target resource request information and combined with the global resource status and a dual-weighting mechanism of credit score and collaborative contribution, the resource scheduling result is obtained; The step of obtaining the resource scheduling result based on the target resource request information combined with the global resource status and a dual-weight mechanism of credit score and cooperation contribution includes: Obtain the global resource status and coordination contribution, wherein the coordination contribution is calculated based on the results of the preset classification model; The coordination contribution is normalized to obtain the processed coordination contribution. The load rate is calculated based on the target resource request information and the global resource status. The overload determination result is obtained by comparing the load rate with the preset safety threshold; A credit score is obtained based on the overload determination result; The resource request priority is calculated based on the credit score and the processed coordination contribution. The target resource request information is sorted according to the resource request priority to determine the resource allocation order; The resource scheduling result is obtained by combining the resource allocation order with the target resource request information and the global resource status.

2. The method as described in claim 1, characterized in that, The step of classifying the feature vector using a cloud-edge collaborative clustering algorithm to obtain user classification labels includes: Initialize cluster centers to obtain a preset number of initial cluster center points; The feature vectors are classified locally to obtain local cluster centers; The local cluster center is encrypted and uploaded to the cloud-based intelligent agent, so that the cloud-based intelligent agent can perform global aggregation calculation on the local cluster center to obtain the global cluster center; The average cluster center is calculated based on the global cluster center. The average cluster centers are subjected to global balanced classification processing according to a preset load level standard to generate calibrated classification labels. The calibrated classification labels are added to the feature vector to obtain the enhanced feature vector; The initial cluster centers are updated based on the enhanced feature vector to obtain the updated cluster centers; Based on the updated cluster centers, user category labels are obtained.

3. The method as described in claim 1, characterized in that, Before the step of obtaining the resource request strategy by inputting the user classification tags into the preset classification model, the following steps are included: Acquire historical resource scheduling data and construct an initial classification model. The historical resource scheduling data includes users' historical resource usage records, execution status of scheduling tasks, resource allocation results, and corresponding energy consumption data. The historical resource scheduling data is preprocessed to obtain the processed data; User behavior characteristics are extracted from the processed data. The user behavior features are input into the initial classification model for training and validation to obtain the preset classification model.

4. The method as described in claim 1, characterized in that, The step of inputting the user classification tags into a preset classification model to obtain resource request information includes: Based on the user classification tags input into the preset classification model, the resource demand type is determined. The resource demand type includes high computing demand, low latency sensitivity, high storage demand, and high bandwidth demand. Based on the resource demand type, a corresponding resource request baseline value is obtained, which is optimized based on historical scheduling data and user behavior patterns. Select the resource request baseline value corresponding to the resource demand type as the basic resource request quantity; Based on the aforementioned basic resource request volume, a random perturbation mechanism is introduced for adjustment to generate a differentiated resource request strategy; Resource request information is obtained based on the differentiated resource request strategy.

5. The method as described in claim 1, characterized in that, The step of obtaining a credit score based on the overload determination result includes: Obtain an initial credit score; When the overload determination result indicates that an overload has occurred, the initial credit score is multiplied by a preset penalty coefficient to obtain the first credit score. When the overload determination result is that no overload situation has occurred, the initial credit score is multiplied by the preset reward coefficient to obtain the second credit score.

6. The method as described in claim 1, characterized in that, After the step of obtaining the resource scheduling result based on the target resource request information combined with the global resource status and the dual-weight mechanism of credit score and cooperation contribution, the following steps are included: Get your current credit score; The target cluster center points in the cloud-edge collaborative clustering algorithm are updated according to a first preset time to obtain the updated cluster center points; The current credit score is ranked and adjusted according to a second preset time to obtain the exploration rate; The target resource request information is optimized based on the updated cluster center points and the exploration rate to obtain optimized target resource request information; Based on the optimized target resource request information, the step of obtaining the resource scheduling result by combining the target resource request information with the global resource status and the dual weighting mechanism of credit score and cooperation contribution is executed.

7. A resource scheduling result generation device based on cloud-edge multi-agent system, characterized in that, The device includes: The acquisition module is used to acquire user resource usage data; The feature extraction module is used to extract features from the user resource usage data to obtain feature vectors, which include feature vectors of the mean, variance, maximum value, and minimum value of CPU, GPU, memory, and storage. The classification module is also used to classify the feature vectors using a cloud-edge collaborative clustering algorithm to obtain user classification labels; The processing module is also used to input the user classification tags into a preset classification model to obtain resource request information; The optimization module is used to optimize the resource request information according to the reinforcement learning strategy to obtain the target resource request information; The result module is used to obtain resource scheduling results based on the target resource request information combined with global resource status and a dual-weight mechanism of credit score and collaboration contribution; it is also used to obtain global resource status and coordination contribution, the coordination contribution being calculated based on the results of the preset classification model; normalize the coordination contribution to obtain a processed coordination contribution; calculate the load rate based on the target resource request information and global resource status; compare the load rate with a preset safety threshold to obtain an overload determination result; obtain a credit score based on the overload determination result; calculate the resource request priority based on the credit score and the processed coordination contribution; sort the target resource request information according to the resource request priority to determine the resource allocation order; and obtain the resource scheduling result according to the resource allocation order combined with the target resource request information and the global resource status.

8. A resource scheduling result generation device based on cloud-edge multi-agent system, characterized in that, The device includes: a memory, a processor, and a cloud-edge multi-agent resource scheduling result generation program stored in the memory and running on the processor, wherein the cloud-edge multi-agent resource scheduling result generation program is configured to implement the steps of the cloud-edge multi-agent resource scheduling result generation method as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a resource scheduling result generation program based on cloud-edge multi-agent systems. When the resource scheduling result generation program based on cloud-edge multi-agent systems is executed by the processor, it implements the steps of the resource scheduling result generation method based on cloud-edge multi-agent systems as described in any one of claims 1-6.

Images