An AI agent-based dynamic load balancing management method, device and medium

Abstract

The application discloses a kind of dynamic load balancing management methods, equipment and medium based on AI intelligent agent, method includes: calculating node correlation coefficient, constructs GCN-GRU hybrid load prediction model, predicts the space-time feature matrix of future node, obtains the predicted node load data;Calculate the normalization weight of each load data, load comprehensive value and load balancing degree;Local agent, coordination agent are deployed, and local agent state space vector is constructed;The mapping relationship of local policy network is constructed using local agent state space vector, and the action space vector and migration amount of node execution load migration action are generated, global value is calculated, and the parameters of local policy network are optimized based on global value, and the optimal action space vector and corresponding migration amount are generated to carry out load balancing migration.Computer equipment includes processor and memory;Computer readable storage medium stores at least one instruction, at least one program, code set or instruction set.

Description

TECHNICAL FIELD

[0001] The application relates to the field of network load, in particular to a dynamic load balancing management method based on an AI agent, equipment and a medium. BACKGROUND

[0002] A distributed system is a system composed of a group of computer nodes that communicate through a network and coordinate work to complete common tasks. Load balancing of a distributed system is the core of ensuring service stability and resource utilization. In the prior art, load prediction only focuses on the time sequence characteristics of a single node, ignoring the spatial correlation between nodes due to data transmission and task dependency, resulting in insufficient prediction accuracy of load mutations (such as load fluctuations caused by task flow between nodes); load evaluation mostly uses a single weighting method (subjective or objective), which cannot balance business demand priority and data objectivity, and the evaluation result deviates greatly from the actual load state; the scheduling strategy is mostly centralized or single-agent decision-making, and the centralized architecture has the problems of high scheduling delay and poor scalability, and the single agent is easy to fall into local optimization and cannot balance local resource efficiency and global load balancing. AI agent is an intelligent system that can independently perceive the environment, make decisions and perform tasks, and is widely used in various fields. In order to address the deficiencies of current distributed system load scheduling, applying AI agent to dynamic load balancing management of distributed system is the current research direction. Based on this, the application provides a dynamic load balancing management method based on AI agent, equipment and medium, which is suitable for edge computing cluster, Internet of Things terminal network, cloud computing data center and other distributed scenarios with dynamic load and node association characteristics. SUMMARY

[0003] In view of the above deficiencies in the prior art, the application provides a dynamic load balancing management method based on AI agent, equipment and medium, which is suitable for edge computing cluster, Internet of Things terminal network, cloud computing data center and other distributed scenarios with dynamic load and node association characteristics.

[0004] In order to achieve the above application purpose, the technical scheme adopted by the application is:

[0005] A dynamic load balancing management method based on AI agent is provided, which comprises:

[0006] Step S1: Collecting load data of nodes in a distributed system, calculating node association coefficients based on load time sequence correlation coefficients and node topological distance;

[0007] Step S2: Based on the load change period of each node, calculating the time feature and space feature of historical load data, and constructing a GCN-GRU hybrid load prediction model to predict the spatio-temporal feature matrix of future nodes, and obtaining the predicted node load data;

[0008] Step S3: Calculate the average and standard deviation of the historical load data, then calculate the normalization weight of each load data, calculate the load comprehensive value of each node using the normalization weight, and calculate the load balancing degree of the distributed system;

[0009] Step S4: Deploy a local agent on each node of the distributed system, deploy a coordination agent globally in the distributed system, and construct a local agent state space vector using the load comprehensive value, real-time available resource data volume, predicted node load data, and average load of adjacent nodes;

[0010] Step S5: Construct the mapping relationship of the local policy network using the local agent state space vector, generate the action space vector and migration amount of the node executing the load migration action, calculate the global value, and optimize the parameters of the local policy network based on the global value to generate the optimal action space vector and the corresponding migration amount for load balancing migration.

[0011] Further, step S1 includes:

[0012] Step S11: Collect the load data of the nodes in the distributed system, and calculate the load time series correlation coefficient and node topology distance between the nodes;

[0013] Step S12: Calculate the node correlation coefficient using the load time series correlation coefficient and node topology distance between the nodes.

[0014] Further, step S2 includes:

[0015] Step S21: Based on the load change period of each node, collect the historical load data and resource data of each node, and calculate the time feature and space feature based on the historical load data;

[0016] The historical load data includes CPU load intensity, memory occupancy, network I / O traffic, and task execution density; the resource data includes CPU rated computing power, memory capacity, network bandwidth upper limit, and task processing threshold;

[0017] The time feature includes load instantaneous change amount and load change trend;

[0018] The space feature includes the average load of adjacent nodes and the recent node load:

[0019] Step S22: Construct a space-time feature matrix using the historical load data, resource data, time feature, and space feature of the node as input data , is the feature dimension of the node, is a real number set;

[0020] The GCN-GRU hybrid load prediction model is constructed, including a GCN spatial feature extraction layer, a GRU time sequence feature extraction layer and a fusion prediction output layer.

[0021] The GCN spatial feature extraction layer extracts 32-dimensional spatial features from the spatio-temporal feature matrix based on the node correlation coefficient by using a first layer graph convolution, extracts 64-dimensional spatial features by using a second layer graph convolution, and obtains a spatial feature sequence.

[0022] The GRU time sequence feature extraction layer extracts time sequence features from the spatial feature sequence and outputs a time sequence feature vector.

[0023] The fusion prediction output layer splices the time sequence feature vector and the spatial feature, outputs a spatio-temporal feature matrix of the node at a future time k, and obtains node load data predicted at a next collection time t+1. .

[0024] Further, the step S3 comprises:

[0025] Step S31: calculating the average value and the standard deviation of each kind of load data according to the historical load data of each node;

[0026] Step S32: calculating the variation data of the historical load data according to the standard deviation of each kind of load data, and calculating the normalized weight of each kind of load data by using the variation data;

[0027] Step S33: calculating the load comprehensive value of each node and the load balancing degree E of the distributed system by using the normalized weight of each kind of load data.

[0028] Further, the step S4 comprises:

[0029] Step S41: deploying a local intelligent agent on each node of the distributed system , and deploying a coordination intelligent agent globally in the distributed system;the local intelligent agent is used for local load monitoring and decision-making, and the coordination intelligent agent is used for global load balancing optimization and conflict coordination;

[0030] Step S42: constructing a node local intelligent agent state space vector by using the load comprehensive value, the real-time available resource data amount, the predicted node load data and the average load of the adjacent nodes. .

[0031] Further, the step S5 comprises:

[0032] Step S51: using the parameters of the local strategy network of the local intelligent agent to calculate the local strategy network output a local agent state space vector a mapping relationship of constructing a local policy network

[0033] Step S52: generating an action space vector of node i based on the mapping relationship of the local policy network , a load migration action of node i to node n;

[0034] a load migration action a migration amount ; wherein, node data predicted for node i, a weight vector of the local policy network and a bias set by a load migration coefficient, historical load data of node i;

[0035] Step S53: constructing a global evaluation network in a coordination agent using all local agent state space vectors and action space vectors a global value evaluation function , calculating a global value ;

[0036] Step S54: setting a global value target value of the distributed system ;

[0037] if , it is determined that the current node i action space vector is optimal, and the nodes in the distributed system perform load balancing migration based on the optimal action space vector and the corresponding migration amount;

[0038] if , it is determined that the current node i action space vector is not optimal, and step S55 is performed;

[0039] Step S55: returning to step S51, correcting the parameters of the local policy network, and the step size of the parameter correction is , and the range of the parameter correction is , , is the correction sliding window length; in the correction process, a new parameter is obtained;

[0040] ​​​Step S56: Input each new parameter into the local policy network, execute steps S51-S54, and calculate the global value corresponding to each new parameter. If the overall value There exists at least one global value. satisfy If the condition is met, proceed to step S57; otherwise, the distributed system will issue a load alarm.

[0041] Step S57: Obtain the global value All of the above satisfy overall value , To meet The total value quantity, For the first One satisfies Global value; filtering global value The maximum value in , to the maximum value The corresponding action space vector is used as the optimal action space vector. Nodes in the distributed system perform load balancing migration based on the optimal action space vector and the corresponding migration amount.

[0042] A computer device is provided, comprising: a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the processor loads and executes the at least one instruction, at least one program, a code set, or an instruction set to implement the above-described dynamic load balancing management method based on AI intelligent agents.

[0043] A computer-readable storage medium is provided, which stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the above-described dynamic load balancing management method based on AI intelligent agents.

[0044] The beneficial effects of this invention are as follows: This invention achieves high-precision prediction of node load by modeling the spatial correlation of node load and using the GCN-GRU hybrid load prediction model for comprehensive prediction. Based on the predicted node load data, the load of the node is migrated and scheduled. The load migration and scheduling strategy of the local agent on the node is optimized by the global value evaluation of multiple agents, avoiding the local optimum of the agent and realizing the collaborative optimization of load balance and resource utilization of the distributed system. Attached Figure Description

[0045] Figure 1 This is a flowchart of a dynamic load balancing management method based on AI intelligent agents. Detailed Implementation

[0046] The specific embodiments of the present application are described below to enable those skilled in the art to understand the present application, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0047] As shown in Figure 1 , an AI agent-based dynamic load balancing management method comprises:

[0048] Step S1: Collecting load data of nodes in a distributed system, and calculating node correlation coefficient based on load time series correlation coefficient and node topology distance. Step S1 specifically comprises:

[0049] The load of n nodes in a distributed system has spatial correlation (for example, the task output of node A is the input of node B, which causes the load change of A to cause the load fluctuation of B), and the spatial correlation is represented by the node correlation coefficient.

[0050] Step S11: Collecting load data of nodes in a distributed system, and calculating load time series correlation coefficient between nodes and node topology distance ;

[0051] ;

[0052] Wherein, are the load data of node i and node j respectively, is the load covariance of node i and node j, are the load standard deviation of node i and node j respectively; node topology distance The network hop number between node i and node j is used to represent the node topology distance, such as 1 hop for direct connection and 2 hops for connection through 1 intermediate node.

[0053] Step S12: Calculating node correlation coefficient based on load time series correlation coefficient between nodes and node topology distance ;

[0054] ;

[0055] Wherein, is the attenuation coefficient based on node topology distance, The greater the network hop number between nodes, the lower the node correlation; is the number of nodes in the distributed system, , which represents the correlation between node i and node j. The larger the value, the stronger the correlation.

[0056] Step S2: Based on the load change cycle of each node, calculate the temporal and spatial characteristics of historical load data, and construct a GCN-GRU hybrid load prediction model to predict the spatiotemporal feature matrix of future nodes, thus obtaining the predicted node load data. Step S2 specifically includes:

[0057] Step S21: Based on the load change cycle of each node, collect historical load data and resource data of each node, and calculate time and spatial characteristics based on the historical load data;

[0058] The historical load data in this embodiment includes CPU load intensity. Memory usage Network I / O traffic and task execution density t represents the collection time of historical load data; resource data includes CPU rated computing power. Memory capacity Network bandwidth limit and task processing threshold .

[0059] Time characteristics include instantaneous changes in load. Load change trend ;

[0060] , ;

[0061] in, The sliding window length is calculated based on the load change trend; in this embodiment, it is taken as... This indicates 10 load change cycles. To collect historical load data at time t, To collect historical load data at time t-1, For collection time Historical load data;

[0062] Spatial characteristics include the average load of neighboring nodes and the load of the nearest node:

[0063] The average load of adjacent nodes is ;in, Let M be the set of nodes adjacent to node i, and M be the set of nodes. The number of nodes within, Historical load data for adjacent nodes. This refers to the historical load data for node i.

[0064] The recent node load is ;

[0065] ;

[0066] Step S22: Constructing a space-time feature matrix with the historical load data, resource data, time characteristics and space characteristics of the nodes as input data , is the feature dimension of the node, is a real set, and in this embodiment, d=8, including four kinds of historical load data, two time characteristics and two space characteristics;

[0067] The GCN-GRU hybrid load prediction model is constructed, including a GCN space feature extraction layer, a GRU time sequence feature extraction layer and a fusion prediction output layer.

[0068] The GCN space feature extraction layer extracts 32-dimensional space features from the space-time feature matrix based on the node correlation coefficient of the first layer graph convolution, and extracts 64-dimensional space features based on the second layer graph convolution, to obtain a space feature sequence; specifically including:

[0069] The node correlation coefficient of each node in the distributed system is used to construct a load correlation adjacency matrix , and an n-order unit matrix is introduced. The load correlation adjacency matrix is modified to obtain a modified adjacency matrix

[0070] The load correlation adjacency matrix is used to describe the degree of mutual influence of the loads of the nodes due to data transmission, task dependence and topological connection, and the larger the node correlation coefficient is, the more significant the influence of the load change of node i on the load fluctuation of node j is. The main diagonal elements of the unit matrix are all 1, which ensures that the load characteristics of each node are preferentially retained during GCN feature extraction, avoiding ignoring the state of the node itself due to excessive attention to neighbor nodes. The modified adjacency matrix solves the problem that the load correlation adjacency matrix may not contain the self-connection information of the node, so that the GCN considers both the "node itself characteristics" and the "neighbor node correlation characteristics" when aggregating the features, improving the completeness of the feature expression.

[0071] The modified adjacency matrix is symmetrically normalized to obtain a symmetric adjacency matrix , is the degree matrix of the modified adjacency matrix ;

[0072] The symmetric adjacency matrix The 32-dimensional spatial feature is extracted from the empty feature matrix in the first layer graph convolution and input into the second layer graph convolution.

[0073] ;

[0074] wherein, are the feature transformation weight matrix and the bias vector of the first layer graph convolution, respectively.

[0075] The second layer graph convolution extracts 64-dimensional spatial features from the spatial features .

[0076] ;

[0077] wherein, is the activation function for spatial feature extraction, are the feature transformation weight matrix and the bias vector of the second layer graph convolution, respectively.

[0078] The spatial feature sequence based on the sliding window length is output.

[0079] The GRU time sequence feature extraction layer extracts time sequence features from the spatial feature sequence, and outputs a time sequence feature vector ; specifically including: a reset gate: ;

[0080] wherein, is the output of the reset gate, are the feature weight matrix, the hidden state weight matrix and the bias vector of the reset gate, respectively, is the hidden state at the collection time , is the activation function for time sequence feature extraction, is the output of the GRU time sequence feature extraction layer, ;

[0081] An update gate: ;

[0082] wherein, is the output of the update gate, are the feature weight matrix, the hidden state weight matrix and the bias vector of the update gate, respectively.

[0083] The candidate hidden state: ;

[0084] wherein, is the candidate hidden state at the collection time t,​ These are the feature weight matrix, hidden state weight matrix, and bias vector of the hidden state, respectively. It is a non-linear activation function;

[0085] Hidden status update:

[0086] ;

[0087] The fusion prediction output layer will combine temporal feature vectors and spatial features The data is concatenated to output the spatiotemporal feature matrix of the nodes predicted at future time k, thus obtaining the node load data predicted at the next data collection time t+1. ;

[0088] ;

[0089] in, These are the feature weight matrix and bias vector of the fused prediction output layer, respectively.

[0090] This invention breaks through the limitations of traditional time series prediction by integrating Graph Convolutional Neural Network (GCN) with Gated Recurrent Unit (GRU) for the first time. By capturing the spatial correlation of load between nodes through GCN and capturing the temporal dependency through GRU, a "spatial-temporal" dual-dimensional prediction model is constructed to solve the problem of insufficient accuracy in load mutation prediction.

[0091] Step S3: Calculate the average and standard deviation of historical load data, then calculate the normalized weight for each type of load data. Use the normalized weights to calculate the overall load value for each node, and calculate the load balance of the distributed system. Step S3 specifically includes:

[0092] Step S31: Calculate the average value of each type of load data based on the historical load data of each node. and standard deviation ;

[0093] ;

[0094] in, Where N represents the type of load data and N represents the number of historically collected load data. Historical load data;

[0095] Step S32: Based on the standard deviation of each load data Calculate the variation data of historical load data ; then utilize the mutated data Calculate the normalized weights for each type of load data. ;

[0096] ;

[0097] Step S33: Utilize the normalized weights of each load data type Calculate the overall load value for each node. And calculate the load balancing degree E of the distributed system;

[0098] ;

[0099] in, For the first Resource limits for various load data This represents the average value of the overall load.

[0100] Step S4: Deploy a local agent on each node of the distributed system and a coordinating agent globally. Construct the state space vector of the local agent using the overall load value, real-time available resource data, predicted node load data, and the average load of neighboring nodes. Step S4 specifically includes:

[0101] Step S41: Deploy local intelligent agents on each node of the distributed system A coordinated intelligent agent is deployed globally in a distributed system. Local intelligent agent Used for local load monitoring and decision-making, coordinating intelligent agents. Used for global load balancing optimization and conflict coordination;

[0102] Step S42: Utilize the overall load value Real-time available resource data volume Predicted node load data and the average load of adjacent nodes Construct the local agent state space vector of the node. ;

[0103] ;

[0104] in, These are the node's available CPU computing power, memory capacity, network bandwidth limit, and task processing threshold, respectively. These are the node's rated CPU computing power, memory capacity, network bandwidth limit, and task processing threshold, respectively.

[0105] Step S5: Construct a mapping relationship for the local policy network using the local agent's state space vector, generate action space vectors and migration amounts for nodes to perform load balancing actions, calculate the global value, and optimize the parameters of the local policy network based on the global value to generate the optimal action space vector and corresponding migration amount for load balancing migration. Step S5 specifically includes:

[0106] Step S51: using the local agent parameters of the local policy network and the local agent state space vector to construct the mapping relationship of the local policy network ;

[0107] ;

[0108] wherein, is the weight vector and bias of the local policy network, ;

[0109] Step S52: generating the action space vector of node i based on the mapping relationship of the local policy network , is the node i to node n load migration action;

[0110] the migration amount of the load migration action ; wherein, is the node i predicted node data, is the load migration coefficient set based on the weight vector and bias of the local policy network; Step S53: using the state space vector

[0111] and the action space vector of all local agents to construct the global value evaluation function of the global evaluation network in the coordination agent , calculate the global value ; ;

[0112] ;

[0113] wherein, is the local agent state space vector of node n, is the action space vector of node n, is the parameter of the global evaluation network;

[0114] Step S54: setting the global value target value of the distributed system ;

[0115] If , it is determined that the current node i action space vector is optimal, and the nodes in the distributed system perform load balancing migration based on the optimal action space vector and the corresponding migration amount;

[0116] If​ Then determine the action space vector of the current node i. This is not optimal; proceed to step S55.

[0117] Step S55: Return to step S51 and adjust the parameters of the local policy network. Make corrections, parameters The correction step size is And the range of parameter correction is , To correct the sliding window length; during the correction process, we obtain A new parameter;

[0118] Step S56: Input each new parameter into the local policy network, execute steps S51-S54, and calculate the global value corresponding to each new parameter. If the overall value There exists at least one global value. satisfy If the condition is met, proceed to step S57; otherwise, the distributed system will issue a load alarm.

[0119] Step S57: Obtain the global value All of the above satisfy overall value , To meet The total value quantity, For the first

[0120] One satisfies Global value; filtering global value The maximum value in , to the maximum value The corresponding action space vector is used as the optimal action space vector. Nodes in the distributed system perform load balancing migration based on the optimal action space vector and the corresponding migration amount.

[0121] A computer device, characterized in that the computer device includes: a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the above-mentioned dynamic load balancing management method based on AI intelligent agents.

[0122] A computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the above-described dynamic load balancing management method based on AI intelligent agents.

[0123] This invention achieves high-precision prediction of node load by modeling the spatial correlation of node load and using a GCN-GRU hybrid load prediction model for comprehensive prediction. Based on the predicted node load data, it performs load migration scheduling on the nodes and optimizes the load migration scheduling strategy of the local agents on the nodes through global value evaluation of multiple agents, avoiding local optima of the ontology agents and achieving collaborative optimization of load balance and resource utilization in the distributed system.

Claims

1. A dynamic load balancing management method based on AI intelligent agents, characterized in that, include: Step S1: Collect load data of nodes in the distributed system, and calculate node correlation coefficient based on load time-series correlation coefficient and node topology distance; Step S1 specifically includes: Step S11: Collect load data of nodes in the distributed system and calculate the load time-series correlation coefficient between nodes. and node topological distance ; ； in, They are nodes i and nodes j Load data, For nodes i With nodes j The load covariance, They are nodes i and nodes j Load standard deviation; node topology distance Use nodes i With nodes j The network hop count between them is represented by the number of network hops; Step S12: Utilize the load timing correlation coefficient between nodes and node topological distance Calculate the node correlation coefficient ; in, The attenuation coefficient is based on the node topological distance. The number of nodes in a distributed system; Step S2: Based on the load change cycle of each node, calculate the temporal and spatial characteristics of historical load data, and construct a GCN-GRU hybrid load prediction model to predict the spatiotemporal feature matrix of future nodes and obtain the predicted node load data. Step S2 includes: Step S21: Based on the load change cycle of each node, collect historical load data and resource data of each node, and calculate time and spatial characteristics based on the historical load data; Historical load data includes CPU load intensity, memory usage, network I / O traffic, and task execution density; resource data includes CPU rated computing power, memory capacity, network bandwidth limit, and task processing threshold. Time characteristics include instantaneous load changes and load change trends; Spatial characteristics include the average load of neighboring nodes and the load of the nearest node: Step S22: Construct a spatiotemporal feature matrix by using the node's historical load data, resource data, temporal features, and spatial features as input data. , For the feature dimensions of the node, It is the set of real numbers; A GCN-GRU hybrid load prediction model is constructed, including: a GCN spatial feature extraction layer, a GRU temporal feature extraction layer, and a fusion prediction output layer; The GCN spatial feature extraction layer uses the first-layer graph convolution to extract features from the spatiotemporal feature matrix based on node correlation coefficients. 32-dimensional spatial features are extracted from the first layer, and 64-dimensional spatial features are extracted using the second layer graph convolution to obtain a spatial feature sequence. The GRU temporal feature extraction layer extracts temporal features from spatial feature sequences and outputs a temporal feature vector. The fusion prediction output layer concatenates temporal feature vectors and spatial features to output future time-series features. k The predicted spatiotemporal feature matrix of the node is used to obtain the next future acquisition time. t +1 Predicted Node Load Data ; Step S3: Calculate the average and standard deviation of historical load data, then calculate the normalized weight of each type of load data, use the normalized weight to calculate the overall load value of each node, and calculate the load balance of the distributed system. Step S4: Deploy a local agent on each node of the distributed system and a coordinating agent globally in the distributed system. Construct the state space vector of the local agent using the comprehensive load value, real-time available resource data, predicted node load data, and average load of neighboring nodes. Step S5: Construct the mapping relationship of the local policy network using the state space vector of the local agent, generate the action space vector and migration amount of the node to perform load migration actions, calculate the global value, optimize the parameters of the local policy network based on the global value, and generate the optimal action space vector and the corresponding migration amount for load balancing migration.

2. The dynamic load balancing management method based on AI intelligent agents according to claim 1, characterized in that, Step S3 includes: averaging and standard deviation ; in, For the types of load data, N The amount of historically collected load data. Historical load data; Step S32: Based on the standard deviation of each load data Calculate the variation data of historical load data ; then utilize the mutated data Calculate the normalized weights for each type of load data. ; ； Step S33: Utilize the normalized weights of each load data type Calculate the overall load value for each node. And calculate the load balancing degree of a distributed system. E ; in, For the first Resource limits for various load data This represents the average value of the overall load.

3. The dynamic load balancing management method based on AI intelligent agents according to claim 2, characterized in that, Step S4 includes: Step S41: Deploy local intelligent agents on each node of the distributed system A coordinated intelligent agent is deployed globally in a distributed system. Local intelligent agent Used for local load monitoring and decision-making, coordinating intelligent agents. Used for global load balancing optimization and conflict coordination; Step S42: Utilize the overall load value, real-time available resource data, and predicted node load data. Construct the local agent state space vector of the node based on the average load of neighboring nodes. .

4. The dynamic load balancing management method based on AI intelligent agents according to claim 3, characterized in that, Step S5 includes: Step S51: Utilize the local intelligent agent Parameters of the local policy network With the local agent's state space vector Constructing the mapping relationship of the local policy network ; Step S52: Mapping relationship based on local policy network Generate nodes i Action space vector , For nodes i To the node n Load migration action; Load migration action migration amount ;in, For nodes i Predicted node data, The weight vector of the local policy network and bias The set load migration factor, For nodes i Historical load data; Step S53: Utilize all local agents state space vector and action space vector Building a coordinated intelligent agent Global value evaluation function of global evaluation network Calculate the global value ; Step S54: Set the global value target value of the distributed system ; like Then determine the current node. i Action space vector In the optimal distributed system, nodes perform load balancing migration based on the optimal action space vector and the corresponding migration amount. like Then determine the current node. i Action space vector This is not optimal; proceed to step S55. Step S55: Return to step S51 and adjust the parameters of the local policy network. Make corrections, parameters The correction step size is And the range of parameter correction is , To correct the sliding window length; during the correction process, we obtain A new parameter; Step S56: Input each new parameter into the local policy network, execute steps S51-S54, and calculate the global value corresponding to each new parameter. If the overall value There exists at least one global value. satisfy If the condition is met, proceed to step S57; otherwise, the distributed system will issue a load alarm. Step S57: Obtain the global value All of the above satisfy overall value , To meet The total value quantity, For the first One satisfies Global value; filtering global value The maximum value in , to the maximum value The corresponding action space vector is used as the optimal action space vector. Nodes in the distributed system perform load balancing migration based on the optimal action space vector and the corresponding migration amount.

5. A computer device, characterized in that, The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the dynamic load balancing management method based on AI intelligent agents as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the dynamic load balancing management method based on AI intelligent agents as described in any one of claims 1 to 4.

Images