A highly reliable resource scheduling optimization method for heterogeneous computing networks based on coded distributed computing and hypergraph neural networks

Abstract

The invention discloses a highly reliable resource scheduling optimization method for heterogeneous computing networks based on coded distributed computing and hypergraph neural networks. First, a multi-level resource representation model is constructed, using a hypergraph neural network to capture high-order correlations between heterogeneous nodes, thereby optimizing the accuracy and efficiency of dynamic resource scheduling. Second, a coded distributed computing strategy is introduced to address edge node failures through redundant load distribution, enhancing system stability. To further improve scheduling optimization, a hybrid algorithm framework combining hypergraph neural networks and particle swarm optimization is proposed. This framework uses hypergraph modeling to capture high-order correlations between heterogeneous nodes in the network. At the same time, combined with the particle swarm optimization algorithm, it can effectively handle discrete and discontinuous problems. Ultimately, the hybrid algorithm improves computational efficiency while ensuring solution quality, achieving efficient resource scheduling optimization and demonstrating its broad application potential in future 6G application scenarios.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular to a high-reliability resource scheduling optimization method for heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks. Background Art

[0002] In the 6G era, the surge in data and data-intensive applications poses unprecedented challenges to current communication and computing networks. The collaborative efforts of cloud computing, edge computing, and the Internet of Things are imperative to process this massive amount of data and ultimately achieve ubiquitous computing and intelligence. "Computing networks," a key direction for future networks, are becoming a hot topic of research in academia and industry. Any network node in these networks can simultaneously route traffic, perform computations, and store data. In many of these computing environments, large amounts of data, including large datasets, machine learning models, or executable code, need to be cached.

[0003] Computing networks integrate distributed computing resources to provide powerful computing support for various applications and services. Resource scheduling optimization is crucial in current computing networks, as it improves overall system performance, reduces energy consumption, and ensures real-time and reliable services. Whether in ubiquitous edge IoT systems, mobile edge computing, the Industrial Internet of Things, federated learning networks, or wireless computing networks, resource scheduling optimization is key to achieving efficient energy utilization, meeting task latency constraints, and maximizing system throughput. Therefore, the design and management strategies for resource representation models in computing networks directly impact network performance and efficiency.

[0004] The existing computing network architecture still faces several challenges and shortcomings in terms of resource integration and dynamic scheduling. First, the delivery model for computing network services has undergone significant changes. The flexible provision of increasingly diverse services is achieved through the agile sharing of smaller and smaller infrastructure units. Traditional network resource models are no longer suitable and cannot directly and clearly demonstrate the interactive relationships between network resources, necessitating a design overhaul. Second, the heterogeneity and time-varying nature of computing networks introduce management complexity. Heterogeneity means that computing nodes exhibit significant differences in computing power, storage capacity, and energy consumption. Time-varying nature reveals that edge nodes may erratically enter and exit the network, leading to performance instability. Therefore, more sophisticated and intelligent algorithms are needed to optimize resource allocation and task scheduling. Summary of the Invention

[0005] To solve the above problems, the present invention discloses a highly reliable resource scheduling optimization method for heterogeneous computing networks based on coded distributed computing and hypergraph neural networks, comprising:

[0006] Step 1: Build a computing network model for collaborative scheduling of heterogeneous cloud-edge resources;

[0007] Step 2: Construct a multi-layer resource representation model of computing power network based on hypernetwork;

[0008] Step 3: Introduce coded distributed computing on the edge side to fully utilize the computing power of the edge layer and provide low-latency and highly reliable computing services for computing tasks;

[0009] Step 4: Taking minimizing the total delay of processing tasks as the optimization goal, establish the joint task offloading and heterogeneous resource allocation problem;

[0010] Step 5: Abstract the cloud-edge heterogeneous computing network into a hypergraph, including classifying and modeling graph-structured data and non-graph-structured data, and constructing hyperedge groups.

[0011] Step 6: Use HyperGNNs’ transformation function to model and optimize high-order dependencies in the network to improve the accuracy and efficiency of resource scheduling;

[0012] Step 7: Use particle swarm optimization algorithm to find the optimal resource scheduling solution.

[0013] The computing network model uses a multi-layered resource representation model to optimize the configuration of cloud, edge, and end resources. Specifically, it includes the following:

[0014] (1) Business equipment layer: equipment uses q i Each device continuously collects business data and generates various computing tasks. Devices can access edge layer APs through heterogeneous wireless networks. Based on the QoS / QoE requirements of applications, each device offloads tasks to the edge layer and cloud center layer for further processing.

[0015] (2) Edge layer: This layer includes multiple APs, task management controllers, and edge computing nodes. In this paper, each AP is equipped with an edge server, denoted as Es, with a total of E. The edge computing node is denoted as n j , j∈{1,2,...,J}. The edge computing node and the edge server have the same functions, and transmit tasks to the task management controller through heterogeneous wireless links to complete the coded distributed computing;

[0016] (3) Cloud Center Layer: The local cloud server is composed of multiple central servers CS, which provide abundant resources to perform large-scale computing. In addition, this layer also provides application management, global control and elastic service functions. The above construction is based on the multi-layer resource representation model of the computing power network of the hypernetwork.

[0017] (1) Characterization of computing power business requirements: The service set B contains the set of all service requests in the network, which is used to represent the distribution of different services among the cloud, edge, and end. iContains its independent demand characteristics and resource constraints.

[0018] It is used to represent the inherent attributes, constraints and performance requirements of the business in the business requirement layer. For example, in Indicates the amount of data for the business. Indicates the amount of floating-point calculations required for the business calculation. is the ratio of the computational effort in the data computation phase to the computational effort required for the entire task, is the ratio of the parallel computing amount in the data computing stage to the entire data computing stage, Indicates business The transmission rate requirement, Give us the maximum allowable delay for a business execution. It is worth noting that you can add collections reasonably according to the specific scenario. Elements in .

[0019] Represents a collection of business dependencies in the business requirements layer. For example, a task must be completed before another task can be executed. Through business dependencies, the system can more effectively coordinate the order of task execution, avoid resource conflicts, and improve scheduling accuracy and task execution reliability.

[0020] (2) Logical Function Resource Characterization: in, The set of all microservices representing a functional resource network is defined as:

[0021]

[0022] V represents the set of virtual network functions (VNFs) of the functional resource network, which is expressed as:

[0023]

[0024] Among them vnf g Indicates a certain VNF type g, for a specific type of virtual network function vnf g The definition is as follows:

[0025] vnf g =Φ(ms g1 ,ms g2 ,...,ms gH )

[0026] Where Φ(·) represents the functional function of the network function VNF, and gH corresponding microservices are selected and combined to implement a specific network function vnfg This microservice-based VNF can better adapt to changes in business needs and achieve rapid iteration and continuous integration.

[0027] is the set of all service function chains in the functional resource network, where each service function chain consists of n VNF instances, represented as

[0028] (3) Physical network resource representation: in A collection of all physical devices in the physical network. Represents the nth physical entity. Assume that the physical entity The identity is defined as:

[0029]

[0030] Where Q represents the type set of physical entities (Q = {q, Es, n, CS}), Represents the entity device's own device information collection, Represents a collection of attribute behaviors of a physical device. It is worth noting that in an actual network, a physical device The type and device information are uniquely identified by the tag. behavior Defined as:

[0031]

[0032] in,{·} top 、{·} per 、{·} res They correspond to topological behavior, performance behavior, and resource behavior respectively. Topological behaviors include Subordinate relationship Adjacency and connectivity etc.; performance behaviors include The transmission rate C n , delay t n and power p n etc.; resource behaviors include The computing resources f n , storage resources n and communication resources β n wait;

[0033] (4) Characterization of control and management resources: in, express The collection of management units (Man), express A collection of control units (Con). Management units correspond directly to computing activities, and control units are directly connected to the functional resource system. Each management unit must be connected to at least one control unit, and a communication relationship exists between the management unit layer and the control unit layer. Each management unit independently assumes responsibility for managing a portion of resources and exchanges information with other units through a collaborative mechanism to achieve optimal resource allocation and scheduling.

[0034] IC is defined to represent the number of control units that a management unit is responsible for, i.e., the internal management capability. The external collaboration capability of a management unit refers to the number of other management units that a management unit directly collaborates with.

[0035] State variable representation: “Task unit - functional resource” allocation variable h ij , "Management Unit - Control Unit" assigns variable x mj , "Management Unit - Task Unit" allocates variable p mi .

[0036]

[0037] From this, we can get the internal management capability IC and external collaboration capability EC of a management unit, which can be expressed as:

[0038]

[0039] EC(Man m ) can be understood as the management unit Man m and Snap-inMan n The number of computing activities completed simultaneously. Task offloading strategy set Ψ: describes the business tasks The strategy of offloading from local devices to edge nodes or cloud servers. The task offloading strategy optimizes the efficiency of resource utilization and ensures the reasonable distribution of services among various computing nodes to meet different constraints such as latency and bandwidth. The TO strategy of multi-layer resource scheduling in this paper's computing network includes the offloading ratio strategy and edge selection strategy in Indicates the local device q i,e Offload tasks to the local edge side Es e The uninstall ratio, For local Es e Select a non-local MEC server The uninstall ratio, Indicates local Es e To the cloud center layer CS c The uninstall ratio, Indicates each local Ese Only one non-local one can be selected To perform collaborative computing on the edge side;

[0040] Networks at all levels form a network according to their respective attributes. Through the collaborative relationship and hierarchical structure between networks, a computing power network matching model is formed. In summary, the computing power network resource representation model based on the super network is defined as:

[0041]

[0042] The introduction of coded distributed computing on the edge side fully utilizes the computing power of the edge layer to provide low-latency and highly reliable computing services for computing tasks;

[0043] Assume that the task offloaded to the edge cloud system is The amount of task data that needs to be processed on the local edge side is When using (n,k) MDS coding, first Divide into k subtasks evenly These subtasks are then encoded into n encoded subtasks (1≤k≤n), and finally send the encoded n subtasks to n edge node servers for calculation, ignoring the time required for task segmentation and encoding; given the encoded subtasks At the edge node n ej The total processing time is recorded as

[0044]

[0045] in, They represent the time it takes for the subtask to be sent from the MEC server to the edge node and the time it takes for the calculation result to be returned from the edge node to the MEC server, respectively. j In wireless channel The number of time slots required for successful data transmission, Indicates wireless channel The set of disconnection probabilities; Represents edge node n e,j The time to process subtasks, f(λ j ) is a random variable, representing the edge node n e,j The random initialization time caused by the discrete problem on j The exponential distribution of With λ={λ1,...,λ j ,...λn} represents the set of rate parameters that determine the random time of edge nodes;

[0046] The time required to complete the encoding distributed calculation is expressed as:

[0047]

[0048] The problem of joint task offloading and heterogeneous resource allocation is established with minimizing the total delay of processing tasks as the optimization goal;

[0049] (1) Local device computing: The latency of data preprocessing and result aggregation is recorded as Expressed as:

[0050]

[0051] The data calculation delays of CPU and GPU are respectively recorded as

[0052]

[0053] Since the serial calculations performed by the CPU and the parallel calculations performed by the GPU are performed simultaneously, the latency of the data calculation phase is recorded as

[0054]

[0055] In time slot t, local device q i The computation queues waiting for CPU and GPU processing are respectively recorded as The waiting time of a task in the computing queue is

[0056]

[0057] In summary, on the local device q i The total delay of task processing Denoted as:

[0058]

[0059] (2) Encoded distributed computing on the local edge side: q at time t i With Es e The signal-to-noise ratio between is:

[0060]

[0061] in For device q i The transmission power, is the maximum transmit power of the device; is the channel gain; σ 2 (W) represents the noise power. Therefore, qi With Es e The transmission rate of the channel between is:

[0062]

[0063] in Indicates that at time t q i With Es e The channel bandwidth allocated to the transmission channel between For Es e The bandwidth resources allocated at time t. Therefore, q i The business Upload to the corresponding Es e The channel transmission delay is:

[0064]

[0065] The other delay calculation methods are the same as (1), so the local Es e The total task processing delay is:

[0066]

[0067] (3) Non-local edge-side coding distributed computing: Total task processing delay

[0068] (4) Cloud center layer computing: central server CS c The total task processing delay is

[0069] The present invention sets the optimization target as the total delay of executing all services within T time slots.

[0070]

[0071] in Expressed as:

[0072]

[0073] Therefore, the optimization problem can be expressed as problem (1):

[0074]

[0075]

[0076] C8:1≤k≤n

[0077] C1 ensures that the sum of the offloading percentages from local ES to non-local ES and CS does not exceed 1, C2 ensures that each local ES can only select one non-local ES to perform collaborative edge computing, C3 ensures that the bandwidth resources allocated to each ES do not exceed its idle bandwidth resources, C4 means that the total energy consumption does not exceed the energy we provide for each device, and C5 ensures Does not exceed the maximum allowable delay for a given business execution C6 and C7 ensure that each q i , n e,j and CS c , in handling business The number of CPUs and GPUs allocated cannot exceed their maximum number, and C8 guarantees the effectiveness of encoding.

[0078] The cloud-edge heterogeneous computing network is abstracted into a hypergraph, including classification modeling of graph-structured data and non-graph-structured data, and construction of hyperedge groups;

[0079] (1) When the data correlation is a graph structure: Let G s =(χ s ,E s ) represents the graph structure, where x i ∈χ s is the vertex, e sij ∈E s is x i and x j The edge between G s The adjacency matrix is ​​denoted as A; the pairwise edge hyperedge group (E pair ): Figure G s Each hyperedge e in sij Connecting two vertices x in the corresponding edge i and x j , E pair It is expressed as follows

[0080] E pair ={{x i ,x j}|(x i ,x j )∈E s}

[0081] k-Hop hyperedge group E hop :E hop The related vertices of the central vertex can be found through the k-Hop reachable positions in the graph structure. s The k-Hop neighborhood of vertex x in is defined as: where k∈[2,n x ],n x It is Figure G sTherefore, the hyperedge group E hop Expressed as:

[0082]

[0083] (2) When data correlation does not have a graph structure: Each node device has two types of data: one is attribute data, and the other is the feature associated with each vertex; the attribute hyperedge group (E attribute ): attribute data of different node devices, a set of hyperedges using neighbors in the attribute space is generated based on the attribute data, where each hyperedge represents an attribute or an available sub-attribute, which connects all nodes with this attribute; the subset of nodes with attribute a is represented as P att (a); Α is the set of all attributes or sub-attributes of the attribute; therefore, E attribute Expressed as:

[0084] E attribute ={P att (a)|a∈Α};

[0085] Node feature hyperedge group (E feature ): Given the features of each vertex, the second type E is generated by finding the neighbors of each vertex in the feature space feature Given a vertex as the centroid, its k nearest neighbors in the feature space are connected by a hyperedge, or all neighbors within a distance d from the centroid or centroid are selected;

[0086]

[0087] The k or d value can be set according to the specific situation during the neighbor discovery process;

[0088] (3) In order to fully integrate and utilize the high-order correlation of multimodal mixtures, an adaptive hyperedge group fusion strategy, namely Adaptive Fusion, is used; each hyperedge group is associated with a trainable parameter that adaptively adjusts the impact of multiple hyperedge groups on the final vertex embedding; it is defined as:

[0089] ω k =copy(sigmoid(w k ),M k )

[0090]

[0091] J=J1||J2||...||J K

[0092] in Is a trainable parameter shared by all hyperedges in a specified hyperedge group k. sigmoid(·) is an element-wise normalization function; vector represents the weight vector of the generated hyperedge group k; copy(a,b) function returns a vector of size b, the value of which is filled by copying ab times. Let M=M1+M2+...+M k represents the sum of hyperedges in all hyperedge groups. Represents the weight matrix of the hypergraph, each entry W ii Represents the corresponding hyperedge e i The weight of J∈{0,1} N×M Represents the incidence matrix of a hypergraph generated by concatenating the incidence matrices of multiple hyperedge groups (·||·).

[0093] The hypergraph convolutional network is used to model and optimize high-order correlations in the network to improve the accuracy and efficiency of resource scheduling;

[0094] The objective function (1) and the constraint function are mostly non-convex, and the optimization variables are mostly discrete variables, which makes it impossible to directly apply gradient-based information to solve the problem. In addition, the resource allocation problem caused by the dynamic and heterogeneous nature of the computing power network is difficult for traditional optimization methods to provide a feasible solution in polynomial time. In order to meet the above challenges, the present invention adopts a method based on hypergraph neural network (HyperGNN) and particle swarm optimization (PSO). HyperGNN effectively solves the complex constraints of the problem by capturing the high-order correlation between heterogeneous nodes in the computing power network. In the present invention, the optimization variables in the replacement problem of problem (1) are The purpose of using these transformation functions is to capture the complex patterns and interdependencies of variables through problem constraints, thereby obtaining better solutions. in is the (continuous) assignment of value to node i, δ=[δ1,δ2,...,δ γ ] is the input embedding of the node. By adopting a parameterized transformation function, the training parameters (ω k ) and the input embedding (δ) to optimize our objective; this alternative optimization problem is described below:

[0095]

[0096] Defined on the combinatorial constrained hypergraph Hyper GNN The augmented (penalty) loss function is:

[0097]

[0098] in λ nis the weight of constraint n. In order to make HGNNConv + The forward pass on GPU / CPU devices is faster, using the matrix format:

[0099]

[0100] D x and D e are the degree matrices of nodes and hyperedges, Θ (l) Is the parameter that layer l needs to learn during training. Apply filter Ω to extract features on the nodes in the hypergraph; after convolution, we can get X (l+1) , which can be used for further learning.

[0101] The particle swarm optimization algorithm is used to find the optimal resource scheduling solution;

[0102] Particle swarm optimization can handle discrete variables and has good robustness in searching for the global optimal solution. + The continuous output is converted into integer node assignment, using HGNNConv + The continuous output generated by the variable is used to deduce a probability distribution over the discrete values ​​{ν1,ν2,...}; this probability distribution is designed in such a way that the probability of the variable taking a certain value is inversely proportional to its distance from the value based on the distance between the continuous output of the variable and a specific value; then we get Take a sample ξ1,...,ξ from the corresponding output distribution Υ ; Use this sample to initialize a group of random particles and find the optimal solution minΤ(ξ i ); In each iteration, the particle is updated by following the two "extreme values". After finding the two optimal values, the particle updates its speed and position using the following formula:

[0103] v i =w×v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i )

[0104] Where w is the inertia factor; v i is the speed of the particle; rand() is a random number between (0, 1); x i is the position of the particle, and x i =x i +v i ; c1 and c2 are learning factors, usually 2.

[0105] Beneficial effects of the present invention:

[0106] The present invention is divided into two stages. In the first stage, a multi-level resource representation model of the computing power network is constructed, and the high-order correlations between heterogeneous nodes are captured through the hypergraph neural network, and dynamic resource scheduling optimization is performed; in the second stage, the particle swarm optimization algorithm is combined to solve the resource scheduling problem, and the distributed computing strategy is encoded to solve the problem of unstable task execution caused by the failure of edge nodes. This method innovatively uses the hypergraph convolutional network to model and process high-dimensional data, significantly improving the accuracy and robustness of resource scheduling. While improving the success rate of task transmission, this method effectively reduces the average latency and energy consumption of the system, and exhibits better performance in an environment with unstable network links. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 Schematic diagram of the CTOHRA algorithm framework based on the combination of HGNN and PSO provided in Example 1 of the present invention;

[0108] Figure 2 This is a diagram of the actual scenario of the cloud-edge heterogeneous computing network provided by the first embodiment of the present invention;

[0109] Figure 3 Schematic diagram of a computing power network resource representation model based on a hypernetwork provided in the first embodiment of the present invention;

[0110] Figure 4 This is a schematic diagram of coding distributed computing in the edge cloud system provided by Example 1 of the present invention. DETAILED DESCRIPTION

[0111] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inward" and "outward" refer to directions toward or away from the geometric center of a particular component, respectively.

[0112] Example 1:

[0113] like Figure 1 As shown, an embodiment of the present invention provides a highly reliable resource scheduling optimization method for a heterogeneous computing network based on coded distributed computing and a hypergraph neural network, comprising the following steps:

[0114] 1. Constructed a computing network for collaborative scheduling of heterogeneous cloud-edge resources, such as Figure 2 shown.

[0115] (1) Business equipment layer: equipment uses q iEach device continuously collects business data and generates various computing tasks. Devices can access edge layer APs through heterogeneous wireless networks. Based on the QoS / QoE requirements of applications, each device offloads tasks to the edge layer and cloud center layer for further processing.

[0116] (2) Edge layer: This layer includes multiple APs, task management controllers, and edge computing nodes. In this paper, each AP is equipped with an edge server, denoted as Es, with a total of E. The edge computing node is denoted as n j , j∈{1,2,...,J}. In this paper, the edge computing nodes and edge servers have the same functions, which transmit tasks to the task management controller through heterogeneous wireless links to complete the coded distributed computing;

[0117] (3) Cloud center layer: The local cloud server consists of multiple central servers (CSs), which provide abundant resources to perform large-scale computing. In addition, this layer also provides application management, global control, and elastic service functions.

[0118] 2. Construct a multi-layer resource representation model of computing power network based on hypernetwork, such as Figure 3 As shown;

[0119] (1) Characterization of computing power business requirements: The service set B contains the set of all service requests in the network, which is used to represent the distribution of different services between the cloud, edge, and end. Each service Bi contains its independent demand characteristics and resource constraints.

[0120] It is used to represent the inherent attributes, constraints and performance requirements of the business in the business requirement layer. For example, in Indicates the amount of data for the business. Indicates the amount of floating-point calculations required for the business calculation. is the ratio of the computational effort in the data computation phase to the computational effort required for the entire task, is the ratio of the parallel computing amount in the data computing stage to the entire data computing stage, Indicates business The transmission rate requirement, Give us the maximum allowable delay for a business execution. It is worth noting that you can add collections reasonably according to the specific scenario. Elements in .

[0121] Represents a collection of business dependencies in the business requirements layer. For example, a task must be completed before another task can be executed. Through business dependencies, the system can more effectively coordinate the order of task execution, avoid resource conflicts, and improve scheduling accuracy and task execution reliability.

[0122] (2) Logical Function Resource Characterization: in, The set of all microservices representing a functional resource network is defined as:

[0123]

[0124] A collection of virtual network functions (VNFs) representing a functional resource network, expressed as:

[0125] V={vnf g |g=1,2,...,G}

[0126] Among them vnf g Indicates a certain VNF type g, for a specific type of virtual network function vnf g The definition is as follows:

[0127] vnf g =Φ(ms g1 ,ms g2 ,...,ms gH )

[0128] Where Φ(·) represents the functional function of the network function VNF, and gH corresponding microservices are selected and combined to implement a specific network function vnf g This microservice-based VNF can better adapt to changes in business needs and achieve rapid iteration and continuous integration.

[0129] is the set of all service function chains in the functional resource network, where each service function chain consists of n VNF instances, represented as

[0130] (3) Physical network resource representation: in A collection of all physical devices in the physical network. Represents the nth physical entity. Assume that the physical entity The identity is defined as:

[0131]

[0132] Where Q represents the type set of physical entities (Q = {q, Es, n, CS}), Represents the entity device's own device information collection, Represents a collection of attribute behaviors of a physical device. It is worth noting that in an actual network, a physical device The type and device information are uniquely identified by the tag. behavior Defined as:

[0133]

[0134] in,{·} top 、{·} per 、{·} res They correspond to topological behavior, performance behavior, and resource behavior respectively. Topological behaviors include Subordinate relationship Adjacency and connectivity etc.; performance behaviors include The transmission rate C n , delay t n and power p n etc.; resource behaviors include The computing resources f n , storage resources n and communication resources β n wait;

[0135] (4) Characterization of control and management resources: in, express The collection of management units (Man), express A collection of control units (Con). Management units correspond directly to computing activities, and control units are directly connected to the functional resource system. Each management unit must be connected to at least one control unit. There is a communication relationship between the management unit layer and the control unit layer. Each management unit independently assumes the management responsibility of a portion of resources and exchanges information with other units through a collaborative mechanism to achieve optimal resource allocation and scheduling. The main tasks of the management and control units include:

[0136] Resource allocation decisions: Allocate computing, caching, and communication resources based on business needs and current network status. Real-time monitoring: Continuously obtain node and link status information to ensure the accuracy and timeliness of resource scheduling. Fault recovery: In the event of node failure or link interruption, quickly adjust resource scheduling strategies to ensure system reliability.

[0137] This article defines IC as the number of control units that a management unit is responsible for, that is, the internal management capability. The external collaboration capability of a management unit refers to the number of other management units that a management unit directly collaborates with.

[0138] State variable representation: “Task unit - functional resource” allocation variable h ij , "Management Unit - Control Unit" assigns variable x mj , "Management Unit - Task Unit" allocates variable p mi .

[0139]

[0140] From this, we can get the internal management capability IC and external collaboration capability EC of a management unit, which can be expressed as:

[0141]

[0142] EC(Man m ) can be understood as the management unit Man m and Snap-inMan n The number of computing activities completed simultaneously. Task offloading strategy set Ψ: describes the business tasks The strategy of offloading from local devices to edge nodes or cloud servers. The task offloading strategy optimizes the efficiency of resource utilization and ensures the reasonable distribution of services among various computing nodes to meet different constraints such as latency and bandwidth. The TO strategy of multi-layer resource scheduling in this paper's computing network includes the offloading ratio strategy and edge selection strategy in Indicates the local device q i,e Offload tasks to the local edge side Es e The uninstall ratio, For local Es e Select a non-local MEC server The uninstall ratio, Indicates local Es e To the cloud center layer CS c The uninstall ratio, Indicates each local Es e Only one non-local one can be selected To perform collaborative computing on the edge side;

[0143] Networks at all levels form a network according to their own attributes. Through the collaborative relationship and hierarchical structure between networks, a computing power network matching model is formed. In summary, the definition of computing power network resource representation based on the super network is

[0144] The model is:

[0145]

[0146] 3. Introducing coded distributed computing at the edge to fully utilize the computing power of the edge layer and provide low-latency and highly reliable computing services for computing tasks;

[0147] like Figure 4 As shown, let the task offloaded to the edge cloud system be The amount of task data that needs to be processed on the local edge side is When using (n,k) MDS coding, first Divide into k subtasks evenly These subtasks are then encoded into n encoded subtasks (1≤k≤n), and finally send the encoded n subtasks to n edge node servers for calculation, ignoring the time required for task segmentation and encoding; given the encoded subtasks At the edge node n ej The total processing time is recorded as

[0148]

[0149] in, They represent the time it takes for the subtask to be sent from the MEC server to the edge node and the time it takes for the calculation result to be returned from the edge node to the MEC server, respectively. j Is in wireless channel l ej The number of time slots required for successful data transmission, Indicates wireless channel The set of disconnection probabilities; Represents edge node n e,j The time to process subtasks, f(λ j ) is a random variable, representing the edge node n e,j The random initialization time caused by the discrete problem on j The exponential distribution of With λ={λ1,...,λ j ,...λ n} represents the rate parameter set that determines the random time of edge nodes;

[0150] The time required to complete the encoding distributed calculation is expressed as:

[0151]

[0152] 4. With the optimization goal of minimizing the total delay of processing tasks, we formulate the joint task offloading and heterogeneous resource allocation problem;

[0153] (1) Local device computing: The latency of data preprocessing and result aggregation is recorded as Expressed as:

[0154]

[0155] The data calculation delays of CPU and GPU are respectively recorded as

[0156]

[0157] Since the serial calculations performed by the CPU and the parallel calculations performed by the GPU are performed simultaneously, the latency of the data calculation phase is recorded as

[0158]

[0159] In time slot t, local device q i The computation queues waiting for CPU and GPU processing are respectively recorded as The waiting time of a task in the computing queue is

[0160]

[0161] In summary, on the local device q i The total delay of task processing Denoted as:

[0162]

[0163] (2) Encoded distributed computing on the local edge side: q at time t i With Es e The signal-to-noise ratio between is:

[0164]

[0165] in For device q i The transmission power, is the maximum transmit power of the device; is the channel gain; σ 2 (W) represents the noise power. Therefore, q i With Es e The transmission rate of the channel between is:

[0166]

[0167] in Indicates that at time t q i With Es eThe channel bandwidth allocated to the transmission channel between For Es e The bandwidth resources allocated at time t. Therefore, q i The business Upload to the corresponding Es e The channel transmission delay is:

[0168]

[0169] The other delay calculation methods are the same as (1), so the local Es e The total task processing delay is:

[0170]

[0171] (3) Non-local edge-side coding distributed computing: Total task processing delay

[0172] (4) Cloud center layer computing: central server CS c The total task processing delay is The optimization objective is set as the total delay of executing the service within T time slots.

[0173]

[0174] in Expressed as:

[0175]

[0176] Therefore, the optimization objective can be expressed as problem (1):

[0177]

[0178] C8:1≤k≤n

[0179] C1 ensures that the sum of the offloading percentages from local ES to non-local ES and CS does not exceed 1, C2 ensures that each local ES can only select one non-local ES to perform collaborative edge computing, C3 ensures that the bandwidth resources allocated to each ES do not exceed its idle bandwidth resources, C4 means that the total energy consumption does not exceed the energy we provide for each device, and C5 ensures Does not exceed the maximum allowable delay for a given business execution C6 and C7 ensure that each q i , n e,j and CS c , in handling business The number of CPUs and GPUs allocated cannot exceed their maximum number, and C8 guarantees the effectiveness of encoding.

[0180] 5. Abstract the cloud-edge heterogeneous computing network into a hypergraph, including classifying and modeling graph-structured and non-graph-structured data and constructing hyperedge groups;

[0181] (1) When the data correlation is a graph structure: Let G s =(χ s ,E s ) represents the graph structure, where x i ∈χ s is the vertex, e sij ∈E s is x i and x j The edge between G s The adjacency matrix is ​​denoted as A; the pairwise edge hyperedge group (E pair ): Figure G s Each hyperedge e in sij Connecting two vertices x in the corresponding edge i and x j , E pair It is expressed as follows

[0182] E pair ={{x i ,x j}|(x i ,x j )∈E s}

[0183] k-Hop hyperedge group E hop :E hop The related vertices of the central vertex can be found through the k-Hop reachable positions in the graph structure. s The k-Hop neighborhood of vertex x in is defined as: where k∈[2,n x ],n x It is Figure G s Therefore, the hyperedge group E hop Expressed as:

[0184]

[0185] (2) When data correlation does not have a graph structure: Each node device has two types of data: one is attribute data, and the other is the feature associated with each vertex; the attribute hyperedge group (E attribute): attribute data of different node devices, a set of hyperedges using neighbors in the attribute space is generated based on the attribute data, where each hyperedge represents an attribute or an available sub-attribute, which connects all nodes with this attribute; the subset of nodes with attribute a is represented as P att (a); Α is the set of all attributes or sub-attributes of the attribute; therefore, E attribute Expressed as:

[0186] E attribute ={P att (a)|a∈Α};

[0187] Node feature hyperedge group (E feature ): Given the features of each vertex, the second type E is generated by finding the neighbors of each vertex in the feature space feature Given a vertex as the centroid, its k nearest neighbors in the feature space are connected by a hyperedge, or all neighbors within a distance d from the centroid or centroid are selected;

[0188]

[0189] The k or d value can be set according to the specific situation during the neighbor discovery process;

[0190] (3) In order to fully integrate and utilize the high-order correlation of multimodal mixtures, an adaptive hyperedge group fusion strategy, namely Adaptive Fusion, is used; each hyperedge group is associated with a trainable parameter that adaptively adjusts the impact of multiple hyperedge groups on the final vertex embedding; it is defined as:

[0191] ω k =copy(sigmoid(w k ),M k )

[0192]

[0193] J=J1||J2||...||J K

[0194] in Is a trainable parameter shared by all hyperedges in a specified hyperedge group k. sigmoid(·) is an element-wise normalization function; vector represents the weight vector of the generated hyperedge group k; copy(a,b) function returns a vector of size b, the value of which is filled by copying ab times. Let M=M1+M2+...+M k represents the sum of hyperedges in all hyperedge groups. Represents the weight matrix of the hypergraph, each entry W ii Represents the corresponding hyperedge ei The weight of J∈{0,1} N×M Represents the incidence matrix of a hypergraph generated by concatenating the incidence matrices of multiple hyperedge groups (·||·).

[0195] 6. Use hypergraph convolutional networks to model and optimize high-order dependencies in the network to improve the accuracy and efficiency of resource scheduling;

[0196] The objective function (1) and the constraint function are mostly non-convex, and the optimization variables are mostly discrete variables, which makes it impossible to directly apply gradient-based information to solve the problem. In addition, the resource allocation problem caused by the dynamic and heterogeneous nature of the computing power network is difficult for traditional optimization methods to provide a feasible solution in polynomial time. In order to meet the above challenges, the present invention adopts a method based on hypergraph neural network (HyperGNN) and particle swarm optimization (PSO). HyperGNN effectively solves the complex constraints of the problem by capturing the high-order correlation between heterogeneous nodes in the computing power network. In the present invention, the optimization variables in the replacement problem of problem (1) are The purpose of using these transformation functions is to capture the complex patterns and interdependencies of variables through problem constraints, thereby obtaining better solutions. in is the (continuous) assignment of value to node i, δ=[δ1,δ2,...,δ γ ] is the input embedding of the node. By adopting a parameterized transformation function, the training parameters (ω k ) and the input embedding (δ) to optimize our objective; this alternative optimization problem is described below:

[0197]

[0198] Defined on the combinatorial constrained hypergraph Hyper GNN The augmented (penalty) loss function is:

[0199]

[0200] in λ n is the weight of constraint n. In order to make HGNNConv + The forward pass on GPU / CPU devices is faster, using the matrix format:

[0201]

[0202] D x and D e are the degree matrices of nodes and hyperedges, Θ (l)Is the parameter that layer l needs to learn during training. Apply filter Ω to extract features on the nodes in the hypergraph; after convolution, we can get X (l+1) , which can be used for further learning.

[0203] 7. Use particle swarm optimization algorithm to find the optimal resource scheduling solution;

[0204] In order to + The continuous output is converted into integer node assignment, using HGNNConv + The continuous output generated by the variable is used to deduce a probability distribution over the discrete values ​​{ν1,ν2,...}; this probability distribution is designed in such a way that the probability of the variable taking a certain value is inversely proportional to its distance from the value based on the distance between the continuous output of the variable and a specific value; then we get Take a sample ξ1,...,ξ from the corresponding output distribution Υ ; Use this sample to initialize a group of random particles and find the optimal solution min Τ(ξ i ); In each iteration, the particle is updated by following the two "extreme values". After finding the two optimal values, the particle updates its speed and position using the following formula:

[0205] v i =w×v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i )

[0206] Where w is the inertia factor; v i is the speed of the particle; rand() is a random number between (0, 1); x i is the position of the particle, and x i =x i +v i ; c1 and c2 are learning factors, usually 2.

[0207] In summary, in the first phase, this embodiment constructs a multi-level resource representation model of the computing power network, captures high-order correlations between heterogeneous nodes through a hypergraph neural network, and performs dynamic resource scheduling optimization; in the second phase, it combines the particle swarm optimization algorithm to solve resource scheduling, and solves the problem of unstable task execution caused by edge node failure by encoding distributed computing strategies.

[0208] The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above-mentioned embodiment, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A highly reliable resource scheduling optimization method for heterogeneous computing networks based on coded distributed computing and hypergraph neural networks, characterized in that: The specific steps include: Step 1: Construct a computing network model for the coordinated scheduling of heterogeneous cloud-edge-end resources; the computing network model in step 1 achieves the optimal configuration of cloud, edge, and end resources through a multi-level resource representation model. the following: (11) Business equipment layer: equipment uses q i Each device continuously collects business data and generates various computing tasks. The devices access the edge layer AP through heterogeneous wireless networks. Based on the QoS / QoE requirements of the application, each device offloads tasks to the edge layer and cloud center layer for further processing. (12) Edge layer: This layer includes multiple APs, task management controllers, and edge computing nodes. Each AP is assumed to be equipped with an edge server, denoted as Es, with a total of E; the edge computing node is denoted as n j , j∈{1,2,...,J}; the edge computing node and the edge server have the same functions, and transmit tasks with the task management controller through heterogeneous wireless links to complete the coded distributed computing; (13) Cloud Center Layer: The local cloud server consists of multiple central servers (CSs), which provide abundant resources to perform large-scale computing. In addition, this layer also provides application management, global control, and elastic service functions. Step 2: Constructing a multi-layer resource representation model of a computing power network based on a hypernetwork; the multi-layer resource representation model of a computing power network based on a hypernetwork constructed in step 2 includes (21) computing power business demand representation, (22) logical function resource representation, (23) physical network resource representation, and (24) control management resource representation; (21) Characterization of computing power business requirements: The service set B contains the set of all service requests in the network, which is used to represent the distribution of different services among the cloud, edge, and end. i Contains its independent demand characteristics and resource constraints; each time slot t, for an AP e The corresponding device q i,e Randomly generate a business record Used to represent the inherent attributes, constraints and performance requirements of the business in the business requirements layer; For example, in Indicates the amount of data for the business. Indicates the amount of floating-point calculations required for the business calculation. is the ratio of the computational effort in the data computation phase to the computational effort required for the entire task, is the ratio of the parallel computing amount in the data computing stage to the entire data computing stage, Indicates business The transmission rate requirement, Give us the maximum allowable delay for business execution; Represents the set of business dependencies in the business requirements layer, (22) Logical Function Resource Characterization: in, The set of all microservices representing a functional resource network is defined as: A collection of virtual network functions representing a functional resource network, expressed as: Among them vnf g Indicates a certain VNF type g, for a specific type of virtual network function vnf g The definition is as follows: vnf g =Φ(ms g1 ,ms g2 ,...,ms gH ) Where Φ(·) represents the functional function of the network function VNF, and gH corresponding microservices are selected and combined to implement a specific network function vnf g This microservice-based VNF can better adapt to changes in business needs and achieve rapid iteration and continuous integration. is the set of all service function chains in the functional resource network, where each service function chain consists of n VNF instances, represented as (23) Physical network resource representation: in A collection of all physical devices in the physical network. Represents the nth physical entity; assuming the physical entity The identity is defined as: Where Q represents the type set of physical entities (Q = {q, Es, n, CS}), Represents the entity device's own device information collection, Represents a collection of attribute behaviors of a physical device; a physical device in an actual network The type and device information are uniquely identified by the tag. behavior Defined as: in,{·} top 、{·} per 、{·} res They correspond to topological behavior, performance behavior, and resource behavior respectively; Topological behaviors include Subordinate relationship Adjacency and connectivity etc.; performance behaviors include The transmission rate C n , delay t n and power p n etc.; resource behaviors include The computing resources f n , storage resources n and communication resources β n wait; (24) Characterization of control and management resources: in, express The collection of management units (Man), express A collection of control units (Con). Management units correspond directly to computing activities one-to-one, and control units are directly connected to the functional resource system. Each management unit must be connected to at least one control unit. There is a communication relationship between the management unit layer and the control unit layer. Each management unit independently assumes the management responsibility of a part of the resources and exchanges information with other units through a collaborative mechanism to achieve optimal allocation and scheduling of resources. IC means the number of control units that a management unit is responsible for, that is, the internal management capability; EC means The external collaboration capability of a management unit refers to the number of other management units that a management unit directly collaborates with. Representation of state variables: "Task unit - functional resource" allocation variable h ij , "Management Unit - Control Unit" allocates variable x mj , "Management Unit - Task Unit" allocates variable p mi ; This results in the internal management capability IC and external collaboration capability EC of a management unit, which can be expressed as: EC(Man m ) is understood as the management unit Man m and Snap-inMan n The number of computing activities completed collaboratively at the same time; Task offloading strategy set Ψ: describes the business task Strategies for offloading from local devices to edge nodes or cloud servers; TO strategies for multi-layer resource scheduling in computing networks include offloading ratio strategies and edge selection strategy in Indicates the local device q i,e Offload tasks to the local edge side Es e The uninstall ratio, For local Es e Select a non-local MEC server The uninstall ratio, Indicates local Es e To the cloud center layer CS c The uninstall ratio, Indicates each local Es e Only one non-local one can be selected To perform collaborative computing on the edge side; Step 3: Introduce coded distributed computing on the edge side to fully utilize the computing power of the edge layer and provide low-latency and highly reliable computing services for computing tasks; Step 4: Taking minimizing the total delay of processing tasks as the optimization goal, establish the joint task offloading and heterogeneous resource allocation problem; Step 5: Abstract the cloud-edge heterogeneous computing network into a hypergraph, including classifying and modeling graph-structured data and non-graph-structured data, and constructing hyperedge groups. Step 6: Use HyperGNNs’ transformation function to model and optimize high-order dependencies in the network to improve the accuracy and efficiency of resource scheduling; Step 7: Use particle swarm optimization algorithm to find the optimal resource scheduling solution.

2. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 1 is characterized in that: In step 2, the networks at all levels form a network according to their respective attributes, and through the collaborative relationship and hierarchical structure between networks, a computing power network matching model is formed. In summary, the computing power network resource representation model based on the super network is defined as:

3. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 1 is characterized in that: The step 3 specifically includes: Assume that the task offloaded to the edge cloud system is The amount of task data that needs to be processed on the local edge side is When using (n,k) MDS coding, first Divide into k subtasks evenly These subtasks are then encoded into n encoded subtasks Finally, the encoded n subtasks are sent to n edge node servers for calculation, ignoring the time required for task segmentation and encoding; the encoded subtasks are given At the edge node n ej The total processing time is recorded as in, They represent the time it takes for the subtask to be sent from the MEC server to the edge node and the time it takes for the calculation result to be returned from the edge node to the MEC server, respectively. j In wireless channel The number of time slots required for successful data transmission, p={p e,1 ,...,p e,j ,...,p e,n } indicates wireless channel The set of disconnection probabilities; Represents edge node n e,j The time to process subtasks, f(λ j ) is a random variable representing the edge node n e,j The random initialization time caused by the discrete problem on j The exponential distribution of With λ={λ1,...,λ j ,...λ n } represents the rate parameter set that determines the random time of edge nodes; The time required to complete the encoding distributed calculation is expressed as:

4. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 1 is characterized in that: The step 4 is specifically as follows: (41) Local device computing: The latency of data preprocessing and result aggregation is recorded as Expressed as: The data calculation delays of CPU and GPU are respectively recorded as Since the serial calculations performed by the CPU and the parallel calculations performed by the GPU are performed simultaneously, the latency of the data calculation phase is recorded as In time slot t, local device q i The computation queues waiting for CPU and GPU processing are respectively recorded as The waiting time of a task in the computing queue is In summary, on the local device q i The total delay of task processing Denoted as: (42) Encoded distributed computing on the local edge side: q at time t i With Es e The signal-to-noise ratio between is: in For device q i The transmission power, is the maximum transmit power of the device; is the channel gain; σ 2 (W) represents the noise power; therefore, q i With Es e The transmission rate of the channel between is: in Indicates that at time t q i With Es e The channel bandwidth allocated to the transmission channel between For Es e The bandwidth resources allocated at time t; so q i The business Upload to the corresponding Es e The channel transmission delay is: The other delay calculation methods are the same as (41), so the local Es e The total task processing delay is: (43) Non-local edge-side coding distributed computing: Non-local Total task processing delay (44) Cloud Center Layer Computing: Central Server CS c The total task processing delay is 5. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 1 is characterized by: The optimization objective is set as the total delay of executing the service within the T time slot; in Expressed as: Therefore, the optimization objective is formulated as problem (1): C8:1≤k≤n C1 ensures that the sum of the offloading percentages from local ES to non-local ES and CS does not exceed 1, C2 ensures that each local ES can only select one non-local ES to perform collaborative edge computing, C3 ensures that the bandwidth resources allocated to each ES do not exceed its idle bandwidth resources, C4 means that the total energy consumption does not exceed the energy we provide for each device, and C5 ensures Does not exceed the maximum allowable delay for a given business execution C6 and C7 ensure that each q i , n e,j and CS c , in handling business The number of CPUs and GPUs allocated cannot exceed their maximum number, and C8 guarantees the effectiveness of encoding.

6. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 1 is characterized in that: The step 5 specifically includes (51) When the data correlation is a graph structure: let G s =(χ s ,E s ) represents the graph structure, where x i ∈χ s is the vertex, e sij ∈E s is x i and x j The edge between G s The adjacency matrix is ​​denoted as A; the pairwise edge hyperedge group (E pair ): Figure G s Each hyperedge e in sij Connecting two vertices x in the corresponding edge i and x j , E pair It is expressed as follows E pair ={{x i ,x j }|(x i ,x j )∈E s } k-Hop hyperedge group E hop :E hop Find the related vertices of the central vertex through the k-Hop reachable position in the graph structure; Graph G s The k-Hop neighborhood of vertex x in is defined as: where k∈[2,n x ],n x It is Figure G s The number of vertices; therefore, the hyperedge group E hop Expressed as: (52) When data dependencies do not have a graph structure: Each node device has two types of data: one is attribute data, and the other is features associated with each vertex; Attribute hyperedge group (E attribute ): attribute data of different node devices, a set of hyperedges using neighbors in the attribute space is generated based on the attribute data, where each hyperedge represents an attribute or an available sub-attribute, which connects all nodes with this attribute; the subset of nodes with attribute a is represented as P att (a); Α is the set of all attributes or sub-attributes of the attribute; therefore, E attribute Expressed as: AND attribute ={P att (a)|a∈Α}; Node feature hyperedge group (E feature ): Given the features of each vertex, the second type E is generated by finding the neighbors of each vertex in the feature space feature Given a vertex as the centroid, its k nearest neighbors in the feature space are connected by a hyperedge, or all neighbors within a distance d from the centroid or centroid are selected; The k or d value is set according to the specific situation during the neighbor discovery process; (53) In order to fully fuse and utilize the high-order correlations of multimodal mixtures, an adaptive hyperedge group fusion strategy, namely Adaptive Fusion, is used; each hyperedge group is associated with a trainable parameter that adaptively adjusts the influence of multiple hyperedge groups on the final vertex embedding; it is defined as: ω k =copy(sigmoid(w k ),M k ) J=J1||J2||...||J K in is a trainable parameter shared by all hyperedges in a given hyperedge group k. sigmoid(·) is an element-wise normalization function; the vector represents the weight vector of the generated hyperedge group k; the copy(a,b) function returns a vector of size b, the value of which is filled by copying ab times; let M = M1 + M2 + ... + M k represents the sum of hyperedges in all hyperedge groups; Represents the weight matrix of the hypergraph, each entry W ii Represents the corresponding hyperedge e i The weight of J∈{0,1} N×M Represents the incidence matrix of a hypergraph generated by concatenating the incidence matrices of multiple hyperedge groups (·||·).

7. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 6 is characterized in that: In step 6: a method based on hypergraph neural network HyperGNN and particle swarm optimization PSO is adopted; HyperGNN captures the high-order correlation between heterogeneous nodes in the computing network; the optimization variables in the replacement problem of problem (1) are is the result of the conversion function based on HyperGNNs; considering in is the continuous assignment of node i, δ=[δ1,δ2,...,δ γ ] is the input embedding of the node; by adopting a parameterized transformation function, the training parameters (ω k ) and the input embedding (δ) to optimize our objective; this alternative optimization problem is described below: Defined on the combinatorial constrained hypergraph Hyper GNN The augmented (penalty) loss function is: in λ n is the weight of constraint n; in order to make HGNNConv + The forward pass on GPU / CPU devices is faster, using the matrix format: D x and D e are the degree matrices of nodes and hyperedges, Θ (l) are the parameters that layer l needs to learn during training; filters Ω are applied to nodes in the hypergraph to extract features; After convolution, we get X (l+1) , for further study.

8. The method for high-reliability resource scheduling optimization of heterogeneous computing power networks based on coded distributed computing and hypergraph neural networks according to claim 1 is characterized in that: The particle swarm optimization in step 7 can handle discrete variables and has good robustness when searching for the global optimal solution; the particle swarm optimization algorithm is used to find the optimal resource scheduling solution. Specifically, in order to convert HGNNConv + The continuous output is converted into integer node assignment, using HGNNConv + The continuous output generated is used to deduce a probability distribution over the discrete values ​​{ν1,ν2,...}; this probability distribution is designed based on the distance between the continuous output of the variable and a specific value, so that the probability of the variable taking a certain value is inversely proportional to its distance from the value; then from Take a sample ξ1,...,ξ from the corresponding output distribution Υ ; Use this sample to initialize a group of random particles and find the optimal solution minΤ(ξ i ); In each iteration, the particle is updated by following the two "extreme values". After finding the two optimal values, the particle updates its speed and position using the following formula: v i =w×v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i ) Where w is the inertia factor; v i is the speed of the particle; rand() is a random number between (0, 1); x i is the position of the particle, and x i =x i +v i ; c1 and c2 are learning factors.

Images