Method and system for scheduling computing power network resources

Abstract

The invention discloses a computing power network resource scheduling method and system, and the method comprises the steps: responding to a service request, carrying out the coding of a network state through an edge attention graph neural network, and enabling the network to integrate an edge feature vector into attention weight calculation in message transmission, so as to sense a link state; based on a coding result, a scheduling decision of joint optimization calculation and network resources is generated through a self-adaptive multi-target reinforcement learning strategy network, and the reward function weight can be dynamically adjusted according to the network state; decisions are executed based on a distributed multi-agent architecture, and all agents realize global coordination through a message coordination mechanism. According to the invention, efficient, adaptive and extensible scheduling of multiple targets in the dynamic computing power network is realized.

Description

Technical Field

[0001] This invention relates to the fields of computing power networks and artificial intelligence technology, and in particular to a method and system for scheduling computing power network resources. Background Technology

[0002] With the rapid development of 5G, IoT, and AI applications, computing networks need to uniformly schedule distributed computing and network resources to meet the real-time service demands of low latency and high bandwidth. Existing AI-based resource scheduling methods still have significant shortcomings: First, in terms of graph neural networks, existing methods mostly use standard Graph Attention Networks (GAT) or Graph Convolutional Networks (GCN), calculating attention or performing convolutions only based on node features, without incorporating edge features representing the real-time state of the link (such as bandwidth, latency, and load) into the weight calculation of message passing, resulting in insufficient perception of network dynamics. Second, in terms of optimization frameworks, existing deep reinforcement learning methods mostly use fixed-weight single-objective or simple weighted multi-objective reward functions, unable to dynamically adjust the priority of each optimization objective (such as latency, bandwidth, and computational efficiency) according to real-time states such as network congestion and uneven load, lacking adaptive capabilities. Third, in terms of system architecture, existing solutions mostly rely on centralized single agents for global decision-making, facing problems such as state space explosion, high decision latency, and single points of failure in large-scale networks, resulting in poor scalability and robustness. Therefore, how to achieve resource joint scheduling that is aware of link status, adaptively balances multiple objectives, and has good scalability in a dynamic and complex computing network environment has become an urgent technical problem to be solved. Summary of the Invention

[0003] To address the problems in existing graph neural networks, such as insufficient link state awareness due to lack of edge features, lack of adaptability in multi-objective optimization due to fixed reward function weights, and poor scalability of centralized single-agent architecture, this invention proposes a computing power network resource scheduling method and system.

[0004] The specific technical solution is as follows: A method for scheduling computing network resources, comprising the following steps:

[0005] In response to a service request, the network state is encoded based on an edge attention graph neural network. During message transmission, the edge feature vector is incorporated into the calculation of attention weights between nodes to dynamically adjust the message transmission strength according to the link state.

[0006] Based on an adaptive multi-objective reinforcement learning policy network, a joint optimization scheduling decision for computational and network resources is generated according to the encoding results; the policy network is optimized using a multi-objective reward function with dynamically adjustable weights, so as to adaptively adjust the priority of each optimization objective according to the real-time state of the network.

[0007] The scheduling decisions are executed based on a distributed multi-agent architecture. Each agent makes independent decisions based on local observations and coordinates globally through a message collaboration mechanism to achieve distributed scheduling of large-scale networks. By organically combining side-attention graph neural networks, adaptive multi-objective reinforcement learning, and the distributed multi-agent architecture, a complete closed-loop scheduling solution is formed. This solution not only accurately perceives the network link status but also dynamically adjusts optimization priorities based on real-time network conditions. Furthermore, distributed collaboration avoids the bottleneck of centralized decision-making, thus significantly improving the overall performance, adaptability, and scalability of resource scheduling in complex and dynamic computing network environments.

[0008] Furthermore, the edge feature vector includes at least the total link capacity, currently available bandwidth, propagation delay, queuing delay, packet loss rate, and link utilization.

[0009] The specific steps for incorporating edge feature vectors into the inter-node attention weight calculation are as follows: For edge (i,j), its feature vector e ij The eigenvector h of node i i The feature vector h of node j j The components are concatenated using a learnable transformation matrix W. n W e After performing a linear transformation and nonlinear activation on the attention vector a, the attention coefficient α is obtained by softmax normalization. ij By clearly defining the specific composition of the edge feature vectors, the graph neural network is able to acquire and process key state information that reflects the true load and quality of the links.

[0010] Furthermore, the attention coefficient α ij The calculation formula is:

[0011] Attention coefficient α ij =softmax(LeakyReLU(a T *[W n *h i ||W n *h j ||W e *e ij ]));

[0012] The activation function LeakyReLU has a negative slope of 0.2, and the transformation matrix W... n W eThe attention vector 'a' is learned through backpropagation during training, and || represents the vector concatenation operation. Embedding edge features into the attention weight calculation allows the strength of message passing to adaptively change with link capacity, latency, load, and other states, thereby making the network state representation encoded by the graph neural network more dynamic and accurate.

[0013] Furthermore, the multi-objective reward function R with dynamically adjustable weights is defined as:

[0014] R=w1*R bandwidth +w2*R latency +w3*R compute +w4*R balance ;

[0015] Among them, R bandwidth For bandwidth satisfaction rate, R latency For delayed gratification, R compute To calculate timeliness, R balance For load balancing;

[0016] Each weight w i The network status is dynamically adjusted according to the following rules: when the average link utilization is higher than the first threshold, increase w2; when the compute node load variance is higher than the second threshold, increase w4; when the request queue length is higher than the third threshold, increase w3; when the bandwidth allocation failure rate is higher than the fourth threshold, increase w1; after adjustment, normalization is performed to make ∑ i w i =1. This enables the system to intelligently respond to real-time conditions such as network congestion and uneven load, and flexibly switch priorities among multiple objectives such as bandwidth guarantee, low latency, fast computing, and load balancing. It achieves a leap from fixed-weight optimization to truly context-aware adaptive multi-objective optimization, better meeting the dynamic needs of diverse services.

[0017] Furthermore, the message coordination mechanism is implemented through an attention mechanism, specifically as follows:

[0018] Each agent will store its own hidden state h i By querying matrix W Q Transform into query vector Q i The hidden state set {h} of other agents j |j≠i} through the bond matrix W K Sum matrix W V Transform them into a set of key vectors K respectively -i Sum value vector set V -i The collaborative message m is obtained through attention calculation. i =Attention(Q i ,K -i V-i );

[0019] The collaborative messages are used to guide local decision-making and, when multiple agents select the same resource, prioritize them based on request urgency, resource demand, and timestamp. An attention mechanism is used to achieve implicit collaboration and information aggregation among agents, enabling each agent to consider the intentions of other agents and the overall resource situation when making local decisions. This effectively reduces decision conflicts, improves global resource utilization efficiency, and ensures the coordination and overall optimality of system decisions in a distributed architecture.

[0020] Furthermore, it also includes a transfer learning step to adapt the model to the new network topology;

[0021] The transfer learning method employs a domain adaptation approach, adding a domain alignment layer to the target topology, and performing feature alignment by minimizing the maximum average difference loss between the source and target topology graph embedding distributions.

[0022] Alternatively, a meta-learning approach can be employed, using a model-independent meta-learning framework to learn a set of initialization parameters. This allows for good performance on new topologies with only a few gradient updates. This enables scheduling models trained for specific topologies to quickly adapt to new network topologies, significantly enhancing the model's generalization ability and practicality.

[0023] Furthermore, the loss function for domain adaptation is:

[0024] L adapt =L task +λ*L domain ;

[0025] Among them, L task For mission losses, L domain The domain alignment loss is λ, which is the balance coefficient, ranging from 0.1 to 0.5.

[0026] The initialization parameter θ of the meta-learning is updated through the outer loop:

[0027] ;

[0028] Where T i For the sampling task, α is the inner loop learning rate. By minimizing the domain alignment loss or rapidly fine-tuning through meta-learning, the model ensures the preservation of core features and rapid convergence during cross-topology transfer, significantly improving deployment efficiency.

[0029] Furthermore, the adaptive multi-objective reinforcement learning policy network is trained using a proximal policy optimization algorithm, and its objective function is:

[0030] ;

[0031] Among them, the strategy ratio r t (θ)=π θ (a t |s t ) / π_θ old (a t |s t A t The advantage function is calculated using generalized advantage estimation. This is the cutting factor, with a value of 0.2;

[0032] Alternatively, a soft actor critic algorithm can be used, whose policy update objective is to maximize the weighted sum of expected reward and policy entropy.

[0033] A computing power network resource scheduling system for implementing the method includes:

[0034] The request and response module is used to receive service requests and trigger the scheduling process;

[0035] The state awareness and graph construction module is used to collect network topology, link status and computing node load information, and construct graph data containing node feature matrices, edge feature matrices and adjacency relationships;

[0036] The edge attention graph neural network encoding module is used to perform multi-layer message passing and feature aggregation on the graph data, and output node embedding and graph embedding that perceive the link state.

[0037] An adaptive multi-objective decision-making module includes a policy network based on proximal policy optimization or soft actor critic algorithm, used to output a joint decision on node selection, path selection and bandwidth allocation based on the embedded representation, and to calculate the reward based on a dynamic weighted multi-objective reward function;

[0038] The multi-agent collaboration and execution module includes agent units deployed at each network entry node. Each agent unit makes decisions based on local observations and collaborative messages generated through an attention mechanism. A conflict resolution unit is provided for priority arbitration and the final scheduling decision is issued for execution.

[0039] The transfer learning adaptation module is used to align feature distributions through the domain adaptation layer or to quickly fine-tune pre-trained model parameters using meta-learning when the network topology changes.

[0040] Furthermore, the edge attention graph neural network encoding module includes at least 3 edge attention layers, each with a hidden dimension of 128, 4 attention heads, a LeakyReLU activation function, and a Dropout layer with a dropout rate of 0.1.

[0041] The adaptive multi-objective decision-making module is equipped with a weight dynamic adjustment unit, which is used to monitor the average link utilization, calculate the node load variance, request queue length and bandwidth allocation failure rate, and dynamically adjust the reward function weight according to a preset threshold.

[0042] In the multi-agent collaboration and execution module, the number of agents is the same as the number of network entry nodes, and each agent maintains an independent policy network with shared parameters.

[0043] The above technical solution has the following advantages or technical effects:

[0044] 1. This invention integrates the real-time state of the link into the message passing of the graph nodes by using an edge attention graph neural network, which solves the problem of insufficient utilization of edge features in existing GNN methods and realizes more accurate and sensitive state perception and encoding of the dynamic topology and link quality of the computing network.

[0045] 2. This invention constructs an adaptive multi-objective reinforcement learning framework by designing a dynamic weight adjustment mechanism triggered by the state and combining it with the policy gradient algorithm (PPO / SAC). This solves the problem of fixed and rigid optimization objectives in existing methods and realizes the autonomous decision-making capability that can intelligently weigh the priorities of multiple objectives based on real-time network congestion and load conditions.

[0046] 3. This invention adopts a distributed multi-agent collaborative architecture and designs an implicit coordination and conflict resolution strategy based on an attention mechanism, which solves the scalability bottleneck and single point of failure risk of centralized single agents, and realizes efficient, robust and scalable distributed resource scheduling in large-scale computing networks.

[0047] 4. This invention integrates domain adaptation and meta-learning techniques, enabling the scheduling model to have rapid migration and generalization capabilities across network topologies. It solves the problems of traditional models requiring "one model per topology" and high retraining costs, significantly improving the practicality and deployment flexibility of this invention in real-world heterogeneous and dynamically changing computing network environments. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention;

[0049] Figure 2 This is a flowchart of the adaptive reward function adjustment process of the present invention. Detailed Implementation

[0050] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1

[0052] like Figure 1As shown, a method for scheduling computing network resources includes the following steps:

[0053] In response to service requests, the network state is encoded based on the edge attention graph neural network. During message transmission, the edge feature vector is incorporated into the calculation of attention weights between nodes to dynamically adjust the message transmission strength according to the link state.

[0054] Based on an adaptive multi-objective reinforcement learning policy network, a joint optimization decision for scheduling computational and network resources is generated according to the encoding results. The policy network is optimized using a multi-objective reward function with dynamically adjustable weights to adaptively adjust the priority of each optimization objective according to the real-time state of the network.

[0055] Scheduling decisions are executed based on a distributed multi-agent architecture. Each agent makes independent decisions based on local observations and coordinates globally through a message collaboration mechanism to achieve distributed scheduling of large-scale networks.

[0056] Specifically, in response to service requests within the computing network, the system first collects real-time network status. Specifically, it collects the features of each node in the network, forming a node feature set. For each node, its feature vector h... i It is 16-dimensional, including node type (e.g., switch, edge server, cloud data center), computing power level (e.g., TFLOPS), current CPU / GPU utilization, memory usage, and the length of the task queue. Simultaneously, it collects features of each link (edge) in the network, forming an edge feature set. For the edge connecting nodes i and j, its feature vector e... ij It is 8-dimensional, including at least the total link capacity, currently available bandwidth, propagation delay, queuing delay, packet loss rate, and link utilization. Based on the network topology connection relationship (adjacency matrix), node features and edge features are constructed into an attribute graph G=(V,E,H,E), where V is the set of nodes, E is the set of edges, H is the node feature matrix, and E is the edge feature matrix.

[0057] The constructed graph G is input into a multi-layered edge-attention graph neural network (Edge-Attention GNN) for encoding. For an edge (i,j) in the graph, the attention coefficient α used when calculating the aggregation of information of node i and its neighbor node j is... ij The calculation is as follows: It calculates the edge feature vector e. ij The eigenvector h of node i i The feature vector h of node j j The components are concatenated and then transformed using a learnable transformation matrix. and attention vector After performing linear transformation and nonlinear activation, the result is obtained by normalizing the calculation on all neighbors of node i using the softmax function.

[0058] More specifically, the attention coefficient α ij The calculation formula is as follows:

[0059] α ij =softmax(LeakyReLU(a T *[W n *h i ||W n *h j ||W e *e ij ]));

[0060] Here, || denotes the vector concatenation operation, and the negative slope of the LeakyReLU activation function is set to 0.2. The learnable parameters W... n W e 'a' and 'a' are obtained through backpropagation during model training. This is achieved by optimizing the real-time link state 'e'. ij Explicitly introducing attention weights allows the strength of message passing to adapt to the quality of the link. For example, when link (i,j) has high latency or heavy load, its corresponding α... ij The value will automatically decrease, so that node i reduces its dependence on node j's information when updating its own state, realizing intelligent bypassing of network congestion areas by information flow, and the encoded graph embedding can more accurately reflect the dynamic bottleneck of the network.

[0061] Node feature updates are achieved by weighted aggregation of neighbor information:

[0062] ;

[0063] Where N(i) represents the set of neighbors of node i. It is a non-linear activation function.

[0064] Multi-head attention mechanism:

[0065] K independent attention heads are used for parallel computation, and the final output is obtained by concatenation or averaging.

[0066] .

[0067] Example 2

[0068] This embodiment is a further extension of Embodiment 1. Specifically, this embodiment uses the Proximal Policy Optimization (PPO) algorithm as the core of the reinforcement learning agent. The agent's policy network π θThe graph embedding representation output from Example 1 is used as the state input. The action output is a composite decision consisting of three parts, which jointly realize the joint scheduling of computing and network resources: (1) Computation node selection: select a target node from the discrete set of all computing nodes; (2) Path selection: select one from the K pre-computed shortest paths (K=3) from the source node to the target computing node; (3) Bandwidth allocation: select a level from 10 discrete bandwidth allocation levels (e.g., 10%, 20%, ..., 100% of the requested bandwidth). This design supports both discrete and continuous decision-making needs.

[0069] To guide the agent in learning excellent scheduling strategies, this invention designs a multi-objective reward function R with dynamically adjustable weights. This function is defined as the weighted sum of the rewards for the four sub-objectives, as shown in the following formula:

[0070] R=w1*R bandwidth +w2*R latency +w3*R compute +w4*R balance ;

[0071] Among them, R bandwidth For bandwidth satisfaction rate, R latency For delayed gratification, R compute To calculate timeliness, R balance For load balancing. Where:

[0072] R bandwidth =B allocated / B requested B allocated For the actual allocated bandwidth, B requested Requested bandwidth;

[0073] R latency =max(0,1-L actual / L threshold );L actual For the actual end-to-end delay, L threshold This is the delay threshold;

[0074] R compute =max(0,1-T compute / T deadline );T compute For calculating the time of the task, T deadline The deadline for the task;

[0075] R balance =1-σ(load) / μ(load); σ(load) is the standard deviation of the load of all computing nodes, and μ(load) is the average load.

[0076] Each weight wi Adjust dynamically based on network status: w i (t+1)=w i (t)+η*Δw i (s t The system continuously monitors a set of key performance indicators (KPIs). At each training time step or fixed time window, the adjustment amount Δw is... i (s t The following adjustment rules will be implemented:

[0077] If the average link utilization is higher than the first threshold (e.g., 80%), then Δw2 = +0.05, which means that the network is becoming congested and the focus should be increased on low latency targets.

[0078] If the variance of the calculated node load is higher than the second threshold, then Δw4 = +0.05, which means that the load is uneven and the importance of load balancing should be increased.

[0079] If the global request queue length is higher than the third threshold, then Δw3 = +0.05, which means that there is a backlog of tasks, and more emphasis should be placed on completing the computation goal quickly.

[0080] If the bandwidth allocation failure rate is higher than the fourth threshold (e.g., 10%), then Δw1 = +0.05, which means that bandwidth resources are scarce and the importance of ensuring bandwidth is increased.

[0081] The adjustment step size η is set to 0.01. After each adjustment, all weights are immediately normalized: w i =w i / Σ j w j , ensure Σ i w i =1. The initial weights can be set to [0.3, 0.3, 0.2, 0.2]. The reward function itself has context-aware capabilities, and can automatically adjust the optimization direction according to the current pain points of the network, transforming the scheduling strategy from a static, potentially conflicting multi-objective trade-off to a dynamic, adaptive optimization that coordinates with the network conditions, such as... Figure 2 As shown:

[0082] R adaptive =α(s)*B(p)-β(s)*L(p)-γ(s)*T c (n);

[0083] Where α(s), β(s), and γ(s) are state-dependent adaptive coefficients, B(p) is the path bandwidth, L(p) is the path delay, and T... c (n) represents the computation time of computation node n.

[0084] The PPO algorithm is used to optimize the policy network parameters θ. The objective function of PPO is as follows:

[0085] ;

[0086] Among them, the strategy ratio r t (θ)=π θ (a t |s t ) / π_θ old (a t |s t A t Let be the advantage function calculated using generalized advantage estimation (GAE, λ=0.95). The pruning factor is γ, with a value of 0.2. This pruning mechanism effectively limits the magnitude of each policy update, greatly enhancing the stability of the training process and avoiding Q-value overestimation and policy oscillation. Simultaneously, its ability to support continuous / discrete mixed action spaces perfectly matches the scheduling decision requirements of this invention. The training hyperparameters are set as follows: discount factor γ = 0.99, learning rate 3 * 10^6. -4 The training rounds consist of 1000 rounds.

[0087] Example 3

[0088] This embodiment is a further extension of Embodiment 1 and / or Embodiment 2, specifically employing a distributed multi-agent architecture. Specifically, an independent scheduling agent is deployed at each entry node of the computing power network (i.e., the access point requested by the client). The number of agents is the same as the number of entry nodes (typically 10-20). All agents share the same set of policy network parameters θ (i.e., parameter sharing), but make decisions based on their respective local observations.

[0089] Each agent i makes decisions based on local observations and cooperative messages:

[0090] a i =π θ (o i ,m i );

[0091] Among them: o i For local observations of agent i, specifically including the request queue status of the local entry node, the load information of neighboring nodes (within 2 hops), and the resource status of locally reachable computing nodes; m i The received collaborative messages specifically include the decision intentions of other agents (target computing node, path selection), global resource utilization summary, and conflict warning information (multiple agents choosing the same resource).

[0092] Each agent will output the hidden state h from its own local GNN encoder. i The query vector Q is generated by transforming the data using three sets of learnable weight matrices. i =W Q *h i , bond vector K- i ={W K *h j} j≠i Sum vector V- i ={W V *h j} j≠i For agent i, its query vector Q is... i The key vector set K of all other agents -i Sum value vector set V -i Perform attention calculations and collaborative messages m i The calculation formula is as follows:

[0093] m i =Attention(Q i ,K -i V -i ).

[0094] Resource conflict occurs when multiple agents still choose overlapping links on the same computing node or network path based on local observations and collaborative messages. At this time, the system initiates a conflict resolution mechanism, arbitrating according to preset priority rules, which are considered in the following order: (1) Request urgency: requests with closer deadlines have higher priority; (2) Resource requirements: tasks requiring more computation or bandwidth have higher priority; (3) Timestamp: requests arriving earlier have higher priority. The beneficial effects of this embodiment are: distributed deployment eliminates the single-point bottleneck and failure risk of the central controller; the attention collaboration mechanism provides a global perspective for distributed decision-making, significantly reducing conflicts; and the arbitration mechanism ensures the service quality of critical services. This architecture enables the overall decision-making capability of the system to scale linearly with the network size.

[0095] Example 4

[0096] This embodiment is a further extension of the above embodiments. Specifically, this embodiment introduces transfer learning to enable the model to generalize across topologies.

[0097] Domain Adaptation method: When it is necessary to adapt an existing domain in the source topology T... src The trained model is deployed to the target topology T. tgtIn this case, domain adaptation techniques are employed. Specifically, the model is further fine-tuned on the dataset with the target topology, and a domain alignment loss term is added to the loss function. The total loss function is: L adapt =L task +λ*L domain L task This is the original task loss of PPO (such as L). CLIP ), L domain This is the loss used to align the feature distributions of the source and target domains; this invention uses the maximum mean difference (MMD) for calculation. λ is a balance coefficient, ranging from 0.1 to 0.5. Typically, we freeze most layers of the edge-attention GNN encoder, only fine-tuning its last few layers and the policy network to preserve generalization features and accelerate adaptation.

[0098] Meta-learning methods: To achieve better and faster adaptation capabilities, meta-learning can be employed during the initial training phase. Specifically, the Model-Independent Meta-Learning (MAML) framework is used to learn initialization parameters that can quickly adapt to new topologies. ; where α is the inner loop learning rate.

[0099] Meta-learning training process: (1) Sample a batch of tasks (topologies of different sizes / structures) from the topological distribution p(T); (2) Perform a small number of gradient updates for each task (inner loop); (3) Calculate the meta-gradient on all tasks and update the initial parameters (outer loop). Good performance can be achieved with only 5-10 rounds of fine-tuning on the new topology, while training from scratch requires more than 500 rounds.

[0100] The test was conducted in a large-scale computing network simulation environment generated by the Barbasi-Albert model, containing 100 nodes, 20 compute nodes, and 15 client nodes.

[0101] Benchmark Comparison: This invention is compared with existing technologies employing DQN+MLP-GNN. Test results show that this invention improves the average reward metric by 38% and the request success rate by 25%.

[0102] Real-time performance: After training is completed, the average time for a single scheduling decision (from encoding to output) is about 2 milliseconds, which meets the requirements of high real-time applications.

[0103] Scalability: When the number of nodes increases linearly from 50 to 200, the decision time of the system increases approximately linearly, which proves the good scalability of the multi-agent distributed architecture.

[0104] Ablation experiments: Verifying the contribution of core technology components by removing them one by one.

[0105] Full model: Average reward 0.51.

[0106] Removing edge attention (replacing it with standard GAT): average reward 0.38 (a 25% decrease), demonstrating the core value of edge features in attention calculation.

[0107] Removing dynamic weights (fixed reward weights): the average reward decreased to 0.41 (a 20% reduction), demonstrating the necessity of adaptively adjusting multi-objective priorities.

[0108] Removing multi-agent collaboration (instead of centralized single agent): On large-scale networks, the average reward is 0.45 (a decrease of 12%), and latency increases, demonstrating the superiority of the distributed architecture.

[0109] Removing transfer learning capability: When the topology changes randomly, the model performance drops by more than 40%, while the model with transfer learning capability only drops by 8%, demonstrating that generalization capability is a key support for practicality.

[0110] Example 5

[0111] The computing power network resource scheduling system for implementing Embodiments 1 to 4 includes:

[0112] Request and Response Module: As the system entry point, it is used to receive service requests submitted by external applications or users, parse the request content (such as source node, required bandwidth, computing type, computing power requirements, latency requirements, etc.), and trigger the entire subsequent scheduling process.

[0113] State Awareness and Graph Construction Module: This module interfaces with the network control layer or monitoring system to collect real-time global state information of the computing network. Specifically, it collects: network topology connections, real-time characteristics of each link (including total link capacity, current available bandwidth, propagation delay, queuing delay, packet loss rate, and link utilization), and load characteristics of each computing node (CPU / GPU utilization, memory usage, and queue length). Subsequently, this module standardizes this information and constructs a computer-processable attribute graph data structure, namely, graph data G containing node feature matrices, edge feature matrices, and adjacency relationships.

[0114] The edge-attention graph neural network encoding module receives graph data G from the graph construction module. This module performs multi-layer message passing and feature aggregation on the input graph data, and finally outputs node embeddings and graph embeddings that are aware of link states.

[0115] Adaptive Multi-Objective Decision Module: This module contains a pre-trained reinforcement learning policy network. The policy network is constructed based on the Proximal Policy Optimization (PPO) or Soft Actor Critic (SAC) algorithm. This module receives the graph embedding output from the encoding module and outputs a joint scheduling decision based on it, which includes computation node selection, path selection, and bandwidth allocation. Simultaneously, this module embeds a reward calculation unit that calculates the immediate reward for each decision based on a dynamic weighted multi-objective reward function, for possible online learning or performance evaluation.

[0116] Multi-Agent Coordination and Execution Module: This module adopts a distributed architecture, including agent units deployed at various network entry nodes. Each agent unit receives global decision suggestions from the decision-making module and combines its own local observations (acquired through an independent state-aware submodule) with collaborative messages generated through an attention mechanism to make the final decision confirmation. The module includes a conflict resolution unit to arbitrate priorities based on request urgency, resource demand, and timestamps when multiple agent units compete for the same resource. After arbitration, the module translates the final, conflict-free scheduling decision into specific control commands and issues them to the data plane of the computing network (such as the SDN controller and compute node manager) to execute resource allocation.

[0117] Transfer learning adaptation module: This module is activated when the system needs to be deployed to a new network topology environment, or when the existing topology changes significantly. It provides two adaptation methods: one is to introduce a domain adaptation layer to fine-tune the model on the target topology data to minimize the difference in feature distribution between the target and source topologies (domain alignment); the other is to directly load a set of highly generalizable model initialization parameters pre-trained through a meta-learning framework, and then perform rapid fine-tuning on the new topology in a very small number of steps.

[0118] Hardware configuration of the side attention graph neural network encoding module: To balance computational efficiency and model expressiveness, the side attention graph neural network in this module preferably adopts the following configuration: it contains at least 3 side attention layers, with the hidden layer dimension of each layer set to 128, and uses 4 independent attention heads (K=4) for parallel computation. The activation function is uniformly LeakyReLU (negative slope 0.2). To prevent overfitting during model training, a Dropout operation is introduced after each attention layer computation, with a dropout rate set to 0.1.

[0119] The intelligent unit of the adaptive multi-objective decision-making module includes an independent dynamic weight adjustment unit. This unit continuously obtains four key performance indicators from the state awareness module: average link utilization, computing node load variance, global request queue length, and recent bandwidth allocation failure rate. The unit has a built-in comparator that compares these indicators with preset thresholds (as described in Example 2) in real time. Once an indicator exceeds the threshold, the unit generates a weight adjustment signal according to predetermined rules (e.g., increasing the weight of the corresponding sub-objective by Δw=0.05) and sends it to the reward calculation unit, thereby achieving dynamic and automated adjustment of the reward function weights.

[0120] Deployment rules for the multi-agent collaboration and execution module: To achieve efficient distributed scheduling, the deployment of this module must follow these principles: the number of agent units in the system should be the same as the number of nodes that actually act as service request entry points in the computing network. For example, if the network has 15 edge access switches that can serve as entry points, then 15 agent units should be deployed. Each agent unit maintains a policy network instance locally for fast inference based on local information. All these local policy network instances share the same set of global parameters, which are periodically synchronized and updated by the central training server.

[0121] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for scheduling computing network resources, characterized in that the steps include: include: In response to a service request, the network state is encoded based on an edge attention graph neural network. During message transmission, the edge feature vector is incorporated into the calculation of attention weights between nodes to dynamically adjust the message transmission strength according to the link state. Based on an adaptive multi-objective reinforcement learning policy network, a joint optimization scheduling decision for computational and network resources is generated according to the encoding results; the policy network is optimized using a multi-objective reward function with dynamically adjustable weights, so as to adaptively adjust the priority of each optimization objective according to the real-time state of the network. The scheduling decisions are executed based on a distributed multi-agent architecture. Each agent makes independent decisions based on local observations and coordinates globally through a message collaboration mechanism to achieve distributed scheduling of large-scale networks.

2. The computing power network resource scheduling method according to claim 1, characterized in that, The edge feature vector includes at least the total link capacity, currently available bandwidth, propagation delay, queuing delay, packet loss rate, and link utilization. The specific steps for incorporating edge feature vectors into the inter-node attention weight calculation are as follows: For edge (i,j), its feature vector e ij The eigenvector h of node i i The feature vector h of node j j The components are concatenated using a learnable transformation matrix W. n W e After performing a linear transformation and nonlinear activation on the attention vector a, the attention coefficient α is obtained by softmax normalization. ij .

3. The computing power network resource scheduling method according to claim 2, characterized in that, The attention coefficient α ij The calculation formula is: Attention coefficient α ij =softmax(LeakyReLU(a T *[W n *h i ||W n *h j ||W e *e ij ])); The activation function LeakyReLU has a negative slope of 0.2, and the transformation matrix W... n W e The attention vector a is learned through backpropagation during training, and || represents the vector concatenation operation.

4. The computing power network resource scheduling method according to claim 1, characterized in that, The multi-objective reward function R, whose weights can be dynamically adjusted, is defined as follows: R=w1*R bandwidth +w2*R latency +w3*R compute +w4*R balance ; Among them, R bandwidth For bandwidth satisfaction rate, R latency For delayed gratification, R compute To calculate timeliness, R balance For load balancing; Each weight w i The network status is dynamically adjusted according to the following rules: when the average link utilization is higher than the first threshold, increase w2; when the compute node load variance is higher than the second threshold, increase w4; when the request queue length is higher than the third threshold, increase w3; when the bandwidth allocation failure rate is higher than the fourth threshold, increase w1; after adjustment, normalization is performed to make ∑ i w i =1.

5. The computing power network resource scheduling method according to claim 1, characterized in that, The message coordination mechanism is implemented through an attention mechanism, specifically as follows: Each agent will store its own hidden state h i By querying matrix W Q Transform into query vector Q i The hidden state set {h} of other agents j |j≠i} through the bond matrix W K Sum matrix W V Transform them into a set of key vectors K respectively -i Sum value vector set V -i The collaborative message m is obtained through attention calculation. i =Attention(Q i ,K -i V -i ); The collaborative message is used to guide local decision-making and, when multiple agents select the same resource, prioritizes the arbitration based on request urgency, resource demand, and timestamp.

6. The computing power network resource scheduling method according to claim 1, characterized in that, It also includes a transfer learning step to adapt the model to the new network topology; The transfer learning method employs a domain adaptation approach, adding a domain alignment layer to the target topology, and performing feature alignment by minimizing the maximum average difference loss between the source and target topology graph embedding distributions. Alternatively, a meta-learning method can be adopted, which learns a set of initialization parameters through a model-independent meta-learning framework, so that good performance can be obtained with a small number of gradient updates on the new topology.

7. A method for scheduling computing network resources according to claim 6, characterized in that, The loss function for domain adaptation is: L adapt =L task +λ*L domain ; Among them, L task For mission losses, L domain The domain alignment loss is λ, which is the balance coefficient, ranging from 0.1 to 0.

5. The initialization parameter θ of the meta-learning is updated through the outer loop: ； Where T i For the sampling task, α is the inner loop learning rate.

8. The computing power network resource scheduling method according to claim 1, characterized in that, The adaptive multi-objective reinforcement learning policy network is trained using a proximal policy optimization algorithm, and its objective function is: ； Among them, the strategy ratio r t (θ)=π θ (a t |s t ) / π_θ old (a t |s t ), A t The advantage function is calculated using generalized advantage estimation. This is the cutting factor, with a value of 0.2; Alternatively, a soft actor critic algorithm can be used, whose policy update objective is to maximize the weighted sum of expected reward and policy entropy.

9. A computing power network resource scheduling system, used to implement the method as described in any one of claims 1 to 8, characterized in that, include: The request and response module is used to receive service requests and trigger the scheduling process; The state awareness and graph construction module is used to collect network topology, link status and computing node load information, and construct graph data containing node feature matrices, edge feature matrices and adjacency relationships; The edge attention graph neural network encoding module is used to perform multi-layer message passing and feature aggregation on the graph data, and output node embedding and graph embedding that are aware of the link state. An adaptive multi-objective decision-making module includes a policy network based on proximal policy optimization or soft actor critic algorithm, used to output a joint decision on node selection, path selection and bandwidth allocation based on the embedded representation, and to calculate the reward based on a dynamic weighted multi-objective reward function; The multi-agent collaboration and execution module includes agent units deployed at each network entry node. Each agent unit makes decisions based on local observations and collaborative messages generated through an attention mechanism. A conflict resolution unit is provided for priority arbitration and the final scheduling decision is issued for execution. The transfer learning adaptation module is used to align feature distributions through the domain adaptation layer or to quickly fine-tune pre-trained model parameters using meta-learning when the network topology changes.

10. A computing power network resource scheduling system according to claim 9, characterized in that, The edge attention graph neural network encoding module contains at least 3 edge attention layers, each with a hidden dimension of 128, 4 attention heads, and LeakyReLU activation function, and includes a Dropout layer with a dropout rate of 0.

1. The adaptive multi-objective decision-making module is equipped with a weight dynamic adjustment unit, which is used to monitor the average link utilization, calculate the node load variance, request queue length and bandwidth allocation failure rate, and dynamically adjust the reward function weight according to a preset threshold. In the multi-agent collaboration and execution module, the number of agents is the same as the number of network entry nodes, and each agent maintains an independent policy network with shared parameters.

Images