A semantic perception routing scheduling method and system for a computing power network

Abstract

The application discloses a semantic perception routing scheduling method and system for computing power network, and belongs to the technical field of computing power network and communication. The application aims at the problems that the traditional host-centered routing scheduling mechanism cannot perceive service semantics and computing power state, resource supply and demand are not matched, and multi-computing power instance optimization capability is insufficient, and proposes an OSAF-SORS paradigm: by introducing a service tag (STag) decoupled from location and home as a new type of addressing routing identifier, an open service available priority (OSAF) routing scheduling mechanism is realized; based on deep reinforcement learning and multi-layer graph attention network, a service-oriented routing scheduling (SORS) algorithm is designed to jointly capture resource dependency and queue dynamics, complete multi-computing power instance optimization and path collaborative optimization. The application can significantly improve the service level agreement (SLA) achievement rate, resource utilization rate and load balancing effect, and is suitable for wide-area high-performance computing, immersive video, distributed AI training and other low-latency high-reliability business scenarios.

Description

Technical Field

[0001] This invention relates to computing power networks, routing scheduling, and deep reinforcement learning technologies, and in particular to a routing scheduling paradigm that shifts from host-centric to semantically aware, applicable to next-generation network scenarios such as 6G, industrial internet, distributed computing, and immersive services. Background Technology

[0002] The deep integration of networks and computing power has driven the development of Computing Networks (CPNs), with routing optimization playing a crucial role in improving resource utilization efficiency and ensuring end-to-end service delivery. Traditional IP networks typically employ host-centric routing mechanisms, such as OSPF and RIP, which primarily rely on IP addresses for network connectivity maintenance, topology discovery, and shortest path calculation. While these methods offer good stability and low control overhead in static, homogeneous network environments, their routing decisions revolve around address reachability, lacking awareness of service semantics, service level requirements, and computing resource status. Furthermore, since IP headers only carry location semantics, they struggle to express higher-level service intentions and cannot differentiate between different service flows. Therefore, traditional host-centric routing mechanisms are ill-suited to the requirements of collaborative optimization of network, computing power, and service needs in CPNs.

[0003] Traditional routing mechanisms have the following drawbacks: (1) The message only carries location semantics and cannot perceive service semantics and computing power status, making it difficult to match business needs; (2) Routing and scheduling are decoupled, scheduling ignores network dynamics and routing ignores computing power load, resulting in resource supply and demand mismatch; (3) It does not support the optimal selection of multiple computing power instances and cannot achieve optimal task allocation in distributed heterogeneous nodes; (4) It has poor scalability and robustness when facing dynamic load, node failure and business fluctuations.

[0004] Existing related technologies include reinforcement learning-based route optimization, computationally aware routing, and cooperative scheduling, but they generally suffer from the following shortcomings: 1) Lack of unified service semantic abstraction identifiers, resulting in semantic separation between the data plane and the control plane; 2) The algorithm is solvable when only considering network conditions, but its complexity increases significantly once computational power is taken into account. 3) Unable to maintain stable convergence and load balancing in large-scale dynamic computing network environments; 4) It does not support the optimal selection of multiple computing power instances, making it difficult to meet the deterministic guarantee of SLA.

[0005] Therefore, there is an urgent need for a unified routing and scheduling scheme that can perceive service semantics, computing power status and network dynamics, and support multi-instance optimization. Summary of the Invention

[0006] Purpose of the invention: To address the problems of traditional host-centric routing scheduling mechanisms failing to perceive service semantics and computing power status, resource supply and demand mismatch, and insufficient ability to optimize multiple computing power instances, this invention aims to provide a semantic-aware routing scheduling method and system for computing power networks.

[0007] Technical solution: A semantic-aware routing scheduling method for computing power networks, comprising: (1) Introduce a service tag STag, which is decoupled from the physical host location and ownership, as an addressing and routing identifier in the computing power network, and build a routing scheduling mechanism that prioritizes open services; The service tag STag is independently encapsulated in the IPv6 extension header, and each field bridges the packet-level identifier with service semantics, including fields reserved for future scalability considerations; (2) Based on multi-layer graph attention network and deep reinforcement learning, a service-oriented routing scheduling (SORS) algorithm is designed. The algorithm jointly models the network resource status, computing load, and queue dynamics to achieve optimal selection of distributed multi-computing instances and collaborative decision-making of routing paths. The process includes: (2.1) A multi-layer graph attention neural network is used to dynamically capture the multi-dimensional resource interdependencies between different network nodes, including the network and computing resource status and the queuing dynamics of service scheduling, and uses it as the semantic abstraction engine of the routing and scheduling SORS algorithm to convert topology and scheduling data into hierarchical embeddings for joint routing and scheduling decisions. (2.2) Extracting topological relevance based on multi-layer graph attention neural network, including aggregating multi-head outputs by splicing or averaging, each layer contains residual connections, random deactivation and layer normalization to ensure stability and generalization ability, and finally deriving attention weights from the neighbor set to avoid explicit adjacency matrix; (2.3) The routing planner RPL is used as the decision core of the route scheduling SORS algorithm. It performs semantic-aware routing by jointly evaluating computing power, network and queuing conditions. The SLA configuration file guides the multi-objective optimization in the routing and scheduling process.

[0008] Furthermore, the service tag STag is encapsulated in the form of an IPv6 extension header with a total length of 128 bits, including: STag identifier bit 8 bits, service identifier type 2 bits, control flag bit 6 bits, first reserved bit 16 bits, STag value 32 bits, traffic indication 32 bits, and second reserved bit 32 bits.

[0009] Furthermore, the routing and scheduling mechanism that prioritizes the availability of open services is specifically as follows: The control plane is based on the STag management service SLA configuration file and computing power-network resource scheduling strategy; The data plane uses STag to perform packet forwarding and service instance matching; By decoupling service logic from physical host location, service continuity can be maintained during resource elasticity adjustments, including computing instance migration, dynamic creation, or release.

[0010] Furthermore, the route scheduling SORS algorithm includes three steps: network resource state modeling, route planning, and route optimization training. The multi-layer graph attention network MMGAT is used to extract features and model topological dependencies of node computing power utilization, bandwidth utilization, latency factor, load balancing, arrival queue length, computing queue length, and service device type vector. Computational power-aware paths are generated based on Dueling DDQN, network-aware paths are generated based on a latency-bandwidth-queue hybrid cost function, and the optimal node-path pairs are output through a multi-objective scoring function. The routing strategy is iteratively optimized based on the Dual-Depth Q Network (DDQN) and the Adaptive Exploration Policy (AEPE).

[0011] In the above scheme, the multi-objective scoring function comprehensively weights network utilization, computing power utilization, latency, normalized arrival queue, normalized computation queue, and reinforcement learning Q-value.

[0012] Furthermore, the reward function of the route scheduling SORS algorithm is composed of a weighted average of path efficiency, resource utilization, route failure penalty, and queue delay penalty.

[0013] Furthermore, the adaptive exploration strategy AEPE adopts an exponentially decaying ε-greedy mechanism, combining softmax sampling with temperature parameter adjustment and uniform sampling to balance exploration and utilization.

[0014] Based on the implementation of the above method, the present invention also provides a semantic-aware routing scheduling system for computing power networks, the semantic-aware routing scheduling system comprising: The STag encapsulation module is used to embed the service tag STag into the IPv6 extension header to generate semantically aware packets; The OSAF control module is used for service registration, requirement parsing, STag mapping, and generation and distribution of computing power network joint routing scheduling policies. The SORS execution module is used for network resource state modeling, route planning, reinforcement learning training, and multi-computing instance optimization decision-making. The data forwarding module is used to perform packet forwarding and service connection establishment based on STag and routing decisions.

[0015] Furthermore, the SORS execution module includes: a resource feature extraction unit, a multi-layer graph attention network unit, a duel dual-depth Q network unit, and an adaptive exploration training unit.

[0016] Beneficial Effects: Compared with existing technologies, the method described in this invention implements the OSAF-SORS paradigm: by introducing service tags (STags) decoupled from location and affiliation as novel addressing routing identifiers, an Open Service Availability First (OSAF) ​​routing scheduling mechanism is realized; based on deep reinforcement learning and multi-layer graph attention networks, a service-oriented routing scheduling (SORS) algorithm is designed to jointly capture resource dependencies and queue dynamics, completing the optimal selection of multiple computing instances and path collaborative optimization. This invention can significantly improve Service Level Agreement (SLA) achievement rate, resource utilization, and load balancing effect, and is suitable for low-latency, high-reliability business scenarios such as wide-area high-performance computing, immersive video, and distributed AI training. Attached Figure Description

[0017] Figure 1 It is a process of modeling the structure and resource state of multi-layer graph attention networks; Figure 2 This is the operating mechanism of the Route Planner (RPL); Figure 3 This is the overall training flowchart for the Service-Oriented Routing and Scheduling (SORS) algorithm. Detailed Implementation

[0018] To illustrate the technical solution provided by this invention in detail, further description is provided below with reference to the accompanying drawings.

[0019] (1) Service Tag

[0020] In computer science and software development, STag (Service Tag) typically refers to a unique identifier associated with a service, particularly in hardware devices and service management systems. Specifically, in certain servers, storage devices, and network equipment, STag is a serial number used to uniquely identify hardware assets.

[0021] The STag carried in the data plane acts as a lightweight abstraction of computing and network service capabilities. It can express fine-grained SLA requirements within the operations and management domains, such as guaranteeing a video service with 20 Mbit / s bandwidth or a rendering task constrained by 10 ms latency. The STag does not directly embed detailed SLA parameters into the packet header; instead, it acts as an index linking packets to the SLA semantics and routing policies maintained by the control plane. On the control plane, routing and scheduling policies are generated based on SLA profiles and the current state of network and computing resources, and are indexed by the STag. These policies are distributed to service gateways to guide traffic forwarding. During ingress processing, user requests specify the target service through their ID and SLA attributes, and the service gateway maps them to specific computing instances and network paths according to the policy table. This process enables coordinated optimization of computing power allocation and path selection across heterogeneous resources.

[0022] Table 1. STag's Independent Encapsulation Format in the IPv6 Extended Header

[0023] The independent encapsulation format of STag in the IPv6 extended header is shown in Table 1. Each field bridges packet-level identification with service semantics: the STag indicator identifies the header type, the SID type specifies the service class, and the control flag enables lightweight signaling for options such as load balancing or latency preference. The STag value and traffic hint fields support instance identification and class-based scheduling, while reserved fields ensure future scalability. This design enables efficient resolution in the data plane while deferring the interpretation of SLA semantics to the control plane.

[0024] STag operates as a bidirectional interface: horizontally bridging the network and computing systems to facilitate comprehensive resource awareness, and vertically coupling service layer requirements with the underlying network to achieve SLA-compliant coordination. STag itself does not carry detailed SLA parameters, but rather acts as a mapping interface linking the data plane and control plane. Its lightweight and controllable architecture maintains secure isolation and imposes minimal hardware overhead on the service gateway.

[0025] (2) Service-oriented routing and scheduling process

[0026] (2.1) A multi-layer graph attention neural network (MMGAT) is used to dynamically capture the multi-dimensional resource dependencies between different network nodes, including network and computing resource states and queuing dynamics of service scheduling. It serves as the semantic abstraction engine for SORS, transforming topology and scheduling data into hierarchical embeddings for joint routing and scheduling decisions. Specifically: For each node First, the utilization rate of computing resources With recognition and processing capabilities:

[0027] In the formula, and Representing nodes respectively Available computing power resources and total computing power resources.

[0028] Bandwidth utilization This is to reflect the risk of congestion on aggregated links between adjacent nodes.

[0029] in Represents a node The set of adjacent nodes, and Representing links respectively Total bandwidth and available bandwidth.

[0030] Define delay factor To quantify nodes The composite delay, taking into account both the local processing delay and the average transmission delay to the neighbor:

[0031] in, It is a node Processing latency, It is a node Transmission delay, and This represents the minimum and maximum combined delay among all nodes.

[0032] Through load balancing To measure the uniformity of local resource distribution to guide global routing and scheduling:

[0033] In the formula, .

[0034] At the node The number of tasks waiting to be scheduled is represented by the arrival queue length:

[0035] In the formula, It is a node The set of unscheduled tasks It is the maximum arrival queue length.

[0036] Calculating queue length is used to measure nodes Number of tasks waiting in the queue:

[0037] In the formula, It is a node A set of tasks awaiting scheduling. It is the maximum computation queue length.

[0038] Service device type vector is defined as a vector of nodes Service capabilities are encoded as one-hot or embedded vectors:

[0039] In the formula, Represents a node Does it support sets? The first in Types of services (e.g., CPU-intensive, GPU-enabled, high-memory, storage-oriented).

[0040] The above indicators are combined to construct an initial feature vector. This allows the capture of the resource, queuing, and service type status of a specific node. Subsequently, topological dependencies between nodes are learned through MMGAT's hierarchical attention layer.

[0041] (2.2) MMGAT multi-layer attention structure to extract topological relevance:

[0042] The multi-head outputs are aggregated through concatenation or averaging. The first two layers emphasize 1-hop and 2-hop neighborhoods, while the last layer captures 3-hop global semantic dependencies. Each layer includes residual connections and random deactivation. Layer normalization is used to ensure stability and generalization ability.

[0043] The final embedding is represented as Expand it into a state vector: Combining the process in step (2.2), the operating mechanism of the route planner (RPL) is as follows: Figure 2 As shown, MMGAT jointly encodes the node resource status, queuing dynamics, and learned topology. MMGAT does this by analyzing the neighbor set... The attention weights are derived to avoid explicit adjacency matrices, thus enhancing adaptability.

[0044] (2.3) Route planning process

[0045] The Route Planner (RPL), as the decision-making core of SORS, performs semantic-aware routing by jointly evaluating computing power, network conditions, and queuing criteria. Upon receiving a route with a STag... ( When a service request is made for (a set of service types), RPL retrieves the corresponding SLA configuration file, which is constructed as a spatial state representation: in This indicates the target computing power, bandwidth, and latency constraints. and Indicates arrival and computation queue constraints, The required service equipment types are coded. The SLA configuration file guides multi-objective optimization during routing and scheduling.

[0046] Further integration Figure 2 As shown, for each candidate compute instance, RPL constructs two routing paths: a compute-aware path and a network-aware path. These paths are further evaluated by incorporating queuing latency to ensure that the target instance is both resource-sufficient and schedulable. Finally, the highest-scoring path and its corresponding instance are determined to optimize the distributed multi-compute instance architecture.

[0047] For perceptual path generation, a deep learning algorithm based on the DQN architecture is used to perform global value estimation and decouple local decision-making. In each iteration, the agent selects the node with the highest Q value.

[0048] in Capture global patterns, such as underutilized clusters, short queue lengths, or low-latency regions, and It focuses on the relative suitability of each neighbor for the current task.

[0049] For network-aware paths, RPL uses a hybrid latency-bandwidth-queue metric to minimize transmission costs:

[0050] In the formula, It is an influencing factor, dynamically adjusted according to service type, reducing computationally intensive services. Prioritize bandwidth and increase latency-sensitive services To emphasize low latency. Parameters Nodes with long arrival queues are penalized to prevent congestion at the scheduling level.

[0051] RPL selects the optimal node by calculating node scores in the following manner:

[0052] Link congestion, node load, latency, and queue length are weighted based on the dynamic Q-value of the self-learning agent, where, , , , and These are the corresponding variable weights.

[0053] (3) Route optimization training

[0054] ROT (Robotic Array Optimizer) promotes Resource Utilization (RPL) by optimizing distributed multi-computing instances through a feedback-driven iterative decision loop. ROT also dynamically enhances routing and scheduling strategies through deep reinforcement learning. Traditional mathematical methods often struggle to reconcile conflicting objectives, such as reducing latency while maintaining load fairness and queuing efficiency. Figure 3 As shown, we develop a ROT using a dual deep Q-neural network (DDQN), incorporating multi-objective rewards and experience replay to dynamically optimize the path selection algorithm. This closed-loop process ensures continuous adaptation to network dynamics, improving resource efficiency, service success rate, and task scheduling fairness.

[0055] (3.1) Defining the Reward Function: The reward function aims to guide DDQN to maintain compatibility with the SORS multi-objective optimization framework while balancing path efficiency, resource utilization, and queue management. It integrates four components through weighted summation:

[0056] in, Quantify path efficiency. Indicates the assessment of resource utilization, This indicates that the penalty route failed. This indicates the delayed item in the management queue. Weight , , and By dynamically adjusting based on the service semantics extracted from STag, SORS can prioritize computing load balancing, network efficiency, and task scheduling for different service types.

[0057] Path efficiency Defined as path length The reciprocal:

[0058] The above formula can encourage agents to choose shorter paths, while avoiding excessive penalties for slightly longer paths through normalization.

[0059] Resource utilization combines node-level computing power load and link-level bandwidth utilization to ensure that routing decisions consider both computing power and bandwidth resources, expressed as:

[0060] in, Represents a node The computing load, Reflection Link Bandwidth utilization. By utilizing all nodes and link set Calculate the average value to obtain .

[0061] The routing failure penalty is applied only when there is no valid path during the routing failure. ).

[0062] Queue latency penalty: Defined as the additional latency incurred by a task in the arrival queue and computation queue. Ensuring routing decisions are based not only on resource availability but also on task scheduling feasibility is expressed as:

[0063] in and It is a weight that balances the impact of arrival queue and computation queue delays.

[0064] Compound rewards Directly contributing to the Q-value update process, the improved DQN can formulate policies corresponding to the dual objectives of efficient path selection and fair resource allocation in SORS, while ensuring scheduling fairness. Using DDQN for Q-value updates: ROT leverages DDQN to promote adaptive routing in dynamic network environments, improving upon traditional DQN by separating action selection from evaluation. This strategy mitigates Q-value overestimation bias while remaining compatible with the multi-objective optimization framework of SORS.

[0065] (3.2) Construct a dual deep Q neural network (DDQN) framework.

[0066] The Dual Deep Q-Network (DDQN) framework comprises two independent Q-networks to enhance training stability. Experience replay buffer. Used to preserve past transfers ,in Indicates the current state. Indicates the selected action. Indicates an immediate reward. Indicate the subsequent state, and Used as a termination flag. During training, ROT is retrieved from the experience replay buffer. Randomly select small batches to disrupt the temporal correlation between subsequent data, thereby improving learning stability and generalization ability.

[0067] In traditional deep Q-neural networks (DQNs), the target Q-value is determined by directly using the target network to select and evaluate the optimal action for the next state, expressed as: It is the target network for optimization. These represent the optimal state and action, respectively, and are highly prone to gradient explosion.

[0068] DDQN decouples action selection from evaluation, and the main network... Choose the optimal action: , Target Network Evaluate its Q value: , This separation mitigates overly optimistic value assessments and greatly reduces the risk of gradient explosion, which is crucial for semantically aware routing in SORS, as seemingly ideal paths may mask hidden obstacles such as uneven computing loads or temporary connection congestion.

[0069] To further stabilize training, ROT integrates two complementary mechanisms: soft parameter synchronization and a robust loss function. The soft parameter synchronization strategy... The target network parameters are updated gradually, replacing sudden hard updates with a smoother transition to mitigate target value drift. Additionally, the Huber loss function... , Accuracy is maintained by exhibiting mean squared error when there are small deviations, while transitioning to linear loss when there are large errors to prevent gradient explosion during backpropagation, thus balancing gradient sensitivity.

[0070] The ROT continuously adjusts the routing and scheduling strategies of SORS based on fluctuating resource conditions and task queue dynamics, striking a balance between exploring unknown paths and utilizing established high-value routes. The closed-loop interaction between Q-value updates and RPL ensures continuous alignment with service semantics and network dynamics.

[0071] Adaptive Exploration and Policy Evolution (AEPE): Adaptive Exploration and Policy Evolution (AEPE) improves SORS training by dynamically balancing the trade-off between exploration and exploitation as it adapts to changing network conditions. AEPE enhances traditional SORS training by incorporating a hybrid exploration strategy. - Greedy strategy, which integrates temperature regulation action selection and adaptive decay scheduling, enables SORS to effectively identify the globally optimal routing path while ensuring the stability of heterogeneous environments.

[0072] In each action selection step, ROT determines the current state. and action space Establish an execution strategy. The probability of an agent participating in exploration is determined by the probability of the path being discovered, which is further divided into two complementary modes to enhance path discovery. Initially, this is based on a probability of 0.3. The probability is obtained using a softmax weighted sampling strategy: , Among them temperature parameters Adjust the response to Q-value differences, prioritizing high-value actions while maintaining random variability. Secondly, use 0.7... The probability is uniformly sampled throughout the action space to ensure global exploration, thus avoiding premature convergence to a local optimum. When no exploration is performed (probability 1- SORS selects the action that maximizes the current Q-value estimate. To implement a greedy strategy.

[0073] Exploration rate Dynamic decay based on an exponential timeline throughout the training process:

[0074] This facilitates the active exploration of unknown paths in the early stages, while gradually transitioning to utilizing identified high-value trajectories as training progresses. This decay mechanism is consistent with the two exploration modes: initial uniform sampling broadly covers the action space, while subsequent softmax sampling focuses on promising regions.

[0075] Temperature parameters Adaptability is further enhanced by adjusting action selection entropy. Lower A higher Q value reinforces the preference for actions with the highest Q value, while a higher Q value promotes exploration diversity. Empirical verification shows that... = 0.5 can best coordinate these objectives in the routing environment of SORS, because over-reliance on short-lived high-Q paths may overlook potential bottlenecks in dynamic networks.

[0076] By integrating the aforementioned components, AEPE enhances ROT's ability to formulate routing and scheduling strategies, enabling it to adapt to constantly changing network conditions and task scheduling. As an inherent component of ROT, AEPE synchronizes with DDQN's Q-value updates and reward-based feedback, dynamically modifying exploration strategies based on real-time resource conditions. This synergy ensures that SORS's routing and scheduling strategies gradually align with solutions that comprehensively improve path efficiency and computational load distribution while strictly adhering to the inherent service constraints within STags.

Claims

1. A semantic-aware routing scheduling method for computing power networks, characterized in that, include: (1) Introduce a service tag STag, which is decoupled from the physical host location and ownership, as an addressing and routing identifier in the computing power network, and build a routing scheduling mechanism that prioritizes open services; The service tag STag is independently encapsulated in the IPv6 extension header, and each field bridges the packet-level identifier with service semantics, including fields reserved for future scalability considerations; (2) Based on multi-layer graph attention network and deep reinforcement learning, a service-oriented routing scheduling (SORS) algorithm is designed. The algorithm jointly models the network resource status, computing load, and queue dynamics to achieve optimal selection of distributed multi-computing instances and collaborative decision-making of routing paths. The process includes: (2.1) A multi-layer graph attention neural network is used to dynamically capture the multi-dimensional resource interdependencies between different network nodes, including the network and computing resource status and the queuing dynamics of service scheduling, and uses it as the semantic abstraction engine of the routing and scheduling SORS algorithm to convert topology and scheduling data into hierarchical embeddings for joint routing and scheduling decisions. (2.2) Extracting topological relevance based on multi-layer graph attention neural network, including aggregating multi-head outputs by splicing or averaging, each layer contains residual connections, random deactivation and layer normalization to ensure stability and generalization ability, and finally deriving attention weights from the neighbor set to avoid explicit adjacency matrix; (2.3) The routing planner RPL is used as the decision core of the route scheduling SORS algorithm. It performs semantic-aware routing by jointly evaluating computing power, network and queuing conditions. The SLA configuration file guides the multi-objective optimization in the routing and scheduling process.

2. The semantic-aware routing scheduling method for computing power networks according to claim 1, characterized in that, The service tag STag is encapsulated in the form of an IPv6 extension header, with a total length of 128 bits, including: STag identifier bit 8 bits, service identifier type 2 bits, control flag bit 6 bits, first reserved bit 16 bits, STag value 32 bits, traffic indication 32 bits, and second reserved bit 32 bits.

3. The semantic-aware routing scheduling method for computing power networks according to claim 1, characterized in that, The specific routing and scheduling mechanism that prioritizes availability for open services is as follows: The control plane is based on the STag management service SLA configuration file and computing power-network resource scheduling strategy; The data plane uses STag to perform packet forwarding and service instance matching; By decoupling service logic from physical host location, service continuity can be maintained during resource elasticity adjustments, including computing instance migration, dynamic creation, or release.

4. The semantic-aware routing scheduling method for computing power networks according to claim 1, characterized in that, The SORS (Scheduled Routing) algorithm comprises three steps: network resource state modeling, route planning, and route optimization training. The multi-layer graph attention network MMGAT is used to extract features and model topological dependencies of node computing power utilization, bandwidth utilization, latency factor, load balancing, arrival queue length, computing queue length, and service device type vector. Computational power-aware paths are generated based on Dueling DDQN, network-aware paths are generated based on a latency-bandwidth-queue hybrid cost function, and the optimal node-path pairs are output through a multi-objective scoring function. The routing strategy is iteratively optimized based on the Dual-Depth Q Network (DDQN) and the Adaptive Exploration Policy (AEPE).

5. The semantic-aware routing scheduling method for computing power networks according to claim 4, characterized in that, The multi-objective scoring function comprehensively weights network utilization, computing power utilization, latency, normalized arrival queue, normalized computation queue, and reinforcement learning Q-value.

6. The semantic-aware routing scheduling method for computing power networks according to claim 4, characterized in that, The reward function of the SORS routing algorithm is composed of a weighted average of path efficiency, resource utilization, routing failure penalty, and queue delay penalty.

7. The semantic-aware routing scheduling method for computing power networks according to claim 4, characterized in that, The adaptive exploration strategy AEPE employs an exponentially decaying ε-greedy mechanism, combining softmax sampling with temperature parameter adjustment and uniform sampling to balance exploration and utilization.

8. A semantic-aware routing and scheduling system for computing power networks, characterized in that, The system is used to perform the method as described in any one of claims 1-7, wherein the semantically aware routing scheduling system comprises: The STag encapsulation module is used to embed the service tag STag into the IPv6 extension header to generate semantically aware packets; The OSAF control module is used for service registration, requirement parsing, STag mapping, and generation and distribution of computing power network joint routing scheduling policies. The SORS execution module is used for network resource state modeling, route planning, reinforcement learning training, and multi-computing instance optimization decision-making. The data forwarding module is used to perform packet forwarding and service connection establishment based on STag and routing decisions.

9. The semantic-aware routing and scheduling system for computing power networks according to claim 8, characterized in that, The SORS execution module includes: a resource feature extraction unit, a multi-layer graph attention network unit, a duel dual-depth Q network unit, and an adaptive exploration training unit.

Images