Knowledge-aware routing method in semantic edge network based on Lyapunov optimization

Abstract

The invention provides a knowledge perception routing method in a semantic edge network based on Lyapunov optimization. The method comprises the following steps: constructing a virtual queue for tracking accumulated deviation of knowledge migration flow and budget in a semantic edge network, setting a shadow price corresponding to the virtual queue, and calculating a Lagrange dominant function by using the shadow price at the current moment to obtain a comprehensive dominant function; the comprehensive path weight of each adjacent link between the starting node and each neighbor node is converted into a probability distribution value of each adjacent link, a deterministic forwarding strategy is generated, a route matching rule is generated according to the forwarding strategy, the route matching rule is executed through a data forwarding plane, and a route matching result is obtained; and optimizing and updating the forwarding strategy according to the comprehensive dominant function. According to the method, the long-term average knowledge migration bandwidth is strictly ensured not to exceed the preset budget mathematically through the virtual queue tracking mechanism, and the problem of network congestion caused by excessive consumption of the knowledge migration bandwidth is effectively avoided.

Description

A knowledge-aware routing method for semantic edge networks based on Lyapunov optimization Technical Field

[0001] This invention relates to the field of semantic communication technology, and in particular to a knowledge-aware routing method in semantic edge networks based on Lyapunov optimization. Background Technology

[0002] Semantic communication, as a novel communication method that transcends the traditional bit-by-bit transmission paradigm, is gradually becoming one of the core technologies of future intelligent networks. Semantic communication extracts semantic features from data through deep learning models for transmission, rather than traditional bit-by-bit transmission. This significantly reduces the amount of data transmitted while ensuring task completion quality, thereby significantly improving spectral and energy efficiency. However, in real-world edge network environments, semantic communication services face several major challenges. First, there is knowledge base dependency. The encoding and decoding process of semantic communication relies on a shared knowledge base (usually a pre-trained deep learning model) between the sender and receiver. These knowledge base files are enormous (typically hundreds of megabytes to gigabytes), while the semantically compressed data stream is only a few hundred kilobytes, a difference of 2-3 orders of magnitude. When a task is routed to an edge node lacking the required knowledge base, knowledge transfer must be performed first. The backhaul traffic generated in this process far exceeds the semantic data itself, severely offsetting the bandwidth gain brought by semantic compression. Second, there is the heterogeneity of knowledge distribution. Due to storage capacity limitations and differences in task history, the computing nodes in edge networks have different combinations of cached knowledge bases, resulting in a highly heterogeneous knowledge distribution. Traditional network routing protocols select paths solely based on network topology and link status, completely ignoring the knowledge base distribution information of nodes, leading to a severe disconnect between routing decisions and knowledge requirements. Secondly, there are dynamic task requirements. Different semantic communication tasks (such as image recognition, natural language processing, and video analysis) rely on different combinations of knowledge bases, and task arrival is random and sporadic, making knowledge requirements highly dynamic in both time and space. This dynamism requires routing strategies to perceive the knowledge distribution status in real time and make rapid decisions. Finally, there are bandwidth constraints and stability guarantees. The backhaul link bandwidth of edge networks is usually limited and needs to be shared among multiple services; frequent knowledge migration can lead to bandwidth exhaustion and network congestion. Therefore, how to achieve knowledge-aware intelligent routing while meeting strict bandwidth budget constraints has become a core technical challenge for semantic communication networks.

[0003] One existing semantic-aware routing method based on deep reinforcement learning involves optimizing packet routing paths through online learning and adjusting network resource allocation in real time to meet the latency and bandwidth requirements of semantic communication tasks. However, it still faces problems such as training instability and insufficient constraint guarantee capabilities. This method handles bandwidth constraints through soft reward shaping, providing only probabilistic constraint satisfaction and failing to mathematically guarantee that long-term average bandwidth consumption will not exceed a preset budget. It is prone to failure when faced with sudden traffic spikes or network fluctuations. Furthermore, this method does not consider the knowledge base distribution of edge nodes and lacks explicit modeling of knowledge transfer costs in routing decisions, potentially leading to a large amount of unnecessary knowledge transfer traffic in heterogeneous edge network environments.

[0004] The disadvantages of the semantic-aware routing method based on deep reinforcement learning in the above-mentioned prior art include: (1) Insufficient constraint guarantee capability. The soft reward shaping method is used to handle bandwidth constraints, which can only provide probabilistic constraint satisfaction. It cannot mathematically guarantee that the long-term average bandwidth consumption does not exceed the preset budget. When facing sudden traffic or network fluctuations, constraint violations are likely to occur, leading to system instability; (2) Lack of knowledge perception capability. The routing decision process does not consider the knowledge base distribution status of edge nodes and lacks explicit modeling of knowledge transfer costs. In the knowledge heterogeneous edge network environment, a large amount of unnecessary knowledge transfer traffic may be generated, which seriously offsets the bandwidth gain brought by semantic compression; (3) Slow training convergence speed. Deep reinforcement learning algorithms require a large amount of exploration and trial and error process to converge to a better strategy, which is difficult to adapt quickly in the dynamically changing network environment.

[0005] Another existing network routing optimization method based on graph neural networks includes: using graph neural networks to aggregate network topology information and link state features through message passing mechanism, learning end-to-end routing strategies, and exhibiting good performance under complex network topologies.

[0006] The disadvantages of another network routing optimization method based on graph neural networks in the above-mentioned prior art include: (1) Incomplete state representation, only considering network attributes of the physical layer (such as bandwidth, latency, packet loss rate, etc.), without involving the semantic layer features of the node knowledge base in the semantic communication scenario, and unable to perceive the impact of knowledge distribution state on routing decisions; (2) Lack of constraint supervision mechanism, without introducing constraint supervision mechanism to handle hard constraints such as bandwidth budget, which may lead to constraint violation and system instability in resource-constrained edge network environments; (3) Separation of routing and knowledge distribution optimization, treating data packet routing and knowledge distribution as independent problems to be optimized separately, unable to achieve joint optimization of routing decisions and knowledge distribution, and difficult to achieve global optimum in dynamic edge networks. Summary of the Invention

[0007] The embodiments of the present invention provide a knowledge-aware routing method in semantic edge networks based on Lyapunov optimization, so as to effectively avoid network congestion caused by excessive bandwidth consumption during knowledge migration.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] A knowledge-aware routing method for semantic edge networks based on Lyapunov optimization includes:

[0010] The semantic edge network is abstracted as a directed graph, and the physical state of the links and the knowledge state matrix of each node in the semantic edge network are obtained.

[0011] Based on the physical state of the link, obtain the physical state vector of the node and its associated link, and then concatenate and fuse the physical state vector of the node and its associated link with the knowledge state matrix of the node to generate the initial comprehensive feature representation of the node.

[0012] A virtual queue is constructed in the semantic edge network to track the cumulative deviation between knowledge transfer traffic and budget. A shadow price is set for the virtual queue, which is a dynamic penalty factor for the backlog of the virtual queue.

[0013] A dual-commentator architecture is used to train the graph neural network, and the Lagrange dominance function is calculated using the shadow price at the current time to obtain the comprehensive dominance function;

[0014] The physical score of the link between the starting node and its neighboring nodes is calculated based on the physical state information contained in the comprehensive features of the starting node and its candidate next-hop neighboring nodes. The knowledge base hash signature required for the current semantic communication task request is compared with the knowledge state matrix of the target node. The knowledge recovery cost is obtained based on the comparison result. The comprehensive path weight of each adjacent link between the starting node and each neighboring node is calculated based on the physical score, the knowledge recovery cost, and the shadow price.

[0015] The comprehensive path weight of each adjacent link is transformed into the probability distribution value of each adjacent link. Action sampling is performed based on the probability distribution value of the adjacent links and the comprehensive advantage function to generate a deterministic forwarding policy. Routing matching rules are generated based on the forwarding policy. The routing matching rules are executed through the data forwarding plane. The forwarding policy is optimized and updated based on the comprehensive advantage function.

[0016] Preferably, the step of abstracting the semantic edge network into a directed graph and obtaining the physical state of the links and the knowledge state matrix of each node in the semantic edge network includes:

[0017] Abstracting the semantic edge network into a directed graph. , Represents a set of links. Representing a set of nodes, the state-aware module uses telemetry technology to acquire the physical state of links in the semantic edge network. This physical state This includes bandwidth, latency, queue depth, and the knowledge state matrix of each edge node. , elements in Represents a node Knowledge components have been cached. .

[0018] Preferably, the step of obtaining the physical state vectors of the node and its associated links based on the physical state of the link, and concatenating and fusing the physical state vectors of the node and its associated links with the knowledge state matrix of the node to generate an initial comprehensive feature representation of the node includes:

[0019] The routing decision module receives the initial comprehensive feature representation of the node from the physical state and knowledge state:

[0020]

[0021] in, Represents a node The initial comprehensive feature vector representation that integrates physical state and knowledge semantic distribution; This represents a multilayer perceptron network used to process physical layer state features; For nodes The physical state vector of its associated links; For nodes The knowledge state matrix represents the current cached semantic knowledge base components of the node;

[0022] Continuous physical characteristics With discrete knowledge semantic embedding The nodes are spliced ​​and fused to generate an initial comprehensive feature representation. .

[0023] Preferably, the virtual queue used to track the cumulative deviation between knowledge transfer traffic and budget in the constructed semantic edge network is configured with a shadow price, which is a dynamic penalty factor for the backlog of the virtual queue, including:

[0024] A virtual queue is maintained through the constraint and supervision module. Virtual queue The elements in the table represent the cutoff time step. The system accumulated bandwidth exceeding the preset budget during the knowledge transfer process. The difference;

[0025] Based on the actual migration traffic fed back from the data plane in the previous cycle Update virtual queue :

[0026]

[0027] in, In time step Within the network, the total size of knowledge migration data streams actually triggered due to nodes lacking knowledge bases, i.e., the actual bandwidth resources consumed; A strict upper limit for the knowledge transfer bandwidth budget set for each time period of the system;

[0028] Calculate shadow prices , Based on the current virtual queue backlog The mapping generates a dynamic Lagrange multiplier, and this shadow price signal will be passed as a penalty factor to subsequent routing algorithms.

[0029] Preferably, the step of calculating the link physical score between the starting node and its neighboring nodes based on the physical state information contained in the comprehensive features of the starting node and the candidate next-hop neighboring nodes, comparing the knowledge base hash signature required for the current semantic communication task request with the knowledge state matrix of the target node, obtaining the knowledge recovery cost based on the comparison result, and calculating the comprehensive path weight of each adjacent link between the starting node and each neighboring node based on the link physical score, the knowledge recovery cost, and the shadow price includes:

[0030] Constructing the comprehensive path weights The calculation function:

[0031]

[0032] The link physical score is achieved by leveraging the node's learnable attention kernel. and nodes Comprehensive characteristics and The physical state information contained therein is used for calculation; The cost of knowledge recovery represents the node. Missing model data;

[0033] The specific update process of the comprehensive features of a node is as follows: For the first node of a graph neural network... Layers, nodes First, collect and aggregate its one-hop neighbor node set. The feature information of all neighboring nodes is aggregated using a specific aggregation function. Layer features The nodes are fused into neighbor context feature vectors. Its own first Layer features The feature vector is concatenated with the context feature vector of the neighbor and then passed through a non-linear activation function. Learnable weight matrix Performing linear transformations and nonlinear mappings, we obtain the first... Layer feature representation :

[0034]

[0035] in, AGG represents the vector concatenation operation. ) represents the feature aggregation function, after which... After forward propagation and multi-hop message passing in the graph convolutional network, each node learns the network's physical topology and knowledge distribution within its surrounding multi-hop range based on its features. The network outputs the first... Layer features are nodes The final comprehensive characteristic is denoted as .

[0036] Preferably, the method of training the graph neural network using a dual-commentator architecture and calculating the Lagrange dominance function using the shadow price at the current moment to obtain the comprehensive dominance function includes:

[0037] Construct rewarded commentator networks separately and Cost Commentator Network ,in Used to estimate long-term semantic utility Used to estimate long-term migration costs, combined with the shadow price at the current moment. Calculate the Lagrange dominance function and synthesize the composite dominance function. :

[0038]

[0039] in and These represent the semantic reward advantage and migration cost calculated using generalized advantage estimation, respectively.

[0040] Preferably, the step of converting the comprehensive path weight of each adjacent link into the probability distribution value of each adjacent link, sampling actions based on the probability distribution value of the adjacent links and the comprehensive advantage function, generating a deterministic forwarding policy, generating routing matching rules based on the forwarding policy, and executing the routing matching rules through the data forwarding plane includes:

[0041] For nodes All candidate next-hop neighbor nodes The comprehensive path weight of each adjacent link is calculated. Input into the Softmax normalization function:

[0042]

[0043] Indicates that the data packet comes from the node Forward to neighboring nodes The probability; determined by the SDN controller based on the probability distribution. Action sampling is performed, and a deterministic forwarding strategy is generated by selecting the optimal path with the highest probability or by performing multi-path probability load balancing.

[0044] The data forwarding module executes routing matching rules based on the programmable switch pipeline, parses and extracts semantic metadata from the routing matching rules, then performs affinity checks, queries the local exact match table, and if the result is Hit, it directly forwards the data packet to the calculation queue; if the result is Miss, it marks the data packet as a migration flow, triggers the hardware counter to increment, and finally provides feedback by reporting the migration count value within the period to the constraint supervision module for the next round of virtual queue updates, forming a closed-loop control.

[0045] Preferably, the optimization and updating of the forwarding strategy based on the comprehensive advantage function includes:

[0046] The comprehensive advantage function As a guiding direction for policy gradient descent, when When the value is 0, it indicates that the current routing action is better than the average expectation in terms of balancing semantic reward and bandwidth constraints. The system increases the probability of this action being selected in the future through gradient updates; conversely, when the value is 0, it indicates that the current routing action is better than the average expectation. When the value is 0, it indicates that the current routing action is below average expectations in terms of balancing semantic rewards and bandwidth constraints. The system reduces the probability of this action being selected in the future through gradient updates.

[0047] As can be seen from the technical solutions provided by the embodiments of the present invention described above, the present invention constructs a constraint supervision framework based on Lyapunov optimization, mathematically guarantees that the long-term average knowledge migration bandwidth does not exceed the preset budget through a virtual queue tracking mechanism, and dynamically adjusts the weight of bandwidth cost in routing decisions through shadow pricing. Compared with the soft constraint methods in the prior art, the present invention can provide deterministic bandwidth constraint guarantees, effectively avoiding network congestion problems caused by excessive knowledge migration bandwidth consumption.

[0048] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 is a schematic diagram of a semantic communication knowledge-aware routing architecture for edge network environments provided by an embodiment of the present invention;

[0051] Figure 2 is a flowchart of a knowledge-aware routing method based on Lyapunov optimization provided in an embodiment of the present invention;

[0052] Figure 3 is a logic diagram of a routing decision module provided in an embodiment of the present invention;

[0053] Figure 4 is a schematic diagram of a data plane pipeline provided by an embodiment of the present invention. Detailed Implementation

[0054] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0055] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0056] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0057] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0058] Lyapunov optimization is a stochastic optimization theory that optimizes routing performance by constructing virtual queues to track constraint deviations and using Lyapunov drift plus penalty techniques to mathematically guarantee long-term bandwidth constraints.

[0059] This invention provides a knowledge-aware routing method based on Lyapunov optimization. By introducing a virtual queue tracking mechanism and a Lyapunov constraint supervision framework, it mathematically guarantees that the long-term average knowledge migration bandwidth does not exceed a preset budget, providing deterministic system stability assurance. A state-aware module acquires the node knowledge base distribution status in real time, and introduces knowledge recovery cost as the core optimization objective in routing decisions, effectively reducing unnecessary knowledge migration traffic. A Lagrange modulation attention mechanism couples shadow prices and knowledge recovery costs to the path selection weight calculation of the graph neural network, achieving joint optimization of routing decisions and knowledge distribution. A dual-commentator reinforcement learning architecture estimates immediate rewards and long-term cumulative rewards respectively, enhancing the algorithm's training stability and convergence speed. Thus, under strict bandwidth constraints, it maximizes the inference accuracy and task completion rate of semantic communication services while reducing task execution latency.

[0060] This invention first constructs a semantic communication knowledge-aware routing architecture for edge network environments, aiming to achieve efficient semantic communication services under heterogeneous knowledge base distribution conditions. Second, it proposes a constraint supervision framework based on Lyapunov optimization, in which the system can strictly guarantee that the long-term average knowledge migration bandwidth does not exceed the preset budget through a virtual queue tracking mechanism. Finally, it designs a knowledge-aware routing decision mechanism based on graph neural networks, which couples the shadow price and knowledge recovery cost to the path selection weight calculation through a Lagrange modulation attention mechanism, thereby improving the system's routing decision-making ability and resource utilization efficiency in dynamic edge network environments.

[0061] Figure 1 shows a schematic diagram of a semantic communication knowledge-aware routing architecture for edge network environments provided by an embodiment of the present invention. This architecture adopts a three-layer edge semantic communication routing framework, mainly composed of a semantic communication terminal device (as a task source), edge computing nodes, a Software Defined Network (SDN) controller, and a knowledge base management module.

[0062] Based on the system architecture shown in Figure 1, the processing flow of a knowledge-aware routing method based on Lyapunov optimization provided in this embodiment of the invention is shown in Figure 2, including the following processing steps:

[0063] Step S10: System modeling and state acquisition.

[0064] Abstracting the semantic edge network into a directed graph. , Represents a set of links. This represents a set of nodes. The state-aware module uses telemetry technology to acquire the physical state of links in the semantic edge network. (Bandwidth, latency, queue depth) and the knowledge state matrix of each edge node .in, elements in Represents a node Knowledge components have been cached. .

[0065] Step S20: Dual-channel feature embedding. Figure 3 is a logic diagram of a routing decision module provided in an embodiment of the present invention. As shown on the left side of Figure 3, the routing decision module receives physical state and knowledge state. To solve the problem of heterogeneous feature fusion, a dual-channel embedding layer is adopted:

[0066]

[0067] in, Represents a node The initial comprehensive feature vector representation that integrates physical state and knowledge semantic distribution; This represents a multilayer perceptron network used to process physical layer state features; For nodes The physical state vectors of its associated links (such as available bandwidth, transmission delay, queue depth, etc.); For nodes The knowledge state matrix represents the current cached semantic knowledge base components of the node.

[0068] Continuous physical characteristics With discrete knowledge semantic embedding The nodes are spliced ​​and fused to generate an initial comprehensive feature representation. .

[0069] The acquisition process can be achieved by viewing the entire edge network topology as a graph, with each node's knowledge state matrix... As initial features of the graph nodes, nodes are aggregated through the message passing mechanism of the graph neural network. The knowledge cache state of its multi-hop neighbor nodes. The initial comprehensive feature representation of the node for subsequent use. This will serve as the core input feature for the "Lagrange Modulation Attention Calculation" in step S40. It is used to calculate the physical matching score between adjacent nodes and the final integrated path weight.

[0070] Step S30: Constraint supervision and shadow price calculation.

[0071] Virtual queues and shadow prices: Based on the Lyapunov optimization principle, virtual queues are used to track the cumulative deviation between knowledge migration traffic and budget; shadow prices are a dynamic penalty factor calculated from queue backlog, which increases when migration exceeds the budget, and is used to penalize high-cost paths in routing decisions.

[0072] To address the hard constraint on knowledge transfer bandwidth, the constraint supervision module maintains a virtual queue. Elements in the virtual queue (i.e., queue length) () indicates the time step up The system accumulated bandwidth exceeding the preset budget during the knowledge transfer process. The difference.

[0073] This module is based on the actual migration traffic fed back from the data plane in the previous cycle. Update virtual queue :

[0074]

[0075] in, In time step Within the network, the total size of knowledge migration data streams actually triggered due to nodes lacking knowledge bases, i.e., the actual bandwidth resources consumed; This is a strict upper limit on the knowledge migration bandwidth budget set for the system for each time period.

[0076] Then, calculate the shadow price. .in, Based on the current virtual queue backlog A dynamic Lagrange multiplier generated by the mapping.

[0077] Shadow price This means that when a large number of sudden migrations occur in the network, causing bandwidth to exceed the limit, virtual queue backlog increases, and shadow pricing increases. The price then increases. This shadow price signal will be passed as a penalty factor to subsequent routing algorithms.

[0078] Step S40: Lagrange modulation attention calculation. The routing decision module uses a graph neural network to calculate node attention. Route the task to the neighbor node The present invention constructs the following comprehensive path weights, which are based on the tendency (attention weights) of the path. The function for calculating the attention coefficient is as follows:

[0079]

[0080] The link physical score can be obtained through a learnable attention kernel with shared weights (e.g., via the query vector and key vector). (dot product), using nodes and nodes Comprehensive characteristics and It is calculated from the physical state information contained therein.

[0081] The specific update process of the comprehensive features of a node is as follows: For the first node of a graph neural network... Layers, nodes First, collect and aggregate its one-hop neighbor node set. The characteristic information of all neighboring nodes is obtained through a specific aggregation function. Layer features The features are then merged into a neighbor context feature vector. Subsequently, the nodes... Its own first Layer features The feature vector is concatenated with the context feature vector of the neighbor and then passed through a non-linear activation function. Learnable weight matrix Perform linear transformations and nonlinear mappings to obtain the first... Layer feature representation The feature update formula is as follows:

[0082]

[0083] in, AGG represents the vector concatenation operation. ) represents the feature aggregation function (such as mean aggregation, max pooling, or graph attention aggregation). After After forward propagation and multi-hop message passing in a graph convolutional network, each node's features can learn the network's physical topology and knowledge distribution within its surrounding multi-hop range. The final network output is the [missing information - likely a specific type of output]. Layer features are nodes The final comprehensive characteristic is denoted as .

[0084] This score reflects the link Suitability of conventional physical transmission in terms of latency and bandwidth. To reduce the cost of knowledge recovery, the knowledge base hash signature required for the current semantic communication task can be used in conjunction with the node. Cached knowledge state matrix Perform a comparison. Calculate the nodes. Total size of missing model data packets that must be retrieved via the backhaul link. Among them, For knowledge recovery costs (i.e., nodes) (Missing model data). The adaptive adjustment mechanism works as follows: when... When resources are plentiful, the weighting primarily depends on the physics score. The algorithm degenerates into a load balancing strategy; when When resources are scarce, penalties apply. Dominant. For any missing knowledge ( The weight of nodes will be significantly reduced.

[0085] Step S50, Original-Dual Strategy Optimization and Model Update: To ensure the parameters of the graph neural network converge to the optimal solution, this invention employs a dual-commentator architecture for online or offline training. A dual-commentator network is constructed, and reward-commentator networks are built separately. and Cost Commentator Network ,in Used to estimate long-term semantic utility Used to estimate long-term migration costs. Combined with the shadow price at the current moment. Calculate the Lagrange dominance function and synthesize the composite dominance function. :

[0086]

[0087] in and These represent the semantic reward advantage and migration cost calculated using generalized advantage estimation, respectively.

[0088] The expected cumulative reward for the commentator network to evaluate the current state is calculated using the generalized advantage estimation method. The timing difference error is calculated by combining the immediate environmental feedback after actual routing execution (such as improved inference accuracy and low latency), and then accumulated to obtain... Similarly, cost commentators' networks assess expected cumulative costs. The actual knowledge transfer flow was calculated to obtain .

[0089] The above comprehensive advantage function Primarily used in the policy network parameter update process, serving as a guiding direction for policy gradient descent: when When the value is 0, it indicates that the current routing action is better than the average expectation in balancing semantic reward and bandwidth constraints, and the system will increase the probability of this action being selected in the future through gradient updates; conversely, when the value is 0, it indicates that the current routing action is better than the average expectation in balancing semantic reward and bandwidth constraints. If the value is 0, the probability is reduced.

[0090] Step S60: Policy Generation and Distribution. Based on the attention weights calculated in step S40, the routing probability distribution is obtained through a normalization function. The controller generates specific routing matching rules based on this probability and sends them to the data forwarding module.

[0091] For nodes All candidate next-hop neighbor nodes The comprehensive path weight of each adjacent link is calculated. Enter to In the normalization function:

[0092]

[0093] This transforms the attention weights into a probability distribution that sums to 1. This indicates that the data packet originated from the node. Forward to neighboring nodes The SDN controller samples actions based on the above probability distribution (such as selecting the optimal path with the highest probability, or performing multi-path probabilistic load balancing) to generate a deterministic forwarding strategy. The above comprehensive advantage function... Primarily used in the policy network parameter update process, serving as a guiding direction for policy gradient descent: when When the value is greater than 0, it indicates that the current routing action is better than the average expectation in terms of balancing semantic reward and bandwidth constraints, and the system will increase the probability of this action being selected in the future through gradient updates; conversely, when the value is greater than 0, it indicates that the current routing action is better than the average expectation in terms of balancing semantic reward and bandwidth constraints. When the value is 0, it indicates that the current routing action is below average expectations in terms of balancing semantic rewards and bandwidth constraints. The system will reduce the probability of this action being selected in the future through gradient updates.

[0094] Since the above probability distribution is derived from step S50 The continuously optimized graph neural network outputs, combined with dynamic shadow pricing for a joint evaluation of semantic reward advantages and knowledge transfer costs, enable a deterministic forwarding strategy generated based on this distribution sampling. This strategy can adaptively select the routing path that maximizes global semantic utility while strictly ensuring the knowledge transfer bandwidth budget. Subsequently, the routing path is converted into routing matching rules compatible with protocols such as OpenFlow or P4 and distributed. These rules contain two parts: 1) a matching field, including source / destination IP addresses, semantic task identifiers, and knowledge base hash values; and 2) an action field, including the next-hop output port, MAC address modification, whether to trigger the "fetch missing knowledge" flag, and queuing priority.

[0095] Step S70: Data plane execution. Figure 4 is a schematic diagram of a data plane pipeline provided in an embodiment of the present invention. As shown in Figure 4, the data forwarding module executes routing matching rules based on the programmable switch pipeline. First, it parses and extracts semantic metadata (such as knowledge hash and semantic importance) from the routing matching rules, then performs an affinity check and queries the local exact matching table. If the result is Hit, it is directly forwarded to the calculation queue. If the result is Miss, the data packet is marked as a migration flow, triggering the hardware counter to increment automatically. Finally, feedback is provided, and the migration count value within the period is reported to the constraint supervision module for the virtual queue update in the next round of step S20, forming a closed-loop control.

[0096] The Lyapunov optimization process mainly involves the coordinated actions of steps S20, S40, and S50. Step S20 forms the basis for virtual queue tracking and the generation of the penalty term (shadow price) in Lyapunov optimization; step S40 introduces this penalty term into the drift plus penalty instantaneous decision-making of action execution; and step S50 uses this penalty term to guide the long-term policy gradient update of the reinforcement learning network. Together, these three steps complete the implementation of Lyapunov optimization theory in a deep reinforcement learning architecture.

[0097] Example of effect description:

[0098] Assuming the network is under high load ( (Higher latency). Path A: Low physical latency, but the target node lacks a knowledge base (50MB migration required). Path B: Slightly higher physical latency, but the target node has a cached knowledge base (0MB migration required).

[0099] According to step S40 of the present invention, path A is due to =50 and If the score is high, its overall weight will be severely penalized. The knowledge transfer / recovery cost of this path refers to the cost of knowledge transfer required to enable the relevant nodes along this path to achieve usable semantic capabilities after selecting this path. Path B, however, is... =0, the penalty term is 0, and the final weight is higher. The system will automatically select path B, avoiding the congestion that path A may cause. This is consistent with the logic of "dynamically adjusting traffic allocation according to weight" described in the reference document, but this invention introduces specific shadow prices and knowledge costs as core variables for weight.

[0100] In summary, the embodiments of the present invention provide a unified model for semantic communication services in edge network scenarios, incorporating the knowledge base distribution status, network link resources, and edge computing resources into the same routing decision process to achieve collaborative optimization of knowledge transfer and task execution. Compared with existing technologies that mainly focus on the inference optimization of a single node or simple load balancing, the present invention is more suitable for edge semantic communication environments with heterogeneous knowledge base distribution, limited resources, and highly dynamic states.

[0101] This invention introduces graph neural networks into the deep reinforcement learning routing decision-making process, utilizes multi-layer graph convolution operations to aggregate network topology information and knowledge base distribution information of multi-hop neighbors, and couples bandwidth constraints and knowledge recovery costs to path selection through a Lagrange modulation attention mechanism. Unlike existing routing methods that rely on heuristic rules or ignore the knowledge distribution state, this invention can directly perceive the global knowledge distribution state and make intelligent routing decisions, making it more suitable for large-scale edge semantic communication scenarios with dynamically updated knowledge bases.

[0102] This invention optimizes task completion time and inference accuracy while satisfying bandwidth constraints through a multi-objective joint reward design and a dual-commenter network training mechanism, and ensures policy convergence through a stable training mechanism. Therefore, in edge semantic communication service routing scenarios with heterogeneous knowledge base distribution and dynamic network state changes, this invention outperforms existing technologies in terms of bandwidth utilization efficiency, task completion latency, and system stability.

[0103] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0104] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0105] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0106] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A knowledge-aware routing method for semantic edge networks based on Lyapunov optimization, characterized in that, include: The semantic edge network is abstracted as a directed graph. The physical state of the links in the semantic edge network and the knowledge state matrix of each node are obtained. The physical state vector of the node and its associated links is obtained according to the physical state of the links. The physical state vector of the node and its associated links is concatenated and fused with the knowledge state matrix of the node to generate the initial comprehensive feature representation of the node. A virtual queue is constructed in the semantic edge network to track the cumulative deviation between knowledge migration traffic and budget. A shadow price is set for the virtual queue, which is a dynamic penalty factor for the backlog of the virtual queue. A graph neural network is trained using a dual-commentator architecture. The Lagrange dominance function is calculated using the shadow price at the current moment to obtain the comprehensive dominance function. The physical score of the link between the starting node and the neighboring nodes is calculated based on the physical state information contained in the comprehensive features of the starting node and the candidate next-hop neighboring nodes. The knowledge base hash signature required for the current semantic communication task request is compared with the knowledge state matrix of the target node. The knowledge recovery cost is obtained based on the comparison result. The comprehensive path weight of each adjacent link between the starting node and each neighboring node is calculated based on the physical score of the link, the knowledge recovery cost, and the shadow price. The comprehensive path weight of each adjacent link is transformed into the probability distribution value of each adjacent link. Action sampling is performed based on the probability distribution value of the adjacent link and the comprehensive dominance function to generate a deterministic forwarding strategy. Routing matching rules are generated based on the forwarding strategy. The routing matching rules are executed through the data forwarding plane. The forwarding strategy is optimized and updated based on the comprehensive dominance function.

2. The method according to claim 1, characterized in that, The process of abstracting the semantic edge network into a directed graph and obtaining the physical state of the links and the knowledge state matrix of each node in the semantic edge network includes: abstracting the semantic edge network into a directed graph. ， Represents a set of links. Representing a set of nodes, the state-aware module uses telemetry technology to acquire the physical state of links in the semantic edge network. This physical state This includes bandwidth, latency, queue depth, and the knowledge state matrix of each edge node. ， elements in Represents a node Knowledge components have been cached. 。 3. The method according to claim 2, characterized in that, The step of obtaining the physical state vectors of a node and its associated links based on the physical state of the link, and then concatenating and fusing the physical state vectors of the node and its associated links with the knowledge state matrix of the node to generate an initial comprehensive feature representation of the node, includes: the routing decision module receiving the physical state and knowledge state to generate the initial comprehensive feature representation of the node. in, Represents a node The initial comprehensive feature vector representation that integrates physical state and knowledge semantic distribution; This represents a multilayer perceptron network used to process physical layer state features; For nodes The physical state vector of its associated links; For nodes The knowledge state matrix represents the current cached semantic knowledge base components of the node; it also represents the continuous physical features. With discrete knowledge semantic embedding The nodes are spliced ​​and fused to generate an initial comprehensive feature representation. 。 4. The method according to claim 2, characterized in that, The construction of a virtual queue in the semantic edge network for tracking the cumulative deviation between knowledge transfer traffic and budget, and the setting of a shadow price corresponding to the virtual queue, wherein the shadow price is a dynamic penalty factor for the backlog of the virtual queue, includes: maintaining a virtual queue through a constraint supervision module. Virtual queue The elements in the table represent the cutoff time step. The system accumulated bandwidth exceeding the preset budget during the knowledge transfer process. The difference; based on the actual migration traffic fed back from the data plane in the previous cycle. Update virtual queue ： in, In time step Within the network, the total size of knowledge migration data streams actually triggered due to nodes lacking knowledge bases, i.e., the actual bandwidth resources consumed; A strict knowledge transfer bandwidth budget limit is set for the system for each time period; shadow prices are calculated. ， Based on the current virtual queue backlog The mapping generates a dynamic Lagrange multiplier, and this shadow price signal will be passed as a penalty factor to subsequent routing algorithms.

5. The method according to claim 4, characterized in that, The process of calculating the link physical score between the starting node and its neighboring nodes based on the physical state information contained in the comprehensive features of the starting node and its candidate next-hop neighboring nodes, comparing the knowledge base hash signature required for the current semantic communication task request with the knowledge state matrix of the target node, obtaining the knowledge recovery cost based on the comparison result, and calculating the comprehensive path weight of each adjacent link between the starting node and each neighboring node based on the link physical score, knowledge recovery cost, and shadow price includes: constructing the comprehensive path weight. The calculation function: The link physical score is achieved by leveraging the node's learnable attention kernel. and nodes Comprehensive characteristics and The physical state information contained therein is used for calculation; The cost of knowledge recovery represents the node. The amount of missing model data; the specific update process of the comprehensive features of the nodes is as follows: For the graph neural network's... Layers, nodes First, collect and aggregate its one-hop neighbor node set. The feature information of all neighboring nodes is aggregated using a specific aggregation function. Layer features The nodes are fused into neighbor context feature vectors. Its own first Layer features The feature vector is concatenated with the context feature vector of the neighbor and then passed through a non-linear activation function. Learnable weight matrix Performing linear transformations and nonlinear mappings, we obtain the first... Layer feature representation ： in, AGG represents the vector concatenation operation. ) represents the feature aggregation function, after which... After forward propagation and multi-hop message passing in the graph convolutional network, each node learns the network's physical topology and knowledge distribution within its surrounding multi-hop range based on its features. The network outputs the first... Layer features are nodes The final comprehensive characteristic is denoted as 。 6. The method according to claim 5, characterized in that, The method described above involves training a graph neural network using a dual-commentator architecture, calculating the Lagrange dominance function using the shadow price at the current time step, and obtaining the comprehensive dominance function. This includes: constructing reward-rewarding commentator networks separately. and Cost Commentator Network ,in Used to estimate long-term semantic utility Used to estimate long-term migration costs, combined with the shadow price at the current moment. Calculate the Lagrange dominance function and synthesize the composite dominance function. ： in and These represent the semantic reward advantage and migration cost calculated using generalized advantage estimation, respectively.

7. The method according to claim 6, characterized in that, The process of converting the comprehensive path weights of each adjacent link into probability distribution values ​​for each adjacent link, sampling actions based on the probability distribution values ​​and the comprehensive advantage function, generating a deterministic forwarding policy, generating routing matching rules based on the forwarding policy, and executing the routing matching rules through the data forwarding plane includes: for nodes... All candidate next-hop neighbor nodes The comprehensive path weight of each adjacent link is calculated. Input into the Softmax normalization function: Indicates that the data packet comes from the node Forward to neighboring nodes The probability; determined by the SDN controller based on the probability distribution. Action sampling is performed, and a deterministic forwarding strategy is generated by selecting the optimal path with the highest probability or performing multi-path probabilistic load balancing. The data forwarding module executes the routing matching rules based on the programmable switch pipeline, parses and extracts the semantic metadata in the routing matching rules, and then performs affinity checks, queries the local exact matching table. If the result is Hit, it is directly forwarded to the calculation queue; if the result is Miss, the data packet is marked as a migration flow, triggers the hardware counter to increment, and finally provides feedback, reporting the migration count value within the period to the constraint supervision module for the next round of virtual queue updates, forming a closed-loop control.

8. The method according to claim 7, characterized in that, The aforementioned optimization and update of the forwarding strategy based on the comprehensive advantage function includes: adjusting the comprehensive advantage function... As a guiding direction for policy gradient descent, when When the value is 0, it indicates that the current routing action is better than the average expectation in terms of balancing semantic reward and bandwidth constraints. The system increases the probability of this action being selected in the future through gradient updates; conversely, when the value is 0, it indicates that the current routing action is better than the average expectation. When the value is 0, it indicates that the current routing action is below average expectations in terms of balancing semantic rewards and bandwidth constraints. The system reduces the probability of this action being selected in the future through gradient updates.

Images