Large-scale optical network routing optimization method and system, electronic equipment and storage medium
By constructing a state-space matrix in a large-scale optical network and optimizing the node shrinkage strategy using a deep reinforcement learning model, the problems of computational time consumption and experience-dependent node sorting in traditional algorithms are solved, achieving efficient route preprocessing and query optimization.
Patent Information
- Application Number
- CN202511496865.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-20
- Publication Date
- 2026-01-23
AI Technical Summary
Traditional Dijkstra's algorithm is computationally time-consuming in large-scale optical networks, making it difficult to complete the calculation of new paths for all services before the next failure occurs. Furthermore, traditional node sorting methods rely on fixed weights, making it difficult to balance local and global performance, resulting in an imbalance between preprocessing performance and query results.
A state space matrix is constructed based on a deep reinforcement learning model. Local topological embeddings and global dependencies are extracted through GNN and Transformer layers. The node shrinkage probability distribution is output. The node with the highest shrinkage probability is selected for shrinkage operation. Shortcut edges are added between adjacent nodes to update the topology structure. Finally, the node shrinkage order sequence is output.
It significantly improves the efficiency of route preprocessing, reduces the density of the graph, shortens the query time, enhances the algorithm's generalization ability under various network structures and scales, and reduces reliance on human experience.
Smart Images
Figure CN121397394A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer technology, and in particular to a method, system, electronic device and storage medium for large-scale optical network routing optimization. Background Technology
[0002] With the booming development of emerging services such as 5G, IoT, and cloud computing, and the rapid rise of edge computing, higher demands are being placed on network bandwidth, latency, and connectivity. To meet users' needs for proximity access and low-latency service experiences, the deployment of Optical Transport Network (OTN) technology to the edge of metropolitan area networks has become an important direction for future network evolution. As OTN extends from the core layer to the access layer (such as base stations and edge data centers), the site size increases approximately tenfold with each layer. Coupled with the exponential growth of 5G base station and MEC (Mobile Edge Computing) deployments, the number of nodes is jumping from hundreds to thousands to support massive local traffic processing and wavelength-level scheduling. This trend leads to a dramatic expansion of network coverage and a significant increase in the number of nodes. Future network scale will reach the level of thousands of nodes, consequently increasing the risk of link failures and posing unprecedented challenges to network operation and maintenance management.
[0003] In large-scale optical networks, single-link failures can occur consecutively within a short period. When a single-link failure occurs, because backbone fiber links typically employ dense wavelength division multiplexing (DWDM) technology, a single fiber can carry hundreds of high-speed wavelength channels (e.g., 100G / 400G), each transmitting independent service flows. Therefore, new protection or recovery paths must be recalculated for thousands of service flows affected by this failure. The traditional Dijkstra algorithm, due to its non-directional search, is computationally time-consuming and struggles to complete new path calculations for all services before the next failure occurs. The Contraction Hierarchy (CH) algorithm significantly improves path calculation efficiency through preprocessing, and by modifying it, it can be adapted to the specific constraints of the optical network for fast routing calculations. However, node ordering, as a key preprocessing step in the CH algorithm, is crucial to its overall efficiency. Therefore, determining the optimal node order for a given network topology is a core problem that urgently needs to be solved to improve the efficiency of the CH algorithm. Summary of the Invention
[0004] In view of this, the purpose of this application is to propose a method, system, device and medium for large-scale optical network routing optimization to solve the above-mentioned technical problems.
[0005] To achieve the above objectives, the first aspect of this application provides a method for optimizing routing in large-scale optical networks, comprising: Based on the priority terms and topological characteristics of each node in the current optical network topology, a state space matrix representing the current optical network topology is constructed. The state space matrix is input into the pre-trained deep reinforcement learning model, and the shrinkage probability distribution of the remaining nodes in the current optical network topology is output. Select the node with the highest shrinkage probability as the current shrinkage target and perform node shrinkage operation. Remove the current shrinkage target and its adjacent edges, and add shortcut edges between the adjacent nodes of the shrinkage target. Update the optical network topology after node shrinkage; After all nodes have shrunk, output the complete node shrinkage sequence.
[0006] In one embodiment, the state space matrix has a dimension of N×F; where N represents the number of remaining nodes in the current optical network topology, F is the feature dimension of each node, and each row of the state space matrix corresponds to a feature vector of a node, which encapsulates the priority term and topological features of the corresponding node.
[0007] In one implementation, the step of inputting the state space matrix into a pre-trained deep reinforcement learning model and outputting the shrinkage probability distribution of the remaining nodes in the current optical network topology includes: The state space matrix is input into a pre-trained deep reinforcement learning model. Local topological embeddings are extracted from the state space matrix and edge indices through a GNN layer. Global dependencies between nodes are captured through a Transformer layer. The shrinkage probability distribution of the remaining nodes in the current optical network topology is output through a linear projection and a Softmax output layer.
[0008] In one implementation, before constructing the state space matrix characterizing the current optical network topology based on the priority terms and topological features of each node in the current optical network topology, the following steps are included: The deep reinforcement learning model is pre-trained using the following objective reward function to obtain a pre-trained deep reinforcement learning model: ; ; ; ; in, For the target reward function, The immediate reward is a weighted sum of priority items; This is the edge difference value. To reduce costs, To reduce the number of neighbors, As weight; This is a network state reward used to reflect the density of the network during the contraction process; It is a scaling factor. The congestion level of the current optical network topology is defined as the edge density of the current optical network topology.
[0009] In one implementation, after updating the shrunk optical network topology of the nodes; and before outputting the complete node shrinkage sequence after all nodes have completed shrinkage, the method further includes: If there are still un-shrunken nodes in the shrunken optical network topology, then continue to construct a state space matrix representing the shrunken optical network topology based on the priority terms and topological features of each node in the shrunken optical network topology; otherwise, determine that all nodes have completed the shrunken process.
[0010] In one embodiment, the topological attributes include one or more of the following: node number, degree centrality, connection density, clustering coefficient, betweenness centrality, proximity centrality, and PageRank score; The priority items include one or more of the following: edge difference value, shrinkage cost, number of neighboring nodes to be shrunk, and search space depth.
[0011] In one implementation, after all nodes have completed shrinking and the complete node shrinking sequence is output, the method further includes: The complete node shrinkage sequence is used for node preprocessing in the CH algorithm.
[0012] Based on the same inventive concept, a second aspect of this application provides a large-scale optical network routing optimization system, comprising: The state space matrix construction module is used to construct a state space matrix representing the current optical network topology based on the priority terms and topological characteristics of each node in the current optical network topology. The shrinkage probability distribution output module is used to input the state space matrix into the pre-trained deep reinforcement learning model and output the shrinkage probability distribution of the remaining nodes in the current optical network topology. The node shrinking module is used to select the node with the highest shrinking probability as the current shrinking target, perform node shrinking operation, remove the current shrinking target and its adjacent edges, and add shortcut edges between the adjacent nodes of the shrinking target. The update module is used to update the optical network topology after node shrinkage. The sequence output module is used to output the complete node shrinkage sequence after all nodes have finished shrinking.
[0013] Based on the same inventive concept, a third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the large-scale optical network routing optimization method described in the first aspect above.
[0014] Based on the same inventive concept, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the large-scale optical network routing optimization method described in the first aspect.
[0015] As described above, the large-scale optical network routing optimization method provided in this application constructs a state space matrix representing the current optical network topology based on the priority term and topological features of each node in the current optical network topology. The state space matrix is input into a pre-trained deep reinforcement learning model, which outputs the shrinkage probability distribution of the remaining nodes in the current optical network topology. The node with the highest shrinkage probability is selected as the current shrinkage target, and node shrinkage is performed. The current shrinkage target and its adjacent edges are removed, and shortcut edges are added between the adjacent nodes of the shrinkage target. The optical network topology after node shrinkage is updated. After all nodes have completed shrinkage, the complete node shrinkage sequence is output. This application overcomes the limitations of traditional methods in balancing local and global performance. Through the trial-and-error learning mechanism of DRL (Deep Reinforcement Learning), an adaptive optimization shrinkage strategy is achieved, thereby minimizing the graph density during node shrinkage, improving routing preprocessing efficiency, and significantly shortening query time. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 A flowchart for preprocessing existing CH algorithms; Figure 2 A diagram illustrating the addition of a shortcut to the existing CH algorithm; Figure 3 This is a schematic diagram of optical network fault recovery based on the CH algorithm; Figure 4 A flowchart illustrating a large-scale optical network routing optimization method provided in an embodiment of this application; Figure 5 An architecture diagram of a large-scale optical network routing optimization method provided in an embodiment of this application; Figure 6 A state space matrix provided in one embodiment of this application; Figure 7 A schematic diagram of a PPO strategy model provided in an embodiment of this application; Figure 8 A schematic diagram of a PPO value model provided in an embodiment of this application; Figure 9 An architecture diagram of a large-scale optical network routing optimization method provided in another embodiment of this application; Figure 10 A schematic diagram of node shrinkage provided for another embodiment of this application; Figure 11 A schematic diagram of a large-scale optical network routing optimization system provided in another embodiment of this application; Figure 12 This is a schematic diagram of an electronic device according to another embodiment of this application. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.
[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0020] Currently, the CH algorithm consists of two parts: preprocessing and querying. In the preprocessing stage, nodes are shrunk sequentially according to their order. When deleting a node, necessary shortcut edges are added to its neighbors to construct a hierarchical graph containing both the original edges and shortcut edges. In the querying stage, a bidirectional Dijkstra's search is used based on this hierarchical graph, expanding upwards from the start and end points along higher-level edges until both search queues are empty, thus quickly obtaining the shortest path.
[0021] The basic idea of preprocessing is that, given a topology G(V, E), it is expected to find a total order relationship among all nodes: < (for example, u < v means that the order of node u is before v), and contract the nodes in this order successively, as Figure 1 shown.
[0022] Among them, such edges added in the third step are called shortcut edges or simply shortcuts, see Figure 2 , new shortcuts, node order: u < v < w. They only represent the paths that already exist in the current graph and are used to preserve the shortest paths when removing node u and all its adjacent edges.
[0023] For a path P = <v,..., w> ≠ <v, u, w> as a path between v and w, if w(P) ≤ w(<v, u, w>), then P is called a witness path of the triple <v, u, w>. This name stems from the fact that such paths prove that <v, u, w> is not the shortest path or not the only shortest path and allow the omission of shortcuts.
[0024] The entire step of finding the witness path and adding a shortcut for node u (steps 3 and 4) is called the contraction of u. Because after this step, for all shortest s - t paths P in graph G, if s, t > u and path P contains node u, then there exists a shortest s - t path P0 that does not contain node u internally.
[0025] The tuple (G = (V, E), <) composed of the graph G generated by the preprocessing algorithm and the node sequence < is called the contraction hierarchy. In the contraction of node u, the edge (v, u) ∈ E is called the incoming edge of u, and (u, w) is called the outgoing edge of u. Nodes > u are called remaining nodes, the edges adjacent to them are called remaining edges, and the graph composed of remaining nodes is called the remaining graph.
[0026] In the query phase, the bidirectional Dijkstra algorithm is used, and it only expands along the "upward" edges (connected to nodes with a higher order). Searches are started from the source point and the target point respectively and advanced alternately; when the search queues on both sides are empty, the shortest path distance can be determined. Finally, through the shortcut edges generated in the preprocessing phase, the shortcuts are restored to the complete shortest path and output.
[0027] The CH algorithm can be modified to adapt to the specific constraints of the optical network for fast routing calculation, and its process is as Figure 3 shown. Before any failure occurs, the order of all nodes will be determined first, and a working path and a protection path will be configured for each service. When a single - link failure occurs in the network, first the preprocessing information is updated, and then a fast protection path query is performed based on the updated data. And so on.
[0028] Traditional node ranking methods are based on a weighted sum of multiple heuristic metrics (priority terms) to calculate the "importance" of each node. However, this method suffers from problems such as reliance on experience for weight settings and a lack of unified standards. Manual parameter tuning is required for different network types and sizes, making it difficult to adapt. Secondly, linear models cannot capture the non-linear interactions between metrics. Complex trade-offs, such as low-cost shrinkage with large edge differences or high-cost shrinkage with minor impacts, may be ignored, potentially leading to "greedy" shrinkage: focusing only on the current local optimum at the expense of overall efficiency. Furthermore, this approach cannot assess the long-term impact of node shrinkage on the entire preprocessing stage, easily causing an imbalance between preprocessing performance and query results, increasing the algorithm's routing computation time.
[0029] The purpose of this application is to address the limitations of traditional CH algorithms, which rely on fixed weights for node sorting and struggle to balance local and global performance. The aim is to find a way to leverage deep reinforcement learning to automatically learn the complex nonlinear relationships between node features without manual parameter tuning, and to optimize the shrinkage strategy based on the comprehensive reward function. This would maximize preprocessing quality, shorten query time, enhance the algorithm's generalization ability across various network structures and scales, and reduce reliance on human experience.
[0030] Reference Figure 4-5 As shown, one embodiment of this application provides a large-scale optical network routing optimization method, which includes the following steps: Step S10: Based on the priority terms and topological characteristics of each node in the current optical network topology, construct a state space matrix representing the current optical network topology; Step S20: Input the state space matrix into the pre-trained deep reinforcement learning model and output the shrinkage probability distribution of the remaining nodes in the current optical network topology. Step S30: Select the node with the highest shrinkage probability as the current shrinkage target and perform node shrinkage operation. Remove the current shrinkage target and its adjacent edges, and add shortcut edges between the adjacent nodes of the shrinkage target. Step S40: Update the optical network topology after node shrinkage; Step S50: After all nodes have shrunk, output the complete node shrinkage sequence.
[0031] The large-scale optical network routing optimization method provided in this embodiment constructs a state space matrix representing the current optical network topology based on the priority term and topological features of each node. The state space matrix is input into a pre-trained deep reinforcement learning model, which outputs the shrinkage probability distribution of the remaining nodes in the current optical network topology. The node with the highest shrinkage probability is selected as the current shrinkage target, and node shrinkage is performed. The current shrinkage target and its adjacent edges are removed, and shortcut edges are added between the adjacent nodes of the shrinkage target. The optical network topology after node shrinkage is updated. After all nodes have completed shrinkage, the complete node shrinkage sequence is output. This application overcomes the limitations of traditional methods in balancing local and global performance. Through the trial-and-error learning mechanism of DRL (Deep Reinforcement Learning), an adaptive optimization shrinkage strategy is achieved, thereby minimizing graph density, improving routing preprocessing efficiency, and significantly shortening query time.
[0032] In one embodiment, the state space matrix has a dimension of N×F; where N represents the number of remaining nodes in the current optical network topology, and F is the feature dimension of each node, as shown in the figure. Figure 6 As shown, each row of the state space matrix corresponds to the feature vector of a node, encapsulating the priority term and topological features of the corresponding node. These metrics are standardized (e.g., min-max scaling or z-score normalization) to ensure numerical stability and model convergence.
[0033] Specifically, each column represents a predefined node attribute, including topological attributes and priority terms. Topological attributes include metrics such as node number, degree centrality, connection density, clustering coefficient, betweenness centrality, proximity centrality, and PageRank score. These metrics are calculated from the graph's adjacency matrix or shortest path and reflect the local and global structural features of the corresponding node. Priority terms include metrics such as edge difference, shrinkage cost, number of neighboring nodes to be shrunk, and search space depth. These metrics reflect the immediate benefits of shrinking the corresponding node. This matrix serves as the initial input to the policy model in a deep reinforcement learning environment. Combined with edge indices, it aggregates neighborhood information through a message-passing mechanism to generate an embedded representation to guide agent decisions. This representation not only preserves the individual characteristics of nodes but also implicitly encodes topological dependencies, supports end-to-end training, and improves the model's generalization ability on variable-length graphs.
[0034] In one embodiment, before step S10, which involves constructing a state space matrix representing the current optical network topology based on the priority terms and topological features of each node in the current optical network topology, the following steps are included: S01. Pre-train the deep reinforcement learning model using the following objective reward function to obtain the pre-trained deep reinforcement learning model: ; ; ; ; in, For the target reward function, The immediate reward is a weighted sum of priority items; This is the edge difference value. To reduce costs, To reduce the number of neighbors, As weight; This is a network state reward used to reflect the density of the network during the contraction process; It is a scaling factor. The congestion level of the current optical network topology is defined as the edge density of the current optical network topology.
[0035] Specifically, this application adopts the Proximal Policy Optimization (PPO) framework as the core algorithm for deep reinforcement learning. The advantages of PPO lie in its sample efficiency and robustness, effectively handling continuous decision-making processes, such as the selection of node shrinkage order, while reducing the risk of gradient explosion and improving convergence speed. In this task, the PPO framework guides the agent to automatically learn the optimal policy under complex graph structures based on a custom reward function (such as a combination of immediate reward and network state penalty), achieving adaptive optimization without human intervention.
[0036] In one embodiment, step S20, inputting the state space matrix into a pre-trained deep reinforcement learning model and outputting the shrinkage probability distribution of the remaining nodes in the current optical network topology, includes: The state space matrix is input into a pre-trained deep reinforcement learning model. Local topological embeddings are extracted from the state space matrix and edge indices through the GNN layer. Global dependencies between nodes are captured through the Transformer layer. The shrinkage probability distribution of the remaining nodes in the current optical network topology is output through linear projection and the Softmax output layer.
[0037] Specifically, refer to Figure 7As shown, the PPO policy model first extracts local topological embeddings from the node feature matrix and edge indices through a GNN layer. Specifically, the GNN aggregates information about each node and its neighbors through a message-passing mechanism, thereby generating an embedding vector for each node. This mechanism allows the model to automatically learn the topological role of nodes, such as identifying highly connected nodes or core nodes in dense regions, thus providing a foundation for subsequent global processing. Subsequently, these embeddings are fed into a Transformer layer, which uses a self-attention mechanism to capture global dependencies between nodes. Finally, a linear projection and a Softmax output layer generate a shrinkage probability distribution for each remaining node, thereby achieving adaptive decision optimization for the PPO agent.
[0038] Reference Figure 8 As shown, the value model in the PPO framework is designed similarly to the policy model, but focuses on evaluating the value V(s) of the current graph state. First, local topological embeddings are extracted from the node feature matrix and edge indices through a GNN layer. Then, these embeddings are processed as a sequence using a Transformer layer. The global dependencies between nodes are captured by a self-attention mechanism. Finally, the value scalar V(s) of the current state is generated through a global pooling and linear projection output layer, thereby realizing the stable optimization of the advantage function calculation and policy training of the PPO agent.
[0039] In one embodiment, after step S40, updating the optical network topology after node shrinkage; and before step S50, outputting the complete node shrinkage sequence after all nodes have finished shrinking, the following steps are further included: Step S41: Detect whether there are any un-shrunken nodes in the shrunken optical network topology; if so, continue to construct a state space matrix representing the shrunken optical network topology based on the priority term and topological features of each node in the shrunken optical network topology; otherwise, determine that all nodes have completed the shrinkage.
[0040] In one embodiment, after step S50, which outputs the complete node shrinkage sequence after all nodes have finished shrinking, the following steps are also included: Step S60: Use the complete node shrinking sequence for node preprocessing in the CH algorithm. This significantly reduces query time.
[0041] Reference Figure 9 The diagram illustrates a large-scale optical network routing optimization method in a specific embodiment. A node ranking method based on deep reinforcement learning is applied to the software-defined networking (SDN) controller to optimize the node shrinking order in the optical network, minimizing graph density and improving routing query efficiency. The SDN controller, as the central management entity, integrates a deep reinforcement learning model and is responsible for real-time monitoring of the network topology, calculating node characteristics, and generating the optimal ranking. Figure 10This is an example of the node contraction process. In the figure, the node order is d < b < c < a, and the contraction order is: d, b, c, a.
[0042] This application proposes a routing optimization method for large-scale optical networks. This method constructs a custom reward function and jointly trains a policy network and a value network for collaborative training to achieve an efficient node sorting strategy. The method pays special attention to the modeling of the state space. By integrating the topological structure features of nodes and the contraction priority item indicators, a node feature matrix is constructed as the representation of the current state. To enhance the state representation ability, a graph neural network is introduced for node embedding to fully capture local structure information. On this basis, the policy network further introduces a Transformer layer to process variable-length node sequences and models the global dependencies between nodes through the self-attention mechanism, and outputs the optimal next contraction node in the current step. The value network is used to estimate the long-term reward of the current state under the policy and assist in policy optimization. The reward function design comprehensively considers the immediate and long-term benefits brought by node contraction, guides the model to automatically learn the complex non-linear relationships between nodes, and does not require parameter tuning relying on manual experience.
[0043] This method overcomes the limitation of traditional methods that it is difficult to balance local and global performance. Through the trial-and-error learning mechanism of DRL, it realizes adaptive optimization of the contraction strategy, thereby minimizing the graph density to the greatest extent, improving the routing preprocessing efficiency, and significantly shortening the query time.
[0044] It should be noted that the method of the embodiment of this application can be executed by a single device, such as a computer or a server, etc. The method of this embodiment can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In this case of a distributed scenario, one of the multiple devices can only execute one or more steps of the method of the embodiment of this application, and these multiple devices will interact with each other to complete the method.
[0045] It should be noted that some embodiments of this application have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order from those in the above embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0046] Based on the same inventive concept, corresponding to the method of any of the above embodiments, this application also provides a routing optimization system for large-scale optical networks. Referring to Figure 11 as shown, the system includes the following modules: The state space matrix construction module is used to construct a state space matrix representing the current optical network topology based on the priority terms and topological characteristics of each node in the current optical network topology. The shrinkage probability distribution output module is used to input the state space matrix into the pre-trained deep reinforcement learning model and output the shrinkage probability distribution of the remaining nodes in the current optical network topology. The node shrinking module is used to select the node with the highest shrinking probability as the current shrinking target, perform node shrinking operation, remove the current shrinking target and its adjacent edges, and add shortcut edges between the adjacent nodes of the shrinking target. The update module is used to update the optical network topology after node shrinkage. The sequence output module is used to output the complete node shrinkage sequence after all nodes have finished shrinking.
[0047] This application overcomes the limitations of traditional methods in balancing local and global performance. By using the trial-and-error learning mechanism of DRL (Deep Reinforcement Learning), it achieves an adaptive optimization shrinkage strategy, thereby minimizing graph density, improving routing preprocessing efficiency, and significantly shortening query time.
[0048] The large-scale optical network routing optimization system in this embodiment has the beneficial effects of the above-described method embodiments, which will not be repeated here.
[0049] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the large-scale optical network routing optimization method of any of the above embodiments.
[0050] Figure 12 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1101, a memory 1102, an input / output interface 1103, a communication interface 1104, and a bus 1105. The processor 1101, memory 1102, input / output interface 1103, and communication interface 1104 are interconnected internally via the bus 1105.
[0051] The processor 1101 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.
[0052] The memory 1102 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1102 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1102 and is called and executed by the processor 1101.
[0053] Input / output interface 1103 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.
[0054] The communication interface 1104 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).
[0055] Bus 1105 includes a pathway for transmitting information between various components of the device, such as processor 1101, memory 1102, input / output interface 1103, and communication interface 1104.
[0056] It should be noted that although the above-described device only shows the processor 1101, memory 1102, input / output interface 1103, communication interface 1104, and bus 1105, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.
[0057] The electronic devices described above are used to implement the corresponding large-scale optical network routing optimization methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0058] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the large-scale optical network routing optimization method as described in any of the above embodiments.
[0059] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.
[0060] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the large-scale optical network routing optimization method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0061] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.
[0062] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.
[0063] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.
[0064] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.
Claims
1. A method for optimizing routing in large-scale optical networks, characterized in that, include: Based on the priority terms and topological characteristics of each node in the current optical network topology, a state space matrix representing the current optical network topology is constructed. The state space matrix is input into the pre-trained deep reinforcement learning model, and the shrinkage probability distribution of the remaining nodes in the current optical network topology is output. Select the node with the highest shrinkage probability as the current shrinkage target and perform node shrinkage operation. Remove the current shrinkage target and its adjacent edges, and add shortcut edges between the adjacent nodes of the shrinkage target. Update the optical network topology after node shrinkage; After all nodes have shrunk, output the complete node shrinkage sequence.
2. The large-scale optical network routing optimization method according to claim 1, characterized in that, The state space matrix has a dimension of N×F; where N represents the number of remaining nodes in the current optical network topology, F is the feature dimension of each node, and each row of the state space matrix corresponds to the feature vector of a node, which encapsulates the priority term and topological features of the corresponding node.
3. The large-scale optical network routing optimization method according to claim 1, characterized in that, The step of inputting the state space matrix into a pre-trained deep reinforcement learning model and outputting the shrinkage probability distribution of the remaining nodes in the current optical network topology includes: The state space matrix is input into a pre-trained deep reinforcement learning model. Local topological embeddings are extracted from the state space matrix and edge indices through a GNN layer. Global dependencies between nodes are captured through a Transformer layer. The shrinkage probability distribution of the remaining nodes in the current optical network topology is output through a linear projection and a Softmax output layer.
4. The large-scale optical network routing optimization method according to claim 1, characterized in that, Before constructing the state space matrix representing the current optical network topology based on the priority terms and topological features of each node in the current optical network topology, the following steps are included: The deep reinforcement learning model is pre-trained using the following objective reward function to obtain a pre-trained deep reinforcement learning model: ； ； ； ； in, For the target reward function, The immediate reward is a weighted sum of priority items; This is the edge difference value. To reduce costs, To reduce the number of neighbors, As weight; The network state reward reflects the density of the network during the contraction process; α>0 is the scaling factor; congestion is the congestion level of the current optical network topology, defined as the edge density of the current optical network topology.
5. The large-scale optical network routing optimization method according to claim 1, characterized in that, After updating the shrunk optical network topology; and before outputting the complete node shrinking sequence after all nodes have finished shrinking, the process further includes: If there are still un-shrunken nodes in the shrunken optical network topology, then continue to construct a state space matrix representing the shrunken optical network topology based on the priority terms and topological features of each node in the shrunken optical network topology; otherwise, determine that all nodes have completed the shrunken process.
6. The large-scale optical network routing optimization method according to claim 1, characterized in that, The topological attributes include one or more of the following: node number, degree centrality, connection density, clustering coefficient, betweenness centrality, proximity centrality, and PageRank score; The priority items include one or more of the following: edge difference value, shrinkage cost, number of neighboring nodes to be shrunk, and search space depth.
7. The large-scale optical network routing optimization method according to claim 1, characterized in that, After all nodes have completed shrinking and the complete node shrinking sequence is output, the process also includes: The complete node shrinkage sequence is used for node preprocessing in the CH algorithm.
8. A large-scale optical network routing optimization system, characterized in that, include: The state space matrix construction module is used to construct a state space matrix representing the current optical network topology based on the priority terms and topological characteristics of each node in the current optical network topology. The shrinkage probability distribution output module is used to input the state space matrix into the pre-trained deep reinforcement learning model and output the shrinkage probability distribution of the remaining nodes in the current optical network topology. The node shrinking module is used to select the node with the highest shrinking probability as the current shrinking target, perform node shrinking operation, remove the current shrinking target and its adjacent edges, and add shortcut edges between the adjacent nodes of the shrinking target. The update module is used to update the optical network topology after node shrinkage. The sequence output module is used to output the complete node shrinkage sequence after all nodes have finished shrinking.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the large-scale optical network routing optimization method as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the large-scale optical network routing optimization method according to any one of claims 1 to 7.