A large language model assisted cross-layer optical network routing optimization method

Abstract

The application relates to a large language model assisted cross-layer optical network routing optimization method. The method comprises the following steps: constructing the architecture of the cross-layer optical network, collecting network physical node composition, link resource parameters and service request characteristics; designing an auxiliary graph model, converting the cross-layer network routing optimization problem into a minimum cost path calculation problem in a virtual topology; building a large language model enabled auxiliary graph model weight optimization framework, and preliminarily generating the edge weight of the auxiliary graph model by the large language model; calculating the minimum cost path based on the initial edge weight of the auxiliary graph, evaluating the network performance index corresponding to the path, inputting the evaluation result into the large language model as feedback data, iteratively adjusting the edge weight until the performance index converges to the optimum; and completing the routing planning and wavelength allocation of all service requests in the auxiliary graph model according to the optimal edge weight. The application combines the large language model and the auxiliary graph model framework, improves the flexibility and multi-objective adaptation capability of the routing optimization, and realizes global optimization.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a cross-layer optical network routing optimization method assisted by a large language model. Background Technology

[0002] Cross-layer optical networks (CLOOR) are a collaborative communication network architecture consisting of an IP layer and an optical layer. Routing optimization in CLOOR refers to optimization techniques that improve service quality by jointly considering the routing mechanisms of the IP and optical layers to reduce resource consumption and transmission latency.

[0003] However, there is currently no technical solution that combines large language models with auxiliary graph frameworks to provide a systematic solution for routing optimization goals in cross-layer optical networks. This fails to fully leverage the potential of large language models in global optimization and dynamic adaptation, resulting in cross-layer networks still experiencing problems such as high resource consumption, poor latency control, and low optical path management efficiency when facing multiple optimization requirements. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a cross-layer optical network routing optimization method assisted by a large language model.

[0005] The technical solution adopted in this invention is: a cross-layer optical network routing optimization method assisted by a large language model, comprising the following steps: Step 1: Construct the architecture of the cross-layer optical network, and collect information on the composition of network physical nodes, link resource parameters, and service request characteristics; Step 2: Design an auxiliary graph model to transform the cross-layer network routing optimization problem into a minimum cost path calculation problem in a virtual topology; Step 3: Build an auxiliary graph model weight optimization framework empowered by a large language model. Input the background information of the cross-layer optical network, the basic features of the cross-layer optical network and multiple optimization objectives into the large language model. The large language model will initially generate the edge weights of the auxiliary graph model. Step 4: Calculate the minimum cost path based on the edge weights of the initial auxiliary graph, evaluate the network performance index corresponding to the path, input the evaluation results as feedback data into the large language model, and iteratively adjust the edge weights until the performance index converges to the optimal level. Step 5: Based on the optimal edge weights, complete the routing planning and wavelength allocation for all service requests in the auxiliary graph model. The allocation process satisfies the wavelength remaining capacity constraint of each fiber link.

[0006] Furthermore, in step 1, the physical node integrates an IP router and a reconfigurable optical add-drop multiplexer; the link resource parameters include the number of fiber links, the total number of wavelengths per link, and the capacity per wavelength; the service request characteristics cover source-destination node pairs and capacity requirements.

[0007] Furthermore, in step 2, the auxiliary graph model includes an access layer, an optical path layer, and a wavelength layer. The access layer maps IP nodes through virtual input / output nodes, the optical path layer marks existing optical connections through virtual input / output nodes, and the wavelength layer corresponds to available wavelengths directly connected between optical nodes through virtual nodes.

[0008] Furthermore, in step 3, the background information of the cross-layer optical network includes the cross-layer network scenario and the principle of the auxiliary graph model; the basic characteristics of the cross-layer optical network include service request data and topology data; and the multiple optimization objectives include wavelength minimization, end-to-end delay minimization, or optical path minimization.

[0009] Furthermore, in step 4, the performance indicators include wavelength consumption, average end-to-end delay, and number of active optical paths.

[0010] Furthermore, set wavelength occupancy thresholds and new request quantity thresholds, monitor network operation status in real time, and repeat steps 3-5 when the number of service requests changes or the link resource status is updated to dynamically update the routing and wavelength allocation scheme.

[0011] Furthermore, the large language model is selected from the GPT series, DeepSeek-R1, LLaMA / LLaMA-2, and ChatGLM.

[0012] The present invention has the following beneficial effects: (1) This invention combines a large language model with an auxiliary graph model framework. By utilizing the contextual reasoning and adaptive decision-making capabilities of the large language model, it dynamically generates auxiliary graph model edge weights that adapt to multiple optimization objectives, thereby improving the flexibility and multi-objective adaptability of route optimization. It does not require long-term training and can effectively avoid local optima problems, thus achieving global optimization. (2) This invention supports multiple optimization objectives such as minimizing wavelength resource consumption, minimizing end-to-end service latency, and minimizing the number of open optical paths. It can be flexibly switched according to network and service requirements, which significantly improves network resource utilization and service quality. (3) The present invention has dynamic optimization capability. By real-time monitoring of network status to trigger re-optimization, it can adapt to the dynamic changes of service requests and link resources, ensure long-term stable operation of the network, reduce network operation and maintenance costs, and reduce resource consumption and latency. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the topology of a cross-layer optical network according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an auxiliary graph model construction in an embodiment of the present invention; Figure 3This is a schematic diagram illustrating a method for routing optimization using an auxiliary graph model to enable a large language model, as described in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1: This example presents a multi-route optimization method for cross-layer networks based on a large language model-assisted graph framework, with wavelength resource consumption minimization (MinWL) as the optimization objective. The specific steps are as follows: Step 1: As Figure 1 As shown, the basic architecture for constructing a cross-layer optical network is constructed. The network includes 9 physical nodes (numbered 1-9), 12 bidirectional optical fiber links, each link is 20 kilometers long, each optical fiber link provides 3 wavelengths, and the capacity of a single wavelength is 10Gbps; 200 service requests are generated, each request includes a source-destination node pair (such as 1-7, 2-9, 3-8, etc.) and a capacity requirement (range: 5-10Gbps). The capacity of a single service request does not exceed 10Gbps, and the collected data is stored in the database of the network management backend.

[0016] Step 2: As Figure 2 As shown, design the auxiliary diagram (i.e., the auxiliary diagram model). The access layer contains 9 pairs of virtual input / output nodes (corresponding to 9 IP nodes, which can be labeled IP1-IN / IP1-OUT to IP9-IN / IP9-OUT). The optical path layer has several virtual input / output nodes (labeled LP1-IN / LP1-OUT to LP9-IN / LP9-OUT). The wavelength layer contains 3 sub-layers (corresponding to 3 wavelengths λ1, λ2, and λ3), and each sub-layer has 9 virtual nodes (labeled λ1-1-IN / λ1-1-OUT to λ1-9-IN / λ1-9-OUT; the same applies to the λ2 and λ3 sub-layers). Configure 8 types of edges and connect them according to their functions: GrmE (grooming-edge): Connects IP1-IN to IP1-OUT, IP2-IN to IP2-OUT, etc., a total of 9 edges; MUXE (mux-edge): Connects IP1-OUT to LP1-IN, IP2-OUT to LP2-IN, etc., a total of 9 edges; DmxE (demux-edge): Connects LP1-OUT to IP1-IN, LP2-OUT to IP2-IN, etc., a total of 9 edges; LPE (lightpath-edge): Connects LP1-OUT to LP2-IN, LP2-OUT to LP3-IN, etc., configured according to the existing number of optical connections; TxE (transmitter-edge): connects λ1-1-IN to IP1-OUT, λ2-1-IN to IP1-OUT, λ3-1-IN to IP1-OUT, etc., for a total of 27 lines; RxE (receiver-edge): Connects IP1-IN to λ1-1-OUT, IP1-IN to λ2-1-OUT, IP1-IN to λ3-1-OUT, etc., for a total of 27 lines; WLE (wavelength-link-edge): connects λ1-1-OUT to λ1-2-IN, λ1-2-OUT to λ1-3-IN, etc., with 8 lines per sub-layer, for a total of 24 lines; WBE (wavelength-bypass-edge): connects λ1-1-IN to λ1-1-OUT, λ1-2-IN to λ1-2-OUT, etc., with 9 edges per sublayer, for a total of 27 edges.

[0017] Step 3: Build a weight optimization framework for the auxiliary graph model assisted by the large language model, such as... Figure 3 As shown. This embodiment uses DeepSeek-R1 as the large language model, and inputs prompt information to the large language model through the API interface: Background information: Cross-layer optical networks consist of an IP layer and an optical layer. The IP layer performs service routing through IP routers (IP nodes), while the optical layer processes optical switching through reconfigurable optical add-drop multiplexers (optical nodes). The auxiliary graph model includes an access layer, an optical path layer, and a wavelength layer. GrmE enables service routing, MUXE / DmxE enables inter-layer transmission, LPE carries optical connections, TxE / RxE enables interaction between the wavelength layer and the access layer, and WLE / WBE enables wavelength transmission and bypass.

[0018] Parameter hints: Topology data is provided as a node matrix, with rows / columns corresponding to nodes 1-9. Element "1" indicates that there is a link between nodes, and "0" indicates that there is no link. The link length is 20 kilometers. Business request data is provided in list form, which includes the source-destination node pair and capacity for each request.

[0019] Task hint: The optimization objective is to minimize the number of wavelengths turned on (MinWL). This requires adjusting the edge weights of the auxiliary graph model to optimize the number of wavelengths turned on. Feedback message: Initial auxiliary graph model edge weights: GrmE=0.2, MUXE=0.15, DmxE=0.15, LPE=0.2, TxE=0.08, RxE=0.08, WLE=0.07, WBE=0.07; Initial number of wavelengths enabled: 86. The initial output of the large language model is the edge weights of the auxiliary graph model: GrmE=0.22, MUXE=0.14, DmxE=0.14, LPE=0.21, TxE=0.07, RxE=0.07, WLE=0.06, WBE=0.06.

[0020] Step 4: Based on the initial weights, the shortest path algorithm is used to calculate the minimum cost path for 200 business requests in the auxiliary graph model, and the wavelength consumption is calculated to be 82. The result is used as feedback input to the large language model: "The wavelength consumption under the current weight is 82, which needs to be further reduced." The large language model adjusts the weights to: GrmE=0.23, MUXE=0.13, DmxE=0.13, LPE=0.22, TxE=0.06, RxE=0.06, WLE=0.05, WBE=0.05; the wavelength consumption is calculated again to be 79; the iteration continues, and the wavelength consumption is 78 after the next weight adjustment. The difference between the two is approximately (79-78) / 79≈1.27%<2%, the performance index converges, and the weight at this time is determined to be the optimal weight.

[0021] Step 5: Based on the optimal weights, recalculate the minimum cost paths for the 200 service requests and map them as physical routes in the cross-layer network. For example, the path for request "1-7" is "Node 1 → Node 2 → Node 5 → Node 7", and the corresponding auxiliary graph model path is "IP1-IN → IP1-OUT (GrmE) → LP1-IN (MUXE) → LP2-OUT (LPE) → IP2-IN (DmxE) → IP2-OUT (GrmE) → LP2-IN (MUXE) → LP5-OUT (LPE) → IP5-IN (DmxE) → IP5-OUT (GrmE) → LP5-IN (MUXE) → LP7-OUT (LPE) → IP7-IN (DmxE) → IP7-OUT". Allocate wavelength λ1 to this path to ensure that λ1 of the links 1-2, 2-5, and 5-7 is not occupied. Similarly, complete the routing and wavelength allocation for all requests, and finally calculate the wavelength consumption as 78.

[0022] Step 6: Monitor network status in real time, set wavelength occupancy threshold to 80%, and new request threshold to 20; if 15 new service requests are detected within 1 hour, which is below the threshold, no further optimization is needed; when the number of new requests reaches 20, repeat steps 3-5, readjust the weights of the large language model, and update the routing and wavelength allocation scheme.

[0023] Tests showed that the wavelength consumption of this embodiment was 78, which is 9.3% less than that of the heuristic method (86) and 4.9% less than that of the DRL method (82), thus achieving the MinWL target.

[0024] Example 2: The difference between this example and Example 1 is that the optimization objective is to minimize end-to-end latency (MinLat). The specific steps are adjusted as follows: In step 3, the task prompt is changed to: "The optimization target is MinLat. The edge weights of the auxiliary graph model need to be adjusted to reduce the average end-to-end latency of 200 business requests." The feedback prompt is changed to: "The initial edge weights of the auxiliary graph model are: GrmE=0.18, MUXE=0.16, DmxE=0.16, LPE=0.18, TxE=0.09, RxE=0.09, WLE=0.07, WBE=0.07; the average end-to-end latency of the business under the initial edge weights is 33.4ms." The initial output of the large language model is the edge weights of the auxiliary graph model: GrmE=0.19, MUXE=0.17, DmxE=0.17, LPE=0.19, TxE=0.08, RxE=0.08, WLE=0.06, WBE=0.08 (increasing the WBE weight to promote bypass transmission and reduce latency).

[0025] In step 4, the performance metric is evaluated as the average end-to-end latency: the average latency is calculated to be 31.2ms under the initial weights. After feeding back to the large language model, the weights are adjusted, and the average latency finally converges to 30.5ms (the difference between the two iterations is 0.3ms, accounting for 0.98% < 2%).

[0026] In step 5, routing allocation prioritizes paths containing WBE. For example, the path for requesting "3-8" is "node 3 → node 5 → node 8". The corresponding auxiliary graph model path includes "λ2-3-IN → λ2-3-OUT (WBE) → λ2-5-IN → λ2-5-OUT (WBE) → λ2-8-IN → IP8-IN (RxE)", reducing forwarding nodes and lowering latency.

[0027] Tests showed that the average end-to-end latency in this embodiment was 30.5ms, which is 8.7% lower than the heuristic method (33.4ms) and 9.4% lower than the DRL method (33.68ms), achieving the MinLat target.

[0028] Example 3: The difference between this example and Example 1 is that the optimization objective is optical path minimization (MinLP). The specific steps are adjusted as follows: In step 3, the task prompt is changed to: "The optimization goal is MinLP. The edge weights of the auxiliary graph model need to be adjusted to reduce the number of active optical paths that satisfy 200 service requests." The feedback prompt is changed to: "Initial auxiliary graph model edge weights: GrmE=0.21, MUXE=0.16, DmxE=0.16, LPE=0.22, TxE=0.07, RxE=0.07, WLE=0.06, WBE=0.05; Initial number of active optical paths: 60." The initial output of the large language model is the edge weights of the auxiliary graph model: GrmE=0.23, MUXE=0.15, DmxE=0.15, LPE=0.24, TxE=0.06, RxE=0.06, WLE=0.05, WBE=0.05 (the LPE weight is increased to prioritize the use of existing optical paths).

[0029] In step 4, the performance evaluation index is the number of optical paths opened: the number of optical paths is calculated to be 48 under the initial weight. After feeding back to the large language model, the weight is adjusted, and finally converges to the number of optical paths of 42 (the difference between the two iterations is 1, accounting for 2.33%. After adjusting the threshold to 2.5%, convergence is determined).

[0030] In step 5, the routing allocation prioritizes the LPE corresponding to the existing optical path. For example, it requests that "4-9" and "5-9" share the existing optical path of "node 7-9" to avoid establishing a new optical path.

[0031] Tests showed that the number of active optical paths in this embodiment is 42, which is 30% less than the existing heuristic method (60) and 10.6% less than the DRL method (47), thus achieving the MinLP target.

[0032] In summary, this invention enables the efficient achievement of routing optimization goals in cross-layer optical networks by empowering an auxiliary graph model framework with a large language model. It demonstrates superior performance compared to existing methods under different optimization objectives, can flexibly adapt to the operational needs of cross-layer optical networks, and possesses good practicality and scalability.

Claims

1. A large language model-assisted cross-layer optical network routing optimization method, characterized in that, Includes the following steps: Step 1: Construct the architecture of the cross-layer optical network, and collect information on the composition of network physical nodes, link resource parameters, and service request characteristics; The physical node integrates an IP router and a reconfigurable optical add-drop multiplexer; the link resource parameters include the number of fiber links, the total number of wavelengths per link, and the capacity per wavelength; the service request characteristics cover source-destination node pairs and capacity requirements; Step 2: Design an auxiliary graph model to transform the cross-layer network routing optimization problem into a minimum cost path calculation problem in a virtual topology. The auxiliary graph model includes an access layer, an optical path layer, and a wavelength layer. The access layer maps IP nodes through virtual input / output nodes. The optical path layer marks existing optical connections through virtual input / output nodes. The wavelength layer corresponds to available wavelengths directly connecting optical nodes through virtual nodes. Step 3: Build an auxiliary graph model weight optimization framework empowered by a large language model. Input the background information of the cross-layer optical network, the basic features of the cross-layer optical network, and multiple optimization objectives into the large language model. The large language model will initially generate the edge weights of the auxiliary graph model. The background information of the cross-layer optical network includes the cross-layer network scenario and the principle of the auxiliary graph model. The basic features of the cross-layer optical network include service request data and topology data. The multiple optimization objectives include wavelength minimization, end-to-end latency minimization, or optical path minimization. in, Background information: Cross-layer optical networks consist of an IP layer and an optical layer. The IP layer performs service routing through IP routers, while the optical layer processes optical switching through reconfigurable optical add-drop multiplexers. GrmE enables service routing, MUXE / DmxE enables inter-layer transmission, LPE carries optical connections, TxE / RxE enables interaction between the wavelength layer and the access layer, and WLE / WBE enables wavelength transmission and bypass. The basic characteristics of cross-layer optical networks are: topology data is provided in the form of a node matrix, where the matrix rows / columns correspond to IP router nodes, and the element "1" indicates that there is a link between nodes, "0" indicates that there is no link, and the link length is specified; service request data is provided in the form of a list, which includes the source-destination node pair and capacity for each request. Multiple optimization objectives: The optimization objective is to minimize the number of wavelengths turned on, which requires adjusting the edge weights of the auxiliary graph model to optimize the number of wavelengths turned on. Step 4: Calculate the minimum cost path based on the edge weights of the initial auxiliary graph, evaluate the network performance metrics corresponding to the path, input the evaluation results as feedback data into the large language model, and iteratively adjust the edge weights until the performance metrics converge to the optimal level; the performance metrics include wavelength consumption, average end-to-end delay, and number of active optical paths. Step 5: Based on the optimal edge weights, complete the routing planning and wavelength allocation for all service requests in the auxiliary graph model. The allocation process satisfies the wavelength remaining capacity constraint of each fiber link.

2. The cross-layer optical network routing optimization method assisted by a large language model according to claim 1, characterized in that, The large language model mentioned is selected from any one of the GPT series, DeepSeek-R1, LLaMA / LLaMA-2, and ChatGLM.

3. The large language model-assisted cross-layer optical network routing optimization method according to claim 1 or 2, characterized in that, Set wavelength occupancy thresholds and new request quantity thresholds, monitor network operation status in real time, and repeat steps 3-5 when the number of service requests changes or the link resource status is updated to dynamically update the routing and wavelength allocation scheme.

Images