Cross-environment generalization network traffic optimization method based on large model agent

Abstract

The application relates to the field of network traffic optimization, and particularly relates to a cross-environment generalization network traffic optimization method based on a large model agent. The technical scheme comprises the following steps: data collection, collecting a historical traffic matrix, network topology and capacity information and a natural language instruction from a network administrator; the collected historical traffic matrix, network topology and capacity information are sent to a feature decoupling module for cross-environment feature decoupling, and the natural language instruction is sent to a large model agent for strategy understanding and embedding; then, a pre-trained network traffic engineering head decoding is used to generate network routing allocation; finally, the routing allocation is issued through a network model context protocol of the large model, and network feedback and performance indexes are monitored, which are used as inputs of the next round of optimization to form a closed loop control. The application is suitable for network traffic optimization.

Description

Technical Field

[0001] This invention relates to the field of network traffic optimization, and specifically to a cross-environment generalized network traffic optimization method based on a large model agent. Background Technology

[0002] In practical applications, the performance of large-scale intelligent agents in downstream tasks is often limited by the scarcity of high-quality data. This is particularly true in the networking domain, where network traffic data is largely distributed locally and is highly private. Even with some open-source data from data centers, significant differences in network topology and data representation make cross-environment migration and effective utilization difficult, limiting the potential for large models to improve performance through richer training data. Furthermore, the vast and dynamically changing state space of network traffic makes designing a compatible, efficient, and practical large-scale model input / output architecture a highly challenging task.

[0003] The main challenge facing current network traffic engineering is that existing methods generally lack sufficient intelligence and architectural flexibility, making it difficult to achieve an effective balance between improving routing quality, ensuring system scalability, and achieving cross-environment generalization. This lack of intelligence limits the traffic engineering controller's ability to perceive and adapt to complex network topologies, and also affects its ability to respond and migrate quickly in dynamic and changing network environments, thus restricting the improvement of overall network performance and the expansion of application scope.

[0004] Traditional optimizer-based traffic engineering controllers (such as those based on linear programming or mixed-integer programming) can perform global traffic optimization, but they lack the ability to flexibly adapt to complex and ever-changing business needs. Furthermore, their computational complexity is high and convergence is slow on large-scale networks, making it difficult to meet the scalability and real-time response requirements of large-scale networks. Machine learning-based methods improve performance through pattern learning; however, they often require repeated training for specific network environments, have weak generalization ability, struggle to maintain effectiveness in new topologies or heterogeneous environments, lack "out-of-the-box" capability, and require continuous monitoring and maintenance. Simultaneously, existing systems have insufficient processing capabilities for natural language commands, unable to directly and intelligently parse and execute complex semantic commands, increasing the operational difficulty for network administrators and reducing overall work efficiency.

[0005] Therefore, the key to meeting the needs of future large-scale, dynamic network environments lies in overcoming the aforementioned bottlenecks, integrating the advantages of large-scale intelligent agents, and improving the intelligence and adaptability of network traffic engineering. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a cross-environment generalized network traffic optimization method based on large model intelligent agents, thereby improving the intelligence and adaptability of network traffic engineering.

[0007] This invention achieves the above objectives by adopting the following technical solution: This invention provides a cross-environment generalized network traffic optimization method based on a large model agent, comprising:

[0008] S1. Data acquisition and input, including historical traffic matrix, network topology and capacity information, and natural language commands from network administrators;

[0009] S2. Cross-environment feature decoupling: The collected historical traffic matrix, network topology and capacity information are sent to the feature decoupling module for cross-environment feature decoupling.

[0010] S3. Input natural language instructions into the large model agent for policy understanding and embedding;

[0011] S4. Generate network route allocation by decoding the pre-trained network traffic engineering header;

[0012] S5. Routing assignments are distributed through the network model context protocol of the large model, and network feedback and performance indicators are monitored as inputs for the next round of optimization, forming a closed-loop control.

[0013] Furthermore, step S1 specifically includes:

[0014] S101, Historical Traffic Matrix Acquisition, Historical Traffic Acquisition Window ,in Indicates from the source node to the destination node The traffic demand is represented by t, where t represents the time index and L represents the length of the historical traffic window.

[0015] S102. Obtaining network topology and capacity information, including obtaining a network topology map. Where V is the set of nodes and E is the set of links. For the link capacity function; for each source-destination pair Generate a set of candidate paths ;

[0016] S103, Natural Language Policy Instruction Reception: Receives natural language instructions from the network administrator, and the instructions are sent to the large model agent for semantic parsing.

[0017] Furthermore, in step S2, the cross-environment feature decoupling specifically includes:

[0018] S201. Calculate the relative flow matrix and normalize it:

[0019] , This represents the normalized traffic value in the traffic demand matrix. Extract the flow distribution pattern, eliminate the influence of absolute magnitude, and make the large model robust to changes in the absolute magnitude of flow.

[0020] S202. Calculate the link capacity and normalize it: Normalized capacity value Extracting capacity patterns that only contain relative relationships allows large models to adapt to networks of different sizes, from Gbps-level wide area networks to Tbps-level data centers.

[0021] S203, Generation Capacity-Flow Rate Relative Coefficient, Calculation Scale Factor: Preserve scale information to support multi-objective optimization beyond maximizing link utilization.

[0022] Furthermore, step S3 specifically includes:

[0023] S301, Network Model Context Protocol-Driven Agent Architecture Initialization;

[0024] A pre-trained large language model is used as the inference engine. The agent interacts with the network system in a structured manner through the network model context protocol, enabling the agent to have network abstract perception, tool invocation, multi-step reasoning and feedback learning capabilities.

[0025] S302, Three-dimensional strategy embedded spatial mapping;

[0026] Natural language strategies are decomposed and mapped to three-dimensional vectors. A three-dimensional vector contains three dimensions:

[0027] Historical dependence To control the sensitivity of response to historical traffic changes, ,in This indicates that historical fluctuations are ignored. This indicates a high degree of responsiveness to historical changes, reserving a margin for sudden traffic surges;

[0028] Global robustness Controlling the resilience to unforeseen traffic surges ,in To do one's best is a strategy. For worst-case robustness;

[0029] Cost Sensitivity The preference for balancing performance and cost. ,in Prioritizing performance over cost To prioritize cost, a certain degree of performance degradation is acceptable;

[0030] S303 and Few-shot small sample hints provide example-guided policy translation for large language models, guiding step-by-step reasoning through thought chain technology: identifying optimization objectives, extracting constraints, quantifying objective weights, mapping to a three-dimensional policy space, and generating policy embedding vectors;

[0031] S304. Policy Verification and Transmission: The generated policy embedding is validated for reasonableness. If ambiguity or conflict is found, a multi-round dialogue is triggered for clarification. After successful validation, the policy embedding vector is transmitted. The reasonableness verification, which is issued to the pre-trained network traffic engineering head, includes range checking, conflict detection, and historical consistency checking.

[0032] Furthermore, step S4 specifically includes:

[0033] S401, Feature Fusion and Encoding;

[0034] Capacity feature encoding is performed using a graph neural network, starting with edge feature encoding. ,in It is an adjacency matrix. The normalized capacity matrix is ​​then used for path feature aggregation. p represents the path, outputs a decoupled feature representation, and integrates policy embedding, decoupled traffic embedding, and relative amplitude coefficient as input to the iterative hybrid expert decoder in the network traffic engineering header;

[0035] S402, Multi-expert iterative decision generation;

[0036] Employing multi-round iterative routing decisions, for each iteration step... arrive Perform the following steps:

[0037] Encoding global congestion state ;

[0038] For each source-destination pair Build expert input ,in Embedded as a path;

[0039] Top- The sparse activation strategy selects only the highest weighted activation values. For each expert, the weights of the selected experts are renormalized: Activated expert parallel reasoning ;

[0040] Weighted Fusion Expert Output ;

[0041] Softmax normalization generates routing probabilities Output Final routing decision;

[0042] S403, Constraint Satisfaction and Output Generation;

[0043] The flow conservation constraint is automatically satisfied by Softmax: Calculate the actual load on each link, and use the maximum link utilization (MLU) as the primary optimization objective: Output the final traffic allocation matrix. For each SD pair and path This indicates the proportion of traffic carried by this path.

[0044] Furthermore, in step S5, the routing allocation is issued through the large model network model context protocol, and network feedback and performance indicators are monitored, specifically including:

[0045] S501, issues commands through the network model context protocol;

[0046] The configuration is distributed to the network controller via the network model context protocol, and the controller then distributes the configuration to the network devices via the OpenFlow or NETCONF protocol.

[0047] S502, Real-time performance monitoring;

[0048] Continuously collect performance metrics, including link utilization, end-to-end latency, packet loss rate, maximum link utilization, and average link utilization. Anomaly detection includes congestion detection, link failure detection, and traffic mutation detection.

[0049] S503, Asynchronous Strategy Adjustment and Continuous Optimization;

[0050] A dual-time-scale control architecture is adopted: the frequency and delay of the fast control loop are set, and the frequency and delay of the slow control loop are set. The fast control loop and the slow control loop run independently to decouple the time scales. The adaptive strategy is adjusted to periodic triggering or event triggering. The large language model analyzes the feedback data to generate new strategy embeddings. The event triggering includes performance anomalies, topology changes, and traffic drift.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention is based on a large model intelligent agent for dynamic semantic understanding and routing strategy control, realizing the intelligent conversion of natural language instructions into routing strategies.

[0053] This invention innovatively designs a feature decoupling module to achieve separate modeling of network traffic demand and link capacity, generating relative traffic, relative capacity, and capacity-traffic relative coefficients, thereby improving the model's robustness in dealing with heterogeneous data scenarios and its cross-environment generalization ability.

[0054] This invention proposes a pre-trained traffic engineering header asynchronous fusion strategy that embeds and decouples features to efficiently form intelligent traffic allocation decisions. It also uses a hybrid expert architecture to improve the model's compatibility and decision quality in the face of heterogeneous routing strategies. Attached Figure Description

[0055] Figure 1 This is a flowchart of the cross-environment generalized network traffic optimization method based on large model intelligent agents provided by the present invention;

[0056] Figure 2 This is a diagram of the cross-environment generalized network traffic optimization system based on large model intelligent agents provided by the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0058] This invention provides a cross-environment generalized network traffic optimization method based on large model agents, such as... Figure 1 As shown, it specifically includes:

[0059] S1. Data acquisition and input, including historical traffic matrix, network topology and capacity information, and natural language commands from network administrators;

[0060] S101. Traffic Matrix Acquisition: Acquire traffic demand matrix. Historical traffic windows can be selectively collected. ,in Indicates from the source node to the destination node The traffic demand is defined by V, which represents the set of nodes, R, which represents the real number field, D, which is a matrix of the number of nodes multiplied by the size of the number of nodes, t, which is the time index, and L, which is the length of the historical traffic window.

[0061] S102. Obtaining network topology and capacity information, including obtaining a network topology map. Where V is the set of nodes and E is the set of links. For the link capacity function; for each source-destination pair Generate a set of candidate paths ;

[0062] S103, Natural Language Policy Instruction Reception: Receives natural language instructions from the network administrator (e.g., "Prioritize low latency while avoiding link congestion"). The instructions are sent to the large model agent for semantic parsing.

[0063] S2. Cross-environment feature decoupling: The collected historical traffic matrix, network topology and capacity information are sent to the feature decoupling module for cross-environment feature decoupling.

[0064] S201. Calculate the relative flow matrix and normalize it: Normalized flow value Extract the flow distribution pattern, eliminate the influence of absolute magnitude, and make the large model robust to changes in the absolute magnitude of flow.

[0065] S202. Calculate the relative capacity matrix and normalize the link capacity: Normalized capacity value Extract capacity patterns that contain only relative relationships, enabling large models to adapt to networks of different sizes, from Gbps-level wide area networks to Tbps-level data centers;

[0066] S203, Generation Capacity-Flow Rate Relative Coefficient, Calculation Scale Factor: To preserve scale information to support multi-objective optimizations beyond MLU (max link utilization, utilization of the most congested link) (such as cost-sensitive, latency-optimized optimization), in practice... .

[0067] S3. Input natural language instructions into the large model agent for policy understanding and embedding;

[0068] S301. The intelligent agent architecture is initialized with network model context protocol driven by a pre-trained large language model (such as Claude-3.5-Sonnet, GPT-4) as the inference engine. The intelligent agent interacts with the network system in a structured manner through the network model context protocol, enabling the intelligent agent to have the ability of network abstract perception, tool invocation, multi-step inference and feedback learning.

[0069] S302, Three-dimensional policy embedding spatial mapping, decomposes and maps natural language policies to three-dimensional vectors. It includes three dimensions:

[0070] Historical dependence To control the sensitivity of response to historical traffic changes, ,in This indicates that historical fluctuations are ignored. This indicates a high degree of responsiveness to historical changes, reserving a margin for sudden traffic surges;

[0071] Global robustness Controlling the resilience to unforeseen traffic surges ,in To do one's best is a strategy. For worst-case robustness (near-equivalent multi-path routing);

[0072] Cost Sensitivity The preference for balancing performance and cost. ,in Prioritizing performance over cost To prioritize cost, a certain degree of performance degradation is acceptable;

[0073] S303 and the Few-shot small sample hinting project provide a small number of examples to guide policy translation for LLM. The chain-of-thought technology guides step-by-step reasoning: identifying optimization objectives, extracting constraints, quantifying objective weights, mapping to a three-dimensional policy space, and generating policy embedding vectors.

[0074] S304. Policy Verification and Transmission: The generated policy embedding is validated for reasonableness (scope check, conflict detection, historical consistency check). If ambiguity or conflict is found, a multi-round dialogue clarification is triggered. After successful verification, the policy embedding vector is transmitted. Issued to the network traffic engineering head.

[0075] S4. Generate network route allocation by decoding the pre-trained network traffic engineering header;

[0076] S401. Feature Fusion and Encoding: Capacity Feature Encoding via Graph Neural Networks: First, edge feature encoding is performed. ,in It is an adjacency matrix. The normalized capacity matrix is ​​then used; then path feature aggregation is performed. The output decoupled feature representation is fused with policy embedding, decoupled traffic embedding, and relative amplitude coefficient as input to the iterative hybrid expert decoder in TE-Head;

[0077] S402, Multi-expert iterative decision generation: The system employs multiple iterations (typically 3-5 rounds) to progressively refine routing decisions. For each iteration step... arrive Perform the following steps:

[0078] Encoding global congestion state ;

[0079] For each source-destination pair Build expert input ,in Path embedding; using Top- Sparse activation strategy (usually) Only select the one with the highest weight. For each expert, the weights of the selected experts are renormalized: Activated expert parallel reasoning ;

[0080] Weighted Fusion Expert Output ;

[0081] Softmax normalization generates routing probabilities Output Final routing decision.

[0082] S403, Constraint Satisfaction and Output Generation: Flow conservation constraints are automatically satisfied via Softmax. Calculate the actual load on each link, and use the maximum link utilization as the primary optimization objective: Output the final traffic allocation matrix. For each SD pair and path This indicates the proportion of traffic carried by this path.

[0083] S5. Routing assignments are distributed through the network model context protocol of the large model, and network feedback and performance indicators are monitored as inputs for the next round of optimization, forming a closed-loop control.

[0084] In step 5, route allocation is distributed through the network model context protocol of the large model, and network feedback and performance metrics are monitored, including:

[0085] S501: Commands are issued through the Network Model Context Protocol (NMT), and configurations are sent to the network controller (such as ONOS or OpenDaylight) through the NMT. ​​The controller then sends the configurations to the network devices through the OpenFlow / NETCONF protocol.

[0086] S502 Real-time performance monitoring continuously collects key performance indicators: link utilization, end-to-end latency, packet loss rate, maximum link utilization, average link utilization, and abnormal event detection including congestion detection (link utilization continuously exceeds 0.9), link failure detection (packet loss rate exceeds 50% or no traffic), and traffic mutation detection (traffic change exceeds 50% threshold).

[0087] S503, asynchronous strategy adjustment and continuous optimization: The system adopts a dual time scale control architecture: the fast control loop frequency is 10-30 seconds with a delay of less than 2 seconds, and the slow control loop frequency is from minutes to hours with a delay of 5-20 seconds. The fast control loop and the slow control loop run independently and decouple the time scales; the adaptive strategy adjustment is triggered periodically (approximately every 30 minutes) or by events (performance anomalies, topology changes, traffic drift), and the feedback data from the large language model analysis generates new strategy embeddings.

[0088] like Figure 2 As shown, this invention provides a cross-environment generalized network traffic optimization system based on a large model agent, comprising:

[0089] The feature decoupling module receives historical traffic matrices, network topology, and capacity information data, and converts them into relative features (such as relative traffic, relative capacity, and capacity-traffic relative coefficients) that are independent of specific network topologies and absolute quantities. This eliminates environment specificity and significantly improves the model's cross-environment generalization ability.

[0090] The large model agent is responsible for receiving and parsing the natural language instructions from the network administrator. Through the logical reasoning and tool usage capabilities of the large language model, it transforms complex semantic instructions into quantifiable routing policy embedding vectors.

[0091] The pre-trained network traffic header, as the core decision-making unit, asynchronously fuses decoupled features output by the feature decoupling module and policy embeddings output by the large model agent to efficiently generate preliminary traffic allocation intentions. To improve the system's compatibility with heterogeneous routing policies, the traffic engineering header adopts a hybrid expert architecture. The hybrid expert architecture router dynamically selects or weights and activates different "experts" (e.g., latency optimization experts, cost optimization experts, jitter minimization experts) based on the current network state and policy requirements, thereby achieving flexible and high-quality traffic decisions.

[0092] Compared to traditional solutions, this invention achieves significant improvements in key performance indicators of traffic engineering. Extensive experiments under various real-world network topologies and traffic patterns have validated the effectiveness of the system.

[0093] Specifically, this invention achieves a breakthrough in zero-shot generalization capability. Compared with existing machine learning-based traffic engineering controllers (such as DOTE, FIGRETE, HARP, AETHER, etc.), this invention achieves zero-shot deployment capability across network sizes for the first time: a single model trained on a network topology with ≤50 nodes can be directly deployed to a large-scale network with 700+ nodes, with an optimality loss of only 1.3% (relative to the optimal solution of a linear programming solver). Traditional machine learning methods typically require retraining for the target network when deployed across environments, otherwise performance will significantly degrade. In direct comparisons with some generalization methods (such as HARP), this invention shows an average lower normalized MLU index of 0.25 to 1.5 on multiple test topologies, demonstrating stronger generalization capability.

[0094] In terms of computational efficiency, compared to linear programming-based optimizers (such as the Gurobi solver), this invention reduces the response time by 123 times on a 700-node network, from minutes to seconds (<2 seconds), while maintaining near-optimal decision quality. Through a hybrid expert sparse activation mechanism, TELLM only needs to activate 20-30% of the model parameters to complete inference, significantly reducing computational overhead while ensuring decision quality. In large-scale network tests with 754 nodes, the inference time of this invention remained within the second range, verifying the system's scalability.

[0095] In terms of actual network deployment, tests based on a 13-node physical network prototype (21 links, 1 Gbps bandwidth) show that the present invention achieves a significant reduction in packet loss rate (from 16.61% to 3.24%), an improvement in average latency (from 1616μs to 562μs), and a reduction in jitter (from 2.55 to 2.27) compared to the traditional ECMP scheme under an 80 Mbps traffic load, proving the effectiveness of the system in real network environments.

[0096] In multi-dimensional verification of generalization ability, this invention demonstrates good robustness: the system performance remains stable under 2x traffic scaling and 10% sparse noise perturbation; on networks with significantly different topologies (from 4 to 754 nodes), a single pre-trained model can be directly deployed and maintain reasonable performance levels. The system supports policy configuration via natural language commands and can generate expected traffic allocation schemes under various policy preferences tested (historical response, global robustness, cost sensitivity, etc.).

[0097] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A cross-environment generalization network traffic optimization method based on a large model agent, characterized in that, include: S1. Data acquisition and input, including historical traffic matrix, network topology and capacity information, and natural language commands from network administrators; S2. Cross-environment feature decoupling: The collected historical traffic matrix, network topology and capacity information are sent to the feature decoupling module for cross-environment feature decoupling. S3. Input natural language instructions into the large model agent for policy understanding and embedding; S4. Generate network route allocation by decoding the pre-trained network traffic engineering header; S5. Routing assignments are distributed through the network model context protocol of the large model, and network feedback and performance indicators are monitored as inputs for the next round of optimization, forming a closed-loop control. Step S4 specifically includes: S401, Feature Fusion and Encoding; Capacity feature encoding is performed using a graph neural network, starting with edge feature encoding. ,in It is an adjacency matrix. The normalized capacity matrix is ​​then used for path feature aggregation. The output decoupled feature representation is fused with policy embedding, decoupled traffic embedding, and relative amplitude coefficient as input to the iterative hybrid expert decoder in the network traffic engineering header; S402, Multi-expert iterative decision generation; Using multi-round iterative routing decisions, for each iteration step to performing the following steps: Encoding global congestion state ; wherein, denotes the flow allocation proportion of the last round, denotes the current flow matrix, denotes the link capacity function; For each source-destination pair Build expert input ,in For path embedding, Indicates the magnitude scaling factor; Adopt Top- Sparse activation strategy, only select the highest weight of the Experts, the weight of the selected experts is renormalized: , activated experts parallel reasoning ; Weighted fusion of expert outputs ; Softmax normalization produces routing probabilities , output final routing decisions; S403, Constraint Satisfaction and Output Generation; The flow conservation constraint is automatically satisfied by Softmax: Calculate the actual load on each link and set the maximum link utilization (MLU) as the optimization objective: Output the final traffic allocation matrix. For each SD pair and path This indicates the proportion of traffic carried by this path.

2. The large model-based intelligent agent cross-environment generalization network traffic optimization method according to claim 1, characterized in that, Step S1 specifically includes: S101, historical flow matrix collection, collect historical flow window wherein denotes the traffic demand from source node to destination node , t denotes the time index, and L denotes the historical flow window length; S102、network topology and capacity information acquisition, acquire network topology graph where V is the set of nodes, E is the set of links, is the link capacity function; is the set of candidate paths for each source-destination pair generate a candidate path set ​ S103, Natural Language Policy Instruction Reception: Receives natural language instructions from the network administrator, and the instructions are sent to the large model agent for semantic parsing.

3. The large model-based agent cross-environment generalization network traffic optimization method of claim 2, wherein, In step S2, the cross-environment feature decoupling specifically includes: S201. Calculate the relative flow matrix and normalize it: Normalized flow value , Indicates the maximum traffic demand. This indicates the source node corresponding to the maximum traffic demand. This indicates the destination node corresponding to the maximum traffic demand. Indicates from the source node to the destination node The traffic demand is analyzed, the traffic distribution pattern is extracted, and the influence of absolute magnitude is eliminated, so that the large model is robust to changes in the absolute magnitude of traffic. S202. Calculate the link capacity and normalize it: , Represents the normalized capacity value of a link. Extract capacity patterns that contain only relative relationships, enabling large models to adapt to networks of different sizes, from Gbps-level wide area networks to Tbps-level data centers; S203, generating capacity-flow relative coefficient, calculating magnitude scale factor: , retaining scale information to support multi-objective optimization beyond maximum link utilization.

4. The cross-environment generalized network traffic optimization method based on large model agents according to claim 1, characterized in that, Step S3 specifically includes: S301, Network Model Context Protocol-Driven Agent Architecture Initialization; A pre-trained large language model is used as the inference engine. The agent interacts with the network system in a structured manner through the network model context protocol, enabling the agent to have network abstract perception, tool invocation, multi-step reasoning and feedback learning capabilities. S302, Three-dimensional strategy embedded spatial mapping; Decomposing and mapping natural language instructions to three-dimensional vectors The three-dimensional vectors contain three dimensions: Historical dependency Controlling the sensitivity of the response to historical traffic changes, wherein ignoring historical fluctuations, highly responding to historical changes to reserve margin for burst traffic; global robustness resilience to unforeseen traffic bursts, wherein is a best effort policy, is worst case robustness; Cost sensitivity , a trade-off preference between control performance and cost, wherein performance is prioritized over cost, cost is prioritized over a certain performance degradation; S303 and Few-shot small sample hints provide example-guided policy translation for large language models, guiding step-by-step reasoning through thought chain technology: identifying optimization objectives, extracting constraints, quantifying objective weights, mapping to a three-dimensional policy space, and generating policy embedding vectors; S304. Policy Verification and Transmission: The generated policy embedding is validated for reasonableness. If ambiguity or conflict is found, a multi-round dialogue is triggered for clarification. After successful validation, the policy embedding vector is transmitted. The reasonableness verification, which is issued to the pre-trained network traffic engineering head, includes range checking, conflict detection, and historical consistency checking.

5. The cross-environment generalization network traffic optimization method based on large model agents according to claim 2, characterized in that, In step S5, route allocation is issued through the network model context protocol of the large model, and network feedback and performance indicators are monitored, specifically including: S501, issues commands through the network model context protocol; The configuration is distributed to the network controller via the network model context protocol, and the controller then distributes the configuration to the network devices via the OpenFlow or NETCONF protocol. S502, Real-time performance monitoring; Continuously collect performance metrics, including link utilization, end-to-end latency, packet loss rate, maximum link utilization, and average link utilization. Anomaly detection includes congestion detection, link failure detection, and traffic mutation detection. S503, Asynchronous Strategy Adjustment and Continuous Optimization; A dual-time-scale control architecture is adopted: the frequency and delay of the fast control loop are set, and the frequency and delay of the slow control loop are set. The fast control loop and the slow control loop run independently to decouple the time scales. The adaptive strategy is adjusted to periodic triggering or event triggering. The large language model analyzes the feedback data to generate new strategy embeddings. The event triggering includes performance anomalies, topology changes, and traffic drift.

Images