Multi-link global routing method and system based on dynamic policy linkage

Abstract

The application relates to the technical field of software-defined wide area network and discloses a multi-link global routing method and system based on dynamic policy linkage, which is applied to an SD-WAN controller and used for solving the technical limitation problem that in a traditional SD-WAN routing method, decision basis is limited to a current state and policy rules are statically fixed. The method generates a predicted value and a confidence evaluation of future performance of a link through a time series prediction model; based on the confidence evaluation, historical, current and future state information is dynamically fused to form a link quality trend value; and the confidence evaluation is used to dynamically modulate the validity degree of a multi-dimensional service policy weight in evaluation; finally, path selection probability distribution is obtained through iterative optimization, and an optimal path is selected according to a service type for forwarding, so that routing decision is upgraded from static rule response to dynamic optimization based on risk perception, and the foresight, intelligence, stability and adaptive robustness of network routing decision in a complex dynamic environment are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of communication network technology, and particularly to intelligent routing and network optimization technology in software-defined wide area networks (SD-WAN), specifically to a multi-link global routing method and system based on dynamic policy linkage. Background Technology

[0002] Software-defined wide area networks (SD-WAN) achieve flexible scheduling of network resources through a centralized controller, with dynamic path selection being a key mechanism for ensuring service quality. Traditional SD-WAN routing methods (e.g., threshold-triggered switching based on real-time monitoring) primarily rely on current instantaneous link performance metrics (such as latency and packet loss rate) and preset static policy rules (such as fixed service priorities and cost weights) for decision-making. This approach has the following inherent technical limitations: First, its decision-making is limited to the current and historical state, lacking the ability to predict the network's short-term future behavior, resulting in an inability to proactively avoid congestion and causing frequent path switching and delayed responses. Second, its policy rules are statically preset and cannot dynamically adjust the trade-offs between multiple objectives such as quality, security, and cost based on the reliability of the network state. Therefore, in complex environments with rapidly changing traffic patterns and network states, these methods face significant challenges in terms of decision-making timeliness, stability, and robustness in dealing with uncertainty.

[0003] Therefore, how to break through the technical limitations of traditional SD-WAN routing methods, where decision-making is limited to the current state and policy rules are statically fixed, and provide a global path optimization method that can integrate future state predictions and quantitatively perceive and dynamically adapt to prediction uncertainties, thereby achieving a more forward-looking, more stable, and risk-controllable approach, has become an urgent technical problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a multi-link global routing method and system based on dynamic policy linkage, which can significantly improve the foresight, intelligence, stability, and adaptive robustness of network routing decisions when facing multi-objective constraints and uncertainties in complex dynamic environments.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-link global routing method based on dynamic policy linkage, applied to an SD-WAN controller, comprising: acquiring policy data from service, security, and cost management and quantizing it into a multi-dimensional policy weight vector; for each link in the network, generating a predicted value for its future short-term performance and a corresponding confidence assessment based on the link's historical performance time-series data and current real-time performance indicators through a time series prediction model; based on the confidence assessment, converting the historical performance time-series data, the current real-time performance indicators, and the predicted value into a comparable and unified link state feature representation, and dynamically determining the weight of each component in the link state feature representation in the fusion calculation based on the confidence assessment or a comprehensive confidence index derived therefrom, and performing weighted fusion to form the link quality of the link. The status value; based on the multi-dimensional strategy weight vector and the link quality status value of each link, a dynamic strategy-weighted evaluation value for each link is calculated using a preset evaluation function. The preset evaluation function is configured to: use the confidence assessment as an adjustment parameter to dynamically adjust the effectiveness of the multi-dimensional strategy weight vector in the evaluation, in order to adapt to the link selection risk caused by prediction uncertainty, thereby calculating the strategy-weighted evaluation value; based on the strategy-weighted evaluation value, a path selection probability model is established and initialized for each service data flow, and the parameters of the path selection probability model are adjusted by iterative updates to make the output path selection probability positively correlated with the strategy-weighted evaluation value, until the path selection probability model converges to obtain an optimized path selection probability distribution; based on the optimized path selection probability distribution and the type of service data flow, the path with the highest probability is selected for forwarding.

[0006] The present invention also provides a multi-link global routing system based on dynamic policy linkage, applied to an SD-WAN controller. The multi-link global routing system based on dynamic policy linkage corresponds one-to-one with the above-mentioned multi-link global routing method based on dynamic policy linkage, as detailed in the following embodiments.

[0007] The beneficial effects of this invention are as follows: First, by quantifying the policy data of business, security, and cost management into a multi-dimensional policy weight vector, and combining it with the link future performance prediction value and confidence assessment generated based on a time series prediction model, a dynamic link quality status value and a policy-weighted evaluation value for risk perception are constructed. Then, through iterative updates and convergence of the path selection probability model, the output path selection probability is positively correlated with the policy-weighted evaluation value, thereby obtaining a stable and adaptive optimized path selection probability distribution. Finally, the highest probability path is selected for forwarding based on the service type. Thus, by introducing future state prediction and uncertainty quantification, multi-time-dimensional state dynamic fusion based on confidence, dynamic modulation of policy weights, and a probabilistic path optimization learning mechanism, this invention systematically breaks through the technical limitations of traditional SD-WAN routing methods where decision-making is limited to the current state and policy rules are statically fixed. This significantly improves the foresight, intelligence, stability, and adaptive robustness of network routing decisions in complex dynamic environments facing multi-objective constraints and uncertainties. Attached Figure Description

[0008] Figure 1 A flowchart illustrating the multi-link global routing method based on dynamic strategy linkage provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the first sub-process of the multi-link global routing method based on dynamic strategy linkage provided in an embodiment of the present invention; Figure 3 This is a schematic block diagram of a multi-link global routing system based on dynamic strategy linkage provided in an embodiment of the present invention. Detailed Implementation

[0009] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0010] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings.

[0011] Example 1, please refer to Figure 1 , Figure 1 This is a flowchart illustrating a multi-link global routing method based on dynamic policy linkage provided in an embodiment of the present invention. Figure 1 As shown, in this embodiment, the method is applied to an SD-WAN controller, and the method includes the following steps S11-S16: S11: Acquire policy data from business, security, and cost management and quantify it into a multi-dimensional policy weight vector.

[0012] Multidimensional strategy weight vector: A fixed-dimensional vector composed of numerical elements, used to comprehensively represent the relative importance and trade-offs of multidimensional strategy requirements from business, security and cost management in a unified quantization space. Each dimension of the multidimensional strategy weight vector corresponds to a normalized strategy weight value (e.g., the value range is [0,1]). The magnitude of the strategy weight value reflects the priority of the strategy in the current network scheduling decision.

[0013] The implementation is as follows: First, raw policy data is obtained from three sources through the policy data acquisition interface: 1) Business side (such as network management system or configuration interface), obtaining business policy data reflecting application priorities, bandwidth requirements, service quality targets (SLAs), and other characteristics of business operation needs; 2) Security side (such as security policy library or compliance platform), obtaining security policy data regarding access control, encryption requirements, threat protection, and other characteristics of communication security rules; 3) Cost management side (such as billing or resource management system), obtaining cost management policy data representing network operation economic constraints, such as tariffs for each link, cost budget constraints, and link utilization optimization targets. It should be noted that the specific content of the business, security, and cost management policy data here refers to relevant content that can be understood by those skilled in the art based on conventional network management practices, and not all possible sub-items are elaborated here. Secondly, the core data structure defined in this invention, the "multi-dimensional policy weight vector," is a fixed-dimensional numerical vector such as W=[w1,w2,...,wn], where each dimension corresponds to a quantized and normalized policy sub-objective (e.g., w1 represents the weight for ensuring video conferencing quality, w2 represents the weight for enabling high-strength encryption, and w3 represents the weight for prioritizing the use of low-cost links). Based on this, the policy quantization processor parses and quantizes the aforementioned heterogeneous raw policy data. The core of the policy quantization processor lies in the configurable "policy-weight mapping model," whose working process includes the following steps: 1) Extracting key parameters from the raw policy data. For example, parsing "maximum latency required for video conferencing D_max" and "minimum bandwidth required for ERP" from business policy data; extracting "forced encryption identifier E_req" from security policy data; and calculating "the unit cost ratio r_cost of MPLS and Internet links" from cost policy data. 2) Assigning a basic weight w_i_base to each policy dimension according to a preset benchmark weight template related to the business scenario. For example, under the "Ensuring Core Experience" template, the w_i_base for "Video Conferencing Latency" is relatively high. 3) A key "policy reconciliation function" f is introduced. This function f takes the current network context state (such as the real-time load L and available bandwidth B_avail of each link) and the key parameters extracted above as input, and outputs a set of dynamic adjustment factors α_i. The specific implementation of this function f can be based on a set of predefined heuristic rules or a lightweight neural network. For example, when the real-time detection of Internet link latency is close to D_max, function f will output a larger α_latency to increase the base value of the latency weight; when the utilization rate of a high-cost link is extremely low, function f will output a larger α_cost suppression.4) Calculate the final weights w_i = normalize( w_i_base * α_i), where normalize is a normalization function (such as Softmax or scaling) to ensure that the sum of all weights is a fixed value (e.g., 1). This ultimately generates a multi-dimensional policy weight vector W = [w_1, w_2, ..., w_n]. Through the above process, the policy quantization processor transforms the qualitative policy description and dynamic network state into a quantitative, computable weight vector, providing standardized policy input for subsequent intelligent decision-making.

[0014] S12: For each link in the network, based on the historical performance time-series data and current real-time performance indicators of the link, a predicted value of its future short-term performance and the corresponding confidence level assessment are generated through a time series prediction model.

[0015] Link: In this embodiment of the invention, it refers to a communication channel between two nodes in a network whose performance can be measured independently and used for service forwarding. For example, a physical or logical communication channel with independent identification and measurable performance parameters established between a CPE device of an enterprise branch and the corresponding cloud network access point POP, or between two POP points.

[0016] Predicted values ​​and confidence assessments of future short-term performance: The estimated values ​​of link performance for future periods, generated by the time series forecasting model, and the metrics used to represent the reliability of these estimates, are called "confidence assessments." These confidence assessments can be expressed as a series of confidence assessments corresponding one-to-one with the predicted value series (e.g., the variance of the predicted values ​​at each future time step), or they can be further aggregated into a comprehensive confidence scalar (e.g., a weighted average of the series).

[0017] The implementation is as follows: First, for each link in the network topology, a link performance collector (a software agent deployed on network devices or independent probes) continuously works, collecting the link's current real-time performance metrics (such as latency, packet loss rate, etc.) at a fixed sampling period (e.g., 5 seconds). This data is then aligned and preprocessed (e.g., handling missing values, Z-score normalization, and other standard data cleaning steps) with stored historical performance time-series data (performance sequences over a past period) to form a complete, multi-dimensional input time series. Second, this input time series is fed into a pre-trained time series prediction model. In this embodiment, the time series prediction model employs an encoder-decoder architecture based on an attention mechanism (e.g., a temporal Transformer model). The encoder encodes the input time series to extract features; the decoder uses an autoregressive approach to iteratively generate a sequence of performance prediction values ​​for multiple future time steps based on the encoded features. The key improvement lies in the model's design and training to synchronously output confidence assessments. To achieve this, a confidence estimation module is integrated within the model. This module takes the hidden state of the decoder at each future time step as input, passes it through a sub-network (e.g., composed of fully connected layers), and outputs a measure of the uncertainty of the predicted value at that time step, such as the variance of the predicted value or its logarithmic form (log-variance). This uncertainty measure serves as the confidence assessment corresponding to that predicted value. The sequence of predicted values ​​and the sequence of confidence assessments correspond one-to-one in time, together forming a quantitative description of the future state of the link. Finally, the time series prediction model outputs a sequence of predicted values ​​for the short-term future performance of each link, along with the corresponding confidence assessment sequence.

[0018] S13: Based on the confidence assessment, the historical performance time series data, the current real-time performance indicators and the predicted values ​​are converted into comparable and unified link state feature representations. Based on the confidence assessment or the comprehensive confidence index derived therefrom, the weight of each component in the link state feature representation in the fusion calculation is dynamically determined and weighted fusion is performed to form the link quality status value of the link.

[0019] Comparable and unified representation of link state features: This represents the mapping of link performance information from different time scales (historical, current, future) and different data modalities (measured values, predicted values) to a common, fixed-dimensional numerical vector space through a series of mathematical transformations (including but not limited to normalization, moving average calculation, and trend encoding). Each dimension represents a standardized, dimensionless link state "feature", such as: "Z-score of historical average latency", "deviation of current latency from historical mean", and "trend slope of predicted latency".

[0020] The comprehensive confidence index represents a single scalar value obtained through further calculation of the confidence assessment. It is used to comprehensively quantify the reliability of the "predicted value of future short-term performance". Its calculation method can be designed according to specific needs; for example, it can be the arithmetic mean, geometric mean, or weighted average of confidence levels at multiple future time steps, taking into account the distance of the prediction time (the more recent the prediction, the greater its impact on the comprehensive index). This is used to compress confidence assessment information into a single key control parameter.

[0021] Link quality status value: This represents a single scalar value generated by further aggregating the weighted and fused link state feature representation (e.g., calculating the norm of the weighted feature vector or mapping it through a lightweight regression model). It is used to quantitatively characterize the overall quality level and stability of the link after integrating past performance, current status and future expectations.

[0022] The implementation is as follows: First, for each link, three core feature scalars are extracted and standardized from its historical performance time-series data (e.g., the key performance indicator sequence of the past 5 minutes), current real-time performance indicators, and predicted values ​​of future short-term performance to construct a comparable and unified representation of link state features: historical trend scalar H_trend (e.g., the slope obtained by linear fitting of historical time-series data), current instantaneous scalar I_now (e.g., the value after Z-score standardization of the current real-time performance indicator), and comprehensive prediction scalar F_agg (e.g., the value obtained by aggregating the predicted value sequence using a weighted average). This constructs a three-dimensional feature vector V = [H_trend, I_now, F_agg] as a comparable and unified representation of link state features. Second, dynamic weighted fusion is performed: 1) Calculate the comprehensive confidence index C_synth. For example, the confidence assessment sequence output by S12 is aggregated, such as by calculating its time-decayed weighted sum: C_synth = Σ (γ^(t-1) * c_t), where c_t is the confidence at the t-th future time step, and γ is the decay factor (e.g., 0.9). 2) Determine dynamic weights: Based on C_synth, the fusion weights (w_h, w_i, w_f) are calculated using a preset weight determination function g. The preset weight determination function g is designed such that: when C_synth is high, a larger w_f (predicted weight) and a smaller w_h (historical weight) are output, reflecting dependence on prediction; when C_synth is low, a larger w_i (current weight) and a smaller w_f are output, reflecting dependence on current measurements. The preset weight determination function g can be implemented using a piecewise linear function or a smoothing function based on the Sigmoid function to ensure that the weights change continuously and smoothly with C_synth. For example, a simple piecewise linear implementation can be: if C_synth > θ_high, then w_f = W_f_max; if C_synth < θ_low, then w_f = W_f_min; otherwise, linear interpolation is performed in the interval [θ_low, θ_high]. 3) Weighted fusion: Calculate the fusion feature F_fused = w_h * H_trend + w_i * I_now + w_f * F_agg. Finally, the fusion feature F_fused is aggregated to generate a single link quality status value. A specific implementation of the link quality status value Q is to calculate its L2 norm and take its negative: Q = -||F_fused||_2. The larger this value is, the better the overall link quality status. A lightweight neural network can also be used for mapping. Thus, the obtained Q is the link quality status value of the link, which quantifies the overall state of the link after fusing multi-time dimension information.

[0023] S14: Based on the multidimensional strategy weight vector and the link quality status value of each link, calculate the dynamic strategy weighted evaluation value of each link through a preset evaluation function. The preset evaluation function is configured to: use the confidence evaluation as an adjustment parameter to dynamically adjust the effectiveness of the multidimensional strategy weight vector in the evaluation, so as to adapt to the link selection risk caused by prediction uncertainty, thereby calculating the strategy weighted evaluation value.

[0024] Preset evaluation function: This refers to a predefined mathematical function that takes a multi-dimensional strategy weight vector and the link quality status value of each link as input, uses confidence assessment as an adjustment parameter, and outputs a scalar value, namely the strategy weighted evaluation value, after a series of mathematical operations. The preset evaluation function is not a simple linear weighted sum; it embodies the logic of dynamically balancing the three elements of "strategy objective", "link quality" and "predicted risk".

[0025] Strategy-weighted evaluation value: This represents a single numerical score for a specific link, calculated using a preset evaluation function. It comprehensively reflects the extent to which selecting this link can meet the enterprise's multi-dimensional strategic objectives (business, security, cost) under the current network conditions, while also considering the link's quality status and the selection risks caused by predictive uncertainty.

[0026] The implementation is as follows: First, for each candidate link i to be evaluated, three key inputs are obtained simultaneously: 1) a globally unified multi-dimensional policy weight vector W; 2) the link quality status value Q_i of candidate link i itself; 3) the confidence assessment C_i corresponding to the future performance prediction of candidate link i (usually represented by its comprehensive confidence index). Second, the preset evaluation function F(W, Q_i, C_i) is called for calculation. Before the calculation, the link attribute vector A_i of candidate link i needs to be constructed. A_i is a vector with the same dimensions as W, used to quantify the performance of candidate link i in each policy dimension. A specific construction method is as follows: 1) Business dimension score a_biz: The link quality status value Q_i is mapped to the [0,1] interval through a monotonically increasing function a_biz = f_biz(Q_i) (f_biz is a monotonically increasing function, such as Sigmoid, i.e., a_biz = sigmoid(k * Q_i), where k is the scaling factor) as a quantitative score of business experience. The higher Q_i is, the closer a_biz is to 1. 2) Security dimension score a_sec: A basic security score s_base (e.g., between 0 and 1) is set according to the security configuration of the link (such as encryption algorithm strength, access control rule matching degree). It can be further fine-tuned according to historical security event records or real-time threat intelligence to finally obtain a_sec. 3) Cost dimension score a_cost: Calculate a_cost = exp(-λ * c_i), where c_i is the unit traffic cost of the link (which needs to be normalized), and λ is the cost sensitivity coefficient. This function ensures that the lower the cost, the higher the score, and exhibits an exponential decay relationship, which aligns with the intuition of cost optimization. Finally, normalize the scores of a_biz, a_sec, a_cost, etc. (e.g., by dividing by their sum) to form the link attribute vector A_i. Then, calculate the original policy fit degree S_i_raw = W · A_i (i.e., the dot product of vector W and A_i) of candidate link i. The S_i_raw value reflects the degree to which candidate link i meets the policy objective without considering prediction risk. 2) Confidence Assessment-Driven Policy Weight Adjustment: A policy weight adjustment factor α(C_i) is introduced, which is a function of the confidence assessment C_i, with an output value between 0 and 1. The design principle is: the higher C_i, the closer α(C_i) is to 1; the lower C_i, the closer α(C_i) is to 0. One specific implementation is: α(C_i) = C_i^γ, where γ is the sensitivity coefficient (typical value γ = 1.0~2.0). The adjustment process includes: adjusting the original multidimensional policy weight vector W to an effective policy weight vector W_eff = α(C_i) * W, which means that when the predicted confidence assessment C_i is low, α(C_i) is small, and the effect of the policy weight W is significantly weakened ("effectiveness" is reduced); when C_i is high, α(C_i) is large, and the policy weight W is almost completely effective.3) Comprehensive assessment of risk adaptation: The policy adaptation degree of candidate link i is recalculated using the adjusted effective policy weight vector W_eff: S_i_eff = W_eff · A_i. Simultaneously, the link quality status value Q_i is introduced as an independent assessment of the link's basic reliability. Finally, the policy-weighted evaluation value V_i is calculated using the following weighted comprehensive formula: V_i = β * S_i_eff + (1 - β) * Q_i; Here, β is a balancing parameter between 0 and 1 (typical value β = 0.5~0.7), used to set the benchmark weight of policy fit and basic link quality in the final evaluation value. Through the above adjustment, when C_i is low, S_i_eff will become smaller, making V_i more inclined to be dominated by Q_i, that is, more emphasis is placed on the measured or historical comprehensive quality of the link (avoiding risk); when C_i is high, S_i_eff can fully reflect the policy objectives, achieving goal-oriented optimization.

[0027] S15: Based on the strategy-weighted evaluation value, establish and initialize a path selection probability model for each business data stream, and adjust the parameters of the path selection probability model through iterative updates to make the output path selection probability positively correlated with the strategy-weighted evaluation value, until the path selection probability model converges and the optimized path selection probability distribution is obtained.

[0028] A path selection probability model is a mathematical model used to calculate the probability of selecting each available path for a given business data flow. It is typically parameterized, with inputs including the characteristics of the business data flow (optional, such as business flow type) and the current network environment state (reflected by policy-weighted evaluation values, etc.). The output is a probability distribution vector, where each element corresponds to the probability of a path being selected. A path selection probability model can be any parameterized model capable of outputting a path selection probability distribution; for example, it can be a multi-armed gambling machine model based on the Softmax function, a deep policy network (such as an Actor network), or a Bayesian gambling machine model, etc.

[0029] The implementation is as follows: First, instantiate a path selection probability model for each business data stream or business type (such as a video session or file transfer stream). For example, the path selection probability model adopts the Softmax decision model. Given K available paths, for a business data stream f, the path selection probability model maintains a K-dimensional "preference score" vector s_f = [s_f1, s_f2, ..., s_fK]. The probability P_fi of path i being selected is calculated using the Softmax function: P_fi = exp(s_fi / τ) / Σ_{j=1 to K} exp(s_fj / τ); where τ is a temperature parameter (typically initial value τ=1.0), controlling the smoothness of the probability distribution (the larger τ is, the more uniform the distribution). Second, initialize the "preference score" s_fi using the policy weighted evaluation value V_i of each path. A direct method is to let the initial s_fi = V_i to ensure that at the initial moment, the output path selection probability is positively correlated with V_i, that is, the higher the policy weighted evaluation value, the greater the probability of the path being initially selected. Third, the iterative update phase begins. The parameters of the path selection probability model are iteratively updated through an online learning mechanism to dynamically adapt the output path selection probabilities to the network state and policy objectives. The update process can be triggered at fixed intervals or when the network environment changes significantly. The online learning mechanism aims to dynamically adjust the model parameters based on the differences between the actual performance of the paths and the policy evaluation, so that the probability distribution of the model output converges to a stable state that reflects the overall advantages and disadvantages of each path. There are various algorithms for implementing the above online learning, including but not limited to: methods based on policy gradients, methods based on value iteration (such as Q-Learning), and methods based on Bayesian inference (such as Thompson Sampling). This invention does not limit the specific algorithm for implementing the online learning mechanism; any optimization algorithm that can achieve the above objectives is applicable here.

[0030] As an example of iterative updates in online learning, the following details the implementation of the policy gradient optimization method: its core lies in constructing an optimization objective that reflects the degree of matching between the comprehensive evaluation value of the path (characterized by the policy weighted evaluation value) and the actual probability of the path being selected, and adjusting the model parameters through gradient ascent to optimize this objective. The specific implementation is as follows: 1) Update signal calculation. In each iteration, the path selection probability model selects a path for business flow forwarding based on the probability distribution P = [P_1, P_2, ..., P_K] (where P_i is the probability of selecting path i, and K is the total number of paths) output by the current parameters. Based on the actual performance feedback of this forwarding and the policy weighted evaluation value V_i of the corresponding path, an advantage function value A_i is calculated to quantify the superiority or inferiority of selecting path i relative to the average performance. A specific method for calculating the advantage function is: A_i = R_i - V_i, where R_i is the immediate utility value calculated based on actual performance data (such as latency, packet loss rate). 2) Parameter update rules. The path selection probability model parameters (i.e., the "preference score" vector s = [s_1, s_2,..., s_K] for each path) are updated based on the advantage function value. This example uses an update rule based on policy gradient: First, under the current parameters, the expected comprehensive evaluation value for selecting each path is defined as V_avg = Σ_{i=1}^{K} P_i * V_i. Second, the preference score s_i is adjusted according to the following update rule: s_i := s_i + η * (V_i - V_avg) * P_i; Where η is the learning rate, a positive configurable parameter (typically ranging from 0.01 to 0.1). The technical meaning of this parameter update rule is as follows: for path i, if its policy-weighted evaluation value V_i is higher than the current expected value V_avg, its preference score s_i receives a positive increment, thus its selection probability P_i will increase in the next calculation; conversely, if V_i is lower than V_avg, s_i receives a negative increment, thus P_i decreases. The update magnitude is modulated by the current probability P_i. For paths with lower current probabilities, the score adjustment magnitude is relatively larger, which helps to promote the exploration of potentially better paths. 3) Exploration-Utilization Trade-off and Convergence. To balance exploration (trying different paths) and utilization (selecting the current optimal path), a temperature parameter annealing mechanism can be introduced during the iteration process. Specifically, the Softmax function used in calculating the probability distribution includes a temperature parameter τ (P_i = exp(s_i / τ) / Σ_j exp(s_j / τ)). In the early stages of iteration, a larger τ value can be set to make the probability distribution more uniform and encourage exploration. As iteration progresses, the τ value is gradually decreased (for example, updating τ := τ * ρ in each round, where the decay factor ρ is a constant slightly less than 1, such as 0.99), so that the probability distribution gradually focuses on the path with high policy weighted evaluation value, promoting policy convergence. Finally, when the model parameters tend to stabilize after multiple iterations, i.e., when the convergence condition is met, the iteration is stopped. The convergence condition can be determined by one of the following methods: 1) In M consecutive iterations (M is a preset positive integer, such as 5), the change in the preference score vector s (such as the L2 norm) is less than the preset minimum positive threshold ε (for example, ε can be on the order of 1e-4); 2) The probability distribution P output by the model remains stable in multiple consecutive iterations. At this time, the path selection probability model has reached the convergence state, and its stable output probability distribution is the optimized path selection probability distribution.

[0031] As another example of iterative updates in online learning, a Bayesian gambling machine framework can be adopted. In this framework, the policy-weighted evaluation value V_i of each path is considered as the prior mean of its performance, and the reciprocal of the confidence evaluation C_i is associated with the prior uncertainty. After each path is selected, the posterior distribution of the path's performance is updated based on the actual utility R_t. The path selection probability can be calculated based on the posterior distribution (e.g., using Thompson Sampling). This method can also achieve intelligent exploration and utilization based on uncertainty.

[0032] S16: Based on the optimized path selection probability distribution and the type of business data flow, select the path with the highest probability for forwarding.

[0033] The implementation is as follows: First, for the service data flow to be forwarded, the optimized path selection probability distribution P_opt is obtained from the converged path selection probability model. The optimized path selection probability distribution P_opt is a probability vector, where each element corresponds to the probability of a candidate path being selected. Further, the optimized path selection probability distribution P_opt can be adaptively adjusted according to the type of service data flow. For example, through a preset service type strategy, the path selection probability that does not meet the corresponding service quality requirements is set to zero, and the selection probabilities of the remaining paths are re-normalized. Second, based on the finally determined optimized path selection probability distribution P_opt, path selection and data forwarding are performed. In a typical implementation, the path with the highest probability in P_opt is usually selected as the forwarding path for each service data flow, and the network device is controlled to forward the data packets of that service data flow along that path.

[0034] In this embodiment of the invention, firstly, by quantifying policy data related to business, security, and cost management into a multi-dimensional policy weight vector, and combining this with link future performance predictions and confidence assessments generated based on a time series prediction model, a dynamic link quality status value and a policy-weighted evaluation value for risk perception are constructed. Then, through iterative updates and convergence of the path selection probability model, the output path selection probability is made positively correlated with the policy-weighted evaluation value, thereby obtaining a stable and adaptive optimized path selection probability distribution. Finally, the highest probability path is selected for forwarding based on the business type. Thus, by introducing future state prediction and uncertainty quantification, and based on confidence... The multi-time-dimensional dynamic fusion of states, dynamic modulation of policy weights, and probabilistic path optimization learning mechanism systematically break through the technical limitations of traditional SD-WAN routing methods, which are limited to the current state and have statically fixed policy rules. It transforms the traditional static and reactive routing mechanism into an adaptive global intelligent routing control mode based on dynamic quantification of multi-source policies, forward-looking prediction of link states and adaptation to uncertainty risks, and smooth load balancing through probabilistic models. This significantly improves the forward-looking, intelligent, stable, and adaptive robustness of the network's routing decisions in complex dynamic environments facing multi-objective constraints and uncertainties.

[0035] In one embodiment, the time series prediction model is a sequence-to-sequence model based on an attention mechanism. Generating predicted values ​​of future short-term performance and corresponding confidence assessments through the time series prediction model includes: encoding the historical performance time-series data and the current real-time performance index into a feature sequence; using the decoder of the time series prediction model, with the feature sequence as context information and the output of the previous decoding time step as an autoregressive input, iteratively generating a sequence of performance prediction values ​​for N future time steps, where N is a natural number; simultaneously, through the confidence estimation module within the time series prediction model, for each future time step of the performance prediction value generated by the decoder, calculating and outputting the confidence interval or variance of the performance prediction value based on the decoder's hidden state and attention context at the future time step; the confidence intervals or variances output at all future time steps constitute a confidence assessment sequence that corresponds one-to-one with the performance prediction value sequence in time.

[0036] Sequence-to-sequence models based on attention mechanisms are deep learning architectures suitable for predicting link performance time series. Their core lies in mapping the input sequence to a feature sequence through an encoder, and then using a decoder combined with an attention mechanism to generate future prediction sequences in an autoregressive manner. The attention mechanism enables the model to dynamically focus on the most relevant historical information, effectively capturing cycles, trends, and sudden fluctuations. It has significant advantages over traditional models (such as ARIMA and simple RNNs) in modeling long-range dependencies and complex patterns.

[0037] The confidence estimation module is a key substructure integrated within the sequence-to-sequence model, working in parallel and synchronously with the decoding process. It is used to quantify the uncertainty of the model's own predictions. It takes the internal state generated by the decoder at each future time step (i.e., the decoder's hidden state) and the attention context vector obtained by weighted aggregation through the attention mechanism as input. After transformation by a lightweight sub-network (e.g., a multilayer perceptron), it directly outputs a measure of the uncertainty of the predicted value at that time step. In a preferred implementation of the present invention, this measure is expressed as the variance of the predicted value (or its logarithmic form).

[0038] The implementation is as follows: First, preprocessed and aligned historical performance time-series data of the links, along with current real-time performance metrics, are input to the encoder of the sequence-to-sequence model (e.g., composed of multi-layer Transformer modules or bidirectional recurrent neural networks). This encoder encodes the input sequence and outputs a feature sequence that represents the overall information of the input sequence. Second, the decoder is initialized, using a specific start marker and the aforementioned feature sequence as input to generate the performance prediction value for the first future time step. For the t-th future time step (t>1), the decoder combines the prediction value generated at the (t-1)-th future time step (as an autoregressive input) with an attention context vector recalculated through an attention mechanism, focusing on the most relevant historical information at the current time step, to iteratively generate the prediction value for the t-th step. This process is repeated until a performance prediction value sequence of a preset length N (N is a natural number) is generated. Third, at each step of the above decoding process, the confidence estimation module is synchronously activated and receives the hidden state and attention context vector of the decoder at the same future time step. It then calculates the uncertainty measure (e.g., variance) of the current step prediction value through its internal sub-network. Ultimately, for each future time step performance prediction output by the decoder, the confidence estimation module outputs a corresponding uncertainty metric, i.e., the corresponding confidence assessment, thus forming a confidence assessment sequence that is strictly one-to-one with the performance prediction sequence in terms of time steps.

[0039] To optimize the prediction task for network link performance, a specific implementation of a sequence-to-sequence model based on an attention mechanism is provided below: 1) Model structure customization: The encoder and decoder adopt a 2-layer bidirectional gated recurrent unit (Bi-GRU) with 128 hidden units; the attention mechanism is improved by introducing a learnable decay bias term based on time difference into the standard calculation to adaptively balance recent and long-term historical information; the confidence estimation module is implemented by a lightweight fully connected network, which takes the decoder hidden state and attention context as input and outputs the log-variance of the predicted value. 2) Offline training is performed using a dataset containing historical link performance data. The loss function is a composite loss function: Loss = MSE (mean squared error) + λ * NLL (negative log-likelihood loss) to simultaneously optimize prediction accuracy and confidence calibration. The typical value of the balancing hyperparameter λ is 0.5. 3) Example of key parameters: {Prediction length N: typically corresponds to the next 30 seconds to 5 minutes; number of encoder / decoder layers: 2 layers; number of hidden units: 128; λ range: 0.1 to 1.0}.

[0040] In this embodiment of the invention, by employing a sequence-to-sequence model based on an attention mechanism that integrates a confidence estimation module, the accurate prediction of future link performance and simultaneous quantification of uncertainty are achieved. This provides a joint risk input of "predicted value + confidence" for the downstream link state fusion and risk assessment steps, fundamentally enhancing the system's ability to perceive predicted risks and improving the adaptability and robustness of the global routing method.

[0041] In one embodiment, based on the confidence assessment, the historical performance time-series data, the current real-time performance indicators, and the predicted values ​​are converted into comparable, unified link state feature representations. Then, based on the confidence assessment or a comprehensive confidence index derived therefrom, the weights of each component in the link state feature representation in the fusion calculation are dynamically determined, and weighted fusion is performed to form the link quality status value of the link. This includes: inputting the performance prediction value sequence {F(t)} and its corresponding confidence assessment sequence {C(t)} into a preset aggregation function, outputting a comprehensive prediction scalar F_agg and a comprehensive confidence scalar C_agg, wherein the preset aggregation function is configured as: confidence C(t). The lower the predicted value F(t) at a lower time step, the smaller its contribution weight to F_agg. A historical trend scalar H_trend, representing the historical trend, is extracted from the historical performance time-series data. The current real-time performance indicator is used as the current instantaneous scalar I_now. Based on the historical trend scalar H_trend, the current instantaneous scalar I_now, and the comprehensive prediction scalar F_agg, a three-dimensional feature vector is constructed as the link state feature representation. According to the comprehensive confidence scalar C_agg, the weights W_h, Wi, and W_f of each scalar component H_trend, I_now, and F_agg in the fusion calculation are dynamically determined. The link quality status value S is calculated by weighted summation S = W_h * H_trend + Wi * I_now + W_f * F_agg.

[0042] The predefined aggregation function is a mathematical function used to aggregate the sequence of future performance predictions {F(t)} and its corresponding confidence evaluation sequence {C(t)} into a single scalar. Its core design lies in weighting the performance predictions according to their confidence levels. Performance predictions with higher confidence levels have higher weights when calculating the comprehensive prediction scalar F_agg, while performance predictions with lower confidence levels have their weights suppressed. This allows F_agg to more robustly represent future trends less affected by uncertainty. Simultaneously, the predefined aggregation function also outputs a comprehensive confidence scalar C_agg representing the overall prediction reliability (e.g., a weighted average or minimum value of the sequence {C(t)}).

[0043] The historical trend scalar H_trend is a scalar extracted from the historical performance time-series data of the link to quantify its recent trend. A typical calculation method is to perform a linear fit on the historical performance time-series data within a specified historical window (such as the latency values ​​of the past M sampling points), and use the slope of the fitted line as H_trend. For example, a positive slope indicates a deteriorating trend in performance indicators (such as increasing latency), while a negative slope indicates an improving trend.

[0044] The current instantaneous scalar I_now is a scalar that characterizes the current real-time performance of the link. Its value is usually calculated based on one or more recent sampled values ​​(such as the average of the last three samples) and is used to represent the instantaneous state of the link "at this moment".

[0045] The implementation is as follows: First, the performance prediction sequence {F(t)} (e.g., the predicted delay values ​​[162, 164, …, 168] ms for the next 12 time steps) and the corresponding confidence evaluation sequence {C(t)} (e.g., the standard deviation sequence [5, 6, …, 6.3] ms) are input into a preset aggregation function to perform the following calculations: 1) Calculate the comprehensive prediction scalar F_agg: The preset aggregation function calculates the weight w_t for each F(t) based on {C(t)}. One specific implementation is: w_t = exp(k * C(t)) / Σ_{j=1}^{N} exp(k * C(j)), where k is a scaling factor (typically k=5.0). The softmax function ensures that the sum of all weights is 1. Then, the weighted average is calculated: F_agg = Σ_{t=1 to N} (w_t * F(t)) Through this weighting method, low-confidence predictions are automatically assigned low weights, making F_agg robust to prediction noise and more reliably representing the overall future level. 2) Calculate the comprehensive confidence scalar C_agg: C_agg can be the mean, median, or a weighted average considering time decay of the {C(t)} sequence (e.g., higher weight for recent confidence). One specific implementation is: C_agg = Σ_{t=1 to N} (γ^{t-1} * C(t)) / Σ_{t=1 to N} γ^{t-1}, where γ is the decay factor (0<γ≤1, typical value γ=0.9, degenerates into a simple average when γ=1), indicating that the more recent the prediction confidence, the greater the impact on the comprehensive confidence scalar. Secondly, extract the historical trend scalar H_trend from historical performance time-series data (such as the past 30 sampling points). One specific implementation is to perform a first-order linear regression on the historical performance time-series data to obtain the slope β. H_trend can be set to β or some transformation thereof (such as tanh(β) to limit the range). At the same time, the current real-time performance indicator (such as the mean of the last 3 samples) is subjected to the same standardization process (e.g., subtracting the historical mean and dividing by the historical standard deviation) to obtain the current instantaneous scalar I_now. Thirdly, combine the three scalars H_trend, I_now, and F_agg, which are already at a comparable scale (such as all being Z-scores or normalized to [0,1]), into a three-dimensional feature vector V = [H_trend, I_now, F_agg].Fourth, based on the comprehensive confidence scalar C_agg, the weights W_h, Wi, and W_f are dynamically calculated through a preset weight allocation function g(C_agg). The weight allocation function g(C_agg) is designed as follows: when C_agg is high (e.g., >0.7), it indicates that the overall future prediction is reliable, so W_f is given a higher weight (e.g., 0.5), Wi is given a medium weight (0.3), and W_h is given a lower weight (0.2), reflecting the decision-making tendency that depends on the prediction; when C_agg is low (e.g., <0.3), it indicates that the prediction uncertainty is high, so Wi is given the highest weight (e.g., 0.6), W_f is given the lowest weight (0.1), and W_h is given a certain weight (0.3), and the decision depends more on the current measured and historical baseline. The weight allocation function g(C_agg) can be a piecewise linear function or a sigmoid function based on C_agg. Finally, the linear weighted summation is performed according to the formula S = W_h * H_trend + W_i * I_now + W_f * F_agg, and the calculated result S is the link quality status value of the link. The fusion strategy is adaptively adjusted according to the overall reliability of the prediction.

[0046] In this embodiment of the invention, firstly, a preset aggregation function weighted by confidence level effectively suppresses noise caused by high uncertainty in the performance prediction value sequence, making the output comprehensive prediction scalar more reliably characterize the robust trend of the link in the future. Secondly, through a dynamic weight allocation mechanism based on comprehensive confidence level, the system can intelligently switch and weight between three decision modes: "relying on historical baselines," "believing in current measurements," and "adopting future predictions," achieving adaptive adjustment of the fusion strategy to prediction uncertainty. In summary, the generated link quality status value is a high-quality status assessment value that integrates information from multiple time dimensions and whose fusion weights are intelligently adjusted according to prediction risk, providing accurate, robust, and highly adaptable key inputs for subsequent risk assessment and route selection decisions.

[0047] In one embodiment, the preset aggregation function outputs the comprehensive prediction scalar F_agg and the comprehensive confidence scalar C_agg through the following steps: For each future time step t = 1 to N, the aggregate weight ω(t) of the predicted value F(t) of the future time step is calculated based on its confidence assessment value C(t), where ω(t) is a monotonically increasing function of C(t), and the monotonically increasing function is: ω(t) = 1 / (1 + exp(-η * (C(t) - C_th))), where η is a preset scaling factor and C_th is a preset confidence threshold; The comprehensive prediction scalar F_agg is calculated according to the following formula: ; The comprehensive confidence scalar C_agg is calculated according to the following formula: ; where N is the total number of predicted future time steps, ω(t) is the aggregation weight of the t-th future time step calculated based on C(t), F(t) is the performance prediction value of the t-th future time step, and C(t) is the confidence evaluation value of the t-th future time step.

[0048] Monotonic increasing function: used to map the confidence evaluation value C(t) to the aggregation weight ω(t), satisfying that ω(t) increases as C(t) increases. A preferred implementation is the parameterized Sigmoid function: ω(t) = 1 / (1 + exp(-η * (C(t) - C_th))); this function smoothly maps C(t) to the interval (0, 1), and controls the weight distribution behavior through two adjustable parameters: 1) preset scaling factor η (η>0): controls the sensitivity of the weight ω(t) to the change of C(t). The larger the η value, the steeper the function curve near the threshold C_th, and the more rapid the weight change. 2) preset confidence threshold C_th (0<C_th<1): defines the decision boundary of weight distribution. When C(t) = C_th, ω(t) = 0.5; when C(t)>C_th, ω(t)>0.5; otherwise ω(t)<0.5.

[0049] The implementation is as follows: First, preset an aggregation function to obtain a sequence of performance prediction values {F(1), F(2),..., F(N)} with length N and the corresponding sequence of confidence evaluations {C(1), C(2),..., C(N)}. Second, for each future time step t (t ranges from 1 to N) in the sequence, according to its corresponding confidence evaluation value C(t), calculate the aggregation weight ω(t) of the prediction value F(t) through a monotonic increasing function, and the calculation is as follows: ω(t) = 1 / (1 + exp(-η * (C(t) - C_th))). Thus, through the Sigmoid function, C(t) is non-linearly mapped to a weight in the interval (0,1), and its behavior is regulated by the parameters η and C_th. Third, use the weight sequence {ω(1), ω(2),..., ω(N)} calculated in the previous step to perform a weighted average on the original sequence of performance prediction values to obtain a single scalar F_agg representing the overall future performance level of the link. The calculation formula is: F_agg = (ω(1)*F(1) + ω(2)*F(2) +... + ω(N)*F(N)) / (ω(1) + ω(2)+... + ω(N)); This step enables robust aggregation of the performance prediction sequence across multiple time steps: low-confidence (potentially inaccurate) predictions have negligible impact on the final F_agg due to their small weight ω(t); high-confidence predictions dominate the calculation of F_agg, effectively filtering out noise and unreliable parts in the performance prediction sequence, making F_agg more reflective of the link's future reliable performance trend.

[0050] Furthermore, the overall confidence scalar C_agg is calculated: Using the same weight sequence {ω(t)}, the original confidence assessment sequence is weighted and averaged to obtain a single scalar C_agg representing the overall confidence of the prediction over the entire future period. The calculation formula is as follows: C_agg = (ω(1)*C(1) + ω(2)*C(2) + ... + ω(N)*C(N)) / (ω(1) + ω(2)+ ... + ω(N)); This step calculates the weights of the overall confidence scalar C_agg, which, like the weight ω(t) used to calculate the overall prediction scalar F_agg, originates from the same weighting function based on the confidence C(t). Therefore, time steps with higher confidence C(t) have a larger proportion of predicted values ​​F(t) in F_agg, and their confidence values ​​C(t) also have a correspondingly higher weight in C_agg. Thus, C_agg is a self-weighted overall confidence index that is inherently consistent with the prediction aggregation logic; it focuses more on reflecting the overall confidence level represented by high-confidence time steps.

[0051] In this embodiment of the invention, on the one hand, the smoothness of the weighting function ω(t) and the adjustable parameters (η, C_th) are used to suppress low-confidence predictions and strengthen high-confidence predictions, ensuring the robustness and accuracy of F_agg; on the other hand, by using homogeneous weighting, C_agg is made to reflect the confidence level of the high-confidence part more, providing a more sensitive and reliable adjustment signal for subsequent decision-making, effectively solving the technical problem of extracting reliable core information from uncertain predictions.

[0052] In one embodiment, the weights W_h, Wi, and W_f of each scalar component H_trend, I_now, and F_agg in the fusion calculation are dynamically determined based on the comprehensive confidence scalar C_agg, including: Get the preset baseline weight W_h_base corresponding to H_trend, the preset baseline weight Wi_i_base corresponding to I_now, and the preset baseline weight W_f_base corresponding to F_agg; A prediction confidence factor τ is calculated based on C_agg, where τ is a monotonically increasing function of C_agg, τ = (C_agg - C_min) / (C_max - C_min), where C_max is the preset upper bound of the confidence assessment value, C_min is the preset lower bound of the confidence assessment value, and τ∈[0, 1]. Based on τ, adjust the baseline weights W_h_base, Wi_base, and W_f_base as follows: The fusion weight W_h = W_h_base + (1 - τ) * μ of the historical trend scalar H_trend, where μ is a preset compensation coefficient; The fusion weight W_i = W_i_base for the current instantaneous scalar I_now; The fusion weight W_f = W_f_base * τ for the comprehensive prediction scalar F_agg; The adjusted weights W_h, W_i, and W_f are normalized so that their sum is 1.

[0053] Preset baseline weights (W_h_base, Wi_i_base, W_f_base): These represent the pre-set weight values ​​reflecting the fundamental importance of the three components of the three-dimensional feature vector (historical trend scalar H_trend, current instantaneous scalar I_now, and comprehensive prediction scalar F_agg) during system initialization or static configuration. The preset baseline weights represent the default fusion strategy for each component when current prediction uncertainty is not considered. For example, setting W_h_base=0.2, Wi_i_base=0.5, and W_f_base=0.3 indicates that by default, the system relies more on the current instantaneous state.

[0054] Prediction confidence factor (τ): This is a scalar value ranging from [0,1] obtained by linearly mapping the comprehensive confidence scalar C_agg. τ quantifies the overall level of confidence in the future prediction. When C_agg reaches the preset upper bound C_max, τ=1, indicating complete confidence in the prediction; when C_agg drops to the preset lower bound C_min, τ=0, indicating complete distrust in the prediction; in between, τ changes linearly. τ is used to adjust the rate of change of the weights based on the comprehensive confidence C_agg.

[0055] Preset compensation coefficient (μ): This represents a configurable parameter used to dynamically adjust the scalar weights of historical trends. When the prediction confidence is low (τ is small), the system will reduce the prediction weight W_f and increase the weights of historical trends W_h according to (1-τ)*μ. The magnitude of parameter μ determines the strength of the compensation for historical experience weights in this situation. By configuring μ≠1, the weight adjustment is not a strict zero-sum exchange, but rather an asymmetric and adjustable weight redistribution strategy among historical, current, and future information sources, thereby improving the robustness of the system's fusion decision-making in uncertain environments.

[0056] The implementation is as follows: First, the pre-configured baseline weights for the three feature components, W_h_base, Wi_base, and W_f_base, are read from the configuration storage. The sum of the three pre-configured baseline weights is usually 1, reflecting the system's default strategy bias (e.g., favoring the current state or historical experience). Second, the prediction confidence factor τ is calculated: the comprehensive confidence scalar C_agg is obtained, and τ is calculated using a linear mapping formula based on the pre-defined upper and lower bounds of the confidence, C_max and C_min (e.g., C_max=1.0, C_min=0.2). τ = (C_agg - C_min) / (C_max - C_min); Wherein, C_min is a preset lower bound for the confidence level used to define the "completely unreliable" state. Its value can be set according to the statistical quantiles of the historical confidence assessment sequence, with a typical range of 0.1 to 0.3. After calculation, τ is truncated to ensure that τ ∈ [0, 1]. The value of τ directly and linearly reflects the overall confidence level of the current prediction. Third, the weights of each component are dynamically adjusted according to τ: 1) Adjust the scalar weight of the comprehensive prediction: W_f = W_f_base * τ; 2) Adjust the scalar weight of the historical trend: W_h = W_h_base + (1 - τ) * μ; 3) Adjust the scalar weight of the current instantaneous value: W_i = W_i_base (or a small adjustment can be made, such as W_i = W_i_base + κ * (1 - τ), where κ is another small coefficient, but preferably remains unchanged); where the preset compensation coefficient μ is used to control the compensation intensity of the historical weights, with a typical range of 0.1 to 0.3. Fourth, normalization processing: W_h' = W_h / (W_h + W_i + W_f); Wi' = W_i / (W_h + W_i + W_f); W_f' = W_f / (W_h + W_i + W_f); The normalized W_h', Wi', and W_f' are the dynamic fusion weights W_h, Wi, and W_f used to calculate the link quality status value S.

[0057] In this embodiment of the invention, the following technical effects are achieved through the above-mentioned dynamic weight adjustment mechanism based on τ: 1) Prediction risk perception and avoidance: W_f decays linearly with τ, ensuring that the dependence on prediction information is automatically reduced when prediction uncertainty is high, thus avoiding decision-making risks. 2) Enhanced decision robustness: Through the compensation mechanism controlled by μ, the decision weights are intelligently redistributed to historical trend information when the prediction is unreliable, improving the system's fallback capability and stability.

[0058] In one embodiment, please refer to Figure 2 , Figure 2 This is a schematic diagram of the first sub-process of the multi-link global routing method based on dynamic strategy linkage provided in an embodiment of the present invention. Figure 2 As shown, in this embodiment, the preset evaluation function is configured to: use the confidence assessment as an adjustment parameter to dynamically adjust the effectiveness of the multi-dimensional strategy weight vector in the evaluation, so as to adapt to the link selection risk caused by prediction uncertainty, specifically including: S21: calculate the strategy sensitivity modulation coefficient according to the confidence assessment, wherein the strategy sensitivity modulation coefficient is positively correlated with the confidence assessment value; S22: perform a power operation on each weight component in the multi-dimensional strategy weight vector with the strategy sensitivity modulation coefficient to obtain the modulated strategy weight component, wherein the power exponent corresponding to the weight component strongly correlated with future performance is greater than the power exponent corresponding to the weight component strongly correlated with historical or current performance, and the weight component strongly correlated with future performance refers to the strategy weight component whose value directly affects the evaluation result of link latency and jitter performance indicators, and the weight component strongly correlated with historical or current performance refers to the strategy weight component whose value directly affects the evaluation result of link historical availability and current packet loss rate indicators; S23: use the modulated strategy weight component to perform a weighted linear combination on the corresponding link quality status value component to calculate the strategy weighted evaluation value.

[0059] The implementation is as follows: First, obtain the multi-dimensional policy weight vector W = [w_1, w_2, ..., w_m], and obtain the link quality status value S (or its decomposed components corresponding to the policy dimension) for each link, as well as the confidence assessment (preferably using the comprehensive confidence scalar C_agg). Second, calculate the policy sensitivity modulation coefficient β: β is calculated based on the confidence assessment (preferably the comprehensive confidence scalar C_agg), where β ∈ [0, 1] and is positively correlated with C_agg. One implementation is to let β = C_agg, that is, directly use the comprehensive confidence scalar as the modulation coefficient. To enhance robustness, the following formula can also be used: β = max(β_min, C_agg), where β_min is the preset minimum modulation coefficient (typical value 0.05 to 0.15); or the smoothing formula can be used: β = α * β_prev + (1-α) * C_agg, where α is the smoothing factor, with a typical value of 0 to 0.2, and β_prev represents the β value of the previous calculation cycle (or the previous time step). Third, the multidimensional strategy weight vector is modulated by exponential operation: First, for the i-th weight component w_i in the multidimensional strategy weight vector W, a power exponent γ_i is pre-assigned to it (γ_i>= 0, configurable range 0.5 to 3.0). The magnitude of γ_i characterizes the sensitivity of the weight component to prediction uncertainty. Weight components that are strongly correlated with future performance (such as latency optimization weights and jitter optimization weights) are assigned larger γ_i (e.g., γ_i = 2.0 or γ_i = 3.0), while weight components that are strongly correlated with historical or current performance (such as historical availability weights and cost weights) are assigned smaller γ_i (e.g., γ_i = 0.5 or γ_i = 1.0). The cost weight may also have some dependence on prediction (e.g., predicting future congestion leading to cost changes), so its γ_i can be set to an intermediate value, such as 1.0. Secondly, power-law modulation is performed to calculate the modulated weight w_i'. A general formula is: w_i' = w_i * (β)^(γ_i). When the prediction confidence is high (β is close to 1), (β)^(γ_i) is close to 1 for all components, and the modulated weight w_i' is almost equal to the original weight w_i, meaning the strategy is fully effective. When the prediction confidence decreases (β decreases), due to the amplification effect of the power exponent γ_i, the decay rate of future relevant weights (larger γ_i) is much faster than that of historical / current relevant weights (smaller γ_i). For example, if β = 0.5, for a future weight with γ = 3, its modulation factor is 0.125, resulting in a sharp decay; for a historical weight with γ = 0.5, its modulation factor is ~0.707, resulting in a moderate decay. This achieves differentiated and non-linear sensitivity modulation of the strategy weights. Fourth, use the modulated weights to calculate the policy weighted evaluation value: use the modulated policy weight vector W' = [w_1', w_2', ..., w_m'] for evaluation.Specifically, the link quality status value S needs to be decomposed or mapped into an attribute vector A = [a_1, a_2,..., a_m] corresponding to the policy dimension, where a_i represents the link's score on the i-th policy dimension (e.g., latency score, cost score). Then, the policy weighted evaluation value V is calculated through a weighted linear combination: V = w_1' * a_1 + w_2' * a_2 + ... + w_m' * a_m; since w_i' has been modulated according to the prediction uncertainty, the final calculated V value reflects the degree of policy satisfaction after risk adaptation.

[0060] Further, the formula for calculating the policy sensitivity modulation coefficient β is: β = σ / (σ + θ), where σ is the quantized value of the confidence assessment and θ is the preset sensitivity threshold; the power operation is: for the i-th type policy weight w_i that is strongly correlated with future performance, the modulation is w_i' = w_i ^ β; for the j-th type policy weight w_j that is strongly correlated with historical or current performance, the modulation is w_j' = w_j ^ (β * δ), where 0 < δ < 1 is the attenuation factor.

[0061] The implementation is as follows: First, calculate the policy sensitivity modulation coefficient β: Obtain the quantization value σ of the confidence evaluation (for example, σ = C_agg, where C_agg is the comprehensive confidence scalar), read the preset sensitivity threshold θ from the system configuration, and calculate the current policy sensitivity modulation coefficient β according to the formula β = σ / (σ + θ). Second, perform differential power operation modulation: According to the preset mapping relationship, identify which weight components in the vector belong to the i-th type of policy weight (denoted as w_i) that is strongly correlated with future performance, and which belong to the j-th type of policy weight (denoted as w_j) that is strongly correlated with historical or current performance. In the modulation process, as a specific implementation, the weight components w_i and w_j in the multi-dimensional policy weight vector are defined as penalty coefficients, and their value ranges are within the interval (0, 1). The smaller the value, the more important the performance target of the corresponding policy dimension or the less desired it is to occur (for example, for a service that is extremely sensitive to latency, its latency penalty coefficient can be set to 0.1; for a service that is insensitive to cost, its cost penalty coefficient can be set to 0.9). The modulation operation is as follows: 1) For the i-th type of policy weight w_i that is strongly correlated with future performance, modulate it to w_i' = w_i ^ β; Since w_i ∈ (0, 1] and β ∈ (0, 1), after modulation, w_i ^ β > w_i, that is, the modulated penalty coefficient w_i' becomes larger. The increase in the penalty coefficient means that the importance of this policy dimension is reduced. The smaller β (the lower the confidence), the closer w_i' is to 1, and the stronger the suppression effect. 2) For the j-th type of policy weight w_j that is strongly correlated with historical or current performance, modulate it to w_j' = w_j ^ (β * δ); where δ is a decay factor that satisfies 0 < δ < 1. Since β * δ < β, for w_j ∈ (0, 1), w_j ^ (β * δ) < w_j ^ β, that is, the increase amplitude of w_j' is less than the increase amplitude of w_i'. This indicates that under the same low confidence, the historical / current related policies are suppressed less severely. Further, the weight vector after the above modulation can be normalized to ensure that its sum is 1 or a certain fixed value for subsequent weighted combination. Finally, substitute the modulated weights w_i' and w_j' into the weighted linear combination formula (such as V = Σ (w' * corresponding dimension score)) to calculate the final policy weighted evaluation value.

[0062] For example, let the delay weight (future-related, very important) w_i = 0.2, the availability weight (historical-related) w_j = 0.6, σ = 0.7, θ = 0.5, and δ = 0.5. We calculate β = 0.7 / (0.7+0.5) ≈ 0.583. After modulation: w_i' = 0.2^0.583 ≈ 0.39, w_j' = 0.6^(0.583*0.5) = 0.6^0.292 ≈ 0.87. The delay penalty coefficient increases from 0.2 to 0.39 (significantly reducing importance), while the availability penalty coefficient increases from 0.6 to 0.87 (relatively small reduction in importance), thus achieving stronger suppression of future-sensitive strategies.

[0063] In this embodiment of the invention, nonlinear and differentiated risk adaptation of strategy weights is achieved by power-law modulation based on β and γ_i. Specifically, the weight of a high-sensitivity strategy (large γ_i) decreases sharply when β decreases to avoid risk, while the weight of a low-sensitivity strategy (small γ_i) changes gradually to maintain robustness. Thus, the adaptability and safety of route selection decisions are improved through dynamic evaluation that can perceive and predict uncertainty and make intelligent adjustments.

[0064] In one embodiment, the path selection probability model is a multi-armed gambling machine model based on the Softmax function, where each arm corresponds to a candidate path. The parameters of the path selection probability model are iteratively updated and adjusted to ensure that the output path selection probability is positively correlated with the strategy-weighted evaluation value. This includes: in each iteration, selecting a target path for the current business data flow to explore or utilize based on the current path selection probability distribution; collecting actual performance feedback data generated after forwarding data based on the target path; calculating the actual utility value of the target path; calculating the dominance function value based on the difference between the strategy-weighted evaluation value and the actual utility value; and updating the parameters of the Softmax function using a strategy gradient ascent algorithm based on the dominance function value, thereby updating the path selection probability distribution. Paths with better strategy-weighted evaluation values ​​obtain larger positive dominance function values, and their corresponding path selection probability increases more significantly.

[0065] The multi-armed gambler model based on the Softmax function represents a mathematical model for solving sequential decision problems, where each available action (i.e., "arm") corresponds to an available candidate path. The multi-armed gambler model maintains a parameter vector θ = [θ_1, θ_2, ..., θ_K], where K is the number of paths and θ_a represents the "preference score" for choosing path a. The multi-armed gambler model uses the Softmax function to transform the preference score vector θ into a probability distribution π(a) = exp(θ_a) / Σ_bexp(θ_b), which is the path selection probability distribution. The goal of the multi-armed gambler model is to learn and adjust the parameter θ through interaction with the environment (i.e., trying different paths and observing the results) so that the probability distribution π maximizes the long-term cumulative reward (or utility).

[0066] The implementation is as follows: First, initialize a multi-armed gambling machine model based on the Softmax function for each business data flow (or according to business type). Assume there are K candidate paths. Initialize the preference score vector θ as a zero vector or based on the strategy-weighted evaluation value V_i (e.g., θ_i = V_i). Then, the initial path selection probability distribution π is uniformly distributed. Second, during system operation, whenever a path needs to be selected for a newly arriving business data flow, an iteration is triggered: 1) Path selection (exploration / exploitation): Query the path selection probability distribution π output by the current model. According to the preset strategy (e.g., ε-greedy strategy, where ε is the exploration rate, e.g., ε=0.1), select the path with the highest probability (exploitation) with a probability of 1-ε, or randomly select one path from all paths with a probability of ε (exploration). The selected path is denoted as the target path a_t. 2) Execute forwarding and feedback collection: Forward the current business data flow through the target path a_t. During or after the transmission of the business data stream, performance feedback data (e.g., end-to-end latency d, packet loss rate l) is collected on the target path a_t using performance monitoring probes deployed in the network. 3) Calculate the actual utility value and advantage function: Based on the collected performance feedback data, the actual utility value R_t = U(d, l, ...) is calculated according to the preset utility function U(·). For example, R_t = - (α * d + β * l), where α and β are weighting coefficients. 4) Calculate the advantage function value A_t. One specific implementation is to use the policy-weighted evaluation value V_{a_t} as the expected benchmark and compare it with the actual effect, that is, define A_t = R_t - V_{a_t}. Based on this, if the actual utility R_t is higher than the predicted evaluation V_{a_t}, then A_t>0, indicating that the actual performance of the path is better than expected and positive feedback should be given; otherwise, A_t<0, and negative feedback should be given. Thus, learning is achieved based on the deviation between the prior evaluation and the actual feedback. 5) Policy Gradient Update: Using the calculated advantage function value A_t, the model's preference score vector θ is updated through the policy gradient ascent algorithm. Specifically, the gradient of the log probability is calculated: for the selected action a_t, _{θ_{a_t}} log π(a_t) = 1 - π(a_t); For other unselected actions b, _{θ_b} log π(a_t) = 0 - π(b). Parameter updates are performed: θ_{a_t} := θ_{a_t} + η * A_t * (1 - π(a_t)); for all b ≠ a_t, θ_b := θ_b + η * A_t * (0 - π(b)), where η is the learning rate (e.g., η=0.01). Based on this, if A_t>0 (actual performance is better than expected), θ_{a_t} is increased, thus increasing the probability π(a_t) of choosing path a_t next time, while proportionally reducing the probability of other paths. If A_t<0, the effect is the opposite. Therefore, the update magnitude depends not only on the size of A_t but also on π(a_t): for paths with lower current probabilities, if positive feedback is obtained, the score θ increases more significantly (because 1-π(a_t) is larger), thus increasing the probability of exploring potentially less common paths. 6) Probability Distribution Update: Using the updated θ vector, the Softmax function is recalculated to obtain a new path selection probability distribution π'. This distribution will be used for path selection in the next iteration. By repeatedly performing the above iterations, the probability distribution is continuously corrected using actual network feedback, and this distribution has a higher selection probability than the path with the better overall evaluation.

[0067] In this embodiment of the invention, the aforementioned online learning mechanism achieves adaptive optimization of the path selection strategy: 1) balancing exploration and utilization to dynamically discover high-quality paths; 2) integrating prior knowledge and feedback to make the probability distribution approximate the true utility distribution that aligns with the multi-dimensional strategy objectives; and 3) efficient gradient updates to ensure that paths with better comprehensive evaluation and robust performance receive a greater probability boost. Thus, static scoring is transformed into an intelligent decision-making model that can continuously learn and evolve, improving the adaptability and decision quality of the route selection system.

[0068] In one embodiment, the calculation of the actual utility value integrates the actual end-to-end latency, packet loss rate, and unit traffic cost of the business data flow on the target path; the formula for calculating the advantage function value A is: A = U_actual - V_predict + λ * H(π), where U_actual is the actual utility value, V_predict is the policy-weighted evaluation value of the target path, H(π) is the entropy of the current path selection probability distribution, used to encourage exploration, and λ is the entropy reward coefficient.

[0069] The implementation is as follows: In each iteration, after the business data flow completes transmission on the selected target path a, the system executes the following steps: First, collect actual performance feedback data and calculate U_actual: the actual end-to-end latency d_actua and packet loss rate l_actual on the target path. Based on the tariff model of the target path and the data volume vol of this business flow, calculate the unit traffic cost c_unit, or directly calculate the total cost of this flow c_total = c_unit * vol. Second, calculate the actual utility value U_actual according to the preset utility function U(·), such as using the following calculation formula: U_actual = - (α * d_actual + β * l_actual + γ * c_unit); or, U_actual = - (α * d_actual + β * l_actual + γ * c_total); Where α, β, and γ are weight coefficients greater than zero, corresponding to the penalty weights for latency, packet loss, and cost, respectively. The negative sign indicates that these indicators are all "cost" type indicators; the smaller the value (the more negative), the lower the actual utility. The coefficient settings should be coordinated with the business strategy weights to ensure consistency of the evaluation objectives. Third, calculate the advantage function value A: obtain the strategy-weighted evaluation value V_predict of the target path a at this decision moment, and query the probability distribution π output by the current path selection probability model, calculate its entropy H(π) = -Σ π_i log(π_i), and the advantage function value A: A = U_actual - V_predict + λ * H(π); Here, (U_actual - V_predict) is the core bias term, which measures the difference between the actual effect U_actual and the prior prediction V_predict. If the actual effect is better than the prediction, this term is positive, providing positive feedback; otherwise, it is negative. λ * H(π): the entropy reward term, is a reward that is independent of the specific action selected and only related to the state of the current overall policy π. Regardless of which path is chosen, as long as the entropy H(π) of the current policy π is high (i.e., the distribution is relatively uniform), a positive term will be added to the advantage function. Thus, in the policy gradient update, an update component that tends to increase H(π) is generated for the preference score θ of all actions, thereby inhibiting the premature convergence of the probability distribution and encouraging exploration. Here, λ is a preset positive coefficient. Fourth, policy gradient update: Substitute the calculated advantage function value A into the policy gradient update rule (as mentioned above: θ_a := θ_a + η * A * (1 - π(a)), for b ≠ a, θ_b := θ_b + η * A * (0 - π(b)), where η is the learning rate), update the model parameters θ, and then update the path selection probability distribution π. Since A includes the entropy reward term λ * H(π), the update process is based not only on the actual deviation of this path selection, but also on the quantitative incentive for exploration behavior.

[0070] This invention achieves the following technical effects: 1) Goal-oriented comprehensive feedback: U_actual integrates key performance and economic indicators, directly aligning learning objectives with business value. 2) Endogenous adaptive exploration: Introducing the λ * H(π) term into the advantage function allows the model to maintain an appropriate exploration level automatically by optimizing the intrinsic drive of the objective, avoiding premature convergence, without relying on external heuristic exploration strategies (such as ε-greedy). 3) Adjustable optimization process: The entropy reward coefficient λ, as a configurable parameter, provides the system with an intuitive means of adjusting the exploration intensity. 4) Efficient and stable learning: Combining biased learning based on high-quality priors (V_predict) and entropy regularization improves learning speed, stability, and the ability to cope with non-stationary environments.

[0071] The multi-link global routing method based on dynamic strategy linkage described in the above embodiments can be recombined with the technical features included in different embodiments as needed to obtain a combined implementation scheme, but all are within the protection scope claimed by this invention.

[0072] In one embodiment, a multi-link global routing system based on dynamic policy linkage is provided. This system corresponds one-to-one with the multi-link global routing method based on dynamic policy linkage described in the above embodiments. Please refer to [link to relevant documentation]. Figure 3 , Figure 3This is a schematic block diagram of a multi-link global routing system based on dynamic strategy linkage provided in an embodiment of the present invention. Figure 3 As shown, the multi-link global routing system 30 based on dynamic policy linkage is applied to an SD-WAN controller. The multi-link global routing system 30 includes: a first quantization module 31, a first generation module 32, a first fusion module 33, a first calculation module 34, a first acquisition module 35, and a first selection module 36. The detailed descriptions of each of the above functional modules are as follows: The first quantization module 31 is used to acquire policy data from business, security, and cost management and quantize it into a multi-dimensional policy weight vector; the first generation module 32 is used to generate, for each link in the network, a predicted value of its future short-term performance and a corresponding confidence assessment based on the historical performance time-series data and current real-time performance indicators of the link through a time series prediction model; the first fusion module 33 is used to convert the historical performance time-series data, the current real-time performance indicators, and the predicted value into a comparable and unified link state feature representation based on the confidence assessment, and dynamically determine the weight of each component in the link state feature representation in the fusion calculation according to the confidence assessment or the comprehensive confidence index derived therefrom, and perform weighted fusion to form the link quality status value of the link; the first calculation module 34 is used to calculate the link quality status value of the link based on the multi-dimensional policy weight vector and each link. The link quality status value of the path is used to calculate the dynamic policy-weighted evaluation value of each link through a preset evaluation function. The preset evaluation function is configured to: use the confidence evaluation as an adjustment parameter to dynamically adjust the effectiveness of the multi-dimensional policy weight vector in the evaluation, so as to adapt to the link selection risk caused by prediction uncertainty, thereby calculating the policy-weighted evaluation value; the first obtaining module 35 is used to establish and initialize a path selection probability model for each service data flow based on the policy-weighted evaluation value, and adjust the parameters of the path selection probability model through iterative updates, so that the output path selection probability is positively correlated with the policy-weighted evaluation value, until the path selection probability model converges, and obtains an optimized path selection probability distribution; the first selection module 36 is used to select the path with the highest probability for forwarding based on the optimized path selection probability distribution and the type of service data flow.

[0073] In one embodiment, the time series prediction model is a sequence-to-sequence model based on an attention mechanism; the first generation module 32 includes: The first encoding submodule is used to encode the historical performance time-series data and the current real-time performance indicators into a feature sequence; The iterative generation submodule is used to iteratively generate a sequence of performance prediction values ​​for N future time steps through the decoder of the time series prediction model, using the feature sequence as context information and the output of the previous decoding time step as the autoregressive input, where N is a natural number. The first calculation submodule is used to synchronously calculate and output the confidence interval or variance of the performance prediction value generated by the decoder at each future time step through the confidence estimation module in the time series prediction model. Based on the decoder hidden state and attention context at the future time step, the confidence interval or variance output by all future time steps constitutes a confidence evaluation sequence that corresponds one-to-one with the performance prediction value sequence in time.

[0074] In one embodiment, the first fusion module 33 includes: The first output submodule is used to input the performance prediction value sequence {F(t)} and its corresponding confidence evaluation sequence {C(t)} into a preset aggregation function, and output a comprehensive prediction scalar F_agg and a comprehensive confidence scalar C_agg. The preset aggregation function is configured such that the prediction value F(t) at the time step with a lower confidence C(t) has a smaller contribution weight to F_agg. The first extraction submodule is used to extract the historical trend scalar H_trend, which represents the historical trend, from the historical performance time series data. The first determining submodule is used to use the current real-time performance metric as the current instantaneous scalar I_now; The first construction submodule is used to construct a three-dimensional feature vector based on the historical trend scalar H_trend, the current instantaneous scalar I_now, and the comprehensive prediction scalar F_agg, as a representation of the link state features; The first dynamic determination submodule is used to dynamically determine the weights W_h, Wi, and W_f of each scalar component H_trend, I_now, and F_agg in the feature vector in the fusion calculation based on the comprehensive confidence scalar C_agg; The first weighted summation submodule is used to calculate the link quality status value S of the link by weighted summation S = W_h * H_trend + W_i * I_now + W_f * F_agg.

[0075] In one embodiment, the first output submodule includes: The second calculation submodule is used to calculate the aggregate weight ω(t) of the predicted value F(t) of the future time step for each future time step t = 1 to N, based on its confidence evaluation value C(t), where ω(t) is a monotonically increasing function of C(t), and the monotonically increasing function is: ω(t) = 1 / (1 + exp(-η * (C(t) - C_th))) , where η is a preset scaling factor and C_th is a preset confidence threshold; The third calculation submodule is used to calculate the comprehensive prediction scalar F_agg according to the following formula: ; The fourth calculation submodule is used to calculate the comprehensive confidence scalar C_agg according to the following formula: ; Where N is the total number of predicted future time steps, ω(t) is the aggregate weight of the t-th future time step calculated based on C(t), F(t) is the performance prediction value of the t-th future time step, and C(t) is the confidence evaluation value of the t-th future time step.

[0076] In one embodiment, the first dynamic determination submodule includes: The first acquisition submodule is used to acquire the preset baseline weight W_h_base corresponding to H_trend, the preset baseline weight W_i_base corresponding to I_now, and the preset baseline weight W_f_base corresponding to F_agg; The fifth calculation submodule is used to calculate a prediction confidence factor τ based on C_agg, where τ is a monotonically increasing function of C_agg, τ = (C_agg - C_min) / (C_max - C_min), where C_max is the preset upper bound of the confidence assessment value, C_min is the preset lower bound of the confidence assessment value, and τ∈[0, 1]. The first adjustment submodule is used to adjust the baseline weights W_h_base, Wi_i_base, and W_f_base according to τ as follows: The fusion weight W_h = W_h_base + (1 - τ) * μ of the historical trend scalar H_trend, where μ is a preset compensation coefficient; The fusion weight W_i = W_i_base for the current instantaneous scalar I_now; The fusion weight W_f = W_f_base * τ for the comprehensive prediction scalar F_agg; The normalization submodule is used to normalize the adjusted weights W_h, W_i, and W_f so that their sum is 1.

[0077] In one embodiment, the first calculation module 34 includes: a sixth calculation submodule, configured to calculate a policy sensitivity modulation coefficient based on the confidence assessment, wherein the policy sensitivity modulation coefficient is positively correlated with the confidence assessment value; a seventh calculation submodule, configured to perform a power operation on each weight component in the multidimensional policy weight vector with the policy sensitivity modulation coefficient to obtain a modulated policy weight component, wherein the power exponent corresponding to the weight component strongly correlated with future performance is greater than the power exponent corresponding to the weight component strongly correlated with historical or current performance, and the weight component strongly correlated with future performance refers to the policy weight component whose value directly affects the evaluation results of link latency and jitter performance indicators, and the weight component strongly correlated with historical or current performance refers to the policy weight component whose value directly affects the evaluation results of link historical availability and current packet loss rate indicators; and an eighth calculation submodule, configured to use the modulated policy weight component to perform a weighted linear combination on the corresponding link quality status value component to calculate the policy weighted evaluation value.

[0078] In one embodiment, the formula for calculating the policy sensitivity modulation coefficient β is: β = σ / (σ + θ), where σ is the quantized value of the confidence assessment and θ is a preset sensitivity threshold; the exponentiation operation is as follows: for the i-th type policy weight w_i that is strongly correlated with future performance, the modulation is w_i' = w_i ^ β; for the j-th type policy weight w_j that is strongly correlated with historical or current performance, the modulation is w_j' = w_j ^ (β * δ), where 0 < δ < 1 is the attenuation factor.

[0079] In one embodiment, the path selection probability model is a multi-armed gambling machine model based on the Softmax function, where each arm corresponds to a candidate path; the first obtaining module 35 includes: a ninth calculation submodule, used to select a target path for the current business data stream for exploration or utilization in each iteration according to the current path selection probability distribution, and collect the actual performance feedback data generated after forwarding data based on the target path, and calculate the actual utility value of the target path; a tenth calculation submodule, used to calculate the advantage function value based on the difference between the strategy-weighted evaluation value and the actual utility value; and a first update submodule, used to update the parameters of the Softmax function according to the advantage function value through a strategy gradient ascent algorithm, thereby updating the path selection probability distribution, wherein the path with a better strategy-weighted evaluation value obtains a larger positive value of the advantage function, and its corresponding path selection probability increase is also greater.

[0080] In one embodiment, the calculation of the actual utility value integrates the actual end-to-end latency, packet loss rate, and unit traffic cost of the business data flow on the target path; the formula for calculating the advantage function value A is: A = U_actual - V_predict + λ * H(π), where U_actual is the actual utility value, V_predict is the policy-weighted evaluation value of the target path, H(π) is the entropy of the current path selection probability distribution, used to encourage exploration, and λ is the entropy reward coefficient.

[0081] For specific limitations regarding the multi-link global routing system based on dynamic policy linkage, please refer to the limitations of the multi-link global routing method based on dynamic policy linkage mentioned above, which will not be repeated here. Each module in the aforementioned multi-link global routing system based on dynamic policy linkage can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0082] Those skilled in the art will understand that the methods and systems provided in the embodiments of the present invention can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. The methods can also be implemented as a computer program product stored in one or more computer-readable storage media, including but not limited to: disks, optical disks, read-only memory (ROM), random access memory (RAM), flash memory, etc. When the computer program product is executed by one or more data processing devices (such as computers), the devices perform the steps as described in any of the preceding method embodiments.

[0083] Software tools, components, or models not belonging to this company that appear in the embodiments of this invention are merely illustrative examples and do not represent actual use. The data collection methods used in the embodiments of this invention comply with relevant laws and regulations, such as the "Data Security Law of the People's Republic of China," the "Personal Information Protection Law of the People's Republic of China," GDPR (General Data Protection Regulation of the European Union), or information security standards of other countries and regions.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-link global routing method based on dynamic strategy linkage, characterized in that, Applications to SD-WAN controllers include: Acquire policy data from business, security, and cost management and quantify it into a multi-dimensional policy weight vector; For each link in the network, based on the historical performance time-series data and current real-time performance indicators of the link, a time series prediction model is used to generate a predicted value of its future short-term performance and a corresponding confidence level assessment. Based on the confidence assessment, the historical performance time series data, the current real-time performance indicators and the predicted values ​​are converted into comparable and unified link state feature representations. Based on the confidence assessment or the comprehensive confidence index derived therefrom, the weight of each component in the link state feature representation in the fusion calculation is dynamically determined and weighted fusion is performed to form the link quality status value of the link. Based on the multidimensional strategy weight vector and the link quality status value of each link, a dynamic strategy weighted evaluation value for each link is calculated using a preset evaluation function. The preset evaluation function is configured to use the confidence evaluation as an adjustment parameter to dynamically adjust the effectiveness of the multidimensional strategy weight vector in the evaluation, so as to adapt to the link selection risk caused by prediction uncertainty, thereby calculating the strategy weighted evaluation value. Based on the strategy-weighted evaluation value, a path selection probability model is established and initialized for each business data stream. The parameters of the path selection probability model are then adjusted by iterative updates to make the output path selection probability positively correlated with the strategy-weighted evaluation value, until the path selection probability model converges and an optimized path selection probability distribution is obtained. Based on the optimized path selection probability distribution and the type of business data flow, the path with the highest probability is selected for forwarding.

2. The multi-link global routing method based on dynamic strategy linkage as described in claim 1, characterized in that, The time series prediction model is a sequence-to-sequence model based on an attention mechanism; the model generates predicted values ​​of its future short-term performance and corresponding confidence assessments, including: The historical performance time-series data and the current real-time performance indicators are encoded into a feature sequence; The time series prediction model decoder uses the feature sequence as context information and the output of the previous decoding time step as the autoregressive input to iteratively generate a sequence of performance prediction values ​​for N future time steps, where N is a natural number. Simultaneously, through the confidence estimation module within the time series prediction model, for the performance prediction value generated by the decoder at each future time step, based on the decoder's hidden state and attention context at the future time step, the confidence interval or variance of the performance prediction value is calculated and output. The confidence intervals or variances output at all future time steps constitute a confidence evaluation sequence that corresponds one-to-one with the performance prediction value sequence in time.

3. The multi-link global routing method based on dynamic strategy linkage as described in claim 2, characterized in that, Based on the confidence assessment, the historical performance time-series data, the current real-time performance indicators, and the predicted values ​​are converted into comparable, unified link state feature representations. Then, based on the confidence assessment or a comprehensive confidence index derived therefrom, the weights of each component in the link state feature representation in the fusion calculation are dynamically determined, and weighted fusion is performed to form the link quality status value of the link, including: The performance prediction sequence {F(t)} and its corresponding confidence evaluation sequence {C(t)} are input into a preset aggregation function, which outputs a comprehensive prediction scalar F_agg and a comprehensive confidence scalar C_agg. The preset aggregation function is configured such that the prediction value F(t) at the time step with a lower confidence C(t) has a smaller contribution weight to F_agg. Extract the historical trend scalar H_trend, which represents the historical trend, from the historical performance time series data; The current real-time performance metric is used as the current instantaneous scalar I_now; Based on the historical trend scalar H_trend, the current instantaneous scalar I_now, and the comprehensive prediction scalar F_agg, a three-dimensional feature vector is constructed as the link state feature representation. Based on the comprehensive confidence scalar C_agg, the weights W_h, Wi, and W_f of each scalar component H_trend, I_now, and F_agg in the feature vector are dynamically determined in the fusion calculation; The link quality status value S of the link is calculated by weighted summation S = W_h * H_trend + W_i * I_now + W_f * F_agg.

4. The multi-link global routing method based on dynamic strategy linkage as described in claim 3, characterized in that, The preset aggregation function outputs the comprehensive prediction scalar F_agg and the comprehensive confidence scalar C_agg through the following steps: For each future time step t = 1 to N, the aggregate weight ω(t) of the predicted value F(t) of the future time step is calculated based on its confidence assessment value C(t), where ω(t) is a monotonically increasing function of C(t), and the monotonically increasing function is: ω(t) = 1 / (1 + exp(-η * (C(t) - C_th))), where η is a preset scaling factor and C_th is a preset confidence threshold; The comprehensive prediction scalar F_agg is calculated according to the following formula: ; The comprehensive confidence scalar C_agg is calculated using the following formula: ; Where N is the total number of predicted future time steps, ω(t) is the aggregate weight of the t-th future time step calculated based on C(t), F(t) is the performance prediction value of the t-th future time step, and C(t) is the confidence evaluation value of the t-th future time step.

5. The multi-link global routing method based on dynamic strategy linkage as described in claim 3 or 4, characterized in that, Based on the comprehensive confidence scalar C_agg, the weights W_h, Wi, and W_f of each scalar component H_trend, I_now, and F_agg in the feature vector are dynamically determined in the fusion calculation, including: Get the preset baseline weight W_h_base corresponding to H_trend, the preset baseline weight Wi_i_base corresponding to I_now, and the preset baseline weight W_f_base corresponding to F_agg; A prediction confidence factor τ is calculated based on C_agg, where τ is a monotonically increasing function of C_agg, τ = (C_agg - C_min) / (C_max - C_min), where C_max is the preset upper bound of the confidence assessment value, C_min is the preset lower bound of the confidence assessment value, and τ∈[0, 1]. The baseline weights W_h_base, Wi_base, and W_f_base are adjusted according to τ as follows: The fusion weight W_h = W_h_base + (1 - τ) * μ of the historical trend scalar H_trend, where μ is a preset compensation coefficient; The fusion weight W_i = W_i_base for the current instantaneous scalar I_now; The fusion weight W_f = W_f_base * τ for the comprehensive prediction scalar F_agg; The adjusted weights W_h, W_i, and W_f are normalized so that their sum is 1.

6. The multi-link global routing method based on dynamic strategy linkage as described in claim 1, characterized in that, The preset evaluation function is configured to: use the confidence level evaluation as an adjustment parameter to dynamically adjust the effectiveness of the multi-dimensional strategy weight vector in the evaluation, in order to adapt to the link selection risk caused by prediction uncertainty, specifically including: Based on the confidence level assessment, the policy sensitivity modulation coefficient is calculated, wherein the policy sensitivity modulation coefficient is positively correlated with the confidence level assessment value; Each weight component in the multidimensional strategy weight vector is subjected to a power operation with the strategy sensitivity modulation coefficient to obtain the modulated strategy weight component. The power exponent corresponding to the weight component that is strongly correlated with future performance is greater than the power exponent corresponding to the weight component that is strongly correlated with historical or current performance. Furthermore, the weight component that is strongly correlated with future performance refers to the strategy weight component whose value directly affects the evaluation results of link latency and jitter performance indicators, while the weight component that is strongly correlated with historical or current performance refers to the strategy weight component whose value directly affects the evaluation results of link historical availability and current packet loss rate indicators. The modulated policy weight components are used to perform a weighted linear combination of the corresponding link quality status value components to calculate the policy weighted evaluation value.

7. The multi-link global routing method based on dynamic strategy linkage as described in claim 6, characterized in that, The formula for calculating the policy sensitivity modulation coefficient β is: β = σ / (σ + θ), where σ is the quantized value of the confidence assessment and θ is the preset sensitivity threshold; the power operation is: for the i-th type policy weight w_i that is strongly correlated with future performance, the modulation is w_i' = w_i ^ β; for the j-th type policy weight w_j that is strongly correlated with historical or current performance, the modulation is w_j' = w_j ^ (β * δ), where 0 < δ < 1 is the attenuation factor.

8. The multi-link global routing method based on dynamic strategy linkage as described in claim 1, characterized in that, The path selection probability model is a multi-armed gambling machine model based on the Softmax function, where each arm corresponds to a candidate path. The parameters of the path selection probability model are adjusted through iterative updates to ensure that the output path selection probability is positively correlated with the policy-weighted evaluation value, including: In each iteration, based on the current path selection probability distribution, a target path is selected for the current business data flow to explore or utilize. After forwarding data based on the target path, the actual performance feedback data generated is collected, and the actual utility value of the target path is calculated. The advantage function value is calculated based on the difference between the strategy-weighted evaluation value and the actual utility value. Based on the advantage function value, the parameters of the Softmax function are updated using the policy gradient ascent algorithm, thereby updating the path selection probability distribution. The path with a better policy weighted evaluation value has a larger positive value of the advantage function and a greater increase in its corresponding path selection probability.

9. The multi-link global routing method based on dynamic strategy linkage as described in claim 8, characterized in that, The calculation of the actual utility value integrates the actual end-to-end latency, packet loss rate, and unit traffic cost of the business data flow on the target path; the formula for calculating the advantage function value A is: A = U_actual - V_predict + λ * H(π), where U_actual is the actual utility value, V_predict is the policy-weighted evaluation value of the target path, H(π) is the entropy of the current path selection probability distribution, used to encourage exploration, and λ is the entropy reward coefficient.

10. A multi-link global routing system based on dynamic strategy linkage, characterized in that, Applications to SD-WAN controllers include: The first quantization module is used to acquire strategy data from business, security and cost management and quantify it into a multi-dimensional strategy weight vector; The first generation module is used to generate, for each link in the network, a predicted value of its future short-term performance and a corresponding confidence level assessment based on the historical performance time-series data and current real-time performance indicators of the link through a time series prediction model. The first fusion module is used to convert the historical performance time series data, the current real-time performance indicators and the predicted values ​​into comparable and unified link state feature representations based on the confidence assessment, and dynamically determine the weight of each component in the link state feature representation in the fusion calculation according to the confidence assessment or the comprehensive confidence index derived therefrom, and perform weighted fusion to form the link quality status value of the link. The first calculation module is used to calculate the dynamic strategy weighted evaluation value of each link based on the multi-dimensional strategy weight vector and the link quality status value of each link through a preset evaluation function. The preset evaluation function is configured to: use the confidence evaluation as an adjustment parameter to dynamically adjust the effectiveness of the multi-dimensional strategy weight vector in the evaluation, so as to adapt to the link selection risk caused by prediction uncertainty, thereby calculating the strategy weighted evaluation value. The first acquisition module is used to establish and initialize a path selection probability model for each business data stream based on the strategy weighted evaluation value, and adjust the parameters of the path selection probability model through iterative updates so that the output path selection probability is positively correlated with the strategy weighted evaluation value, until the path selection probability model converges and an optimized path selection probability distribution is obtained. The first selection module is used to select the path with the highest probability for forwarding based on the optimized path selection probability distribution and the type of business data flow.

Images