A design method of weight coefficient based on fuzzy rule

Abstract

The application discloses a design method of weight coefficients based on fuzzy rules, comprising the following steps: after system initialization, a plurality of network parameters and a plurality of auxiliary parameters are collected in a measurement window, and the collected plurality of auxiliary parameters are standardized; the plurality of standardized auxiliary parameters are adaptively rewarded and aggregated to obtain a final reward; a target function is established according to the final reward, and a candidate weight coefficient is generated; the candidate weight coefficient is subjected to KL constraint to obtain KL divergence, when the KL divergence is less than or equal to a KL threshold, the next step is performed; when the KL divergence is greater than the KL threshold, step reduction is performed to recalculate the candidate weight coefficient or backtracking is performed to calculate a safe weight coefficient; the candidate weight coefficient or the safe weight coefficient is smoothed to obtain an optimal weight coefficient. The design method can automatically approach a more optimal weight coefficient under different service scenarios, reduce sawtooth effects and periodic degradation caused by multi-flow synchronization, and does not need frequent manual parameter adjustment.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method for designing weight coefficients based on fuzzy rules. Background Technology

[0002] With the rapid development of the Internet and cloud-edge collaboration, the types and number of services carried by networks continue to grow. Real-time video, interactive applications, industrial IoT, and satellite / vehicle-to-everything (V2X) networks place higher demands on low latency, low jitter, and high reliability. Congestion control, as a key component of transmission performance, needs to balance multiple metrics such as throughput, latency, jitter, packet loss, and fairness under complex and time-varying link conditions, while maintaining stable adaptive capabilities across different network scenarios. Therefore, building a congestion control mechanism that can accurately perceive network conditions and quickly adjust accordingly has become a core requirement for improving overall network efficiency and user experience.

[0003] Traditional congestion control primarily relies on bandwidth utilization and packet loss / delay signals: the transmitting end estimates utilization based on the actual transmission rate and the port's rated bandwidth, then adjusts the congestion window or transmission rate additively or subtractively based on the packet loss rate, supplemented by simple queue length or round-trip delay trends as auxiliary references. For example, Chinese invention patent CN120474979B, entitled "A Network Congestion Control Method and Device Based on Fuzzy Rules," introduces fuzzy rules and pre-sets differentiated fuzzy rules according to the characteristics of different network types. It dynamically allocates weight coefficients based on network type, employs a fuzzy inference mechanism to adapt to the nonlinear characteristics of network states, and further uses the adjustment factor ΔF obtained from the fuzzy rules to calculate the size of the new congestion window, thus improving the timeliness and accuracy of network congestion control. However, the parameter settings of this type of method are relatively fixed, and the original weight coefficients are manually preset. Typically, manual experience is required to select control parameters such as additive gain, multiplicative loss, sampling window, and threshold to obtain the desired performance under "typical scenarios."

[0004] In complex and time-varying network environments, the aforementioned methods often exhibit the following characteristics: parameters are extremely sensitive to the scenario and rely on manual experience for maintenance. When faced with jitter and fading in wireless and satellite links, changes in RTT (physical link limit) and packet loss are difficult to distinguish from queuing and congestion, leading to frequent mismatches in adjustment direction and magnitude, resulting in both sluggish and oscillating responses. Furthermore, fixed parameters are difficult to adapt to time-varying networks, and the cross-scenario migration effect is unstable, with oscillations and unpredictability in engineering implementation. Consequently, traditional congestion control is unable to meet the current network's comprehensive requirements for low latency, low jitter, and high utilization. Summary of the Invention

[0005] The purpose of this invention is to provide a design method for weight coefficients based on fuzzy rules, which can automatically approach more suitable weight coefficients in different business scenarios, reduce the sawtooth effect and periodic degradation caused by multi-stream synchronization, and eliminate the need for frequent manual parameter adjustments, making it easy to deploy and maintain at scale.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for designing weight coefficients based on fuzzy rules, wherein the weight coefficients include weight coefficients for bandwidth utilization, ECN tagging rate, and queue length, and the design method includes the following steps:

[0008] After system initialization, multiple network parameters and multiple auxiliary parameters are collected within the measurement window, and the collected auxiliary parameters are standardized. The multiple network parameters include bandwidth utilization, ECN tagging rate, and queue length, while the multiple auxiliary parameters include throughput, round-trip time, latency jitter, packet loss rate, and bandwidth.

[0009] The final reward is obtained by adaptively aggregating multiple standardized auxiliary parameters.

[0010] Establish an objective function based on the final reward and generate candidate weight coefficients;

[0011] Apply KL constraints to the candidate weight coefficients to obtain KL divergence. If the KL divergence is less than or equal to the KL threshold, proceed to the next step. If the KL divergence is greater than the KL threshold, perform step reduction to recalculate the candidate weight coefficients or backtrack to calculate the safe weight coefficients.

[0012] The candidate weight coefficients or safety weight coefficients are smoothed to obtain the final weight coefficients for use in fuzzy rules.

[0013] By introducing an online policy update with KL constraints, the weight coefficients of each network parameter are adaptively adjusted in small steps. When the KL offset from the previous policy cycle exceeds a threshold, step reduction or backtracking is performed. Simultaneously, the endpoint continues to execute fuzzification, rule inference, and defuzzification within the fuzzy rules to obtain the adjustment factor ΔF, which is then used to smoothly adjust the congestion window. This design method eliminates the need for manual parameter tuning, reducing maintenance burdens, and automatically converges to optimal weight coefficients as load, link quality, and service structure change. Furthermore, in highly volatile environments such as wireless / satellite networks, stabilization techniques like KL constraints and entropy regularization enhance the ability to distinguish between congested and non-congested packet loss, mitigating response lag and oscillations, and improving consistency and portability across DCN / WAN scenarios. The weight coefficients of each network parameter in this design method are used as input weights for fuzzy inference within the fuzzy rules. By optimizing only the weight coefficients while maintaining the original membership degrees, fuzzy rules, and congestion window update formulas, it facilitates automatic focusing on more discriminative signals in different environments.

[0014] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for standardizing the collected auxiliary parameters is as follows: the standardized throughput rate is calculated according to the following formula:

[0015] ,

[0016] Where T0 is the currently acquired throughput rate, T is the normalized throughput rate, P5G refers to the 5th percentile of the throughput rate within the measurement window, and P95G refers to the 95th percentile of the throughput rate within the measurement window.

[0017] The standardized round-trip time delay is calculated using the following formula:

[0018] ,

[0019] Where D0 is the currently collected round-trip delay, D is the standardized round-trip delay, P5rtt refers to the 5th percentile of the round-trip delay within the measurement window, and P95rtt refers to the 95th percentile of the round-trip delay within the measurement window.

[0020] The standardized time delay jitter is calculated using the following formula:

[0021] ,

[0022] Where J0 is the currently collected time delay jitter, J is the standardized time delay jitter, P5jit refers to the 5th percentile of the time delay jitter within the measurement window, and P95jit refers to the 95th percentile of the time delay jitter within the measurement window;

[0023] The standardized packet loss rate can be calculated using the following formula:

[0024] ,

[0025] Where L0 is the currently collected packet loss rate, L is the standardized packet loss rate, P5loss refers to the 5th percentile of the packet loss rate within the measurement window, and P95loss refers to the 95th percentile of the packet loss rate within the measurement window.

[0026] The standardized bandwidth is calculated using the following formula:

[0027] ,

[0028] Where E0 is the currently acquired bandwidth, E is the standardized bandwidth, P5Ecn refers to the 5th percentile of the bandwidth within the measurement window, and P95Ecn refers to the 95th percentile of the bandwidth within the measurement window.

[0029] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for adaptively aggregating the standardized auxiliary parameters to obtain the final reward is as follows: first, calculate the basic reward according to the following formula. , where r base Based on the reward, λ T The reward aggregation weight for throughput rate, λ D The reward aggregation weight for round-trip delay, λ J For the reward aggregation weight of latency jitter, λ L The reward aggregation weight for packet loss rate, λ E The bandwidth reward aggregation weight is D, the normalized round-trip time is J, the normalized latency jitter is L, the normalized packet loss rate is E, and the normalized bandwidth is E.

[0030] The final reward is then calculated using the following formula: , where r is the final reward, and the value of the final reward r is in the range of 0-1; k is the tail penalty coefficient, and the value of k is in the range of 0.15-0.25; p is the tail penalty term.

[0031] Furthermore, based on any one or a combination of the aforementioned technical solutions, the tail penalty term is calculated using the following formula:

[0032] ,

[0033] Where τ is the tail penalty trigger threshold, and the value of τ ranges from 0.85 to 0.95; the max function is the maximum value function.

[0034] Furthermore, following any one or a combination of the aforementioned technical solutions, the λ T , λ D , λ J , λ L With λ E The sum of is 1, and λ T , λ D , λ J , λ L , λ E The values ​​of T, D, J, L, and E all range from 0 to 1.

[0035] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for establishing the objective function and generating candidate weight coefficients based on the final reward is as follows: first, establish the objective function according to the following formula. Where J is the optimization objective, r is the final reward, and λ is the final reward. H λ is the entropy regularity coefficient. H The value range is 0.001-0.01; H(w) is the entropy of the weight distribution, and the formula for calculating H(w) is: , where w old,i This represents the weight coefficient of the i-th network parameter in effect in the previous cycle, i=1,2,3; m represents the total number of all network parameters; by introducing entropy regularization, we can prevent the weight coefficient from collapsing rapidly into a single channel.

[0036] Then, calculate the gradient reward of the objective function using the following formula:

[0037] ,

[0038] Among them, g i Let δ represent the reward gradient of the i-th network parameter, and δ be the magnitude of the weight perturbation, with a value ranging from 0.01 to 0.02. It is w old,i The reward obtained by offline recalculation after applying the +δ perturbation. It is w old,i The reward obtained by offline recalculation after applying the -δ perturbation. It is obtained by taking the gradient of J;

[0039] Finally, the candidate weight coefficients are calculated using the following formula: , where w raw w represents the candidate weight coefficients for all network parameters. old η represents the weight coefficients of all network parameters that were effective in the previous cycle, and η is the learning rate, which ranges from 0.1 to 0.5.

[0040] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for applying KL constraints to the candidate weight coefficients to obtain the KL divergence is as follows: the KL divergence is calculated using the following formula:

[0041] ,

[0042] Among them, D KL The KL divergence is given by a KL threshold of 0.01–0.05; w raw,i Let w represent the candidate weight coefficients for the i-th network parameter. Calculate w using KL constraints. raw The KL divergence can limit the magnitude of weight coefficient updates, ensuring that w raw With w old The differences between them are kept within a controllable range to prevent the weight coefficients from being updated too much, thereby ensuring steady-state optimization.

[0043] Furthermore, following any one or a combination of the aforementioned technical solutions, when D KL When D is greater than 2ε, the step is reduced to the step of generating candidate weight coefficients to recalculate the candidate weight coefficients; when D KL When the value is greater than ε and less than 2ε, backtrack to calculate the safety weight coefficient.

[0044] Furthermore, based on any one or a combination of the aforementioned technical solutions, the security weight coefficient is calculated using the following formula: , where w safe Let β be the safety weight coefficient for all network parameters, and let β be the safety coefficient. The value of β is in the range of 0-1 and the value of β must satisfy the condition that the KL divergence obtained after applying KL constraints to the calculated safety weight coefficient is less than the KL threshold.

[0045] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for smoothing the candidate weight coefficients or safety weight coefficients to obtain the final weight coefficients is as follows: the final weight coefficients are calculated using the following formula: , where w new represents the final weight coefficients for all network parameters, α is the EMA smoothing coefficient, and the value of α ranges from 0.2 to 0.4; w0 is selected from w raw or w safe One of them.

[0046] Furthermore, following any one or a combination of the aforementioned technical solutions, the w old Including w old,1 w old,2 and w old,3 The w raw Including w raw,1 w raw,2and w raw,3 The w safe Including w safe,1 w safe,2 and w safe,3 The w new Including w new,1 w new,2 and w new,3 Among them, w old,1 w is the weighting factor for bandwidth utilization in the previous period. old,2 w is the weighting factor for the ECN labeling rate in effect during the previous period. old,3 The weighting factor for the queue length in the previous cycle; w raw,1 Candidate weighting coefficients for bandwidth utilization, w raw,2 Candidate weight coefficients for ECN labeling rate, w raw,3 The candidate weight coefficients for the queue length; w safe,1 w is the security weighting coefficient for bandwidth utilization. safe,2 w is the safety weighting coefficient for the ECN labeling rate. safe,3 w is the safety weight coefficient for the queue length. new,1 The final weighting coefficient for bandwidth utilization, w new,2 The final weighting coefficient for the ECN labeling rate, w new,3 This is the final weighting coefficient for the queue length.

[0047] Furthermore, following any one or a combination of the aforementioned technical solutions, the period of the measurement window is 400-600ms.

[0048] The beneficial effects of the technical solution provided by this invention are as follows:

[0049] a. The design method of this invention can run simultaneously with fuzzy rules without blocking each other, taking into account both fast response and steady-state optimization. When the KL offset between the current cycle and the previous cycle exceeds the KL threshold, the step size is automatically reduced or back off, ensuring that online updates are controllable and eliminating the need for manual parameter tuning. Through online learning, it automatically converges to a better weight configuration under different topology, load and channel conditions, which can significantly reduce parameter maintenance costs and solve the problem of traditional schemes being sensitive to empirical thresholds.

[0050] b. The design method of this invention combines KL constraints for lightweight adaptation, which allows the weight coefficients to be updated in only one short evaluation window (measurement window) per cycle. There is no need to build a network model or introduce additional probe flow, resulting in low learning cost and good performance for existing networks.

[0051] c. The weight coefficients of each network parameter targeted by the design method of the present invention are used as input weighting for fuzzy inference in fuzzy rules. While keeping the original membership degree of the fuzzy rules, the fuzzy rules and the congestion window update formula unchanged, only the weight coefficients are optimized, which is beneficial for automatically focusing on more discriminative signals in different environments.

[0052] d. The solution provided by this invention requires extremely low computational and storage overhead and can be directly embedded into existing switching / routing / host stacks; it automatically approaches more suitable weights in different business scenarios, without the need for frequent manual parameter tuning, and is easy to deploy and maintain at scale. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating a method for designing weight coefficients based on fuzzy rules, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0056] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0057] See Figure 1In one embodiment of the present invention, a method for designing weight coefficients based on fuzzy rules is provided. The weight coefficients obtained by the method of the present invention can be directly applied to the fuzzy inference steps of fuzzy rules to calculate the activation degree, and then substituted into W. new The updated congestion window is calculated in the formula, which can automatically approach a more suitable weight coefficient in different business scenarios, reduce the sawtooth effect and periodic degradation caused by multi-stream synchronization, and eliminate the need for frequent manual adjustment of the weight coefficient. This avoids the problem that the parameters are extremely sensitive to the scenario and rely on manual experience for maintenance, which is difficult to adapt to time-varying networks, as is the case with the preset weight coefficients of each network parameter in the existing technology.

[0058] Specifically, the design method of this embodiment includes the following steps:

[0059] Step 1: System initialization.

[0060] The system performs a global initialization process, namely startup configuration, which is a system-level initialization process. It is responsible for loading all rule bases, parameters, sliding windows and security mechanisms before the algorithm runs, so that the entire subsequent method process and fuzzy rules can start stably.

[0061] Step 2: Collect multiple network parameters and multiple auxiliary parameters within the measurement window, and standardize the collected auxiliary parameters.

[0062] Among them, several network parameters include bandwidth utilization, ECN tagging rate, and queue length, while several auxiliary parameters include throughput, round-trip time, latency jitter, packet loss rate, and bandwidth.

[0063] Specifically, the method for standardizing each auxiliary parameter is as follows:

[0064] The standardized throughput rate can be calculated using the following formula:

[0065] ,

[0066] Where T0 is the currently acquired throughput rate, T is the normalized throughput rate, P5G refers to the 5th percentile of the throughput rate within the measurement window, and P95G refers to the 95th percentile of the throughput rate within the measurement window.

[0067] The standardized round-trip time delay is calculated using the following formula:

[0068] ,

[0069] Where D0 is the currently collected round-trip delay, D is the standardized round-trip delay, P5rtt refers to the 5th percentile of the round-trip delay within the measurement window, and P95rtt refers to the 95th percentile of the round-trip delay within the measurement window.

[0070] The standardized time delay jitter is calculated using the following formula:

[0071] ,

[0072] Where J0 is the currently acquired latency jitter, J is the standardized latency jitter, P5jit refers to the 5th percentile of the latency jitter within the measurement window, and P95jit refers to the 95th percentile of the latency jitter within the measurement window.

[0073] The standardized packet loss rate can be calculated using the following formula:

[0074] ,

[0075] Where L0 is the currently collected packet loss rate, L is the standardized packet loss rate, P5loss refers to the 5th percentile of the packet loss rate within the measurement window, and P95loss refers to the 95th percentile of the packet loss rate within the measurement window.

[0076] The standardized bandwidth is calculated using the following formula:

[0077] ,

[0078] Where E0 is the currently acquired bandwidth, E is the standardized bandwidth, P5Ecn refers to the 5th percentile of the bandwidth within the measurement window, and P95Ecn refers to the 95th percentile of the bandwidth within the measurement window.

[0079] By standardizing the various auxiliary parameters, a sequence of indicators with consistent dimensions can be obtained.

[0080] Step 3: Perform adaptive reward aggregation on the standardized auxiliary parameters to obtain the final reward.

[0081] Step 3.1: Calculate the base reward.

[0082] The formula for calculating the basic reward is as follows: , where r base Based on the reward, λ T The reward aggregation weight for throughput rate, λ D The reward aggregation weight for round-trip delay, λ J For the reward aggregation weight of latency jitter, λ L The reward aggregation weight for packet loss rate, λ E The bandwidth reward aggregates weights, λT , λ D , λ J , λ L With λ E The sum of is 1, and λ T , λ D , λ J , λ L , λ E The values ​​of T, D, J, L, and E all range from 0 to 1. D is the normalized round-trip time, J is the normalized delay jitter, L is the normalized packet loss rate, and E is the normalized bandwidth.

[0083] Step 3.2: Calculate the final reward.

[0084] The formula for calculating the final reward is as follows: , where r is the final reward, and the value of the final reward r is in the range of 0-1; k is the tail penalty coefficient, and the value of k is in the range of 0.15-0.25; p is the tail penalty term.

[0085] In the formula for calculating the final reward mentioned above, the tail penalty term p is calculated using the following formula:

[0086] ,

[0087] Where τ is the tail penalty trigger threshold, and the value of τ ranges from 0.85 to 0.95; the max function is the maximum value function. The introduction of the max function can ensure that the tail penalty term p is always greater than or equal to 0, and there will be no negative penalty.

[0088] By combining the basic reward with the tail penalty to obtain the final reward r, the learning process is guaranteed to improve average performance metrics such as throughput and reduce latency, while effectively suppressing the high percentile tail risk of round-trip latency and packet loss, thus achieving the dual objective of "optimal performance and reliable stability". In addition, the final reward is amplitude-limited to ensure numerical stability, avoid unnecessary oscillations in online learning, and significantly improve the convergence speed and steady-state performance of congestion control.

[0089] Step 4: Establish the objective function based on the final reward and generate candidate weight coefficients.

[0090] Step 4.1: Establish the objective function.

[0091] The objective function is established based on the following formula: Where J is the optimization objective, r is the final reward, and λ is the final reward. H λ is the entropy regularity coefficient. H The value range is 0.001-0.01; H(w) is the entropy of the weight distribution, and the formula for calculating H(w) is: , where w old,i This represents the weight coefficient of the i-th network parameter that took effect in the previous cycle, where i = 1, 2, 3, i.e., w old,1 w is the weighting factor for bandwidth utilization in the previous period. old,2 w is the weighting factor for the ECN labeling rate in effect during the previous period. old,3 The weight coefficients that take effect in the previous cycle for the queue length; m represents the total number of all network parameters.

[0092] Step 4.2: Calculate gradient rewards.

[0093] The formula for calculating gradient rewards is as follows:

[0094] ,

[0095] Among them, g i Let g1 be the reward gradient for the i-th network parameter, i=1,2,3, i.e. g1 is the reward gradient for bandwidth utilization, g2 is the reward gradient for ECN labeling rate, and g3 is the reward gradient for queue length; δ is the magnitude of weight perturbation, and the value of δ ranges from 0.01 to 0.02. It is w old,i The reward obtained by offline recalculation after applying the +δ perturbation. It is w old,i The reward obtained by offline recalculation after applying the -δ perturbation. It is obtained by taking the gradient of J.

[0096] Step 4.3: Calculate the candidate weight coefficients.

[0097] The formula for calculating the candidate weight coefficient is as follows: , where w old w represents the weights of all network parameters that were active in the previous cycle. old Including w old,1 w old,2 and w old,3 ;w raw w represents the candidate weight coefficients for all network parameters. raw Including w raw,1 ( w raw,2 ( ) and w raw,3 ( ), that is, w raw,1 w are candidate weighting coefficients for bandwidth utilization. raw,2 w are candidate weighting coefficients for the ECN labeling rate. raw,3 η is the candidate weight coefficient for the queue length; η is the learning rate, and the value of η ranges from 0.1 to 0.5.

[0098] Step 5: Apply KL constraints to the candidate weight coefficients to obtain the KL divergence.

[0099] Step 5.1: Calculate the KL divergence.

[0100] The KL divergence is calculated by applying KL constraints to the candidate weight coefficients using the following formula:

[0101] ,

[0102] Among them, D KL The KL divergence is given by a KL threshold of 0.01–0.05; w raw,i This represents the candidate weight coefficient for the i-th network parameter, i=1,2,3, i.e., w raw,1 w are candidate weighting coefficients for bandwidth utilization. raw,2 w are candidate weighting coefficients for the ECN labeling rate. raw,3 The candidate weight coefficients are the queue length.

[0103] Step 5.2: Compare the KL divergence with the KL threshold.

[0104] When the KL divergence is less than or equal to the KL threshold, proceed to step 6.

[0105] When D KL When the value is greater than 2ε, the process is abbreviated to the step of generating candidate weight coefficients (step 4.3) to recalculate the candidate weight coefficients. Typically, when recalculating the candidate weight coefficients, the new η value is obtained by directly halving the original η value and recalculating, thereby ensuring that the update step size is small enough to guarantee the stability and controllability of the policy update. After obtaining the new candidate weight coefficients, the subsequent step 5 is executed again with the new candidate weight coefficients.

[0106] When D KL When the value is greater than ε and less than 2ε, a rollback is performed to calculate the safety weight coefficient, which is calculated using the following formula: , where w safe w represents the safety weight coefficients for all network parameters. safe Including w safe,1 ( w safe,2 ( ) and w safe,3 ( ), that is, w safe,1 w is the security weighting coefficient for bandwidth utilization. safe,2 w is the safety weighting coefficient for the ECN labeling rate. safe,3β is the safety weight coefficient for the queue length; β is the safety coefficient, which ranges from 0 to 1 and must satisfy the condition that the KL divergence obtained after applying KL constraints to the calculated safety weight coefficient is less than the KL threshold. safe w also needs to be calculated using KL constraints. safe KL divergence D KL To confirm the safety weighting coefficient w safe If it meets the requirements, then it means that the value of β is inappropriate and a new value of β needs to be selected to calculate the safety weight coefficient.

[0107] Step 6: Smooth the candidate weight coefficients or safety weight coefficients to obtain the final weight coefficients for use in fuzzy rules.

[0108] Specifically, the smoothing applied to the candidate weight coefficients or the safety weight coefficients is EMA smoothing, and the final weight coefficients are calculated using the following formula: , where w new The final weight coefficients for all network parameters can be directly applied to fuzzy rules to calculate and update the congestion window, w. new Including w new,1 w new,2 and w new,3 That is, w new,1 The final weighting coefficient for bandwidth utilization, w new,2 The final weighting coefficient for the ECN labeling rate, w new,3 is the final weighting coefficient for the queue length; α is the EMA smoothing coefficient, with a value ranging from 0.2 to 0.4; w0 is selected from w raw or w safe One of the steps, namely when comparing the KL divergence with the KL threshold in step 5.2, if the comparison result shows that the KL divergence is less than or equal to the KL threshold, the formula for calculating the final weight coefficient in step 6 is as follows: When comparing the KL divergence with the KL threshold in step 5.2, the comparison result is D. KL When the value is greater than ε and less than 2ε, the formula for calculating the final weight coefficient in step 6 is as follows: .

[0109] Finally, the calculated w new The three final weighting coefficients w included new,1 w new,2 and w new,3 In the fuzzy inference steps applied to fuzzy rules, the size of the updated congestion window can be obtained, which can better improve the timeliness and accuracy of network congestion control, while ensuring that congestion control meets the current network's comprehensive requirements for low latency, low jitter and high utilization.

[0110] The weight coefficient design method based on fuzzy rules of this invention adopts a small-step adaptive approach to the weight coefficients of various network parameters by introducing an online policy update with KL constraints. While keeping the original membership degree of the fuzzy rules, the fuzzy rules themselves, and the congestion window update formula unchanged, only the weight coefficients are optimized. This facilitates automatic focusing on more discriminative signals under different environments. Furthermore, the method operates independently of the fuzzy rules, achieving both rapid response and steady-state optimization. The design method eliminates the need for manual parameter tuning, removing the maintenance burden of manual tuning. It automatically converges to optimal weight coefficients as load, link quality, and service structure change. In addition, stabilization techniques such as KL constraints and entropy regularization enhance the ability to distinguish between congested and non-congested packet loss, mitigating response lag and oscillations, and improving consistency and portability across different scenarios.

[0111] Due to the advanced nature of this technical solution, the computational and storage overhead required by the solution provided by this invention is extremely low, and it can be directly embedded into existing switching / routing / host stacks; it automatically approaches more suitable weights in different business scenarios, without the need for frequent manual parameter tuning, and is easy to deploy and maintain at scale.

[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0113] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for designing weight coefficients based on fuzzy rules, characterized in that, The design method includes the following steps: After system initialization, multiple network parameters and multiple auxiliary parameters are collected within the measurement window, and the collected auxiliary parameters are standardized. The multiple network parameters include bandwidth utilization, ECN tagging rate, and queue length, while the multiple auxiliary parameters include throughput, round-trip time, latency jitter, packet loss rate, and bandwidth. The final reward is obtained by adaptively aggregating multiple standardized auxiliary parameters. Establish an objective function based on the final reward and generate candidate weight coefficients; Apply KL constraints to the candidate weight coefficients to obtain KL divergence. If the KL divergence is less than or equal to the KL threshold, proceed to the next step. If the KL divergence is greater than the KL threshold, perform step reduction to recalculate the candidate weight coefficients or backtrack to calculate the safe weight coefficients. The final weight coefficients are obtained by smoothing the candidate weight coefficients or the safe weight coefficients, which are then used for fuzzy rules. The method for establishing an objective function and generating candidate weight coefficients based on the final reward is as follows: First, establish the objective function according to the following formula: Where J is the optimization objective, r is the final reward, and λ is the final reward. H λ is the entropy regularity coefficient. H The value range is 0.001-0.01; H(w) is the entropy of the weight distribution, and the formula for calculating H(w) is: , where w old,i This represents the weight coefficient of the i-th network parameter in the previous cycle, where i = 1, 2, 3; m represents the total number of all network parameters. Then, calculate the gradient reward of the objective function using the following formula: ， Among them, g i Let δ represent the reward gradient of the i-th network parameter, and δ be the magnitude of the weight perturbation, with a value ranging from 0.01 to 0.

02. It is w old,i The reward obtained by offline recalculation after applying the +δ perturbation. It is w old,i The reward obtained by offline recalculation after applying the -δ perturbation. It is obtained by taking the gradient of J; Finally, the candidate weight coefficients are calculated using the following formula: , where w raw w represents the candidate weight coefficients for all network parameters. old η represents the weight coefficients of all network parameters that were effective in the previous cycle, and η is the learning rate, which ranges from 0.1 to 0.

5.

2. The method for designing weight coefficients based on fuzzy rules according to claim 1, characterized in that, The method for standardizing the multiple auxiliary parameters collected is as follows: the standardized throughput rate is calculated according to the following formula: ， Where T0 is the currently acquired throughput rate, T is the normalized throughput rate, P5G refers to the 5th percentile of the throughput rate within the measurement window, and P95G refers to the 95th percentile of the throughput rate within the measurement window. The standardized round-trip time delay is calculated using the following formula: ， Where D0 is the currently collected round-trip delay, D is the standardized round-trip delay, P5rtt refers to the 5th percentile of the round-trip delay within the measurement window, and P95rtt refers to the 95th percentile of the round-trip delay within the measurement window. The standardized time delay jitter is calculated using the following formula: ， Where J0 is the currently collected time delay jitter, J is the standardized time delay jitter, P5jit refers to the 5th percentile of the time delay jitter within the measurement window, and P95jit refers to the 95th percentile of the time delay jitter within the measurement window; The standardized packet loss rate can be calculated using the following formula: ， Where L0 is the currently collected packet loss rate, L is the standardized packet loss rate, P5loss refers to the 5th percentile of the packet loss rate within the measurement window, and P95loss refers to the 95th percentile of the packet loss rate within the measurement window. The standardized bandwidth is calculated using the following formula: ， Where E0 is the currently acquired bandwidth, E is the standardized bandwidth, P5Ecn refers to the 5th percentile of the bandwidth within the measurement window, and P95Ecn refers to the 95th percentile of the bandwidth within the measurement window.

3. The method for designing weight coefficients based on fuzzy rules according to claim 2, characterized in that, The method for adaptively aggregating multiple standardized auxiliary parameters to obtain the final reward is as follows: First, the basic reward is calculated according to the following formula: , where r base Based on the reward, λ T The reward aggregation weight for throughput rate, λ D The reward aggregation weight for round-trip delay, λ J For the reward aggregation weight of latency jitter, λ L The reward aggregation weight for packet loss rate, λ E The bandwidth reward aggregation weight is D, the normalized round-trip time is J, the normalized latency jitter is L, the normalized packet loss rate is E, and the normalized bandwidth is E. The final reward is then calculated using the following formula: , where r is the final reward, and the value of the final reward r is in the range of 0-1; k is the tail penalty coefficient, and the value of k is in the range of 0.15-0.25; p is the tail penalty term.

4. The method for designing weight coefficients based on fuzzy rules according to claim 3, characterized in that, The tail penalty term is calculated using the following formula: ， Where τ is the tail penalty trigger threshold, and the value of τ ranges from 0.85 to 0.95; the max function is the maximum value function.

5. The method for designing weight coefficients based on fuzzy rules according to claim 3, characterized in that, The λ T , λ D , λ J , λ L With λ E The sum of is 1, and λ T , λ D , λ J , λ L , λ E The values ​​of T, D, J, L, and E all range from 0 to 1.

6. The method for designing weight coefficients based on fuzzy rules according to claim 3, characterized in that, The method for applying KL constraints to candidate weight coefficients to obtain KL divergence is as follows: the KL divergence is calculated using the following formula: ， Among them, D KL The KL divergence is given by a KL threshold of 0.01–0.05; w raw,i This represents the candidate weight coefficient for the i-th network parameter.

7. The method for designing weight coefficients based on fuzzy rules according to claim 6, characterized in that, When D KL When D is greater than 2ε, the step is reduced to the step of generating candidate weight coefficients to recalculate the candidate weight coefficients; when D KL When the value is greater than ε and less than 2ε, backtrack to calculate the safety weight coefficient.

8. The method for designing weight coefficients based on fuzzy rules according to claim 7, characterized in that, The safety weight coefficient is calculated using the following formula: , where w safe Let β be the safety weight coefficient for all network parameters, and let β be the safety coefficient. The value of β is in the range of 0-1 and the value of β must satisfy the condition that the KL divergence obtained after applying KL constraints to the calculated safety weight coefficient is less than the KL threshold.

9. The method for designing weight coefficients based on fuzzy rules according to claim 8, characterized in that, The method for smoothing the candidate weight coefficients or safety weight coefficients to obtain the final weight coefficients is as follows: the final weight coefficients are calculated using the following formula. , where w new represents the final weight coefficients for all network parameters, α is the EMA smoothing coefficient, and the value of α ranges from 0.2 to 0.4; w0 is selected from w raw or w safe One of them.

10. The method for designing weight coefficients based on fuzzy rules according to claim 9, characterized in that, The w old Including w old,1 w old,2 and w old,3 The w raw Including w raw,1 w raw,2 and w raw,3 The w safe Including w safe,1 w safe,2 and w safe,3 The w new Including w new,1 w new,2 and w new,3 .

11. The method for designing weight coefficients based on fuzzy rules according to claim 1, characterized in that, The period of the measurement window is 400-600ms.

Images