Link determination method and device, computer equipment and storage medium

Abstract

The invention discloses a link determination method and device, computer equipment and a storage medium. The method comprises the steps that link attribute parameters corresponding to multiple links and packet attribute parameters of a data packet to be transmitted are acquired, the link attribute parameters and the packet attribute parameters of any link of the multiple links are fused, and fusion attribute parameters of any link are obtained, and further based on a link evaluation model and the fusion attribute parameters, link preference scores corresponding to the plurality of links are determined, the link preference scores are used for representing the degree of dependence of the to-be-transmitted data packets on the links, so that a target link is determined from the plurality of links according to the link preference scores, and the target link is used for transmitting the to-be-transmitted data packets. The link evaluation model can evaluate the link from multiple attribute dimensions such as the link, the data packet and the like, so that the link selection is more accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, in particular to a link determination method and device, a computer device and a storage medium. BACKGROUND

[0002] With the increase of business complexity, higher requirements are put forward for the reliability and efficiency of the transmission network in the process of data transmission between SOC (System on Chip) and OpenCPU (Open Central Processing Unit) modules.

[0003] At present, single-link data transmission is generally used to carry out data transmission between SOC and OpenCPU modules. Although it can meet the demand in a simple scenario, in a variable environment, single-link may not perform well when facing large data packet transmission due to insufficient bandwidth, and may not meet the real-time requirement when transmitting high-priority small data packets due to excessive delay and jitter.

[0004] In order to solve the above problems, a multi-link switching strategy is proposed, which aims to realize efficient and reliable data transmission by dynamically monitoring and utilizing the characteristics of multiple links. However, the current multi-link switching strategy usually evaluates based on a single performance indicator when selecting a link, ignoring the influence of other key performance indicators, thereby leading to inaccurate link selection. SUMMARY

[0005] Embodiments of the present application provide a link determination method, device, computer device and storage medium to at least solve the technical problem that when selecting a link, a single performance indicator is usually evaluated, other key performance indicators are ignored, and thus inaccurate link selection is caused.

[0006] According to an aspect of an embodiment of the present application, a link determination method is provided, comprising: obtaining link attribute parameters corresponding to a plurality of links respectively, and packet attribute parameters of a data packet to be transmitted; for any link of the plurality of links, fusing the link attribute parameters of the any link with the packet attribute parameters to obtain fused attribute parameters of the any link; inputting the fused attribute parameters of the plurality of links into a pre-trained link evaluation model respectively, to determine link preference scores corresponding to the plurality of links respectively by the link evaluation model; wherein the link preference score is used to represent the degree of dependence of the data packet to be transmitted on the link; and determining a target link from the plurality of links based on the link preference scores corresponding to the plurality of links respectively.

[0007] In an example embodiment, the link evaluation model comprises a plurality of regression trees; and the inputting the fusion attribute parameters of the plurality of links into the pre-trained link evaluation model respectively to determine the link preference scores corresponding to the plurality of links respectively by the link evaluation model comprises: for any link of the plurality of links, inputting the fusion attribute parameters of the any link into the plurality of regression trees respectively to obtain prediction scores output by the plurality of regression trees; and accumulating the prediction scores output by the plurality of regression trees to determine the link preference score corresponding to the any link.

[0008] In an example embodiment, the inputting the fusion attribute parameters of the any link into the plurality of regression trees to obtain prediction scores output by the plurality of regression trees comprises: inputting the fusion attribute parameters of the any link into the plurality of regression trees sequentially; in a case that a current regression tree in the inputting is a first regression tree, determining an output value of the first regression tree as the prediction score output by the current regression tree; and in a case that the current regression tree in the inputting is not the first regression tree, determining the prediction score of the current regression tree based on the output value of the current regression tree and a previous prediction score of a previous regression tree of the current regression tree.

[0009] In an example embodiment, the link evaluation model comprises a plurality of regression trees to be trained; and before the inputting the fusion attribute parameters of the plurality of links into the pre-trained link evaluation model to determine the link preference scores corresponding to the plurality of links respectively by the link evaluation model, the method further comprises: obtaining sample fusion attribute parameters of the plurality of links; inputting the sample fusion attribute parameters of the plurality of links into a target regression tree of the plurality of regression trees to be trained respectively to determine local ranking parameters and global ranking parameters corresponding to the plurality of links respectively; wherein the target regression tree is a first regression tree of the plurality of regression trees to be trained; the local ranking parameter is used to represent a difference degree between a local link ranking predicted by the target regression tree and a set local link ranking of any two links of the plurality of links; and the global ranking parameter is used to represent a difference degree between an overall link ranking predicted by the target regression tree and a set overall link ranking; for any link of the plurality of links, determining a target loss of the any link based on the local ranking parameter and the global ranking parameter of the any link; and training the plurality of regression trees to be trained based on the target losses of the plurality of links respectively until a training stop condition is met to stop, and obtaining a trained link evaluation model.

[0010] In an example embodiment, the training of the plurality of regression trees based on the target loss corresponding to the plurality of links respectively is stopped until a training stop condition is met, and a trained link evaluation model is obtained, including: determining a negative gradient corresponding to the plurality of links respectively based on the target loss of the plurality of links; fitting each of the negative gradients according to the target regression tree to update a regression tree parameter of the target regression tree, and obtaining a target regression tree after the regression tree parameter is updated; predicting a first sample prediction score of the plurality of links based on the target regression tree after the regression tree parameter is updated; and passing the first sample prediction score and the negative gradient to a next regression tree of the target regression tree to train the next regression tree of the target regression tree until the plurality of regression trees to be trained are all trained, and the trained link evaluation model is obtained.

[0011] In an example embodiment, the passing of the first sample prediction score and the negative gradient to the next regression tree of the target regression tree to train the next regression tree of the target regression tree includes: determining a gradient adjustment strategy according to the first sample prediction score and a set true label, adjusting the negative gradient according to the gradient adjustment strategy to obtain an adjusted negative gradient; fitting the adjusted negative gradient according to the next regression tree to update a regression tree parameter of the next regression tree, and obtaining a next regression tree after the regression tree parameter is updated; and predicting a second sample prediction score of the plurality of links based on the next regression tree after the regression tree parameter is updated to train the next regression tree of the target regression tree.

[0012] In an example embodiment, the fusing of the link attribute parameter of any link of the plurality of links with the packet attribute parameter to obtain a fused attribute parameter of the any link includes: constructing a link performance matrix based on the link attribute parameter of the any link and constructing a packet feature matrix based on the packet attribute parameter for any link of the plurality of links; splicing the link performance matrix and the packet feature matrix to obtain a fused matrix; and determining the fused matrix as the fused attribute parameter of the any link.

[0013] Another aspect of the present application provides a link determination apparatus, comprising: an acquisition module configured to acquire link attribute parameters corresponding to a plurality of links respectively, and a packet attribute parameter of a data packet to be transmitted; a fusion module configured to fuse, for any one of the plurality of links, the link attribute parameter of the any one link with the packet attribute parameter to obtain a fused attribute parameter of the any one link; a first determination module configured to input the fused attribute parameters of the plurality of links into a pre-trained link evaluation model respectively, so as to determine link preference scores corresponding to the plurality of links respectively by the link evaluation model; wherein the link preference score is used to represent a degree of dependence of the data packet to be transmitted on the link; and a second determination module configured to determine a target link from the plurality of links based on the link preference scores corresponding to the plurality of links respectively.

[0014] According to still another aspect of the embodiments of the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program, wherein the computer program is configured to execute the above-mentioned link determination method when running.

[0015] According to still another aspect of the embodiments of the present application, a computer device is also provided, and the computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the above-mentioned link determination method by the computer program.

[0016] The above-mentioned link determination method acquires link attribute parameters corresponding to a plurality of links respectively, and a packet attribute parameter of a data packet to be transmitted, fuses, for any one of the plurality of links, the link attribute parameter of the any one link with the packet attribute parameter to obtain a fused attribute parameter of the any one link, and further determines link preference scores corresponding to the plurality of links respectively based on a link evaluation model and the fused attribute parameter, wherein the link preference score is used to represent a degree of dependence of the data packet to be transmitted on the link, so as to determine a target link from the plurality of links according to the link preference scores, that is, the link evaluation model can evaluate the link from multiple attribute dimensions such as links and data packets, so that the selection of the link is more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which are included to provide a further understanding of the present application, constitute a part of the present application, and the illustrative embodiments of the present application and their description serve to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0018] Figure 1 is a schematic diagram of an application environment of an optional link determination method according to the embodiments of the present application;

[0019] Figure 2 is one of flowcharts of a link determination method according to the embodiments of the present application;

[0020] Figure 3 Figure 2 is a flow chart of another link determination method according to an embodiment of the present application;

[0021] Figure 4 Figure 3 is a flow timing chart of a link determination method according to an embodiment of the present application;

[0022] Figure 5 Figure 4 is a structural block diagram of an optional link determination apparatus according to an embodiment of the present application;

[0023] Figure 6 Figure 5 is a structural schematic diagram of an optional link determination product according to an embodiment of the present application;

[0024] Figure 7 Figure 6 is a structural schematic diagram of an optional computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative work should fall within the protection scope of the present application.

[0026] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to the process, method, product or device.

[0027] According to an aspect of an embodiment of the present application, a link determination method is provided. Optionally, in the present embodiment, the above-mentioned link determination method can be applied to a target device as shown in Figure 1 The target device is deployed with a system on chip (SOC) and an open central processing unit (OpenCPU) module. The OpenCPU and the SOC can be connected through multiple links, such as Figure 1The links 1, 2, …, N, etc. shown are used for data interaction. When determining the links, the control software on the OpenCPU module or the main processor of the SOC can be used to determine the links. The target device can be an Internet of Things device configured with an application program, and can include but is not limited to at least one of the following: a mobile phone (such as an Android phone, an iOS phone, etc.), a notebook computer, a tablet computer, a palm computer, a MID (Mobile Internet Device), a PAD, a desktop computer, a smart television, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, a VR (Virtual Reality) terminal, an AR (Augmented Reality) terminal, an MR (Mixed Reality) terminal, and the like. The target device can also be a server, which can be a single server or a server cluster composed of multiple servers, or a cloud server.

[0028] In combination Figure 1 As shown, the link determination method described above can be performed by a target device, which can be a terminal device or a server. The link determination method described above can be implemented by the terminal device or the server respectively, or by the terminal device and the server together.

[0029] The above is only an example, and the present embodiment is not limited specifically.

[0030] Optionally, as an optional implementation, as shown in Figure 2 As shown, the link determination method described above is applied to a server side, and includes steps S202-S208:

[0031] In step S202, link attribute parameters corresponding to a plurality of links are obtained, and a packet attribute parameter of a data packet to be transmitted is obtained.

[0032] It should be noted that the link attribute parameter is a quantitative index describing the characteristics of the link, and reflects the transmission capability and the current working state of the link. The link attribute parameter can include but is not limited to: bandwidth, delay, packet loss rate, jitter, and current load.

[0033] The packet attribute parameter is a description of the characteristics of the data packet to be transmitted. The packet attribute parameter includes but is not limited to: packet priority, packet size, and data type.

[0034] In some embodiments, the link attribute parameters can be obtained by real-time monitoring of the plurality of links at a communication level. For example, the bandwidth can be estimated by measuring the amount of data transmitted on the link within a certain time; the delay can be obtained by sending probe packets and recording the round-trip time; the packet loss rate and the jitter can be obtained by detecting the link state and analyzing the statistics; and the current load can be read from the link manager or dynamically calculated by monitoring the data flow.

[0035] In some embodiments, the data packet attribute parameters can be relatively directly obtained and determined when the data packet is generated. The data packet priority can be set by the business logic, such as that the data packet of an emergency control instruction has a higher priority than the data packet of a general sensor; the data packet size directly depends on the amount of encapsulated data; and the data type is identified based on the source and destination of the data packet and the type of information carried by the data packet.

[0036] In step S204, for any link of the plurality of links, the link attribute parameters of the any link are fused with the packet attribute parameters to obtain the fusion attribute parameters of the any link.

[0037] The fusion attribute parameters are parameters reflecting the comprehensive performance of the link in processing a specific data packet, which are obtained by comprehensively considering the link attribute parameters and the data packet attribute parameters. The fusion attribute parameters fuse the transmission capability, working state of the link, and the characteristics and requirements of the data packet, and are used to evaluate the suitability and efficiency of the link for transmitting the specific data packet.

[0038] In an exemplary embodiment, for any link of the plurality of links, the link attribute parameters of the any link are fused with the packet attribute parameters to obtain the fusion attribute parameters of the any link, including: for any link of the plurality of links, constructing a link performance matrix based on the link attribute parameters of the any link, and constructing a packet feature matrix based on the packet attribute parameters; splicing the link performance matrix and the packet feature matrix to obtain a fusion matrix; and determining the fusion matrix as the fusion attribute parameters of the any link.

[0039] The link performance matrix is an array used to comprehensively describe the performance characteristics of the network link, and the packet feature matrix is an array used to comprehensively describe the performance characteristics of the data packet.

[0040] Specifically, in order to eliminate the influence of the dimensions of the indicators, the fusion attribute parameters and the packet attribute parameters can be standardized, and then the link performance matrix and the packet feature matrix can be determined. After the link performance matrix and the packet feature matrix are determined, the link performance matrix and the packet feature matrix can be spliced to obtain a fusion matrix, and then the fusion matrix can be determined as the fusion attribute parameters of the any link.

[0041] In the above embodiments, by constructing the link performance matrix and the packet feature matrix, and then splicing them into the fusion matrix, a fusion attribute parameter can be generated for each link, which contains the comprehensive information of the link and the packet, thereby guiding the selection of the link in the subsequent process and ensuring that the packet can be transmitted through the most suitable link.

[0042] In step S206, the fusion attribute parameters of the plurality of links are respectively input into a pre-trained link evaluation model to determine the link preference scores corresponding to the plurality of links respectively through the link evaluation model; wherein the link preference score is used to represent the degree of dependence of the data packet to be transmitted on the link.

[0043] It can be understood that the link preference score is calculated by the link evaluation model according to the fusion attribute parameter obtained by fusing the link attribute parameter and the packet attribute parameter, which can reflect the degree of dependence of the data packet on the link performance and whether the current state of the link can meet the requirements of the data packet transmission.

[0044] In some embodiments, there are three links (links X, Y and Z), and their fusion attribute parameters are as follows: link X: high bandwidth (50 Mbps), low latency (10 ms), low packet loss rate (0.1%), low jitter (5 ms), low load (20%); link Y: medium bandwidth (20 Mbps), medium latency (30 ms), medium packet loss rate (0.5%), medium jitter (10 ms), low load (15%); link Z: low bandwidth (10 Mbps), high latency (50 ms), high packet loss rate (1%), high jitter (20 ms), very low load (5%). In the case where video conference data needs to be transmitted, the characteristics of the video conference data are high priority, extremely sensitive to delay and packet loss rate, and moderate packet size. After fusing these packet attribute parameters and link attribute parameters, the fusion attribute parameter of each link is obtained. The fusion attribute parameter is input into the pre-trained link evaluation model. For link X, due to its excellent performance in bandwidth, latency, packet loss rate and jitter, as well as the low load level, it is expected to get the highest link preference score in the link evaluation. For link Y, although the load is low, it is not as good as link X in terms of latency and packet loss rate, and the preference score is expected to be lower than link X. For link Z, although the load is very low, its bandwidth, latency, packet loss rate and jitter are far inferior to links X and Y, and its link preference score is expected to be the lowest. The high and low of the link preference score reflects the degree of dependence of the video conference data on each link. In the application scenario of video conference, link X is selected as the preferred transmission path based on the link preference score, to ensure low latency and low error rate of data transmission, thereby improving the smoothness and quality of the video conference.

[0045] In step S208, a target link is determined from the plurality of links based on the link preference scores corresponding to the plurality of links respectively.

[0046] It can be understood that the link preference scores of all the links can be sorted, the link preference scores of different links are compared, the link with the highest score is identified, and the higher the score is, the more suitable the link is for the transmission requirement of the current data packet. If there are multiple links with the same highest score, other factors (such as stability, historical performance or resource utilization of the link) can be further considered to make a final selection. After selecting the target link, an evaluation can be performed to confirm whether the link currently meets the basic requirements of data transmission. For example, if the packet loss rate of the target link suddenly rises, even if it has the highest preference score, the link selection can be re-evaluated. If the evaluation result of the target link meets the transmission requirement, the transmission path of the data packet is switched to the link.

[0047] The steps S202-S208 described above, by obtaining the link attribute parameters corresponding to the plurality of links respectively and the packet attribute parameters of the data packet to be transmitted, fusing the link attribute parameters and the packet attribute parameters of any one of the plurality of links to obtain the fusion attribute parameters of any one of the plurality of links, and further determining the link preference scores corresponding to the plurality of links respectively based on the link evaluation model and the fusion attribute parameters, the link preference score is used to represent the dependence of the data packet to be transmitted on the link, so as to determine the target link from the plurality of links according to the link preference score, that is, the link evaluation model can evaluate the link from multiple attribute dimensions such as link and data packet, so that the selection of the link is more accurate.

[0048] In an exemplary implementation, the link evaluation model includes a plurality of regression trees; the fusion attribute parameters of the plurality of links are respectively input into the pre-trained link evaluation model to determine the link preference scores corresponding to the plurality of links respectively by the link evaluation model, including: for any one of the plurality of links, the fusion attribute parameters of any one of the plurality of links are respectively input into the plurality of regression trees to obtain the prediction scores output by the plurality of regression trees; and the prediction scores output by the plurality of regression trees are accumulated to determine the link preference score corresponding to any one of the plurality of links.

[0049] The regression tree is a decision tree model for regression tasks, which can predict continuous value output. In this scenario, the regression tree is used to predict the performance score of the link in processing the data fused with the link attribute and the packet attribute, that is, the link preference score.

[0050] In particular, the link evaluation model includes multiple regression trees, each of which has learned how to predict a corresponding score based on the input fused attribute parameters. For each link, the fused attribute parameters of the link are input into each regression tree of the model. The fused attribute parameters are a combined parameter that includes link performance indicators (such as bandwidth, latency, packet loss rate, etc.) and packet characteristics (priority, size, etc.). Each regression tree independently predicts a score based on the input fused attribute parameters. The score here reflects the expected performance or dependence of the data packet when transmitted over the link. The predicted scores output by all regression trees for the same link are accumulated. The purpose of accumulation is to combine the contributions of all regression trees to the evaluation of the link, so as to obtain a more comprehensive and robust link preference score. The accumulated score is the link preference score of the link, reflecting the overall performance of the data packet using the link. The higher the score, the more suitable the link is for the current data packet transmission requirements.

[0051] In some embodiments, a LightGBM (Light Gradient Boosting Machine) model can be used as the link evaluation model, which can be composed of 100 regression trees, each of which has been trained to be able to evaluate the fused attribute parameters and predict the preference score of the link. For link I, its fused attribute parameters include bandwidth, latency, packet loss rate, packet priority, and packet size. These parameters are sequentially input into the 100 regression trees in the LightGBM model, the first regression tree outputs a predicted score of 0.75, the second regression tree outputs a predicted score of 0.83, and the 100th regression tree outputs a predicted score of 0.78. The predicted scores of all regression trees are accumulated to obtain the link preference score of link I.

[0052] In the above embodiments, by using a regression tree cluster (i.e., a model composed of multiple regression trees) to predict the link preference score, the advantages of ensemble learning of the model can be utilized to improve the accuracy and robustness of the evaluation.

[0053] In one exemplary implementation, the fused attribute parameters of any link are input into multiple regression trees respectively to obtain the predicted scores output by the multiple regression trees, including: sequentially inputting the fused attribute parameters of any link into multiple regression trees, in the case that the current input regression tree is the first regression tree, determining the output value of the first regression tree as the predicted score output by the current regression tree; in the case that the current input regression tree is not the first regression tree, determining the predicted score of the current regression tree based on the output value of the current regression tree and the previous predicted score of the previous regression tree of the current regression tree.

[0054] It should be noted that when the input current regression tree is the first regression tree, the link evaluation model has not accumulated any prediction information of other regression trees. At this time, the output value of the first regression tree is directly determined as the score of the first prediction of the link, that is, the prediction score output by the current regression tree. For regression trees other than the first regression tree, their predictions will be based on the output of the previous regression tree and their own evaluation of the input features, that is, the prediction score of the previous regression tree will be considered, and attempts will be made to correct or enhance the prediction to more accurately reflect the link preference of the input fusion attribute parameters. Specifically, for regression trees other than the first regression tree, the prediction will be based on the input features and the prediction residual of the previous regression tree (i.e. the difference between the true preference score and the prediction score of the previous regression tree). The goal is to learn how to reduce the residual and thus obtain a more accurate final score.

[0055] In the above embodiment, the determination of the prediction score through multiple regression trees not only utilizes the local information of each tree, but also integrates the information of all trees, effectively improving the accuracy of the link preference score.

[0056] In one exemplary implementation, the link evaluation model includes a plurality of regression trees to be trained; before the fusion attribute parameters of the plurality of links are respectively input into the pre-trained link evaluation model to determine the link preference scores corresponding to the plurality of links respectively by the link evaluation model, the link determination method further comprises: obtaining sample fusion attribute parameters of the plurality of links; inputting the sample fusion attribute parameters of the plurality of links into a target regression tree in the plurality of regression trees to be trained respectively to determine local ranking parameters and global ranking parameters corresponding to the plurality of links respectively; wherein the target regression tree is the first regression tree in the plurality of regression trees to be trained; the local ranking parameter is used to represent the difference between the local link ranking predicted by the target regression tree and the set local link ranking of any two links; the global ranking parameter is used to represent the difference between the overall link ranking predicted by the target regression tree and the set overall link ranking; for any link of the plurality of links, the target loss of any link is determined based on the local ranking parameter and the global ranking parameter of any link; the plurality of regression trees to be trained are trained based on the target loss corresponding to the plurality of links respectively, and the training is stopped when the training stop condition is met, and the trained link evaluation model is obtained.

[0057] It can be understood that when training the link evaluation model, sample data of a plurality of links can be collected, which includes performance indicators (such as bandwidth, delay, packet loss rate, etc.) and attributes (such as priority, size, etc.) of data packets of each link. These data constitute sample fusion attribute parameters for training the model.

[0058] The first target regression tree to be trained makes predictions based on sample fusion attribute parameters, calculating the difference between the local link ranking between any two links and the actual set local link ranking. For example, it can determine the ranking between link A and link B, and further obtain the actual ranking between link A and link B, thus calculating the local ranking parameters. The local ranking parameters are used to evaluate the accuracy of the model's ranking prediction within a small range, i.e., between link pairs.

[0059] In addition to local ranking parameters, the link evaluation model can also calculate the accuracy of the overall link ranking, that is, the degree of difference between the overall link ranking predicted by the model and the actual overall link ranking set. For example, if there are three links: link C, link D, and link E, the model can determine the ranking of links C, D, and E, and further obtain the actual ranking of links C, D, and E, thereby calculating the global ranking parameters. This helps to evaluate the consistency of the model in overall performance.

[0060] Specifically, for each link, a weighted calculation is performed based on its local and global ranking parameters to determine the target loss for that link. The local ranking parameters are each assigned a corresponding weight, which can be customized and typically adjusted according to the application scenario to balance local ranking accuracy and global ranking consistency. Further, based on the target losses corresponding to multiple links, the model begins training the first target regression tree. In each iteration, the model attempts to minimize the target loss, adjusting the tree's parameters using optimization algorithms such as gradient descent. After the first regression tree is trained, its predictions are used to train subsequent regression trees. Subsequent regression trees learn to correct the prediction errors of the first regression tree; this process continues until all regression trees are trained.

[0061] In the above embodiments, the link evaluation model can be continuously learned and optimized through the above training process to improve the accuracy and reliability of link evaluation. In particular, when dealing with multi-link switching problems, it can more intelligently predict link performance and make the best link selection decision.

[0062] In one embodiment, such as Figure 3 As shown, the above-mentioned regression trees to be trained are trained based on the target loss corresponding to each of the multiple links until the training stops when the stopping condition is met. The trained link evaluation model can be obtained through the following steps S302-S308:

[0063] S302, based on the target loss of multiple links, determines the negative gradients corresponding to each of the multiple links.

[0064] S304, fitting each negative gradient according to the target regression tree to update the regression tree parameters of the target regression tree, and obtaining the target regression tree after updating the regression tree parameters.

[0065] S306, predicting the first sample prediction scores of the plurality of links based on the target regression tree after updating the regression tree parameters.

[0066] S308, passing the first sample prediction scores and the negative gradients to the next regression tree of the target regression tree to train the next regression tree of the target regression tree until all the regression trees to be trained are trained, and obtaining the trained link evaluation model.

[0067] It should be noted that at the beginning of each iteration, the target loss of each link can be differentiated to determine the negative gradient. The negative gradient reflects the gap between the link preference score predicted by the model and the actual score, which is used to guide the update of the model parameters in the next step. The target regression tree is used to fit the negative gradient, and the target regression tree learns how to reduce these negative gradients, i.e., how to reduce the prediction error, by constructing a decision tree. After fitting the negative gradient, the target regression tree updates its internal parameters. The updated target regression tree can better fit the data and reduce the prediction error. The updated target regression tree is used to predict the training samples to obtain the first sample prediction scores. This is an estimate of the link preference score based on the current model state. The first sample prediction scores and the negative gradients are passed to the next tree of the target regression tree for subsequent regression tree training. The training target of the next tree is to further reduce the error that the previous tree has not completely fitted, i.e., to further fit the negative gradient to improve the overall prediction ability of the model. This process is recursively performed between multiple regression trees, and each tree is trained based on the prediction results of the previous trees to further reduce the prediction error. The training process will continue until one or more stopping conditions are met.

[0068] In some embodiments, the training stopping conditions can include reaching a preset maximum number of iterations, the prediction performance of the model no longer significantly improving, the complexity of the model reaching a preset upper limit, etc. When all the regression trees are trained and the training stopping conditions are met, a trained link evaluation model is obtained. This model includes a collection of all regression trees and can predict the link preference score by inputting the fused attribute parameters.

[0069] In the above embodiments, through this training process, a model capable of intelligently evaluating link preferences is constructed, which can consider link performance and packet attributes to provide accurate link evaluation scores, thereby supporting the decision of multi-link selection and switching strategies. This negative gradient-based training method can ensure that the model effectively reduces errors and improves prediction accuracy in each iteration.

[0070] In one example embodiment, the first sample prediction score and the negative gradient are passed to the next regression tree of the target regression tree to train the next regression tree of the target regression tree, including: determining a gradient adjustment strategy according to the first sample prediction score and the set true label, adjusting the negative gradient according to the gradient adjustment strategy to obtain an adjusted negative gradient; fitting the adjusted negative gradient based on the next regression tree to update the regression tree parameters of the next regression tree, and obtaining the next regression tree after the regression tree parameters are updated; and predicting a second sample prediction score of the plurality of links based on the next regression tree after the regression tree parameters are updated, to train the next regression tree of the target regression tree.

[0071] Wherein, at the beginning of each iteration, a gradient adjustment strategy is determined based on the difference between the first sample prediction score and the true label (actual link preference score), that is, how to adjust the negative gradient. The gradient adjustment strategy can include but is not limited to adjustment of learning rate, application of regularization, etc., to control the magnitude and direction of model update, avoid overfitting or oscillation. The negative gradient is adjusted by applying the gradient adjustment strategy to obtain an adjusted negative gradient. This adjusted negative gradient will be used as a guide signal for the next tree training, helping the model to more accurately fit the data.

[0072] In some embodiments, the adjusted negative gradient can be fitted into the next tree of the target regression tree as a target. The next tree of the target regression tree will learn how to reduce the negative gradient, that is, how to further reduce the prediction error. During the fitting process, the parameters of this tree are updated through the learning mechanism of the decision tree, such as optimizing the split point threshold, adjusting the score of the leaf node, etc., in order to more accurately fit the adjusted negative gradient. After the parameter update is completed, the next tree trained based on the adjusted negative gradient is obtained. Further, the updated next tree can be used to predict the training samples again to obtain a second sample prediction score. This is the latest estimate of the link preference score based on the current model state. The second sample prediction score will reflect the prediction ability of the model on the data after this update, and can be used to evaluate the current state of the model and make further optimization decisions. The above process will be repeated during the entire model construction until all the predetermined number of regression trees are trained or other stopping conditions (such as no significant performance improvement) are reached.

[0073] In the above embodiments, through iteration and optimization, the prediction error can be gradually reduced, and a highly accurate link evaluation model can be finally constructed. In actual link selection, the link evaluation model can make the best link selection and switching decisions according to the real-time state of the link and the attributes of the data packet.

[0074] Obviously, the above-described embodiments are only part of the embodiments of the present application, not all. In order to better understand the above method, the following describes the above process in combination with the embodiments, but is not used to limit the technical solutions of the embodiments of the present application, specifically:

[0075] Among them, the MLS (Multi-Link Switching, multi-link switching) technology realizes load balancing, fault recovery, and optimal utilization of link resources by dynamically switching data flow between different links. However, when performing multi-link switching, the performance of the link is usually evaluated from a single dimension, which may have certain deviation, resulting in problems in link selection and switching.

[0076] Based on this, the present application provides a link determination method, as shown in Figure 4 The flow chart of the link determination method provided by the embodiments of the present application is as follows:

[0077] S401, acquiring link attribute parameters. Collect the performance indicators of each link, including: bandwidth, delay, packet loss rate, jitter and current load. According to the performance indicators of each link, a link performance matrix A is constructed, is expressed as:

[0078]

[0079] Among them, B j represents bandwidth, D j represents delay, L j represents packet loss rate, J j represents jitter, W j represents the current load, n is the total number of links, and j = 1…n.

[0080] S402, acquiring packet attribute parameters. Collect the characteristics of the data packet, such as: data packet priority, data size, etc. According to the characteristics of the data packet, a feature matrix B is constructed, is expressed as:

[0081] B = [P norm S norm ]

[0082] Among them, P norm represents the priority of the data packet, P norm ∈ [1, 2, …, 5]. S norm represents the size of the data packet.

[0083] S403, normalizing the link state. In order to eliminate the influence of the dimension of the indicators, it is necessary to standardize each indicator. For performance indicators such as bandwidth, the larger the better, the formula Normalization is performed. For performance indicators such as delay and packet loss rate, the smaller the better, the formula is used Normalization is performed. By using multiple normalization strategies, the influence of dimension is reduced. For bandwidth type, the larger the better, and for delay type, the smaller the better, different normalization strategies are used to effectively reduce the influence of dimension on calculation. For each link j, the normalized feature A' is:

[0084]

[0085] S404, normalizing the data features. The normalization of the feature matrix B forms the matrix B' and is expressed as:

[0086] B' = [P' norm S' norm ]

[0087] S405, determine the fusion attribute parameter. The link feature matrix A' is spliced with the data link matrix B' to form the input matrix Then is expressed as:

[0088]

[0089] For each link feature row, the spliced data packet features are the same.

[0090] S406, use the link evaluation model to obtain the link score. The input matrix is put into the optimized LightGBM Ranker model, then for each row of an output preference score S j is output, and has:

[0091]

[0092] where γ represents the output weight (scaling factor) obtained by machine learning training, f t (x j ) represents the output of the t-th tree for sample x j .

[0093] S407, select the link with the highest score to transmit data. According to the generated preference score S j , the final output score vector S is formed, and S is expressed as:

[0094] S = [S1, S2, ···, S n ]

[0095] In the output score vector table S, the link with the highest score is selected, that is

[0096]

[0097] S408, updating the link performance matrix. After the completion of a data transmission, the performance data of the current link is obtained through the size of the data packet and the duration of the transmission, and the link performance matrix A is updated with the data.

[0098] Further, the training process of the link evaluation model is described: the input features of the model:

[0099] j = [B' j , D' j , L' j , J' j , W' j , P' norm , S' norm ]

[0100] Wherein, B' j represents the bandwidth, D' j represents the delay, L' j represents the packet loss rate, J' j represents the jitter, W j ' represents the current load of the link, P' norm represents the priority of the data packet, and S' norm represents the size of the data packet.

[0101] Loss function:

[0102]

[0103] Wherein, rel j represents the true relevance of the jth link, S i , S j represents the link score output by the model, Z represents the normalization factor, k represents the number of links, M represents the ranking of the first m links, avoids full calculation, and a represents the weight coefficient, represents the local ranking parameter, represents the global ranking parameter.

[0104] Gradient calculation and negative gradient generation: in each iteration, the negative gradient of the model needs to be calculated for training the next tree. That is:

[0105]

[0106] Wherein, β∈(0,1] is the gradient scaling factor. represents all link pairs (i,j) with higher true relevance but ranked later by the model among the top M highest score link pairs predicted by the current model (i.e. ρTop-M).​

[0107] Further, training a new tree f t Fitting negative gradient, goal is to fit gradient g j :

[0108]

[0109] Update model score:

[0110]

[0111] Where f t (x j ) is determined by training a new tree f t , which fits the negative gradient g j output by the model of the last round. Is the model score of the last tree of the newly trained tree.

[0112] Optimized model score:

[0113]

[0114] The link determination method provided by the application evaluates the link and data from multiple dimensions, so that the link selection is more accurate. The performance of the link is evaluated from multiple dimensions such as bandwidth, delay, and current load, and the size and priority of the data are combined. The selection and switching of the link are more accurate, which can effectively improve the transmission efficiency of the data and ensure the load balancing of the link. The optimized LightGBM Ranker model is used, the parameter amount of the model is reduced by optimizing the loss function and reducing the depth of the tree, the training speed is improved, and the calculation speed and accuracy are guaranteed. At the same time, it is ensured that the model can realize the calculation of the link performance weight coefficient on a low-resource device.

[0115] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all described as a series of action combinations, but those skilled in the art should know that the application is not limited to the action sequence described, because according to the application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the application.

[0116] According to another aspect of the embodiments of the application, a link determination apparatus for implementing the above-mentioned link determination method is also provided. As Figure 5 shown, the apparatus comprises:

[0117] The acquisition module 502 is configured to acquire link attribute parameters corresponding to a plurality of links respectively, and packet attribute parameters of a data packet to be transmitted.

[0118] fusing the link attribute parameter of any one of the plurality of links with the packet attribute parameter to obtain a fused attribute parameter of the any one of the plurality of links;

[0119] The first determining module 506 is configured to input the fused attribute parameters of the plurality of links into a pre-trained link evaluation model respectively, to determine link preference scores respectively corresponding to the plurality of links by the link evaluation model; wherein the link preference score is used to represent a degree of dependence of the data packet to be transmitted on the link.

[0120] The second determining module 508 is configured to determine a target link from the plurality of links based on the link preference scores respectively corresponding to the plurality of links.

[0121] Through the above device, the link attribute parameters respectively corresponding to the plurality of links and the packet attribute parameter of the data packet to be transmitted are obtained, the link attribute parameter of any one of the plurality of links is fused with the packet attribute parameter to obtain the fused attribute parameter of the any one of the plurality of links, and further based on the link evaluation model and the fused attribute parameter, the link preference scores respectively corresponding to the plurality of links are determined, the link preference score is used to represent the degree of dependence of the data packet to be transmitted on the link, so as to determine the target link from the plurality of links according to the link preference score, that is, the link evaluation model can evaluate the link from multiple attribute dimensions such as link and data packet, so that the selection of the link is more accurate.

[0122] In one embodiment, the link evaluation model includes a plurality of regression trees; the first determining module 506 is further configured to input the fused attribute parameter of any one of the plurality of links into the plurality of regression trees respectively to obtain prediction scores output by the plurality of regression trees; and accumulate the prediction scores output by the plurality of regression trees to determine the link preference score corresponding to the any one of the plurality of links.

[0123] In one embodiment, the first determining module 506 is further configured to input the fused attribute parameter of the any one of the plurality of links into the plurality of regression trees in sequence, in a case where a current regression tree of the input is a first regression tree, determine an output value of the first regression tree as the prediction score output by the current regression tree; and in a case where the current regression tree of the input is not the first regression tree, determine the prediction score of the current regression tree based on the output value of the current regression tree and a previous prediction score of a previous regression tree of the current regression tree.

[0124] In an embodiment, the link evaluation model comprises a plurality of regression trees to be trained; the apparatus further comprises a training module; the training module is configured to obtain sample fusion attribute parameters of the plurality of links; input the sample fusion attribute parameters of the plurality of links into a target regression tree of the plurality of regression trees to be trained respectively, to determine local ranking parameters and global ranking parameters corresponding to the plurality of links respectively; the target regression tree is the first regression tree of the plurality of regression trees to be trained; the local ranking parameter is used to represent a difference between a local link ranking of any two links predicted by the target regression tree and a set local link ranking of the any two links; the global ranking parameter is used to represent a difference between an overall link ranking predicted by the target regression tree and a set overall link ranking; for any link of the plurality of links, a target loss of the any link is determined based on the local ranking parameter and the global ranking parameter of the any link; the plurality of regression trees to be trained are trained based on the target losses corresponding to the plurality of links respectively, until a training stop condition is met to stop, and a trained link evaluation model is obtained.

[0125] In an embodiment, the training module is further configured to determine negative gradients corresponding to the plurality of links respectively based on the target losses of the plurality of links; fit each of the negative gradients according to the target regression tree, to update regression tree parameters of the target regression tree, obtain a target regression tree after updating the regression tree parameters, predict first sample prediction scores of the plurality of links based on the target regression tree after updating the regression tree parameters, and pass the first sample prediction scores and the negative gradients to a next regression tree of the target regression tree, to train the next regression tree of the target regression tree, until the plurality of regression trees to be trained are all trained, and a trained link evaluation model is obtained.

[0126] In an embodiment, the training module is further configured to determine a gradient adjustment strategy according to the first sample prediction scores and set true labels, adjust the negative gradients according to the gradient adjustment strategy, obtain adjusted negative gradients, fit the adjusted negative gradients according to the next regression tree, update regression tree parameters of the next regression tree, obtain a next regression tree after updating the regression tree parameters, and predict second sample prediction scores of the plurality of links based on the next regression tree after updating the regression tree parameters, to train the next regression tree of the target regression tree.

[0127] In one embodiment, the fusion module 504 is further configured to construct a link performance matrix based on the link attribute parameters of any link of the plurality of links and construct a packet feature matrix based on the packet attribute parameters for the any link; splice the link performance matrix and the packet feature matrix to obtain a fusion matrix; and determine the fusion matrix as the fusion attribute parameters of the any link.

[0128] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with predetermined functions, and works together with other related parts to achieve predetermined targets, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an integral module or unit that contains the functions of the module or unit.

[0129] As to the apparatus in the above-described embodiments, the specific manners in which the respective modules perform operations have been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0130] According to an aspect of the present application, there is provided a computer program product including a computer program.

[0131] The above-mentioned sequence numbers of the embodiments of the present application are only for description, and do not represent advantages or disadvantages of the embodiments.

[0132] Figure 6 The computer system structure block diagram of the computer device for implementing the embodiments of the present application is schematically shown.

[0133] It should be noted that, Figure 6 The computer system 600 of the computer device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the present application.

[0134] As Figure 6As shown, the computer system 600 includes a central processing unit 601 (CPU), which can perform various suitable actions and processes according to programs stored in a read-only memory 602 (ROM) or loaded from a storage section 608 into a random access memory 603 (RAM). In the random access memory 603, various programs and data required for system operation are also stored. The central processing unit 601, the read-only memory 602, and the random access memory 603 are connected to each other through a bus 604. An input / output interface 605 (I / O interface) is also connected to the bus 604.

[0135] Connected to the input / output interface 605 are an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a local area network card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output interface 605 as necessary. A removable recording medium 66 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 610 as necessary, so that a computer program read out therefrom is installed in the storage section 608 as necessary.

[0136] In particular, according to embodiments of the present application, the processes described in the various method flowcharts can be implemented as a computer software program. For example, embodiments of the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network through the communication section 609, and / or installed from the removable recording medium 66. When the computer program is executed by the central processing unit 601, various functions defined in the system of the present application are performed.

[0137] In such embodiments, the computer program can be downloaded and installed from a network through the communication section 609, and / or installed from the removable recording medium 66. When the computer program is executed by the central processing unit 601, various functions provided by embodiments of the present application are performed.

[0138] According to still another aspect of the embodiments of the present application, a computer device for implementing the above link determination method is further provided, which can be Figure 1 The computer device in the embodiments of the present application can be a terminal device or a server. The embodiments of the present application are described taking the computer device as a terminal device. Figure 7 As shown in FIG. 7, the computer device includes a memory 702 and a processor 704, the memory 702 stores a computer program, and the processor 704 is configured to execute the steps in the above at least one method embodiment through the computer program.

[0139] Optionally, in the embodiments, the above computer device can be located in at least one of the network devices in the computer network.

[0140] Optionally, in the embodiments, the above processor can be configured to execute the methods in the embodiments of the present application through the computer program.

[0141] Optionally, those skilled in the art can understand that the above computer device can be a terminal device or a server, and the above computer device can be a terminal device or a server. Figure 7 As shown in FIG. 7, the structure is only schematic, Figure 7 which does not limit the structure of the above computer device. For example, the computer device can include more or less components (such as a network interface) than those shown in FIG. 7, or have a different configuration from that shown in FIG. 7. Figure 7 Figure 7

[0142] The memory 702 can be used to store software programs and modules, such as program instructions / modules corresponding to the link determination method and device in the embodiments of the present application. The processor 704 executes various function applications and data processing by running the software programs and modules stored in the memory 702, that is, implements the above link determination method. The memory 702 can include a high-speed random access memory, and can further include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 702 can further include a memory remotely arranged with respect to the processor 704, which can be connected to the terminal through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. Specifically, the memory 702 can be but is not limited to used for device information and the like. As an example, as shown in FIG. 7, the memory 702 can include but is not limited to the above link determination device, which will not be described herein again in this example. Figure 7

[0143] ​​​Optionally, the transmission device 706 is configured to receive or send data via a network. Examples of the network can include a wired network and a wireless network. In an example, the transmission device 706 includes a network interface controller (NIC) which can be connected to other network devices and routers through a network cable to communicate with the Internet or a local area network. In an example, the transmission device 706 is a radio frequency (RF) module which is configured to communicate with the Internet in a wireless manner.

[0144] In addition, the computer device further includes a display 708 configured to display data transmitted by the target device, and a connection bus 710 configured to connect various module components in the computer device.

[0145] In other embodiments, the terminal device or the server can be a node in a distributed system, and the distributed system can be a blockchain system. The blockchain system can be a distributed system formed by the plurality of nodes communicating with each other via a network. The nodes can form a peer-to-peer network, and any computing device such as a server, a terminal, or the like can become a node in the blockchain system by joining the peer-to-peer network.

[0146] According to an aspect of the present disclosure, a computer readable storage medium is provided. The processor of the computer device reads the computer instructions from the computer readable storage medium. The processor executes the computer instructions to enable the computer device to perform the link determination method provided in the various optional implementation manners of the link determination aspect.

[0147] Optionally, in the present embodiment, the computer readable storage medium can be configured to store the program for executing the method in the embodiments of the present disclosure.

[0148] Optionally, in the present embodiment, it can be understood by those skilled in the art that all or part of the steps in the various methods of the above embodiments can be completed by a program instructing the hardware of the terminal device. The program can be stored in a computer readable storage medium, and the storage medium can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0149] The serial numbers of the embodiments of the present disclosure are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0150] The integrated units in the above embodiments, if implemented in the form of software function units and sold or used as independent products, can be stored in the above computer-readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions to make one or more computer devices execute all or part of the steps of the methods described in the various embodiments of the present application.

[0151] In the above embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0152] In the several embodiments provided in the present application, it should be understood that the disclosed application program can be implemented in other ways. Among them, the above-mentioned device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, unit or module, and can be electrical or other forms.

[0153] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment.

[0154] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of software function unit.

[0155] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principle of the present application, some improvements and refinements can be made, and these improvements and refinements should be regarded as the protection scope of the present application.

Claims

1. A link determination method, characterized in that, include: Obtain the link attribute parameters corresponding to multiple links, as well as the packet attribute parameters of the data packets to be transmitted; For any link among the multiple links, the link attribute parameters of the any link are fused with the packet attribute parameters to obtain the fused attribute parameters of the any link; The fusion attribute parameters of the multiple links are respectively input into a pre-trained link evaluation model to determine the link preference scores corresponding to the multiple links; wherein, the link preference scores are used to characterize the degree of dependence of the data packet to be transmitted on the link; Based on the link preference scores corresponding to the multiple links, the target link is determined from the multiple links.

2. The method according to claim 1, characterized in that, The link evaluation model includes multiple regression trees; the step of inputting the fusion attribute parameters of the multiple links into the pre-trained link evaluation model to determine the link preference scores corresponding to the multiple links through the link evaluation model includes: For any link among the multiple links, the fusion attribute parameters of any link are input into the multiple regression trees respectively to obtain the prediction scores output by the multiple regression trees; The predicted scores from the multiple regression tree outputs are summed to determine the link preference score corresponding to any given link.

3. The method according to claim 2, characterized in that, The step of inputting the fusion attribute parameters of any link into the plurality of regression trees to obtain the prediction scores output by the plurality of regression trees includes: The fusion attribute parameters of any link are sequentially input into the plurality of regression trees. If the current regression tree is the first regression tree, the output value of the first regression tree is determined as the predicted score output by the current regression tree. If the input current regression tree is not the first regression tree, the prediction score of the current regression tree is determined based on the output value of the current regression tree and the previous prediction score of the previous regression tree.

4. The method according to claim 1, characterized in that, The link evaluation model includes multiple regression trees to be trained; before inputting the fusion attribute parameters of the multiple links into the pre-trained link evaluation model to determine the link preference scores corresponding to the multiple links respectively through the link evaluation model, the method further includes: Obtain the sample fusion attribute parameters of the multiple links; The sample fusion attribute parameters of the multiple links are respectively input into the target regression tree in the multiple regression trees to be trained, so as to determine the local ranking parameters and global ranking parameters corresponding to the multiple links respectively; wherein, the target regression tree is the first regression tree in the multiple regression trees to be trained; the local ranking parameters are used to characterize the degree of difference between the local link ranking of any two links predicted by the target regression tree and the set local link ranking of any two links; the global ranking parameters are used to characterize the degree of difference between the overall link ranking predicted by the target regression tree and the set overall link ranking. For any link among the multiple links, the target loss of any link is determined by weighting the local sorting parameters and the global sorting parameters of that link. Based on the target loss corresponding to each of the multiple links, the multiple regression trees to be trained are trained until the training stops when the training stopping condition is met, thus obtaining the trained link evaluation model.

5. The method according to claim 4, characterized in that, The process of training the multiple regression trees to be trained based on the target losses corresponding to the multiple links until the training stops when the stopping condition is met, thereby obtaining a trained link evaluation model, includes: Based on the target loss of the multiple links, the negative gradients corresponding to the multiple links are determined respectively; Fit each of the negative gradients to the target regression tree to update the regression tree parameters of the target regression tree, and obtain the target regression tree after updating the regression tree parameters; Based on the updated regression tree parameters, predict the first sample prediction score of the multiple links; The predicted score of the first sample and the negative gradient are passed to the next regression tree of the target regression tree to train the next regression tree of the target regression tree until all the regression trees to be trained are trained, and a trained link evaluation model is obtained.

6. The method according to claim 5, characterized in that, The step of passing the predicted score of the first sample and the negative gradient to the next regression tree of the target regression tree to train the next regression tree of the target regression tree includes: Based on the predicted score of the first sample and the set true label, a gradient adjustment strategy is determined, and the negative gradient is adjusted according to the gradient adjustment strategy to obtain the adjusted negative gradient. The adjusted negative gradient is fitted based on the next regression tree to update the regression tree parameters of the next regression tree, thereby obtaining the next regression tree with updated regression tree parameters; Based on the next regression tree after updating the regression tree parameters, predict the second sample prediction score of the multiple links to train the next regression tree of the target regression tree.

7. The method according to claim 1, characterized in that, For any one of the multiple links, the link attribute parameters of that link are fused with the packet attribute parameters to obtain the fused attribute parameters of that link, including: For any one of the multiple links, a link performance matrix is ​​constructed based on the link attribute parameters of the link, and a packet feature matrix is ​​constructed based on the packet attribute parameters. The link performance matrix and the packet feature matrix are concatenated to obtain a fusion matrix; The fusion matrix is ​​determined as the fusion attribute parameter of any link.

8. A link determination device, characterized in that, include: The acquisition module is used to acquire the link attribute parameters corresponding to multiple links, as well as the packet attribute parameters of the data packets to be transmitted. The fusion module is used to fuse the link attribute parameters of any link with the packet attribute parameters for any link among the multiple links to obtain the fused attribute parameters of the any link. The first determining module is used to input the fusion attribute parameters of the multiple links into a pre-trained link evaluation model, so as to determine the link preference scores corresponding to the multiple links through the link evaluation model; wherein, the link preference scores are used to characterize the degree of dependence of the data packet to be transmitted on the link; The second determining module is used to determine the target link from the multiple links based on the link preference scores corresponding to the multiple links respectively.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein the computer program can be executed by a computer device to perform the method described in at least one of claims 1 to 7.

10. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in at least one of claims 1 to 7 via the computer program.

Images