Software-defined multi-path network flow real-time dynamic optimization method

Abstract

The invention relates to the technical field of network communication, and discloses a software-defined multi-path network flow real-time dynamic optimization method, which comprises the following steps of: acquiring an initial application experience intention of a user on a service flow; collecting context information of the service flow, and performing multi-mode user experience quality perception on the service flow to obtain a current user experience quality evaluation result; under specific conditions, performing dynamic intention negotiation with a user or an agent application to obtain a negotiated user intention; generating or adjusting a multi-path network strategy according to the negotiated user intention, the current user experience quality evaluation result, the context information and the state information of each available path of the current network; controlling the transmission of the service flow on a plurality of paths according to a multi-path network strategy; according to the invention, the dynamic change experience requirement of the user can be more accurately met, and the user satisfaction and the network automation level are improved.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and more specifically, to a software-defined method for real-time dynamic optimization of multipath network flows. Background Technology

[0002] In modern network applications, especially for services sensitive to Quality of Experience (QoE) such as video conferencing, online gaming, and real-time interaction, ensuring a good user experience is crucial. Software-defined networking (SDN) separates the control plane from the data plane, bringing greater flexibility and programmability to the network, making it possible to dynamically adjust network resources and optimize traffic paths. Multipath transmission technology utilizes multiple available paths in the network to improve bandwidth, enhance reliability, or reduce latency.

[0003] However, existing technologies for optimizing network flows using SDN and multipathing often rely primarily on monitoring and optimizing key performance indicators (KPIs) at the network layer, such as bandwidth, latency, and packet loss rate. While this approach can improve network performance to some extent, a gap remains between it and the actual user experience. A user's QoE is multi-dimensional, subjective, and dynamically changing; simply optimizing network KPIs does not always directly equate to an improved user experience. For example, even with good network KPIs, inappropriate encoding / decoding parameters or application-layer configurations can lead to a poor user experience.

[0004] Furthermore, user intent is often high-level and vague (e.g., "I want this meeting to be important and not to lag"), and accurately translating such intent into specific, actionable network control policies is a challenge. Existing methods mostly employ static policy mapping or rely on manual configuration by administrators, making it difficult to adapt to dynamic changes in user intent and the real-time evolution of application context (e.g., an important person joins the meeting midway, or a game enters a critical battle phase, at which point the focus of QoE requirements may change).

[0005] Therefore, current technology lacks an effective mechanism that can comprehensively perceive the user's multi-dimensional QoE, dynamically understand and negotiate the user's intent, and evolve multi-path network strategies in real time according to the context. This makes it difficult to truly achieve user-centric intelligent network optimization, affecting user satisfaction and the effective utilization of network resources. Summary of the Invention

[0006] This invention provides a software-defined method for real-time dynamic optimization of multi-path network flows, which solves the technical problems in related technologies such as inaccurate mapping between user intent and network behavior, single QoE perception dimension, and difficulty in adapting static strategies to dynamic requirements.

[0007] This invention provides a software-defined method for real-time dynamic optimization of multipath network flows, comprising the following steps: Obtain the user's initial application experience intent for the business flow; Collect contextual information about the business flow, perform multimodal user experience quality perception on the business flow, and obtain the current user experience quality assessment results; When the initial application experience intent is unclear, or when there is a deviation between the initial application experience intent and the current user experience quality assessment result, or when the context information of the business flow changes, dynamic intent negotiation is carried out with the user or agent application to obtain the negotiated user intent. Based on the negotiated user intent, the current user experience quality assessment results, context information, and the status information of each available path in the current network, generate or adjust the multi-path network strategy. Control the transmission of service flows on multiple paths according to the multi-path network strategy.

[0008] In a preferred embodiment, the step of obtaining the user's initial application experience intent for the service flow includes: Receive the user's initial application experience intent expressed in natural language or declarative strategies through the northbound interface; An intent understanding model is used to analyze the initial application experience intent, and preliminary quantitative parameters of user experience quality expectations are obtained.

[0009] In a preferred embodiment, the intent understanding model is based on natural language processing technology and knowledge graph construction to identify key entities and relationships between entities from natural language intents, and combines application profile database and user historical behavior data to parse the initial application experience intent into preliminary quantified user experience quality expectation parameters.

[0010] In a preferred embodiment, the step of performing multimodal user experience quality perception on the service flow includes: By monitoring probes and application programming interfaces, network layer performance parameters, application layer performance parameters, user feedback information, and context information related to business flows are collected. A multimodal user experience quality assessment model is used to fuse and process the collected parameters and information to obtain the current user experience quality assessment result, which is a multi-dimensional user experience quality vector.

[0011] In a preferred embodiment, the step of dynamically negotiating intent with the user or agent application includes: When preset triggering conditions are met, an interactive negotiation is initiated with the user or agent application through the user interface or application programming interface. The interactive negotiation includes at least one of asking questions, providing options, or displaying a comparison of user experience quality. Receive feedback from users or agent applications and transform the feedback into negotiated user intents, which may include weight adjustments for specific user experience quality dimensions or confirmation of target values ​​for key performance indicators.

[0012] In a preferred embodiment, the step of generating or adjusting the multipath network strategy includes: We adopt a user experience quality-driven strategy generation and evolution model, aiming to maximize the utility function of user satisfaction. By combining negotiated user intent, current user experience quality assessment results, context information, and the status information of each available path in the current network, we can calculate and output the optimal multi-path network strategy in real time.

[0013] In a preferred embodiment, the multi-path network strategy includes at least one of the following: selecting the optimal combination of one or more paths for the service flow, the proportion of traffic distribution and load balancing on the selected multiple paths, and adjusting network resource parameters on the paths.

[0014] In a preferred embodiment, a software-defined real-time dynamic optimization method for multipath network flows further includes: Continuously monitor the user experience quality results after the multi-path network strategy is implemented; When the user experience quality outcome fails to meet the expectations of the negotiated user intent or deteriorates, an alert is triggered, allowing a rollback from the policy version repository to a previously valid policy version.

[0015] In a preferred embodiment, a software-defined real-time dynamic optimization method for multipath network flows further includes: Cases of poor strategy execution, including the negotiation intent, contextual information, network state, executed strategy, and resulting user experience quality, are used to optimize the intent understanding model, the multimodal user experience quality assessment model, and the user experience quality-driven strategy generation and evolution model.

[0016] In a preferred embodiment, a software-defined real-time dynamic optimization system for multipath network flows is used to execute a software-defined real-time dynamic optimization method for multipath network flows, including: The intent acquisition module is used to acquire the user's initial application experience intent for the business flow; The user experience quality perception module is used to collect contextual information of the business flow and perform multimodal user experience quality perception on the business flow to obtain the current user experience quality assessment result. The intent negotiation module is used to dynamically negotiate the user intent with the user or agent application when the initial application experience intent is unclear, or when there is a deviation between the initial application experience intent and the current user experience quality assessment result, or when the context information of the business flow changes, so as to obtain the negotiated user intent. The strategy generation module is used to generate or adjust multi-path network strategies based on the negotiated user intent, the current user experience quality assessment results, context information, and the status information of each available path in the current network. The control execution module is used to control the transmission of service flows on multiple paths according to the multi-path network policy.

[0017] The beneficial effects of this invention are as follows: By introducing dynamic intent negotiation, multimodal QoE awareness, and context-driven policy evolution, this invention can more accurately understand and meet users' dynamically changing real-world experience needs in different application scenarios. Compared to existing optimization methods that rely on static configuration or simple threshold judgments, this invention can improve user satisfaction in complex application scenarios.

[0018] For example, when conducting high-definition video conferencing, this implementation can adjust the multi-path transmission strategy for audio and video streams in real time based on contextual information such as the importance of the meeting, the speaker's identity, and the dynamic changes in the number of participants. It prioritizes ensuring the high definition and low latency of audio and video data of key personnel, providing a better experience even when network resources are relatively scarce.

[0019] For example, in applications that are highly sensitive to network latency, such as online games, this implementation can intelligently adjust the path selection strategy and traffic scheduling priority according to different stages of the game, thereby ensuring that the player's operations at critical moments can be responded to in a timely manner.

[0020] Furthermore, through closed-loop adaptive control and policy version management (including optional automatic backtracking mechanism), the overall flexibility, adaptability, and robustness of network policies in the face of emergencies are further improved. Ultimately, user-centric intelligent multi-path network optimization is achieved, effectively bridging the mapping gap between user intent and specific network behavior, thereby improving the level of network automation management and overall operation and maintenance efficiency. Attached Figure Description

[0021] Figure 1 This is a flowchart of a software-defined real-time dynamic optimization method for multipath network flow according to the present invention. Detailed Implementation

[0022] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses a software-defined real-time dynamic optimization method for multi-path network flows, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the user's initial application experience intent for the business flow; Specifically, the following steps are included: Step 1.1, Receive intent input; The system receives initial application experience intent from users or network administrators via northbound interfaces (e.g., graphical user interface, command line, or API) in the form of natural language (e.g., “I want video conference A to be very smooth and have high-definition picture quality”) or declarative policies (e.g., “Business X: latency <20ms, bandwidth >10Mbps”).

[0024] The input data type is either a text string or a structured strategy description.

[0025] Step 1.2, construct the intent understanding model; This application provides a method for constructing an intent understanding model, for example, the model is constructed by hybrid construction based on Natural Language Processing (NLP) technology and knowledge graph.

[0026] The intent understanding model specifically includes an NLP processing unit and a knowledge reasoning unit.

[0027] NLP processing unit: For example, the BERT (Bidirectional Encoder Representations from Transformers) model combined with the CRF (Conditional Random Fields) layer can be used for sequence labeling to identify key entities from natural language intent (input data), such as business type (e.g., "video conferencing"), QoE descriptors (e.g., "smooth" and "high-definition"), and specific application names (e.g., "application X").

[0028] In some implementations, the unit can also directly parse the parameters of a structured strategy declaration.

[0029] In addition, this unit can also use relation extraction algorithms (such as dependency parsing-based methods or deep learning-based joint extraction models) to identify the relationships between these entities.

[0030] For example, in the statement "I want video conference A to be very smooth and have high-definition video", "video conference A" is identified as an application, and "smooth" and "high-definition" are its QoE expectations.

[0031] Knowledge Reasoning Unit: This unit utilizes a pre-built application profile knowledge graph, where nodes represent application types, QoE dimensions, KPI metrics, etc., and edges represent the relationships between them. For example, "high-definition video" is usually associated with "high resolution", "high frame rate", "low packet loss rate", etc.

[0032] After the NLP processing unit extracts entities and relationships, the knowledge reasoning unit queries and reasons on the knowledge graph, mapping the fuzzy QoE descriptors to specific KPI indicator sets.

[0033] The inference unit can also be combined with a user preference model (e.g., trained by analyzing users' historical selection data) to take into account users' personalized needs during the mapping process.

[0034] Finally, by combining the typical QoE requirement range of this application type in the application profile library with the user's historical preference data (if available), preliminary quantitative QoE expectation parameters are generated.

[0035] For example, in response to the identified "Video Conference A" and "Smooth" and "High Definition", the inference unit might output: , , , .

[0036] The model first performs entity recognition (such as business name and QoE descriptor) and relation extraction on natural language intent. Then, it combines a predefined application profile library (containing common application types and their typical QoE requirement range) and user historical behavior data to initially parse the intent into a set of structured QoE expectation parameters.

[0037] For example, "very smooth and high-definition" can be interpreted as an expectation of low packet loss, low jitter, high resolution, and high frame rate.

[0038] Therefore, the output of this sub-step is a set of preliminary quantified QoE expectations or ranges for this business flow, such as: ; in, The mathematical expression representing the expected quality of user experience is a set of multiple target KPI values; , , They represent the first , , Target values ​​or acceptable ranges for key performance indicators; This represents the total number of KPIs.

[0039] Step 2: Collect contextual information of the business flow, perform multimodal user experience quality perception on the business flow, and obtain the current user experience quality assessment result; Specifically, it includes the following sub-steps: Step 2.1: Configure the QoE monitoring probe and information collection interface; Deploy lightweight QoE monitoring probes on user terminals, network edge nodes, or SDN switches to collect performance parameters related to the network and application layers.

[0040] At the same time, it provides standardized API interfaces, allowing third-party applications to proactively report their internal QoE-related metrics and current business context information.

[0041] The input data types here include raw network packets, application layer logs, structured data reported by APIs, etc.

[0042] Step 2.2: Construct a multimodal QoE evaluation model; The model can integrate data from multiple information sources. The multimodal QoE evaluation model specifically includes a data fusion module, a QoE calculation module, and a context association module.

[0043] Data fusion module: responsible for collecting different types of data (video / audio stream parameters, text comments, contextual information, etc.) from QoE monitoring probes and application API interfaces, preprocessing them (such as normalization and feature extraction), aligning timestamps, and forming a unified feature vector.

[0044] For example, for text comments, sentiment analysis models (such as BERT-based text classification models) can be used to extract sentiment polarity as a feature; For video stream parameters, statistical measures such as the mean, variance, or peak value within a specific time window can be calculated as features.

[0045] QoE Calculation Module: This module receives the feature vector processed by the data fusion module.

[0046] When using a weighted fusion model, this module performs a weighted summation of different features based on preset or dynamically adjusted weights (for example, for video conferencing, video quality may have a higher weight than audio, but the opposite is true in audio conferencing scenarios) to obtain the evaluation value for each QoE dimension.

[0047] When using machine learning-based models (such as Gradient Boosting Tree (XGBoost), Support Vector Regression (SVR), or a small neural network), this module uses a pre-trained model, inputs the fused feature vectors, and directly outputs the evaluation values ​​for each QoE dimension.

[0048] For example, the model can be trained to predict the Mean Opinion Score (MOS score) for videos.

[0049] The model The training data can come from subjective rating data and corresponding objective parameters collected in a laboratory environment, or it can include user feedback data collected from the live network under anonymization.

[0050] Contextualization module: This module analyzes contextual information. (e.g., number of meeting participants) VIP users The potential correlation between the QoE dimension evaluation value output by the QoE calculation module (such as whether the current application is running in the foreground, etc.) and the QoE dimension evaluation value.

[0051] For example, in some implementations, association rule mining (such as the Apriori algorithm) or statistical analysis methods are used to discover rules such as "the probability of video stuttering increases significantly when the network type is a mobile network and the signal strength is below a certain threshold".

[0052] The results of these correlation analyses can be used for subsequent intent negotiation or strategy adjustments. For example, when a user is identified as being in a weak signal environment, a proactive recommendation can be made to switch to audio priority mode.

[0053] For example, for video conferencing services, this model can combine the following inputs: Objective parameters of the video stream: such as average bitrate Average frame rate Total number of lost packets End-to-end delay ; Objective parameters of an audio stream: such as average bitrate jitter buffer delay Packet loss rate ; Text comments provided by users through the app's in-app feedback channels (such as "the sound is intermittent" or "the screen is lagging"); Contextual information reported by the application: such as the number of participants in the current meeting. Is the current speaker a VIP user? wait.

[0054] The model uses, for example, a weighted fusion model or a machine learning-based model (such as gradient boosting trees) to process the aforementioned multi-dimensional input data, and its output is a comprehensive, multi-dimensional QoE vector: ; in, , , They represent the first time. , , Current rating values ​​on each QoE dimension (such as clarity, fluency, and timeliness of interaction); This represents the total number of QoE dimensions; A multidimensional vector representing the current user experience quality.

[0055] Step 2.3, structuring contextual information; The collected contextual information is structured to form contextual data records in a unified format. .

[0056] For example, for game applications, this might include the current game stage (such as "loading", "in battle", "settlement"), the number of players, etc.

[0057] The output of this sub-step is the multi-dimensional QoE vector of the current business flow. and structured application context information .

[0058] Step 3: When the initial application experience intent is unclear, or when there is a deviation between the initial application experience intent and the current user experience quality assessment result, or when the context information of the business flow changes, dynamic intent negotiation is conducted with the user or agent application to obtain the negotiated user intent. Specifically, it includes the following sub-steps: Step 3.1, trigger condition judgment; The system monitors the following conditions to trigger this step: Initial Intent Conflict: Preliminary Quantification of Expected QoE There are conflicting or difficult-to-meet indicators (such as requiring extremely low latency and extremely high bandwidth at the same time, which exceeds the capabilities of existing networks).

[0059] QoE and Intent Deviation: Current QoE Vector With expectations There is a significant bias.

[0060] Application context changes: Application context information Significant changes occur, and these changes are associated with predefined QoE adjustment rules (such as the start of an important meeting or the game entering a critical phase).

[0061] Step 3.2, construct the intent negotiation and clarification model; Construct an intent negotiation and clarification model, which is responsible for initiating interactions with users or agent applications under specific conditions (such as initial intent conflict, QoE and intent deviation, and application context changes) to clarify the true QoE priority.

[0062] The model may specifically include a condition judgment unit, an interaction generation unit, and a feedback processing unit.

[0063] Condition judgment unit: Receives the output of trigger condition judgment step 3.1, and the initial intent. Current QoE vector and application context As input.

[0064] This unit has a built-in set of rules, for example, when The bandwidth required exceeds the maximum total bandwidth of the currently available paths. And the required latency is lower than the path average latency. When this occurs, it is determined to be a conflict of initial intent.

[0065] Or, when Video smoothness rating Below The corresponding expectation When this occurs, it is determined to be a deviation between QoE and intention.

[0066] In some implementations, these thresholds It can be adaptively adjusted, for example, optimized based on historical negotiation success rates. The interactive generation unit is activated when any rule is met.

[0067] Interactive Generation Unit: Once activated, this unit generates interactive content based on specific triggering conditions and relevant input data. If the conflict stems from conflicting intents, such as simultaneously requiring low latency and high bandwidth, the interactive content might be: "Current resources may not be able to simultaneously meet your requirements for extremely low latency and extremely high bandwidth. Would you prioritize (A) the lowest latency or (B) the maximum bandwidth?"

[0068] An option (C) can also be provided to allow users to input more specific compromise requirements. If the change is due to a contextual shift, such as a VIP user joining a video conference, the interaction might be: "A VIP user has been detected joining the conference. Should we upgrade the current call's protection level to the highest level (which may affect the bandwidth of other applications)?", along with an estimated QoE improvement that the upgrade might bring. The interaction will be sent to the user or agent application via the user interface (e.g., pop-ups, in-app messages) or API.

[0069] Feedback processing unit: Receives selections or inputs returned by users or agent applications through interactive interfaces or APIs (e.g., the user selected "A" to prioritize ensuring the lowest latency, or adjusted the preference ratio of different QoE dimensions through a slider).

[0070] This unit transforms this feedback into structured, negotiated user intent. For example, if the user chooses to prioritize ensuring the lowest latency, then The middle will include latency-related KPIs (such as...) Assign higher weight Alternatively, stricter target values ​​may be set, while the requirements for bandwidth KPIs may be appropriately relaxed.

[0071] This model can be a rule-based reasoning system or a simple interactive question-and-answer system. Its input data types include triggering conditions and initial intentions. Current QoE vector and application context .

[0072] Step 3.3: Perform interactive negotiation; If the triggering conditions are met, the model initiates interaction with the user or agent application through the user interface or application API.

[0073] Interaction methods include: Question: For example, "Under current network conditions, do you value the smoothness of video conferencing or the lowest possible latency more?" Offer options: List several different QoE-focused strategy combinations and their expected effects for users to choose from.

[0074] Demonstrates QoE comparison: Explains the potential differences in QoE performance caused by different optimization objectives (such as maximizing bandwidth vs. minimizing latency) under current resource conditions. Users can select or input feedback through the interactive interface.

[0075] Therefore, the output of this sub-step is the clear and prioritized user intent in the current scenario, after negotiation and clarification. .

[0076] This intention may manifest as a weighting adjustment for a specific QoE dimension, or a specific confirmation of certain KPI target values.

[0077] For example: ; in, It represents the user's intent after negotiation and clarification, and is a set that includes the final KPI target value and weights; , These represent the user's views on the first... , Preference weights for each QoE dimension; , Indicates the first , The final confirmed target value or range for each Key Performance Indicator (KPI).

[0078] Step 4: Generate or adjust the multi-path network strategy based on the negotiated user intent, the current user experience quality assessment results, context information, and the status information of each available path in the current network. Specifically, the following steps are included: Step 4.1: Construct a strategy generation and evolution model; The Intent-Based Networking (IBN) system has a built-in QoE-driven policy generation and evolution model.

[0079] The IBN system can be understood as a logical architecture that includes a northbound interface module, an intent management module (including the aforementioned intent understanding model, intent negotiation and clarification model, and the policy generation and evolution model in this step), a QoE awareness module (including the aforementioned multimodal QoE evaluation model), and a southbound control module.

[0080] The QoE-driven policy generation and evolution model is the core of the intent management module, which can be specifically divided into a state evaluation unit, a policy search unit, and a policy output unit.

[0081] State Assessment Unit: This unit receives the negotiated intent. Current QoE vector and application context And the real-time status information of each available path in the current network reported by the SDN controller through the southbound control module. (such as bandwidth utilization of each path) End-to-end delay Packet loss rate ).

[0082] Based on these inputs, this unit calculates the utility values ​​that can be achieved by taking different multipath strategies in the current state.

[0083] Utility function It is used to quantify user satisfaction, for example: ; in, This represents the user's intent after negotiation and clarification, including the final KPI target value and weight set; A multi-dimensional vector representing the current user experience quality, containing evaluation values ​​for each QoE dimension; This represents a utility function used to quantify user satisfaction. This represents summing over all QoE dimensions i; This represents the predicted evaluation value for the i-th QoE dimension under a certain strategy; This represents the final confirmed target value for the i-th key performance indicator; This represents the user's preference weight for the i-th QoE dimension, reflecting the degree of importance the user attaches to different experience dimensions; This function represents the assessment of the relationship between the predicted QoE value and the target KPI value.

[0084] Policy Search Unit: When using evolutionary algorithms based on multi-objective optimization (such as NSGA-II, MOEA / D), this unit uses different path combinations, flow allocation ratios, queue parameters, etc., as parameters to be optimized (solution individuals) in order to maximize the utility function. Iteratively search for other possible optimization objectives (such as minimizing network resource consumption and maximizing network fairness) to generate a set of Pareto optimal policies.

[0085] When reinforcement learning (such as Q-learning, SARSA, DDPG, or A3C algorithms) is used, the state space can be derived from... , , , The discretized or continuous representation is constituted, and the action space is a selectable multi-path control instruction (such as selecting path A, allocating X% of the traffic to path B, and adjusting the queue weight of path C to Y).

[0086] reward function With utility function Or related to the amount of change, for example: ; in, Let t represent the reward function at time t, used to evaluate the quality of the agent's actions; This represents the value of the utility function at time t; This represents the utility function value at time t-1.

[0087] The agent learns the optimal policy function by interacting with a real network or a high-fidelity network simulation environment (as the "environment" for reinforcement learning). or state-action value function .

[0088] Policy output unit: The optimal policy obtained from the policy search unit (or the policy selected from the Pareto optimal set according to specific preferences). These commands are converted into specific network control instructions (e.g., OpenFlow flow table entries, P4 program update instructions, router configuration commands) and sent to the SDN controller or directly to network devices for execution via the southbound control module of the IBN system.

[0089] The core of this model can be an optimization algorithm, such as an evolutionary algorithm based on multi-objective optimization or a reinforcement learning agent. Its input data type is negotiated intent. Current QoE vector and application context And the status information of each available path in the current network (such as bandwidth, latency, packet loss rate, provided by the SDN controller).

[0090] Step 4.2: Dynamically generate and adjust multi-path strategies; The model calculates and outputs the optimal multi-path network strategy in real time based on the input, with the goal of maximizing user satisfaction. .

[0091] This strategy includes: Select the optimal combination of one or more paths for the business flow; The proportion of traffic distribution and load balancing across selected multiple paths; Adjust network resource parameters along the path, such as queue scheduling weight and bandwidth reservation value; Suggest adjustments to encoding / decoding parameters to the application layer (such as reducing the bitrate to accommodate network congestion).

[0092] Step 4.3, Strategy Evolution; Policy generation is a continuous evolutionary process. When the inputs (negotiation intent, current QoE, application context, network state) change, the policy generation and evolution model will recalculate and output an updated policy. .

[0093] This evolutionary process can be abstractly described as follows: ; in, This represents the network policy at time t, including configurations such as path selection, traffic allocation ratio, and network resource parameters. This represents the network policy updated at time t+1, which is the result of policy evolution; This represents the policy optimization function, which is responsible for calculating a new optimization policy based on the current policy and various inputs; This represents the user's intent after negotiation and clarification, including the final KPI target value and weight set; A multi-dimensional vector representing the current user experience quality, containing evaluation values ​​for each QoE dimension; It represents structured application context information, such as game stage, number of meeting participants, etc. This represents the set of real-time status information for each available path in the network at time t, including bandwidth utilization, latency, and packet loss rate.

[0094] Step 5: Control the transmission of service flows on multiple paths according to the multi-path network policy; Specifically, it includes the following sub-steps: Step 5.1, continuous monitoring and feedback; After system deployment, the multimodal QoE perception and context information collection module continues to run, outputting its real-time QoE vector. and application context Feedback is provided to the strategy generation and evolution model, as well as the intent negotiation and demand clarification model, thereby forming a closed-loop control.

[0095] For example, QoE evaluation results can be fed back at predetermined time intervals (such as 5 seconds) or when a significant change in a key QoE metric exceeding a preset threshold is detected.

[0096] Step 5.2, Strategy effectiveness evaluation and version control; The system records each policy. The QoE result after execution.

[0097] If a new strategy After implementation, over a period of time (e.g., a configurable evaluation window) Observed within 1 minute Not achieved The expected, or compared QoE deteriorates significantly (e.g., MOS score of video conferencing drops above the threshold). If the score is 0.5 points, the system will automatically trigger an alarm.

[0098] Meanwhile, the system maintains a policy version repository, which can store the complete configuration parameters of the policy, the policy effective timestamp, the associated intent version number, the typical application context snapshot at that time, and the corresponding QoE performance evaluation vector.

[0099] Step 5.3, Strategy backtracking and model optimization; When a policy fails to perform well, the system administrator, or in some implementations, the authorized user, may choose to roll back to a previously verified, better-performing policy version from the policy repository.

[0100] In some implementations, the system can also be configured to automatically backtrack to the most recent stable version when certain conditions are met (such as QoE remaining below a certain critical alarm threshold for a certain period of time).

[0101] Meanwhile, cases of poor policy execution will be recorded in detail as valuable negative samples or negative reward signals in reinforcement learning. These samples will be used to optimize the intent understanding model (e.g., adjusting the parameters of an NLP model or rules in a knowledge graph), the QoE evaluation model (e.g., retraining a machine learning model or adjusting fusion weights), and the policy generation and evolution model (e.g., updating the value function of a reinforcement learning agent or the policy network) in an online or offline manner to improve the accuracy and robustness of future decisions.

[0102] The output of this sub-step is continuously optimized network performance and user experience, as well as a continuously learning and evolving strategy model and knowledge base.

[0103] Application example: Application Scenario: A senior executive (VIP user) of a company needs to conduct a high-definition video conference with an important client via the company's internal network. The company network has deployed the software-defined multi-path network flow real-time dynamic optimization system described in this application. Multiple paths exist in the network, and the bandwidth, latency, packet loss rate, and other statuses of each path change dynamically. At the beginning of the conference, the network load is low, and several high-quality paths are available. Halfway through the conference, other high-traffic internal services (such as data backup) are initiated, causing a deterioration in the quality of some network paths. Simultaneously, the VIP executive begins a key product demonstration, at which point the requirements for smoothness and clarity of the conference reach their highest level.

[0104] Implementation process example: Example of user intent capture and initial understanding (initialization phase): The enterprise IT administrator sets a general intent template for the "executive video conference" scenario through the system's northbound management interface: "High priority, ensure clear and smooth audio and video, and low latency interaction".

[0105] The template is activated when a VIP executive initiates a video conference designated as a "customer presentation".

[0106] The intent understanding model, combined with the context of "customer demonstration," concretizes the general template into a set of initial QoE expected parameters. Expected video MOS score > 4.0, expected audio MOS score > 4.2, expected interaction latency < 100ms.

[0107] Multimodal QoE Awareness and Contextual Information Collection Example (Meeting in Progress): Ten minutes after the meeting started, the QoE awareness module collected the following representative data: Video stream (path A): average bitrate 3Mbps, frame rate 28fps, packet loss rate 0.2%, end-to-end latency 60ms.

[0108] Audio stream (path A): jitter 40ms, packet loss rate 0.1%.

[0109] User feedback (via app embed): No negative feedback.

[0110] Context information: Meeting type "Customer Presentation", 3 participants, with the current VIP executive as the main speaker. The multimodal QoE assessment model integrates the above inputs and outputs the current QoE assessment. The video MOS value is 4.1, the audio MOS value is 4.3, and the interaction latency is 60ms.

[0111] Example of dynamic intent negotiation and demand clarification (network fluctuations and context reinforcement): When the meeting was 30 minutes in, other departments in the network launched a large-scale software update, which caused the bandwidth of path A to be preempted, the packet loss rate to rise to 2%, and the latency to increase to 120ms.

[0112] Meanwhile, the application API reported a context change: VIP executives began sharing their screens for key PPT presentations.

[0113] The system detected that the current QoE (e.g., video MOS score dropped to 3.5, latency exceeded 100ms) is no longer met. Furthermore, the key context "screen sharing presentation" appears.

[0114] The intent negotiation module was triggered, pushing a negotiation request to the IT administrator's monitoring interface: "VIP executives are currently sharing critical screens, and the current network quality is affecting the user experience. It is recommended to activate the 'Highest Guarantee for Critical Business' policy. Do you want to execute?" The administrator confirmed execution.

[0115] User intent after negotiation Updated to: During “Screen Sharing Presentation”, the weight of video clarity and smoothness is adjusted to the highest level, and the tolerance for interaction latency is further tightened to 80ms.

[0116] QoE-driven multi-path policy generation and evolution example (policy switching): The policy generation and evolution model receives... And the updated QoE assessment (video MOS score 3.5) and path status (e.g., path A: bandwidth 2Mbps / packet loss 2% / latency 120ms); Path B: Bandwidth 5Mbps / Packet Loss 0.3% / Latency 70ms; Path C: Bandwidth 3Mbps / Packet Loss 1% / Latency 90ms).

[0117] The model (e.g., based on a predefined decision tree or a small reinforcement learning agent) determines that path B is the current optimal choice. The output of the new multi-path policy includes: The video conferencing traffic (including audio, video, and screen sharing data streams) of VIP executives will be hard-switched from path A to path B.

[0118] Configure higher queue priority and bandwidth reservation for this VIP traffic on path B. This policy is issued to the relevant network devices by the SDN controller for execution.

[0119] Closed-loop adaptive control and strategy version management example (effect verification and learning): 5 minutes after traffic switches to path B, the QoE perception module re-evaluates and obtains a new QoE assessment. The video MOS score rebounded to 4.2, the audio MOS score to 4.4, and the interaction latency to 70ms, meeting the requirements. .

[0120] The system stores this successful "context awareness - intent negotiation - policy switching - effect improvement" complete event chain (including all relevant input data, decision-making process and output results) in the policy version repository and knowledge base.

[0121] If the switch is ineffective, such as due to user complaints or a persistent QoE below a certain threshold, the system will record this failure case. The administrator can then trigger a rollback to the previous strategy or another backup strategy. The data from this failure case will be used for subsequent iterative optimization of the relevant models.

[0122] Technical effectiveness verification: To verify the technical effectiveness of this implementation method, in the above-mentioned simulated scenario, the QoE performance and network path selection of the optimized system using this implementation method (experimental group) and the traditional fixed-priority multi-path selection strategy (control group) were compared during the key product demonstration phase for VIP executives (i.e., when network fluctuations occurred 30 minutes after the meeting and the context importance increased).

[0123] In the experimental group, due to dynamic intent negotiation and context awareness, the system proactively switched the VIP executive's traffic to the then-better path B after detecting network quality degradation and the presentation context, and may have prioritized their traffic. In the control group, assuming it only relied on the initially set high priority, traffic might still have remained on the degraded path A, or simply switched based on link load, without fully considering the negotiated intent and refined QoE objectives.

[0124] The collected data is described in text form as follows: Key business QoE (video MOS score) performance: In the control group, when the quality of path A deteriorated, the MOS score of VIP executive video conference dropped significantly from the initial 4.1 to 3.5, with noticeable stuttering and blurring.

[0125] Although the system has multiple paths, its switching strategy has failed to respond promptly and effectively to the refined QoE requirements and context changes of specific business flows.

[0126] In the experimental group, after the system completed path switching and strategy optimization (e.g., switching to path B and adjusting resources), the MOS score of the VIP executive video conference recovered and stabilized at around 4.2, effectively ensuring the meeting experience during the key presentation phase.

[0127] Network path selection and resource utilization: In the control group, after the traffic of VIP executives deteriorated on path A, it may have been assigned to path C (latency 90ms, higher than 70ms on path B) due to a simple load balancing strategy, or due to insufficient identification of business importance, their traffic was still competing with other traffic on path A, resulting in a poor experience.

[0128] In the experimental group, the system not only selected Path B (70ms latency, 0.3% packet loss), which currently offers better overall quality, for VIP executives, but also ensured sufficient resources on Path B to support the high-quality needs of VIP meetings by temporarily limiting the rate of low-priority services such as data backup or adjusting the path. This demonstrates the intelligence of resource allocation and the precise response to user intent.

[0129] The simulation results of key technologies are shown in Tables 1 and 2: Table 1: QoE Comparison of Key Demonstration Stages (Video MOS Score);

[0130] Table 2: Comparison of Path Selection and Latency in Key Demonstration Phases;

[0131] This invention constructs a complete methodology and system modules for user intent capture and understanding, multimodal QoE perception, dynamic intent negotiation, QoE-driven multi-path strategy generation and evolution, and closed-loop adaptive control. This enables real-time and accurate mapping of users' dynamic changes in application experience needs and intents into specific network control behaviors.

[0132] Compared with existing technologies, the present invention can more comprehensively perceive user experience and more flexibly adapt to changes in user intent and application context, thereby optimizing and scheduling multi-path network resources more accurately.

[0133] This invention solves the technical problems in the prior art, such as inaccurate mapping between user intent and network behavior, single QoE perception dimension, and static strategies being difficult to adapt to dynamic needs. It achieves beneficial effects such as improving user experience and satisfaction, realizing user-centered intelligent network management, enhancing the flexibility and adaptability of network strategies, and improving the level of network automation and intelligence.

[0134] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A software-defined method for real-time dynamic optimization of multipath network flows, characterized in that, Includes the following steps: Obtain the user's initial application experience intent for the business flow; Collect contextual information about the business flow, perform multimodal user experience quality perception on the business flow, and obtain the current user experience quality assessment results; When the initial application experience intent is unclear, or when there is a deviation between the initial application experience intent and the current user experience quality assessment result, or when the context information of the business flow changes, dynamic intent negotiation is carried out with the user or agent application to obtain the negotiated user intent. Based on the negotiated user intent, the current user experience quality assessment results, context information, and the status information of each available path in the current network, generate or adjust the multi-path network strategy. Control the transmission of service flows on multiple paths according to the multi-path network strategy.

2. The software-defined real-time dynamic optimization method for multi-path network flows according to claim 1, characterized in that, The steps for obtaining the user's initial application experience intent for the business flow include: Receive the user's initial application experience intent expressed in natural language or declarative strategies through the northbound interface; An intent understanding model is used to analyze the initial application experience intent, and preliminary quantitative parameters of user experience quality expectations are obtained.

3. The software-defined real-time dynamic optimization method for multi-path network flows according to claim 2, characterized in that, The intent understanding model is based on natural language processing technology and knowledge graph construction. It is used to identify key entities and relationships between entities from natural language intents, and combine application profiles and user historical behavior data to parse the initial application experience intent into preliminary quantified user experience quality expectation parameters.

4. The software-defined real-time dynamic optimization method for multi-path network flows according to claim 1, characterized in that, The steps for performing multimodal user experience quality perception on the business flow include: By monitoring probes and application programming interfaces, network layer performance parameters, application layer performance parameters, user feedback information, and context information related to business flows are collected. A multimodal user experience quality assessment model is used to fuse and process the collected parameters and information to obtain the current user experience quality assessment result, which is a multi-dimensional user experience quality vector.

5. The software-defined real-time dynamic optimization method for multi-path network flows according to claim 1, characterized in that, The steps for dynamically negotiating intent with the user or agent application include: When preset triggering conditions are met, an interactive negotiation is initiated with the user or agent application through the user interface or application programming interface. The interactive negotiation includes at least one of asking questions, providing options, or displaying a comparison of user experience quality. Receive feedback from users or agent applications and transform the feedback into negotiated user intents, which may include weight adjustments for specific user experience quality dimensions or confirmation of target values ​​for key performance indicators.

6. The software-defined real-time dynamic optimization method for multi-path network flows according to claim 1, characterized in that, The steps for generating or adjusting the multipath network strategy include: We adopt a user experience quality-driven strategy generation and evolution model, aiming to maximize the utility function of user satisfaction. By combining negotiated user intent, current user experience quality assessment results, context information, and the status information of each available path in the current network, we can calculate and output the optimal multi-path network strategy in real time.

7. The software-defined real-time dynamic optimization method for multi-path network flows according to claim 6, characterized in that, The multi-path network strategy includes at least one of the following: selecting the optimal combination of one or more paths for the service flow, allocating traffic and load balancing ratios on the selected multiple paths, and adjusting network resource parameters on the paths.

8. The software-defined real-time dynamic optimization method for multi-path network flow according to claim 1, characterized in that, Also includes: Continuously monitor the user experience quality results after the multi-path network strategy is implemented; When the user experience quality outcome fails to meet the expectations of the negotiated user intent or deteriorates, an alert is triggered, allowing a rollback from the policy version repository to a previously valid policy version.

9. A software-defined real-time dynamic optimization method for multi-path network flows according to claim 8, characterized in that, Also includes: Cases of poor strategy execution, including the negotiation intent, contextual information, network state, executed strategy, and resulting user experience quality, are used to optimize the intent understanding model, the multimodal user experience quality assessment model, and the user experience quality-driven strategy generation and evolution model.

10. A software-defined real-time dynamic optimization system for multi-path network flows, used to execute the software-defined real-time dynamic optimization method for multi-path network flows according to any one of claims 1-9, characterized in that, include: The intent acquisition module is used to acquire the user's initial application experience intent for the business flow; The user experience quality perception module is used to collect contextual information of the business flow and perform multimodal user experience quality perception on the business flow to obtain the current user experience quality assessment result. The intent negotiation module is used to dynamically negotiate the user intent with the user or agent application when the initial application experience intent is unclear, or when there is a deviation between the initial application experience intent and the current user experience quality assessment result, or when the context information of the business flow changes, so as to obtain the negotiated user intent. The strategy generation module is used to generate or adjust multi-path network strategies based on the negotiated user intent, the current user experience quality assessment results, context information, and the status information of each available path in the current network. The control execution module is used to control the transmission of service flows on multiple paths according to the multi-path network policy.

Images