An adaptive congestion control algorithm and system based on large language model

Abstract

The present invention relates to the field of computer network communication technology, and discloses an adaptive congestion control algorithm and system based on a large language model, which is applied to the network optimization of the QUIC protocol. The method collects network status time series data, including round-trip delay, packet loss rate and throughput, converts it into a text sequence, inputs a pre-trained large language model to predict the optimal congestion control parameter α value, and dynamically adjusts the QUIC protocol performance. The model is trained using supervised learning or reinforcement learning, and real-time performance is ensured through model distillation technology. The system includes data collection, preprocessing, LLM prediction engine, verification and parameter application modules. The present invention utilizes the intelligent prediction capability of the large language model to improve the throughput and delay performance of the QUIC protocol in complex network environments, reduce computing overhead, and is suitable for high-performance communication scenarios such as video streaming and online games.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a congestion control technology for the QUIC (Quick UDP Internet Connections) protocol, specifically a method and system for adaptively adjusting congestion control parameters using a large language model (LLM). Background Art

[0002] Modern computer networks, especially those widely adopting the QUIC protocol, face increasingly dynamic network conditions and diverse application requirements. QUIC, with its advantages over UDP, including reliable transmission, reduced connection establishment latency, and support for multiplexing, has been widely adopted in scenarios highly sensitive to throughput and latency, such as video streaming and online gaming. However, its performance is highly dependent on the effectiveness of the adopted congestion control algorithms (CCAs).

[0003] Traditional congestion control algorithms often employ fixed or simple rule-based adjustment strategies. These strategies perform well in relatively stable network environments, but can struggle to adapt quickly to changes in complex and volatile network environments. For example, high packet loss rates in mobile or satellite networks can lead to reduced throughput, increased latency, and low network resource utilization.

[0004] A key feature of the QUIC protocol is that its congestion control logic is primarily implemented in user space, which facilitates the introduction of more flexible and intelligent congestion control methods. Compared to traditional TCP congestion control implemented in the operating system kernel, user-space implementation makes it easier to deploy and iterate new congestion control algorithms or dynamically adjust their parameters.

[0005] In recent years, artificial intelligence, particularly large language models (LLMs), has demonstrated powerful capabilities in natural language processing, pattern recognition, and sequence prediction. Researchers have begun exploring the application of LLMs to network management and automation, such as network operations and maintenance, traffic analysis, and troubleshooting. However, effectively leveraging LLMs' pattern recognition and prediction capabilities for transport layer congestion control, which requires extremely high real-time performance, and specifically for adaptive parameter optimization of the QUIC protocol to address the challenges of complex network environments, remains a critical challenge in the field of network technology.

[0006] Some existing research attempts to apply machine learning to congestion control, but these efforts may suffer from issues such as insufficient model generalization, difficulty ensuring real-time performance, and failure to fully utilize network state timing information. Therefore, developing a method and system that leverages the powerful capabilities of LLM to quickly, accurately, and intelligently adjust QUIC congestion control parameters is crucial for improving the performance of modern networks. Summary of the Invention

[0007] The purpose of the present invention is to provide an adaptive congestion control algorithm and system based on a large language model to solve the problems raised in the above background technology that traditional congestion control algorithms are difficult to adapt to complex network environments, have insufficient performance optimization, and have application limitations of existing AI methods.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an adaptive congestion control algorithm based on a large language model, characterized by comprising:

[0009] Acquire a network status time series data set; the network status time series data at least includes round trip delay (RTT), packet loss rate and throughput.

[0010] The network state time series data is converted into a text sequence; the text sequence is suitable for serving as an input of a large language model (LLM).

[0011] The text sequence is input into a pre-trained large-scale language model prediction engine; the large-scale language model prediction engine predicts and outputs the optimal congestion control parameter α value under the current network state based on its learned pattern recognition and sequence prediction capabilities.

[0012] The α value output by the large language model prediction engine is verified; the verification is intended to ensure the validity and security of the α value.

[0013] The verified α value is applied to the congestion controller of the QUIC protocol to dynamically adjust its congestion control behavior to optimize network transmission performance.

[0014] Preferably, converting the network status time series data into a text sequence includes:

[0015] Preprocessing the network status time series data, including noise reduction and normalization, and converting the preprocessed numerical data into a text sequence using at least one of the following techniques:

[0016] Numeric stringification: convert the numerical values ​​in the time series directly into text strings;

[0017] Chunking and lemmatization: Split the time series into chunks and lemmatize the chunks;

[0018] Symbolic Aggregate Approximation (SAX) or other symbolic representations: convert a continuous sequence of values ​​into a discrete sequence of symbols;

[0019] Quantization / binning: Divide continuous values ​​into predefined discrete intervals and map them to specific tokens;

[0020] Ensure that the converted text sequence can effectively represent the time dependency and numerical scale information of the original data.

[0021] Preferably, the large language model prediction engine is trained by at least one of the following methods:

[0022] Supervised learning: Construct a labeled dataset containing network state text sequences and corresponding optimal α values; the optimal α values ​​are determined based on predefined performance indicators through network simulation or real-world experiments;

[0023] Using the labeled dataset, fine-tuning a large language model to learn a mapping from a network state text sequence to an optimal α value;

[0024] Reinforcement learning: State is defined as a textual sequence of network states; Action is defined as selecting or adjusting the α value; Reward is defined as a function calculated based on network performance indicators such as throughput, latency, and packet loss rate;

[0025] Through interaction with the network environment or simulation environment for training, the large language model learns the α value selection strategy that maximizes the cumulative reward;

[0026] Preferably, the method further comprises adopting a model distillation technique:

[0027] Train a smaller, faster-inference “student” model to mimic the behavior of a larger, better-performing “teacher” large language model.

[0028] The "student" model is deployed as a large-scale language model prediction engine to meet the real-time requirements of QUIC congestion control.

[0029] Preferably, the verification of the α value includes at least one of the following checks:

[0030] Range check: Ensure that the predicted α value is within the reasonable range allowed by the congestion control algorithm;

[0031] Stability and safety constraint check: Evaluate whether the α value may cause network oscillation, congestion collapse, or sharp performance deterioration;

[0032] Rollback mechanism: Monitor network performance after applying a new α value. If significant degradation or instability is observed, roll back to a previously known safe and effective α value.

[0033] Preferably, the congestion control parameter α value is used to adjust the core behavior of the QUIC congestion control algorithm, such as adjusting the window gain rate or target window size.

[0034] Preferably, the fine-tuning of the large language model adopts parameter efficient fine-tuning (PEFT) technology, such as LoRA.

[0035] In another aspect of the present invention, an adaptive congestion control system based on a large language model is provided, characterized in that it includes:

[0036] Data collection module: used to obtain network status time series data sets in real time or quasi-real time, including round-trip delay (RTT), packet loss rate and throughput;

[0037] Preprocessing module: used to convert the network state time series data into a text sequence that can be processed by a large language model;

[0038] LLM prediction engine module: configured to receive the text sequence as input, predict and output the optimal congestion control parameter α value under the current network state through a pre-trained and distillable large language model.

[0039] Verification module: used to verify the validity and security of the α value before applying it to the QUIC protocol;

[0040] Parameter application module: used to apply the verified α value to the congestion controller of the QUIC protocol to dynamically adjust its congestion control behavior.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] Improving network performance and adaptability: This method leverages the powerful pattern recognition and sequence prediction capabilities of large-scale language models to analyze complex network state time series data, more accurately predicting the optimal congestion control parameter α for the current network conditions. Compared to traditional fixed or simple rule-based congestion control algorithms, this method enables the QUIC protocol to more intelligently adapt to dynamically changing network environments, effectively improving throughput, reducing latency, and achieving superior network transmission performance.

[0043] Innovation in Data Conversion: This paper proposes converting numerical network status time series data into text sequences that can be processed by LLM. This conversion process is a critical bridge between network telemetry data and LLM. Through techniques such as numerical stringification, block tokenization, SAX, and quantization, LLM can understand and process network status information. This cross-modal data processing approach offers new insights into using LLM to solve non-textual problems. While facing challenges in information fidelity, it holds great potential.

[0044] Flexible Training and Deployment Mechanism: This invention supports both supervised learning and reinforcement learning training paradigms, allowing for flexible selection based on available data and application scenarios. Supervised learning leverages historical data to rapidly construct mapping relationships, while reinforcement learning learns optimal strategies through interaction with the environment, offering greater adaptability. Furthermore, the introduction of model distillation technology allows for the training of lightweight "student" models for deployment, effectively addressing the high latency of LLM inference and ensuring the real-time requirements of congestion control. Fine-tuning with PEFT technology further reduces training costs and resource requirements.

[0045] Enhanced reliability and security: This invention incorporates a verification module that performs multiple checks on the α values ​​predicted by the LLM, including range checks, stability checks, and safety constraint checks, and incorporates a rollback mechanism. This ensures that the parameters used in the actual network are reasonable and safe, effectively preventing performance degradation or network crashes caused by hallucinations or prediction errors that may occur with the LLM, significantly improving system reliability and security.

[0046] Promoting Intelligent Network Management: This invention applies cutting-edge LLM technology to the core network transmission control layer, a significant exploration of AI-driven network management. It not only provides a new optimization path for QUIC congestion control but also builds experience for applying LLM to a wider range of network protocol optimization and automated management tasks, potentially driving network intelligence to a higher level. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a working principle diagram of the QUIC adaptive congestion control method based on a large language model of the present invention;

[0048] Figure 2 A flowchart for converting network status time series data into a text sequence according to the present invention;

[0049] Figure 3 A schematic diagram of a large language model training method of the present invention;

[0050] Figure 4 This is an architectural diagram of the QUIC adaptive congestion control system based on a large language model in the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] See also Figures 1 to 4 The present invention provides a technical solution: an adaptive congestion control algorithm based on a large language model, the algorithm comprising:

[0053] S101. Actively acquire a time series data set of network status. The data comes from QUIC-connected terminal devices or monitoring nodes deployed on the network path. Key network performance indicators are collected in real-time or near-real-time, including at least round-trip delay, packet loss rate, and throughput.

[0054] S102, converting the originally acquired network status time series data into a text sequence. This step is the core link to achieve compatibility between network telemetry data and LLM;

[0055] S103: Input the converted text sequence into a pre-trained LLM prediction engine, which performs in-depth analysis on the input text sequence, understands the network status patterns hidden in the text sequence, and predicts the most appropriate congestion control parameter α value under the current network status based on these patterns;

[0056] S104. Strictly verify the α value output by the LLM prediction engine to verify the validity and security of the predicted α value, and prevent network performance degradation or even crash due to improper parameter application;

[0057] S105. Apply the verified α value to the congestion controller of the QUIC protocol. Once the α value is confirmed to be safe and valid, it is passed to the congestion control module of the QUIC connection.

[0058] S106. Monitor network performance and form a closed-loop control. After applying the new α value, the system will continue to monitor various performance indicators of the network, forming a closed-loop control system with continuous optimization and dynamic adjustment.

[0059] The present invention will be further described below with reference to Examples 1 to 7.

[0060] Example 1

[0061] In this embodiment, the specific process and technical details of step S102, i.e., converting the network status time series data into a text sequence, will be described in detail. Figure 2shown.

[0062] After successfully obtaining raw time series data such as RTT, packet loss rate, and throughput, the first task is to preprocess this raw data to improve data quality and lay a good foundation for subsequent text conversion and LLM processing.

[0063] Preprocessing typically involves two key steps: noise reduction and normalization. Common noise reduction techniques include moving average filtering, which smooths data by calculating the average value within a certain window size in the time series. Kalman filtering is also an effective method. It can estimate the current state based on the previous state and the current observation value, and is particularly suitable for dealing with noise in dynamic systems.

[0064] Another step in preprocessing is normalization, which unifies network metrics of different scales and units into a comparable range, such as the interval [0, 1] or conforms to a standard normal distribution. This can be achieved through min-max normalization or Z-score normalization.

[0065] After preprocessing, these numerical time series data need to be converted into text sequences that can be processed by LLM. This conversion process can use at least one or a combination of the following techniques to ensure that the converted text sequences can effectively represent the temporal dependence and numerical scale information of the original data.

[0066] Numeric stringification: This is a relatively straightforward method that converts the numerical values ​​in a time series directly into text strings. To distinguish different network metrics, you can add labels or prefixes to the numerical strings. For example, an RTT sequence [10, 12, 15] can be converted to the text "RTT: 10 12 15".

[0067] Chunking and tokenization: This approach divides long time series data into fixed-size "chunks" or "segments." Each chunk can be considered a local snapshot of the network state over a short period of time. Each chunk is then mapped to a specific token.

[0068] Symbolic Aggregate Approximation (SAX): This method converts continuous time series data into a discrete sequence of symbols. It typically involves two steps: first, reducing the dimensionality of the time series through piecewise aggregate averaging. This involves dividing the series into segments of equal length and representing each segment with its average value. Second, the PAA-converted values ​​are mapped to discrete symbols or letters based on pre-set breakpoints. For example, RTT values ​​can be mapped to symbols such as "Low," "Medium," and "High" based on their range.

[0069] Quantization / Binning: Divide continuous values ​​into predefined discrete intervals, with each interval corresponding to a specific token. For example, RTT values ​​in the range of 0-20ms can be defined as the token "RTT_VeryLow", 20-50ms as "RTT_Low", and so on.

[0070] To ensure that the generated text sequence can fully and effectively preserve the temporal dependencies and numerical scale information contained in the original data, timestamp information can be explicitly added to the generated text sequence, or the temporal relationship can be implicitly expressed through the order of the tokens. For example, a carefully designed text sequence might look like "RTT_Low LOSS_None TP_High Time_Delta_5ms RTT_Low LOSS_None TP_High Time_Delta_5ms RTT_Increasing LOSS_Low TP_Decreasing..."

[0071] Example 2

[0072] In this embodiment, the LLM prediction engine is trained by supervised learning. Figure 3 The core step is to learn a mapping function from labeled training data so that the model can predict the output based on the input.

[0073] First, a high-quality annotated dataset is constructed. This dataset contains a large number of samples, each of which consists of two parts: one is a text sequence of network status, which is the LLM input after the transformation in Example 1; the other is the optimal congestion control parameter α corresponding to the network status.

[0074] The optimal α value is determined through extensive experimentation and simulation. Experiments are conducted using professional network simulation platforms or in real physical network environments. By precisely controlling or simulating a variety of complex network conditions, such as varying link bandwidths, propagation delays, packet loss patterns, and the type and intensity of background traffic, the QUIC protocol is tested across a range of different α values ​​for each specific combination of network conditions.

[0075] The QUIC transmission performance at the current α value is then evaluated based on predefined performance metrics. These metrics may include, but are not limited to, maximizing throughput, minimizing round-trip latency, minimizing packet loss, or some weighted combination of these metrics. For each network state, the α value that yields the best performance is selected and labeled the "optimal α value" for that state.

[0076] By repeating this process, a large number of "network status text sequence - optimal α value" data pairs are collected to form the final labeled dataset. For example, a training sample may have an "input" of "RTT_Low RTT_Stable LOSS_None TP_HighBDP_High..." and a corresponding output of α value of "1.2".

[0077] After constructing the annotated dataset, select a suitable large pre-trained model as the base model. These pre-trained models have been trained on massive amounts of text data. Fine-tune this pre-trained LLM using the annotated dataset. This fine-tuning allows the LLM to further refine the specific mapping from the input network state text sequence to the output optimal α value based on its existing knowledge.

[0078] In order to improve the efficiency of fine-tuning and reduce the consumption of computing resources, the parameter efficient fine-tuning technology LoRA is adopted.

[0079] Through the above supervised learning and fine-tuning process, the LLM prediction engine is able to predict a suitable α value based on the text description of the network status under various network conditions to optimize the transmission performance of QUIC.

[0080] Example 3

[0081] In this embodiment, the LLM prediction engine is trained by reinforcement learning. Figure 3 In reinforcement learning, an agent learns how to make a series of optimal decisions to maximize a certain cumulative reward by interacting with a dynamic environment. In this invention, the reinforcement learning framework mainly includes the following core components:

[0082] State (S). In this invention, the state is a description of the current environment and is a textually processed sequence of network states, as described in Example 1. It may include a tokenized sequence of metrics such as the RTT mean, RTT standard deviation, packet loss rate, and throughput mean within a recent time window. The current application's alpha value can also be input to the LLM as part of the state. For example, the state could be text such as "CURRENT_ALPHA: 1.0 RTT_Avg_LowRTT_Std_Low LOSS_None TP_High..."

[0083] Action (A): An action is a decision that the LLM can make based on the current state. In this invention, an action is selecting a new α value or adjusting the current α value by some amount. The action space can be discrete, for example, selecting from a predefined set of candidate α values ​​or selecting a predefined adjustment step size such as {-0.2, -0.1, 0, 0.1, 0.2}. The α value can also be continuous, for example, directly predicting a floating-point α value within the allowed range.

[0084] Reward (R) is a scalar feedback signal used to evaluate the performance of the LLM after taking a specific action in a specific state. The design of the reward function is crucial to the performance of reinforcement learning, as it directly guides the LLM to learn the desired behavior. The reward function is usually calculated based on network performance indicators. For example, a reward function that comprehensively considers throughput, latency, and packet loss rate can be designed, such as .in, , , is a weight coefficient used to balance the importance of different performance indicators. High throughput and low latency will be positively rewarded, while high packet loss rate will be penalized.

[0085] The environment (E) is the object with which the LLM interacts. It can be a real physical network environment or a high-fidelity network simulator. The LLM outputs a value α to the environment, which updates its state based on the action and feeds back the new state and corresponding reward to the LLM.

[0086] The training process is usually as follows:

[0087] The agent is in the initial state Next, choose an action based on its current strategy , which is a new α value. This action Applied to the QUIC congestion controller and executed in the network environment.

[0088] After an RTT period or a fixed time interval, the network state changes to a new state , and the environment calculates a reward value based on a predefined reward function . LLM receives the new status and rewards .

[0089] Then, LLM uses reinforcement learning algorithms such as Q-Learning, SARSA, and Policy Gradient to update its internal parameters so that it can choose actions in the future that tend to maximize long-term cumulative rewards.

[0090] Through a continuous learning cycle of trial and error with the environment and leveraging existing experience, LLM can eventually learn to select the α value strategy that can bring the best long-term network performance under different network state text sequences.

[0091] Ultimately, by learning through direct interaction with the environment, LLM is able to adapt to very dynamic and complex network conditions without the need for manual pre-definition of optimal behaviors in all cases.

[0092] Example 4

[0093] While large models demonstrate powerful capabilities in pattern recognition and sequence prediction, a significant challenge is their typically large number of parameters. This results in high computational resources and long latency when performing inference on large models. To meet millisecond-level real-time requirements and rapidly respond to changes in network conditions, this embodiment employs a model distillation technique.

[0094] During the model distillation process, a "student model" with smaller parameters, simpler structure, and faster reasoning speed is trained to imitate a larger "teacher model" with better performance but slower reasoning.

[0095] First, a high-performance "teacher model" is required. This can be a large language model carefully trained using Example 2 or Example 3. This model has billions or even trillions of parameters, and its inference latency may far exceed real-time requirements. However, its accuracy and performance in predicting the optimal α value are proven.

[0096] Next, we need to design a "student model" with excellent inference speed. The student model is structurally simpler than the teacher model, with fewer network layers, fewer attention heads, and smaller hidden layer dimensions. This results in faster inference speed and lower computational overhead.

[0097] Next, the process of the teacher model training the student model is model distillation. The student model is trained not only to learn the labels of the original dataset, but also to learn to imitate the output behavior of the teacher model.

[0098] Specifically, a teacher model can be used to predict a large number of network state text sequences to obtain the teacher model's output. Then, the student model is trained to make its output as close as possible to the teacher model's output. This is achieved by minimizing the difference between the student model's output and the teacher model's output.

[0099] For example, if the teacher model outputs a continuous α value, you can use the mean squared error (MSE) as the loss function; if the teacher model outputs a probability distribution, you can use the KL divergence to measure the difference between the two probability distributions and use it as part of the loss function.

[0100] After student model training is complete, the lightweight "student model" is deployed as the final LLM prediction engine. Due to its small number of parameters and simple structure, the student model significantly improves inference speed and consumes less computing resources. This makes the student model more suitable for running on resource-constrained edge devices or for use in network control loops that require high-throughput real-time decision-making.

[0101] Through model distillation technology, the present invention can effectively overcome the real-time challenges brought about by the high inference delay of large language models while utilizing the powerful capabilities of large language models, making the QUIC congestion control method based on large language models more practical.

[0102] Example 5

[0103] In this embodiment, the specific process of verifying the α value output by the LLM prediction engine in step S104 will be described in detail.

[0104] The verification module is a key component to ensure the stability and security of the LLM-based QUIC adaptive congestion control system. Although LLMs have strong prediction capabilities, they may sometimes produce "hallucinations" or make predictions that are inconsistent with the current network conditions.

[0105] Therefore, before the α value predicted by LLM is actually applied to the QUIC congestion controller, it must be subjected to a series of rigorous checks to ensure its security and effectiveness. This verification process mainly includes the following check mechanisms:

[0106] Range checking: Different congestion control algorithms typically have a predefined valid range for their parameter α. For example, one algorithm may require that the α value must be within the range 0.5 ≤ α ≤ 2.0. Range checking ensures that the α value predicted by the LLM falls within this predefined acceptable range. If the predicted value falls outside this range, the system can take various actions. For example, it can directly reject the predicted value and retain the current α value, or it can "clamp" the predicted value to the nearest limit of the allowed range. For example, if the predicted value is 2.5 and the upper limit is 2.0, it will be corrected to 2.0.

[0107] Stability and safety constraint checking: This check evaluates whether the α value predicted by the LLM could lead to a sharp deterioration in network performance, such as network oscillation or congestion collapse, where network throughput plummets to near zero, leading to unacceptable performance. This is achieved through pre-defined heuristic rules or a simple auxiliary prediction sub-model.

[0108] For example, a rule could be set: If RTT has been observed to increase rapidly and packet loss has remained high over a recent period, avoid choosing an overly aggressive α value, even if the LLM's original prediction suggests so. Alternatively, a small model could be trained based on historical empirical data to quickly determine the potential risk of a particular α value. If a predicted α value fails to pass stability and safety constraints, it will be rejected.

[0109] Rollback mechanism: After applying a new α value that passes the aforementioned checks, the system does not fully trust its effectiveness. Instead, it continues to closely monitor changes in network performance indicators. If a significant, unexpected drop in network performance or drastic performance fluctuations are observed within a short period of time after applying the new α value, the system will automatically trigger a rollback mechanism.

[0110] The rollback mechanism immediately restores the QUIC congestion controller's α value to a previously known, verified, and effective α value. This mechanism effectively prevents LLM prediction errors that may not be detected by the pre-check and cause lasting damage to the network, thereby enhancing the robustness and reliability of the entire system.

[0111] Through these multi-level verification methods, the present invention aims to ensure that the congestion control parameter α value predicted by LLM can not only improve performance but also ensure network stability and security in practical applications, thereby making the LLM-based adaptive congestion control method more mature and practical.

[0112] Example 6

[0113] This embodiment details how the verified α value is applied to the QUIC protocol congestion controller in step S105, and explains the role of the α value in the congestion control algorithm. The α value is typically a key adjustable parameter in the congestion control algorithm, directly influencing the algorithm's core behavior and decision-making logic, thereby changing how the QUIC connection responds to network congestion.

[0114] The specific meaning and function of the congestion control parameter α value will vary depending on the type of congestion control algorithm used. Taking the α parameter introduced in the TCPTuner research as an example, the α value can be used to adjust the size of the target window, that is, it is used to scale the , that is, the adjusted = This adjusted This will further affect the subsequent congestion window growth curve.

[0115] When the LLM predicts a large α value, such as α > 1, this typically means that the network condition is considered good based on the text sequence analysis of the current network state. In this case, the congestion controller will adopt more aggressive strategies, such as increasing the congestion window at a faster rate or setting a higher target window value, to better utilize the available network bandwidth and thus improve throughput.

[0116] Conversely, if the LLM predicts a smaller α value, such as α < 1, it indicates that the network condition may be poor. In this case, the congestion controller will adopt a more conservative growth strategy, such as slowing the growth rate of the congestion window or lowering the target window value, to avoid further aggravating network congestion and prioritize connection stability.

[0117] In other types of congestion control algorithms, such as those based on utility functions, the value of α may directly affect the degree to which the algorithm pursues the throughput target, its sensitivity to delay, or the trade-off between throughput and delay.

[0118] For example, the α value can be used as a coefficient to adjust the weight of the throughput term or the latency term in the utility function. By dynamically adjusting this α value through LLM, the QUIC protocol can intelligently switch the emphasis of its congestion control behavior based on current network conditions and application requirements, in order to optimize overall transmission performance.

[0119] The parameter application module in this invention communicates the LLM's intelligent decisions to the actual congestion control logic in the QUIC stack. Once the verification module confirms the safety and validity of a new α value, it is passed to the parameter application module, which then calls the corresponding API function or interface to update the α parameter value currently in use within the QUIC congestion controller in real time. This allows the congestion controller's behavior to adapt to the new α value, enabling adaptive response to network changes.

[0120] Example 7

[0121] This embodiment will describe in detail the overall architecture of the QUIC adaptive congestion control system based on a large language model proposed in the present invention and the functions of each component module, such as Figure 4 As shown in Figure 4, the system aims to dynamically adjust the congestion control parameters of the QUIC protocol in an intelligent manner to adapt to the ever-changing network environment, thereby optimizing network transmission performance. The system mainly includes the following core modules, which work together to form a closed-loop adaptive control system 400:

[0122] Data collection module 410 is the sensing unit of the entire system, responsible for capturing critical network status information from the network in real time or near real time. It can be deployed on the client or server side of a QUIC connection, or on monitoring nodes along the network path. This module continuously collects various metrics that form time series data, primarily including round-trip latency, packet loss rate, and throughput. Depending on the complexity of the implementation, other relevant information may also be collected, such as jitter and bandwidth estimates.

[0123] Preprocessing module 420 receives raw network state time series data from the data collection module. Its main function is to clean and transform this raw data to make it suitable as input for large-scale language models. The specific workflow is as described in Example 1, including data noise reduction, data normalization, and, most importantly, converting numerical time series data into text sequences that can be processed by the LLM. The converted text sequences need to be able to effectively represent the temporal dependencies and numerical scale information of the raw data.

[0124] The LLM prediction engine module 430 is the intelligent core of the entire system. This module embeds a pre-trained LLM. It receives as input the network status text sequence generated by the preprocessing module. LLM uses its powerful pattern recognition and sequence prediction capabilities to analyze and understand the input text sequence, thereby predicting and outputting the optimal congestion control parameter α value under the current network status. The deployment location of the LLM prediction engine module can be flexibly selected. Depending on the requirements for inference latency, computing power, and data privacy, it can run on a cloud server or be deployed on an edge computing node closer to the data source. In some cases, if the model is lightweight enough, it can even be directly integrated into the terminal device.

[0125] Verification module 440 is responsible for verifying the validity and security of the α value output by the LLM prediction engine module, as described in Example 5. Before the predicted α value is actually applied to the QUIC congestion controller, the verification module performs a series of checks, such as range checks, stability and safety constraint checks, and provides a rollback mechanism. This module is a key line of defense for ensuring system reliability and network security, preventing the negative impact of LLM mispredictions on the network.

[0126] Parameter application module 450 is responsible for applying the α value verified by the verification module and deemed safe and valid to the QUIC protocol congestion controller, as described in Example 6. It dynamically updates the current congestion control parameter α value by interacting with the user-state congestion control logic in the QUIC protocol stack. The congestion control behavior of the QUIC connection can be adjusted in real time based on the intelligent decision-making of the LLM to adapt to the current network status, thereby optimizing transmission performance.

[0127] These modules work closely together to form a complete closed-loop adaptive control system, encompassing data collection, preprocessing, intelligent prediction, security verification, and parameter application. This system continuously monitors network status, analyzes data, makes predictions, verifies decisions, and adjusts parameters, enabling the QUIC protocol to more intelligently and efficiently navigate the complex and ever-changing modern network environment, enhancing the user experience.

[0128] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0129] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An adaptive congestion control algorithm based on a large language model, characterized in that: include: Acquire a network status time series data set; the network status time series data includes at least round-trip delay (RTT), packet loss rate and throughput; Converting the network state time series data into a text sequence; wherein the text sequence is suitable for serving as an input to a large language model (LLM); The text sequence is input into a pre-trained large-scale language model prediction engine; the large-scale language model prediction engine predicts and outputs the optimal congestion control parameter α value under the current network state based on its learned pattern recognition and sequence prediction capabilities; wherein the congestion control parameter α value is used to adjust the core behavior of the QUIC congestion control algorithm to adjust the window gain rate or target window size; the large-scale language model prediction engine is trained through supervised learning, including: Constructing a labeled dataset containing network status text sequences and corresponding optimal α values; the optimal α values ​​are determined based on predefined performance indicators through network simulation or real-world experiments; Fine-tuning a large language model using the labeled dataset so that the large language model learns a mapping relationship from a network state text sequence to an optimal α value; The large language model prediction engine is also trained through reinforcement learning, including: The state is defined as a textual sequence of network states; the action is defined as selecting or adjusting the α value; and the reward is defined as a reward function calculated based on the network performance indicators of throughput, latency, and packet loss rate. The reward function is calculated as follows: in, , , is the weight coefficient, Throughout represents throughput, Latency represents delay, and Loss_Rate represents packet loss rate; Training the large language model through interaction with a network environment or a simulation environment, so that the large language model learns an α value selection strategy that maximizes the cumulative reward; Verifying the α value output by the large language model prediction engine; the verification is intended to ensure the validity and security of the α value; The verified α value is applied to the congestion controller of the QUIC protocol, and the congestion control behavior of the congestion controller is dynamically adjusted to optimize network transmission performance.

2. The adaptive congestion control algorithm based on a large language model according to claim 1, characterized in that: The converting the network status time series data into a text sequence includes: Preprocessing the network status time series data, including noise reduction and normalization; Convert the preprocessed numeric data into text sequences using at least one of the following techniques: numeric stringification, chunking and tokenization, symbolic aggregation approximation (SAX) or other symbolic representation, quantization / binning; Ensure that the converted text sequence can effectively represent the time dependency and numerical scale information of the original data.

3. The adaptive congestion control algorithm based on a large language model according to claim 1, characterized in that: The adaptive congestion control algorithm based on the large language model also includes the use of model distillation technology: Train a smaller, faster-inference "student" model to mimic the behavior of a larger, better-performing "teacher" large language model. The "student" model is deployed as a large-scale language model prediction engine to meet the real-time requirements of QUIC congestion control.

4. The adaptive congestion control algorithm based on a large language model according to claim 1, characterized in that: The fine-tuning of the large language model using the labeled dataset further includes adopting a parameter efficient fine-tuning technology.

5. The adaptive congestion control algorithm based on a large language model according to claim 1, characterized in that: The verification of the α value output by the large language model prediction engine includes at least one of the following checks: Range check: Ensure that the predicted α value is within the reasonable range allowed by the congestion control algorithm; Stability and safety constraint check: Evaluate whether the α value may cause network oscillation, congestion collapse, or sharp performance deterioration; Rollback mechanism: Monitor network performance after applying a new α value. If significant degradation or instability is observed, roll back to a previously known safe and effective α value.

6. An adaptive congestion control system based on a large language model, characterized in that include: Data collection module: used to obtain network status time series data sets in real time or quasi-real time, the data including at least round-trip delay (RTT), packet loss rate and throughput; Preprocessing module: used to convert the network state time series data into a text sequence that can be processed by a large language model; An LLM prediction engine module is configured to receive the text sequence as input, predict and output the optimal congestion control parameter α value under the current network state through a pre-trained and distilled large language model; wherein the congestion control parameter α value is used to adjust the core behavior of the QUIC congestion control algorithm to adjust the window gain rate or the target window size; the large language model in the LLM prediction engine module is trained through supervised learning, including: Constructing a labeled dataset containing network status text sequences and corresponding optimal α values; the optimal α values ​​are determined based on predefined performance indicators through network simulation or real-world experiments; Fine-tuning a large language model using the labeled dataset so that the large language model learns a mapping relationship from a network state text sequence to an optimal α value; The large language model in the LLM prediction engine module is also trained through reinforcement learning, including: The state is defined as a textual sequence of network states; the action is defined as selecting or adjusting the α value; and the reward is defined as a reward function calculated based on the network performance indicators of throughput, latency, and packet loss rate. The reward function is calculated as follows: in, , , is the weight coefficient, Throughout represents throughput, Latency represents delay, and Loss_Rate represents packet loss rate; Training the large language model through interaction with a network environment or a simulation environment, so that the large language model learns an α value selection strategy that maximizes the cumulative reward; Verification module: used to verify the validity and security of the α value before applying the α value to the QUIC protocol; Parameter application module: used to apply the verified α value to the congestion controller of the QUIC protocol to dynamically adjust the congestion control behavior of the congestion controller.

7. The adaptive congestion control system based on a large language model according to claim 6, characterized in that: The pre-processing module is suitable for: Preprocessing the network status time series data, including noise reduction and normalization; Convert the preprocessed numeric data into text sequences using at least one of the following techniques: numeric stringification, chunking and tokenization, symbolic aggregation approximation (SAX) or other symbolic representation, quantization / binning; Ensure that the converted text sequence can effectively represent the time dependency and numerical scale information of the original data.

8. The adaptive congestion control system based on a large language model according to claim 6, characterized in that: The distilled large language model used by the LLM prediction engine module is a "student" model with smaller parameters and faster inference speed. The "student" model is trained by imitating the behavior of a larger and better-performing "teacher" large language model.

9. The adaptive congestion control system based on a large language model according to claim 6, characterized in that: The verification module is adapted to perform at least one of the following checks: Range check: Ensure that the predicted α value is within the reasonable range allowed by the congestion control algorithm; Stability and safety constraint check: Evaluate whether the α value may cause network oscillation, congestion collapse, or sharp performance deterioration; Rollback mechanism: After the parameter application module applies the new α value, the network performance is monitored. If a significant degradation or instability is found, a rollback to the previously known safe and effective α value is triggered.

10. An electronic device comprising at least one processor; and a memory communicatively connected to the at least one processor; characterized in that: The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the algorithm according to any one of claims 1 to 5.

Images