A burst traffic allocation method based on deep learning and cooperative game

Abstract

A burst traffic allocation method based on deep learning and cooperative game theory, belonging to the field of communication network traffic engineering technology, solves the problems of resource optimization and user experience fairness under burst traffic in dynamic networks. The core of this invention lies in deeply coupling network-side traffic engineering (QoS optimization) with user-side resource allocation (QoE optimization) to form a closed-loop collaborative optimization framework. In burst traffic scenarios, this method first allocates traffic to alternative paths according to the path allocation ratio calculated by the model, reducing maximum link utilization, avoiding network congestion, and enhancing the network's robustness to burst traffic. Then, the available bandwidth on the allocated paths is used as a resource and distributed among multiple users, aiming to maximize the overall experience of all users while ensuring fairness among users, achieving collaborative optimization of service quality and user experience quality in a dynamic network environment.

Description

Technical Field

[0001] This invention belongs to the field of communication network traffic engineering technology, specifically relating to a burst traffic allocation method based on deep learning and cooperative game theory. Background Technology

[0002] In recent years, data center networks and wide area networks (WANs) have faced immense traffic pressure from applications such as video-on-demand, online conferencing, and cloud services, leading to numerous network challenges. During peak network periods or major events, traffic from various applications exhibits significant bursts. Instantaneous traffic requests can surge to several times or even tens of times the baseline level, creating massive traffic spikes. This sudden traffic surge acts like a "tsunami" in the network, rapidly impacting critical nodes in the transmission path, causing core link utilization to jump sharply from a stable state to an overload threshold in a very short time, resulting in network congestion. Furthermore, as internet applications evolve from "connectivity above" to "experience is king," applications such as video streaming, online conferencing, and interactive games have become the mainstay of network traffic. These applications require stable bandwidth guarantees to support large-scale data throughput, directly determining user retention rates and the core competitiveness of service providers, making user experience quality the ultimate benchmark for measuring network service capabilities. Furthermore, in modern network environments, there are significant differences in traffic characteristics between different source-destination node pairs. Some SD pairs (such as video service clusters) have large traffic fluctuations, while some SD pairs (such as backup storage nodes) have smaller traffic fluctuations. At the same time, the network topology also exhibits high heterogeneity.

[0003] To address these issues, existing research can be broadly categorized into two perspectives: one is traffic allocation for bursty traffic. Existing burst-aware TE (Transmission-Oriented) schemes typically handle bursty traffic at the expense of network performance under normal conditions. The other is TE solutions that neglect route optimization to handle the worst-case scenario of all traffic demands. Traditional research methods mainly follow a two-stage "prediction-optimization" paradigm, such as COPE, which predicts traffic demand and optimizes routes to minimize maximum link utilization. However, the inherent dynamics and unpredictability of real-world network traffic impose significant limitations on this approach. Therefore, a new class of TE methods has been proposed. For example, some methods provide robustness guarantees by optimizing for the worst-case scenario, but these methods are often too conservative, severely sacrificing network performance under normal conditions. Google's "Hedging" mechanism, introduced in its Jupiter data center network, enhances robustness by limiting path sensitivity, but still imposes uniform constraints on all SD (Site-Dependent) pairs, failing to fully consider the significant differences in traffic stability between different SD pairs.

[0004] Second, there's the issue of resource allocation geared towards user experience quality. Traditional resource allocation mechanisms primarily employ methods based on fixed rules or static optimization. While these methods are simple to implement, they struggle to adapt to dynamic changes in the network environment. Some client-driven delivery schemes and heuristic algorithms attempt to optimize QoE by adapting to dynamic bandwidth, but they lack system-wide scheduling and management of concurrent requests, performing poorly particularly when facing sudden traffic spikes and heterogeneous multi-user demands. However, most existing research treats network traffic allocation and user resource allocation as two independent processes, lacking system-level collaborative optimization. This disconnect between network-side traffic engineering optimization and user-side experience improvement goals leads to inefficient resource allocation and makes it difficult to achieve global optimum in dynamic and uncertain environments. Summary of the Invention

[0005] To address the problems existing in the aforementioned background technologies, this invention proposes a burst traffic allocation method based on deep learning and cooperative game theory. First, an intelligent traffic allocation mechanism is designed to dynamically and robustly allocate network traffic to candidate paths in scenarios with burst traffic caused by a surge in user requests, thereby fundamentally avoiding network congestion and laying a resource foundation for subsequent user experience optimization. After optimizing the allocation of network-side path resources through the first-stage model, these discrete path resources are fairly and efficiently allocated to multiple user requests arriving simultaneously with heterogeneous needs, maximizing the overall user experience. A joint optimization framework that tightly couples network resource optimization and user experience improvement is constructed. A cooperative game theory approach is used to solve for the Nash equilibrium solution to allocate resources, achieving a balance between fairness and efficiency. This invention achieves closed-loop collaboration between network-side optimization and user-side protection, breaking down the barriers between network traffic engineering and user resource allocation in traditional research. It does not simply connect the two modules but strives to build a deeply coupled joint optimization framework, enabling network-side bandwidth allocation decisions to accurately serve the goal of improving user experience.

[0006] Addressing the bursty nature of network traffic, the differences in traffic characteristics between different source-destination node pairs, and the heterogeneity of network topology, this invention constructs an end-to-end joint optimization framework based on a deep learning-based traffic allocation model. This framework establishes a direct mapping from historical traffic data to optimal routing configurations, bypassing the prediction stage in the traditional "prediction-optimization" paradigm and reducing the amplification of downstream solutions by upstream prediction biases. Furthermore, a multi-objective loss function considering network topology and link importance is designed to achieve differentiated traffic scheduling for different source-destination pairs. This drives the model to learn a routing strategy that is both efficient and inherently robust, thereby achieving precise scheduling and efficient utilization of network resources. This intelligent resource allocation strategy can significantly reduce Maximum Link Utilization (MLU), fundamentally preventing network congestion.

[0007] In situations with limited resources, simple "allocation on demand" may lead to resource waste, while "equal allocation" can severely damage the experience of high-demand users. This invention introduces a user resource allocation model based on cooperative game theory, shifting the optimization objective from purely network-level metrics (such as MLU) to a user-centric quality of experience (QoE) metric. Addressing the efficiency and fairness issues of resource allocation in multi-user competitive environments, this invention finds the optimal balance between maximizing global user experience and fairness among users while strictly satisfying network resource constraints. By establishing a cooperative bargaining game model, the network resource allocation problem is transformed into a global user experience quality optimization problem. Reasonable user utility functions and resource constraints are designed to optimize user experience and fairness while meeting network resource limitations. Furthermore, the alternating direction multiplier method is used to solve the convex optimization problem, ensuring efficient resource utilization and fair satisfaction of user needs in complex network environments. This not only improves service satisfaction for high-demand users but also guarantees the basic experience of all users, achieving a balance between fairness and efficiency in resource allocation.

[0008] A burst traffic allocation method based on deep learning and cooperative game theory, characterized by the following steps:

[0009] Step 1: The network controller inputs the topology information of the target network and prepares several shortest paths as a set of candidate paths. At the same time, it continuously collects real-time traffic data of each link to construct a traffic demand matrix as data input.

[0010] Step 2: Preprocess the traffic demand matrix by converting the alternative paths, nodes, and link capacities into different formats; extract data from the past T time slices from the historical database to construct a training set; the input for each sample is a sequence of historical traffic matrices for L consecutive time slices, the data is standardized, and divided into training and test sets;

[0011] Step 3: Construct a deep neural network model. The input is the historical traffic demand matrix, alternative paths, and network topology. The output is the traffic allocation ratio for each source-destination pair on all candidate paths.

[0012] Step 4: During the real-time network operation phase, the historical traffic data within the previous time window is input into the trained DNN model for forward inference. The model outputs a set of optimized traffic allocation ratios. Based on these ratios, the network controller dynamically allocates the real-time traffic to each candidate path according to source and destination pairs, completing the routing optimization on the network side and calculating the maximum link utilization on each link as a performance indicator.

[0013] Step 5: Based on the traffic allocation results, calculate the actual traffic fp carried on each path p, and allocate the bandwidth resource B of each path p. p Summing gives the total bandwidth resources B available for allocation to users across the entire network. S ;

[0014] Step 6: Model the multi-user resource allocation problem as a cooperative bargaining game (CBG), and define the logarithmic sum of the QoE of each user as the optimization objective. The objective is to find a set of bit rate allocation schemes that maximize the global QoE function while satisfying the total bandwidth constraint.

[0015] Step 6: Assuming the utility function and data size are linearly related to the bit rate, the objective is transformed into allocating a suitable bit rate q to each user. m By introducing auxiliary variable z and Lagrange multipliers λ, an augmented Lagrange function is constructed, and the alternating direction multiplier method is used for distributed iterative solution to obtain the bit rate q allocated to each user. m ; Update user local variables, global auxiliary variables, and Lagrange multipliers in parallel; Repeat the iteration until both the original residual and the dual residual are less than the preset convergence threshold ε, and output the final optimal bit rate allocation scheme;

[0016] Step 7: The network controller allocates corresponding bandwidth resources to each user's data stream according to the optimal bit rate allocation scheme obtained from the solution; at the same time, it continuously monitors the network status after implementation and calculates the user's single QoE and total QoE as performance indicators as the standard for evaluating the model.

[0017] Compared with the prior art, the significant advantages of this invention are as follows:

[0018] (1) This invention proposes an intelligent traffic allocation mechanism that combines high robustness and high performance, effectively addressing sudden traffic surges. It directly learns the mapping from historical traffic to the optimal allocation ratio, bypassing the explicit traffic prediction stage and fundamentally eliminating the problem of prediction errors being amplified during the optimization phase. The innovative composite loss function introduces traffic variance, path influence, and link centrality as penalty terms, enabling the model to adaptively apply differentiated robust constraints to source-destination pairs with different stability. It limits the path concentration of highly volatile traffic and allows for more efficient allocation of stable traffic. This intelligent traffic allocation method reduces the average MLU by 5.2%-21.0%, fundamentally avoiding network congestion, improving network resource utilization efficiency, and providing key technical support for building high-performance network infrastructure.

[0019] (2) On the user side, this invention constructs resource allocation as a Nash bargaining cooperative game problem and uses the alternating direction multiplier method for efficient solution. This ensures that the overall system experience is maximized while meeting the minimum experience threshold for each user. It fully considers the heterogeneous needs and service priorities of users, overcomes the drawbacks of resource waste caused by "on-demand allocation" and the damage to the user experience of high-demand users caused by "average allocation", and achieves the best balance between user experience and fairness, truly taking into account both fairness and efficiency. Compared with traditional allocation strategies, the system's global QoE is improved by more than 10%, and the resource allocation result is smoother and fairer.

[0020] (3) The invention addresses the problem of single optimization objectives in traditional methods by fundamentally breaking down the barrier between network layer optimization and application layer experience. This allows network layer routing decisions to directly serve the application layer's experience objectives, achieving a fundamental shift from "resource allocation as a service to experience." It systematically achieves global optimization of network resource utilization and end-user satisfaction, providing a complete solution with empirical advantages for building a truly experience-centric future network. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the method in a specific embodiment of the present invention.

[0022] Figure 2 This is a flowchart illustrating the training process of the DNN model in a specific embodiment of the present invention.

[0023] Figure 3 This is a flowchart of the ADMM iterative update process in a specific embodiment of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this invention proposes a burst traffic allocation method based on deep learning and cooperative game theory, with the following specific steps:

[0026] Step 1: The network controller inputs the topology information of the target network, including the node set V, edge set E, and link capacity C, and prepares K shortest paths as a candidate path set P. Simultaneously, a data acquisition module is deployed to continuously collect real-time traffic data from each link to construct a traffic demand matrix, which serves as the data input.

[0027] Step 2 involves flattening the traffic demand matrix, providing alternative paths, and converting network topology information such as node sets, link capacities, and edge sets into corresponding JSON formats. A training set is constructed by extracting data from the past T time slices from the historical database. Each sample's input is a sequence of historical traffic matrices spanning t consecutive time slices. The traffic matrix is ​​flattened and divided into training and test sets.

[0028] Step 3, construct a deep neural network model, such as Figure 2 As shown, the input consists of the historical traffic demand matrix of the training set, candidate paths, and network topology. The output is the traffic allocation ratio for each source and destination across all candidate paths. The model is trained using a custom composite loss function, and the network parameters are optimized through backpropagation. The model employs a multilayer perceptron architecture, containing multiple hidden layers and ReLU activation functions. The output layer uses a Softmax function to ensure that the sum of the allocation ratios for each source and destination across all paths is 1. The model is trained using the training set with the goal of minimizing the following composite loss function. :

[0029]

[0030] The first term is the maximum link utilization loss, which aims to minimize the maximum link utilization (MLU) under the existing configuration; the second term is the fine-grained robustness constraint loss, whose parameters include the historical traffic variance of a specific source-destination pair. Link centrality of path p and the maximum path influence predicted by the model. Defined as the traffic allocation ratio The capacity of the minimum link on path p The ratio of λ to λ is the balancing weight. The model parameters are optimized using backpropagation until the model converges on the validation set.

[0031] Among these parameters, a larger historical traffic variance indicates greater traffic fluctuation and instability, resulting in stricter penalties. Link centrality represents the number of times traffic passes through a link in the network; a larger value indicates the path's criticality in the network topology. The critical links differ across networks, and this parameter allows for the application of different constraints based on the network topology. The maximum traffic influence on a path, if large, indicates a significant amount of traffic concentrated on a potentially low-capacity path, signifying high traffic dependence on that path—a high-risk behavior, thus penalizing this parameter. This model design reduces the maximum link utilization in the network while applying different fine-grained constraints to different source-destination pairs (SD pairs), ensuring the robustness of the entire network. Through this innovative loss function design, the model is infused with prior knowledge of network topology and traffic dynamics during training, driving it to learn a routing strategy that is both efficient and inherently robust.

[0032] Step 4: During the real-time network operation phase, historical traffic data within a time window *t* preceding the current moment in the test set is input into the trained DNN model for forward inference. The model directly outputs a set of optimized traffic allocation ratios. Based on these ratios, the network controller dynamically allocates the real-time arriving traffic to each candidate path according to source-destination pairs, completing network-side routing optimization and calculating the MLU (Mean Integration Limit) for each link as a performance metric.

[0033] Maximum Link Utilization (MLU) of Path p p The calculation formula is:

[0034]

[0035] Among them, f p c represents the traffic transmitted along path p. p This represents the minimum link capacity on path p.

[0036] Step 5: Based on the traffic allocation results, calculate the actual traffic f transmitted on each path p. p The bandwidth resource B for each path p p Summing gives the total bandwidth resources B available for allocation to users across the entire network. S .

[0037] The available bandwidth in the network is:

[0038]

[0039] in, The proportion of traffic allocated to path p. This represents the traffic volume between SD pairs at the current time.

[0040] Step 6: Model the multi-user resource allocation problem as a cooperative bargaining game (CBG), defining the logarithmic sum of QoE for each user as the optimization objective. The goal is to find a set of bit rate allocation schemes that, while satisfying the total bandwidth constraint, maximize the following global QoE function:

[0041]

[0042]

[0043] Where M represents the set of user requests, m represents the user's request, k represents the bitrate level allocated to the user, and n m This indicates the priority of the m-th user's request. This indicates the data size of request m at bit rate level k. The data size of all requests m cannot exceed the bandwidth allocated on the path in the first step.

[0044] Step 7: Assuming the utility function and data size are linearly related to the bit rate, the objective is transformed into allocating a suitable bit rate q to each user. m The cooperative game optimization problem is solved by introducing an auxiliary variable z and a Lagrange multiplier λ, constructing its augmented Lagrange function, and then using the alternating direction multiplier method for distributed iterative solution to obtain the bit rate q allocated to each user. m The process involves parallel updates of user-local variables, global auxiliary variables, and Lagrange multipliers. This iteration is repeated until both the original residual and the dual residual are less than a preset convergence threshold ε, at which point the final optimal bit rate allocation scheme is output.

[0045] S71, in order to find the Nash solution, it is assumed that the utility function and data size are linearly related to the bit rate. , Linear parameters to be set:

[0046]

[0047]

[0048] S72, assuming a linear relationship, sets a minimum utility requirement for the user. To ensure the lowest possible user experience, the objective function will be transformed into the following form. The appropriate bit rate q to allocate to each user needs to be calculated based on this objective function. m :

[0049]

[0050] S73, based on the transformed objective function, to solve the Nash solution, first construct the Lagrange augmented function:

[0051]

[0052] Since the objective function is a convex optimization problem, it is suitable to use the Alternating Direction Multiplier Method (ADMM) to solve it.

[0053] S74 first fixes the global variable and updates the user local variable in parallel. Each user independently updates its requested bit rate q based on the current "price signal" (Lagrange multiplier). m To optimize its local objective, for each q, first calculate the derivative and set it to 0, then calculate the gradient, perform gradient descent update, project it onto the feasible region, and obtain the result:

[0054]

[0055] in For projection operators, Step size, The gradient is calculated using the following formula:

[0056]

[0057] in , Here, j represents the user's summation index in the current step, and is the weight parameter. For linear parameters, This represents the user's bit rate at the current step.

[0058] Because the utility function is convex, and the projected gradient method is particularly suitable for convex optimization problems, it avoids directly solving nonlinear equations, is computationally simple and stable, has a small computational cost, and converges faster than traditional methods.

[0059] S75, then globally constrain and coordinate the update of slack variables to ensure that the total resource consumption of all users does not exceed the total path capacity. The resulting closed-form solution is:

[0060]

[0061] S76, the final multiplier update updates the Lagrange multiplier based on the degree of constraint violation, which serves as the "price" for adjusting user behavior in the next iteration.

[0062]

[0063] S77, finally check the stopping criteria, and output the result after iterative update, which is the bit rate q allocated to each user:

[0064]

[0065] Where r k and s k denoted as the original residual and the dual residual, respectively.

[0066] Repeat steps S74 through S77 until both the original residual (degree of constraint violation) and the dual residual (degree of change in the solution) are less than the preset convergence threshold. Output the final optimal bit rate allocation scheme.

[0067] Step 8: The network controller allocates corresponding bandwidth resources to each user's data stream according to the optimal bit rate allocation scheme obtained from the solution. Simultaneously, the system continuously monitors the network status after implementation (such as actual link utilization and user QoE feedback), calculating the user's single QoE and total QoE as performance indicators and standards for evaluating the model.

[0068] The QoE index is then calculated based on the bit rate obtained from the ADMM iteration.

[0069] The formula for calculating the QoE for each user u is:

[0070]

[0071] in, Let be the average bit rate utility function. The bit rate allocated to user u For bit rate switching penalty function, To switch bit rates, , These are their respective weight parameters.

[0072] The logarithm of QoE for a single user and the overall QoE are used to measure the overall QoE. A higher QoE indicates a better user experience and a more ideal model performance. For the entire user set U, the total QoE is:

[0073]

[0074] After the model calculation outputs, the maximum link utilization (MLU) in the first stage and the QoE of each user and the total QoE of the user in the second stage can be calculated according to the above formula. The comparison scheme compares the MLU results with existing traffic allocation schemes (such as COPE, Oblivious, etc.) and the QoE with traditional resource allocation schemes (such as average allocation, proportional allocation, etc.). The results show that the present invention ensures a high QoE for users while reducing the maximum link utilization, achieving a balance between load balancing and fairness.

[0075] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A burst traffic allocation method based on deep learning and cooperative game theory, characterized in that: The method includes the following steps: Step 1: The network controller inputs the topology information of the target network and prepares several shortest paths as a set of candidate paths. At the same time, it continuously collects real-time traffic data of each link to construct a traffic demand matrix as data input. Step 2: Preprocess the traffic demand matrix by converting the alternative paths, nodes, and link capacities into different formats; extract data from the past T time slices from the historical database to construct a training set; the input for each sample is a sequence of historical traffic matrices for L consecutive time slices, the data is standardized, and divided into training and test sets; Step 3: Construct a deep neural network model. The input is the historical traffic demand matrix, alternative paths, and network topology. The output is the traffic allocation ratio for each source-destination pair on all candidate paths. Step 4: During the real-time network operation phase, the historical traffic data within the previous time window is input into the trained DNN model for forward inference. The model outputs a set of optimized traffic allocation ratios. Based on these ratios, the network controller dynamically allocates the real-time traffic to each candidate path according to source and destination pairs, completing the routing optimization on the network side and calculating the maximum link utilization on each link as a performance indicator. Step 5: Based on the traffic allocation results, calculate the actual traffic fp carried on each path p, and allocate the bandwidth resource B of each path p. p Summing gives the total bandwidth resources B available for allocation to users across the entire network. S ; Step 6: Model the multi-user resource allocation problem as a cooperative bargaining game (CBG), and define the logarithmic sum of the QoE of each user as the optimization objective. The objective is to find a set of bit rate allocation schemes that maximize the global QoE function while satisfying the total bandwidth constraint. Step 7: Assuming the utility function and data size are linearly related to the bit rate, transform the objective into a bit rate q allocated to each user. m By introducing auxiliary variable z and Lagrange multipliers λ, an augmented Lagrange function is constructed, and the alternating direction multiplier method is used for distributed iterative solution to obtain the bit rate q allocated to each user. m ; Update user local variables, global auxiliary variables, and Lagrange multipliers in parallel; Repeat the iteration until both the original residual and the dual residual are less than the preset convergence threshold ε, and output the final optimal bit rate allocation scheme; Step 8: The network controller allocates corresponding bandwidth resources to each user's data stream according to the optimal bit rate allocation scheme obtained from the solution; at the same time, it continuously monitors the network status after implementation and calculates the user's single QoE and total QoE as performance indicators as the standard for evaluating the model.

2. The burst traffic allocation method based on deep learning and cooperative game theory according to claim 1, characterized in that: In step 3, the model is trained using a custom composite loss function, and the network parameters are optimized through backpropagation. The model employs a multilayer perceptron structure, containing multiple hidden layers and a ReLU activation function. The output layer uses a softmax function to ensure that the sum of the allocation ratios of each source and destination along each path is 1. The model is trained using a training set, with the goal of minimizing the following composite loss function. : The first term is the maximum link utilization loss, which aims to minimize the maximum link utilization (MLU) under the current configuration; the second term is the fine-grained robustness constraint loss, whose parameters include the historical traffic variance of the source and destination pairs. Link centrality of the path and the maximum path influence predicted by the model. Defined as the traffic allocation ratio and path capacity The ratio of λ to λ is used; λ is the balancing weight; the model parameters are optimized through backpropagation until the model converges on the validation set.

3. The burst traffic allocation method based on deep learning and cooperative game theory according to claim 2, characterized in that: In step 4, the maximum link utilization (MLU) of path p is... p The calculation formula is: Among them, f p c represents the traffic transmitted along path p. p This represents the minimum link capacity on path p.

4. The burst traffic allocation method based on deep learning and cooperative game theory according to claim 3, characterized in that: In step 5, the available bandwidth in the network is: in, The proportion of traffic allocated to path p. This represents the traffic volume between the source node S and the destination node D at the current time.

5. The burst traffic allocation method based on deep learning and cooperative game theory according to claim 4, characterized in that: In step 6, the global QoE function is: Where M represents the set of user requests, m represents the user's request, k represents the bitrate level allocated to the user, and n m This indicates the priority of the m-th user's request. This indicates the data size of request m at bit rate level k. The data size of all requests m cannot exceed the bandwidth allocated on the path in the first step.

6. The burst traffic allocation method based on deep learning and cooperative game theory according to claim 5, characterized in that: Step 7 includes the following steps: S71, in order to find the Nash solution, it is assumed that the utility function and data size are linearly related to the bit rate. , Linear parameters to be set: S72, assuming a linear relationship, sets a minimum utility requirement for the user. To ensure the lowest possible user experience, the objective function is transformed into the following form. Based on this objective function, the appropriate bit rate q to allocate to each user is calculated. m : S73, based on the transformed objective function, to solve the Nash solution, first construct the Lagrange augmented function: The alternating direction multiplier method (ADMM) is used to solve the function; S74 first fixes the global variable and updates the user local variable in parallel. Each user independently updates its requested bit rate q based on the current Lagrange multipliers. m To optimize its local objective; for each q, first calculate the derivative and set the derivative to 0, then calculate the gradient, perform gradient descent update, project onto the feasible region, and obtain the result: in For projection operators, Step size, The gradient is calculated using the following formula: in , Here, j represents the user's summation index in the current step, and is the weight parameter. For linear parameters, The bit rate for the user in the current step; S75, then globally constrain and coordinate the update of slack variables to ensure that the total resource consumption of all users does not exceed the total path capacity; the resulting closed-loop solution is: S76, Update the Lagrange multipliers based on the degree of constraint violation, as the Lagrange multipliers for adjusting user behavior in the next iteration: S77, finally check the stopping criteria, and output the result after iterative update, which is the bit rate q allocated to each user: Where r k and s k represent the original residual and the dual residual, respectively; Repeat steps S74 through S77 until both the original residual and the dual residual are less than the preset convergence threshold. Output the final optimal bit rate allocation scheme.

7. The burst traffic allocation method based on deep learning and cooperative game theory according to claim 6, characterized in that: In step 8, the QoE calculation formula for each user u is: in, Let be the average bit rate utility function. The bit rate allocated to user u For bit rate switching penalty function, To switch bit rates, , For their respective parameters; The logarithm of QoE for a single user and the overall QoE are used to measure the overall QoE. A higher QoE indicates a better user experience and a more ideal model performance. For the entire user set U, the total QoE is: 。

Images