AI-driven real-time network optimization algorithm

Abstract

The application relates to the technical field of network optimization, and discloses an AI-driven real-time network optimization algorithm, which comprises the following steps: data optimization: using a generative adversarial network to preprocess network state data, extracting optimization features and generating enhanced data; topology modeling: using a graph neural network to model network topology and update node features; resource optimization: using deep reinforcement learning to dynamically adjust computing, storage and bandwidth resources; global scheduling: based on a multi-objective optimization method, optimizing resource allocation; secondary optimization: combining the Lagrange relaxation method and graph optimization technology to further optimize the scheduling strategy; adaptive learning: improving the adaptability of the system in a dynamic environment through self-supervised learning and meta-learning. Through matrix operation technology based on normal equations and in combination with an efficient data preprocessing process, the speed of solving a regression model is improved, the problem of slow calculation of a traditional method on a large data set is solved, and fast and accurate regression analysis is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of network optimization, in particular to an AI-driven real-time network optimization algorithm. BACKGROUND

[0002] Currently, artificial intelligence (AI) technology is increasingly widely applied in the field of network optimization. Many optimization algorithms are used for key tasks such as traffic scheduling, resource allocation and data transmission to improve the performance and stability of network systems. Among them, gradient-based optimization algorithms such as gradient descent and its variants are widely used in parameter adjustment and optimization calculation in complex network environments. These methods gradually optimize network performance by iteratively adjusting parameters and have achieved good results in many application scenarios. At the same time, mathematical tools such as matrix calculation and regression analysis also play an important role in network optimization, helping models better adapt to changing network environments.

[0003] However, with the continuous expansion of network size and the surge of data traffic, traditional gradient optimization methods face certain challenges in terms of computational efficiency and stability. First, iterative optimization methods usually require multiple training rounds, which are computationally intensive and can affect response speed in real-time application scenarios. Second, such methods are sensitive to hyperparameters, such as the choice of learning rate, which directly affects the convergence speed and optimization effect. If the parameters are not properly set, it can cause training process oscillation or slow convergence. In addition, in complex data environments, gradient optimization methods are affected by gradient vanishing or gradient explosion, reducing the stability of the algorithm. Therefore, how to improve computational efficiency, reduce dependence on hyperparameters, and ensure the stability of the optimization process has become a direction that needs to be broken through in current network optimization technology. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides an AI-driven real-time network optimization algorithm, which solves the problems of low computational efficiency, strong dependence on hyperparameters and unstable convergence of traditional optimization algorithms in real-time network optimization scenarios.

[0005] To achieve the above purpose, the application is implemented by the following technical scheme: an AI-driven real-time network optimization algorithm, comprising the following steps:

[0006] S1, preprocessing network state data by generating an adversarial network, extracting optimized network feature data, and generating high-quality enhanced data;

[0007] S2, modeling the network topology based on the enhanced data using a graph neural network, capturing the dependency relationship between nodes, and updating the node features through a graph convolution network;

[0008] S3. Based on node features, a deep reinforcement learning algorithm is used to dynamically optimize and adjust computing, storage, and bandwidth resources, and adaptively allocate resources.

[0009] S4. Input the resource allocation scheme into the multi-objective optimization method to further minimize latency and packet loss rate, maximize throughput, optimize energy efficiency, and finally output a global optimized resource scheduling strategy.

[0010] S5. Combining the Lagrange relaxation method and graph optimization techniques, further optimize the global resource scheduling strategy, or the final scheduling scheme;

[0011] S6. The final scheduling scheme improves the adaptability of the network model in dynamic environments through self-supervised learning and meta-learning, ensuring that the optimization system operates efficiently under different network conditions.

[0012] Preferably, in step S1, the generative adversarial network is trained using a conditional generative adversarial network, specifically including:

[0013] The generator produces enhanced data based on network status inputs such as network traffic, node load, and bandwidth utilization.

[0014] The discriminator evaluates the quality of the generated data and provides feedback to optimize the generator, thereby improving the authenticity and validity of the data.

[0015] Preferably, in step S2, the graph neural network performs multi-layer feature propagation through a graph convolutional network, updates node features, and fuses neighboring node information to optimize resource allocation strategies, specifically including:

[0016] A K-layer graph convolutional network is used to update node features layer by layer, so that the computing power, bandwidth utilization and storage resource information of each node are gradually propagated in the adjacency relationship, thereby enhancing the perception of the global network state.

[0017] By assigning different weights to adjacent nodes through an attention mechanism, the accuracy of resource allocation decisions is increased.

[0018] Optimized node embedding vectors are generated and used as input to a reinforcement learning algorithm to improve the decision-making ability for resource allocation.

[0019] Preferably, in step S3, the deep reinforcement learning optimization of network resource allocation specifically includes:

[0020] The reward function is calculated based on network latency, throughput, packet loss rate, and energy efficiency performance indicators, and the weights are dynamically adjusted according to the current network status.

[0021] Combine deep Q-networks for offline training and leverage proximal policy optimization for online optimization to improve policy adjustment capabilities;

[0022] The state inputs of the reinforcement learning include node computing power, storage utilization, bandwidth load and topology information, and the action space includes resource allocation strategy adjustment, using features generated by graph neural networks to make policy decisions.

[0023] Preferably, in step S4, the multi-objective optimization employs multi-task learning, specifically including:

[0024] By dynamically adjusting weights, the optimization objectives of latency, throughput, packet loss rate, and energy efficiency are balanced under different network conditions.

[0025] The conflicts between the optimization objectives are addressed by dynamically adjusting the weights using the Pareto optimization method, and by combining reinforcement learning to adaptively adjust the weight coefficients of different optimization objectives.

[0026] Preferably, step S5 further includes the following steps:

[0027] Global resource scheduling is optimized through graph optimization, and global resource allocation and coordination are achieved through local information transmission in a distributed environment.

[0028] Lagrange multipliers are used to optimize local resource allocation so that overall resource scheduling satisfies global constraints.

[0029] The local information includes, but is not limited to:

[0030] Computing resource status: CPU load, GPU load, memory usage, and storage resources;

[0031] Network resource status: bandwidth utilization, link load, packet loss rate, and network latency;

[0032] Task scheduling information: current task queue, task execution priority, and task execution status;

[0033] Energy consumption information: power consumption of a single computing node, thermal management, and battery power status;

[0034] Load balancing information: local node load status, available computing resources, and task migration information;

[0035] Global constraints include, but are not limited to, computing resource constraints, network resource constraints, latency constraints, energy consumption constraints, global optimization of task scheduling, and reliability and fault tolerance constraints.

[0036] Preferably, step S6 further includes the following steps:

[0037] Self-supervised learning automatically learns network state features through unsupervised pre-training;

[0038] Meta-learning uses a small amount of sample data to quickly adjust model parameters, enabling optimization strategies to adapt to different network environments.

[0039] The sample data includes computing resource status, network status, task scheduling information, energy consumption data, and user demand data.

[0040] The model parameters include computational resource allocation weights, task priority adjustment parameters, reinforcement learning reward functions, and deep learning optimization parameters.

[0041] Preferably, in step S6, the network state is analyzed in real time using a self-attention mechanism and a Transformer model, specifically as follows:

[0042] By improving policy computation efficiency through self-attention mechanisms, Q-value calculation and policy updates can be accelerated, thereby enhancing the real-time performance of network optimization.

[0043] Preferably, the algorithm has a fault-tolerant mechanism that dynamically adjusts in the event of network failure, specifically including:

[0044] By monitoring bandwidth utilization, response time, and packet loss rate through fault detection, faulty nodes can be identified.

[0045] Automatically perform load balancing to ensure uninterrupted task execution, and select the optimal node to take over based on resource availability;

[0046] Dynamic resource reallocation optimizes bandwidth, computing resources, and storage space allocation, reducing the impact of faults on network performance.

[0047] Preferably, the fault tolerance mechanism further includes a preventative fault tolerance strategy based on predictive analytics, specifically including:

[0048] By analyzing historical data through machine learning models, potential failure points are predicted, and the probability of failure is calculated based on long-term trends.

[0049] Adjust resource allocation in advance within a 5-10 time window before a failure occurs to reduce the impact of future failures;

[0050] Once the faulty node recovers, resource allocation is adaptively adjusted to optimize global resource utilization, and historical data is used to update the prediction model to improve the accuracy of future predictions.

[0051] This invention provides an AI-driven real-time network optimization algorithm. It has the following beneficial effects:

[0052] 1. This invention improves the speed of solving regression models by using matrix operation techniques based on normal equations and combining them with an efficient data preprocessing process. Compared with traditional iterative algorithms, this invention solves the problem of slow computation on large datasets by reducing the amount of computation and avoiding multiple optimization steps, thus achieving fast and accurate regression analysis.

[0053] 2. This invention successfully solves the training instability caused by manual parameter tuning in the prior art by combining matrix inversion operation with optimization algorithm. This technical solution is not only highly automated, but also ensures the adaptability and reliability of the regression model on different datasets through stable mathematical derivation, thereby improving the accuracy of the training process.

[0054] 3. By optimizing the regression solution strategy, this invention makes regression analysis not only more efficient, but also stable in various complex data environments. Compared with existing gradient descent-dependent schemes, this invention avoids gradient vanishing and hyperparameter adjustment problems, achieves a smoother convergence process, and effectively improves the processing capacity of large-scale datasets.

[0055] 4. This invention reduces redundant calculations and improves data utilization through matrix operation optimization, solving the problems of high memory consumption and low efficiency in traditional algorithms. Combined with automated model training, this invention provides a new method for regression analysis that is both efficient and accurate, significantly improving work efficiency and analysis quality. Attached Figure Description

[0056] Figure 1 This is a flowchart of the algorithm optimization steps of the present invention. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see the appendix Figure 1 This invention provides an AI-driven real-time network optimization algorithm, comprising the following steps:

[0059] S1. Preprocess the network state data using a generative adversarial network, extract optimized network feature data, and generate high-quality augmented data.

[0060] S2. Based on augmented data, a graph neural network is used to model the network topology, capture the dependencies between nodes, and update node features through a graph convolutional network.

[0061] S3. Based on node features, a deep reinforcement learning algorithm is used to dynamically optimize and adjust computing, storage, and bandwidth resources, and adaptively allocate resources.

[0062] S4. Input the resource allocation scheme into the multi-objective optimization method to further minimize latency and packet loss rate, maximize throughput, optimize energy efficiency, and finally output a global optimized resource scheduling strategy.

[0063] S5. Combining the Lagrange relaxation method and graph optimization techniques, further optimize the global resource scheduling strategy, or the final scheduling scheme;

[0064] S6. The final scheduling scheme improves the adaptability of the network model in dynamic environments through self-supervised learning and meta-learning, ensuring that the optimization system runs efficiently under different network conditions.

[0065] The algorithm has a fault-tolerant mechanism that dynamically adjusts itself in the event of network failures, specifically including:

[0066] By monitoring bandwidth utilization, response time, and packet loss rate through fault detection, faulty nodes can be identified.

[0067] Automatically perform load balancing to ensure uninterrupted task execution, and select the optimal node to take over based on resource availability;

[0068] Dynamic resource reallocation optimizes bandwidth, computing resources, and storage space allocation, reducing the impact of faults on network performance;

[0069] The fault tolerance mechanism further includes preventative fault tolerance strategies based on predictive analytics, specifically including:

[0070] By analyzing historical data through machine learning models, potential failure points are predicted, and the probability of failure is calculated based on long-term trends.

[0071] Adjust resource allocation in advance within a 5-10 time window before a failure occurs to reduce the impact of future failures;

[0072] Once the faulty node recovers, resource allocation is adaptively adjusted to optimize global resource utilization, and historical data is used to update the prediction model to improve the accuracy of future predictions.

[0073] In step S1, the generative adversarial network is trained using a conditional generative adversarial network, specifically including:

[0074] The generator produces enhanced data based on network status inputs such as network traffic, node load, and bandwidth utilization.

[0075] The discriminator evaluates the quality of the generated data and provides feedback to optimize the generator, thereby improving the authenticity and validity of the data.

[0076] Specifically, the main objective of step S1 is to preprocess the network state data using a generative adversarial network (GAN) to extract optimized network feature data and generate high-quality augmented data. This step lays the foundation for subsequent modeling processes based on augmented data. In particular, GANs can generate realistic data samples through adversarial training, thereby improving the performance and robustness of the network model when handling complex network states. It is important to note that step S1 is closely related to subsequent steps; the optimized network feature data directly affects the modeling effect of the graph neural network and the optimization of resource allocation strategies in step S2.

[0077] In this embodiment, the generative adversarial network (GAN) is trained using a conditional GAN. As one possible implementation, the conditional GAN, by introducing network traffic, node load, and bandwidth utilization as conditional inputs, helps generate data samples that better reflect the actual network environment. Specifically, the conditional GAN ​​consists of a generator and a discriminator. The generator produces enhanced data based on network state inputs (e.g., network traffic, node load, bandwidth utilization), while the discriminator evaluates the quality of the generated data and optimizes the generator through feedback, thereby improving the realism and effectiveness of the data.

[0078] Specifically, when generating network state data, the generator first extracts feature information from parameters such as network traffic, node load, and bandwidth utilization, and then generates enhanced data based on these features. Here, network traffic can be represented as... Node load can be represented as Bandwidth utilization can be expressed as The generator uses these conditional data to generate augmented data through a deep neural network. The goal of the generator is to minimize the gap between the generated data and the actual network data, thereby improving the realism of the generated data.

[0079] While the generator generates data, the discriminator receives the generated data and actual network data as input to evaluate their authenticity. The discriminator's output is a scalar.

[0080] The discriminator continuously optimizes its judgment ability, aiming to maximize its recognition rate of real data and minimize its misclassification rate of generated data during training. During training, the discriminator provides feedback to the generator with its score, and the generator adjusts its parameters to improve the quality of its generated data.

[0081] In some embodiments, the optimization process of the generative adversarial network can be achieved by minimizing the following loss function:

[0082] ;

[0083] in, This represents the actual distribution of network data. Let be the loss function, representing the total loss of the generative adversarial network; To obtain from real data distribution Real data from mid-sample ; This is the expectation operator, representing the expected value of the data distribution; For discriminator For real data The output; For discriminator For generated data The output; To generate data distribution Mid-sampled generated data ; To generate a data distribution, and These represent the discriminator's judgment results for real data and generated data, respectively. The generator's goal is to minimize the difference between generated data and real data, while the discriminator's goal is to maximize its ability to distinguish real data and minimize its misclassification rate for generated data.

[0084] Training generative adversarial networks (GANs) can, to some extent, compensate for the lack of real-world data by generating high-quality augmented data through adversarial training. This augmented data can effectively supplement the scarce data samples that may exist in real-world network environments, thereby improving the adaptability of network optimization models under various network conditions.

[0085] In some implementations, the enhanced output data of the generative adversarial network is further used in the subsequent step S2 for network topology modeling based on the graph neural network. The optimized network feature data is then provided as input to the graph neural network to capture the dependencies between network nodes, and the node features are updated through a graph convolutional network, thereby providing strong support for subsequent dynamic optimization and adjustment.

[0086] Through the preprocessing in step S1, the network state data is optimized and enhanced data is generated. This data not only improves the accuracy of network state modeling, but also effectively enhances the efficiency of subsequent resource scheduling optimization.

[0087] In step S2, the graph neural network performs multi-layer feature propagation through a graph convolutional network, updates node features, and fuses neighboring node information to optimize resource allocation strategies. Specifically, this includes:

[0088] A K-layer graph convolutional network is used to update node features layer by layer, so that the computing power, bandwidth utilization and storage resource information of each node are gradually propagated in the adjacency relationship, thereby enhancing the perception of the global network state.

[0089] By assigning different weights to adjacent nodes through an attention mechanism, the accuracy of resource allocation decisions is increased.

[0090] Optimized node embedding vectors are generated and used as input to reinforcement learning algorithms to improve resource allocation decision-making capabilities.

[0091] Specifically, in step S2, based on the augmented data generated in step S1, the network topology is further modeled, and the resource allocation strategy is optimized. Through graph neural networks, we can effectively capture the complex dependencies between network nodes, improving the accuracy and flexibility of the model in the network optimization process. This process provides strong support for subsequent dynamic resource scheduling and network traffic management. It is worth noting that step S2 directly relies on the augmented data generated in step S1, and the quality of this data determines the performance of the graph neural network in network topology modeling.

[0092] In this embodiment, the input to the graph neural network is enhanced data processed by a generative adversarial network. This data includes network node state information, link bandwidth, traffic requirements, etc., and is processed by the graph neural network to capture the connections and mutual influences between nodes. Specifically, in the structure of the graph neural network, each node represents a physical or logical node in the network (such as a switch, router, etc.), and edges represent connections between nodes. To improve the performance of the graph neural network, in some embodiments, the input features of each node in the network not only include the basic information of the node, but also combine the enhanced data generated in the previous step S1, which helps to make the network topology modeling more accurate.

[0093] In one possible implementation, each layer of the graph neural network updates the state information of its nodes through graph convolution operations. The update rule for the graph convolutional layer can be expressed as:

[0094] ;

[0095] in, Represents a node In the Characteristics of the layer; Adjacent nodes In the Characteristics of the layer; It is a node The set of adjacent nodes; It is the normalization coefficient; and These represent the weights and biases of the graph convolutional layer, respectively. For activation functions; Represents a node In the figure A graph neural network updates the features of its nodes through this process, enabling each node in the network topology to learn richer information from its neighboring nodes.

[0096] In graph neural network processing, the generated data is passed through convolutional layers to progressively update the features of the nodes. Each graph convolution operation takes into account the connections between nodes, which helps to capture complex network dependencies. In some embodiments, this process can further enhance the model's ability to learn long-distance dependencies through multiple layers of graph convolution.

[0097] Specifically, after multi-layer graph convolution calculations, the resulting node features will be used for subsequent resource allocation and optimization decisions. These features will influence the formulation of network resource scheduling strategies, thereby optimizing network flow control, bandwidth allocation, and other aspects. Alternatively, the output of the graph neural network can be used to construct a decision model for network resource scheduling. This model makes optimal decisions based on network state information, thereby improving the overall efficiency and resource utilization of the network.

[0098] It is important to emphasize that during the training process of a graph neural network, the parameters... and The model is updated using the backpropagation algorithm. By continuously optimizing these parameters, the model can better fit data in real-world network environments, improving its robustness and adaptability.

[0099] In step S3, deep reinforcement learning optimizes network resource allocation, specifically including:

[0100] The reward function is calculated based on network latency, throughput, packet loss rate, and energy efficiency performance indicators, and the weights are dynamically adjusted according to the current network status.

[0101] Combine deep Q-networks for offline training and leverage proximal policy optimization for online optimization to improve policy adjustment capabilities;

[0102] The state inputs for reinforcement learning include node computing power, storage utilization, bandwidth load, and topology information. The action space includes resource allocation strategy adjustments, and policy decisions are made using features generated by graph neural networks.

[0103] Specifically, in step S3, resource scheduling optimization is further performed based on the network topology model constructed in step S2. Through reinforcement learning algorithms, the resource allocation strategy in the network is dynamically adjusted to achieve optimal network resource utilization under different load conditions. The dynamic nature of the network environment makes static optimization methods difficult to adapt; therefore, a deep reinforcement learning-based approach is used to achieve adaptive optimization. The key to this step is utilizing a reinforcement learning agent to learn the optimal strategy through continuous interaction, thereby improving network performance.

[0104] In this embodiment, the state input of the reinforcement learning agent is provided by the graph neural network model in step S2. State information includes node traffic load, bandwidth usage, link availability, etc., which constitute the environmental state. In some embodiments, the agent at each time step It is necessary to consider the environmental conditions. Extract key features and apply them according to the current strategy. Choose an action This refers to adjusting the resource allocation of a certain link or optimizing the data flow routing.

[0105] In one possible implementation, the agent learns using a policy gradient-based method, and the loss function can be expressed as:

[0106] ;

[0107] in, Represents the intelligent agent at time step Execution strategy The trajectory experienced; For the agent in time step The rewards received; Indicates a final time step; These are the parameters of the policy network. Generally, the reward... It is calculated by comprehensively considering factors such as the network's overall throughput, link utilization, and latency.

[0108] Specifically, the policy network is implemented using a deep neural network (DNN) and employs parameterized functions. To approximate the optimal policy, the agent continuously samples state information from the environment during training, selects actions based on the policy network, and then updates parameters according to the reward function. In some embodiments, the agent employs an Actor-Critic architecture, where the Actor is responsible for selecting actions, and the Critic selects actions based on a value function. Evaluate the value of the current state to improve the stability of policy optimization.

[0109] As an option, the value function Recursively updated using the following Bellman equation:

[0110] ;

[0111] in, For time steps The state of time; This is the desired state; In time step The reward received at that time; The value function for the state at the next time step; This is a discount factor used to control the degree to which future rewards influence current decisions; a larger one... This makes the agent more focused on long-term gains, and less focused on smaller gains. This tends to favor short-term optimization.

[0112] During training, the agent optimizes the policy network using the policy gradient method, and the direction of gradient update is given by the following formula:

[0113] ;

[0114] in, The objective function is... For the strategy The generated trajectory To determine the expected value; for the trajectory, which typically consists of a series from the initial state to the final state; for the state... At that time, take action The conditional probability is given by the parameter The control strategy is given; For time step From 0 to Summation; To about The gradient; for time step The rewards received.

[0115] In some possible implementations, the agent can employ a proximal policy optimization algorithm, which improves training stability and convergence speed by limiting the magnitude of policy updates and preventing drastic changes in the policy network during training.

[0116] Specifically, the objective function for optimizing the near-end strategy is as follows:

[0117] ;

[0118] in, It is the objective function of the PPO algorithm; This indicates the expectation for a time step or sample. It is an operation that takes the minimum value; It is usually the probability ratio of the current policy to the old policy, used to measure the magnitude of the policy update; It is the advantage function, which represents the additional benefit of an action relative to the baseline value in the current state; Then the ratio is limited to Within the range, prevent excessive policy updates; It is a hyperparameter used to control the cropping range; These are the upper and lower limits for trimming.

[0119] In one possible implementation, the training data for the reinforcement learning model is generated from a simulation environment built on a real-world network topology and simulating different traffic patterns to improve the model's generalization ability. Once trained, the reinforcement learning agent can adaptively adjust network resources in the real-world environment, thereby optimizing network performance under varying load conditions.

[0120] In step S4, multi-objective optimization employs multi-task learning, specifically including:

[0121] By dynamically adjusting weights, the optimization objectives of latency, throughput, packet loss rate, and energy efficiency are balanced under different network conditions.

[0122] To resolve conflicts between optimization objectives, the Pareto optimization method is used to dynamically adjust the weights, and reinforcement learning is combined to adaptively adjust the weight coefficients for different optimization objectives.

[0123] Specifically, the core of step S4 lies in executing the optimized resource scheduling strategy and dynamically adjusting the overall system operation. This step directly relies on the strategy model obtained in step S3 and is validated in a real network environment. The optimization goals include not only efficient utilization of network resources but also adaptability under different network states. Due to the uncertainty of network states, relying solely on static allocation is insufficient to meet actual needs; therefore, real-time evaluation and adjustment of the scheduling strategy are required.

[0124] In this embodiment, the scheduling system allocates execution resources based on the output of the reinforcement learning model. Specifically, the system periodically samples the current network state. And through the trained policy network Choosing a suitable scheduling scheme. As one option, scheduling decisions are based not only on the current state but also on the historical state sequence. To enhance the ability to perceive long-term trends.

[0125] In some embodiments, the system evaluates the effectiveness of scheduling decisions and calculates the reward value of the current strategy. The reward value is calculated by weighting multiple indicators such as throughput, latency, and packet loss rate.

[0126] ;

[0127] in, Indicates network throughput; Indicates the average delay; Indicates packet loss rate; These are hyperparameters used to balance the impact of different performance metrics. Generally, adjusting these weighting parameters can adapt to different optimization objectives, such as low-latency applications or high-throughput scenarios.

[0128] In one possible implementation, the system uses a rolling window method to adjust the scheduling strategy online, with a window size of [missing information]. Calculate the cumulative reward at the end of each window:

[0129] ;

[0130] in, It is a time step from to Summation; It is a discount factor used to measure the weight of future rewards in current value; Indicates time step Rewards are based on the time difference Discounts should be offered proportionally. It is a time step The rewards received; As an overall profitability metric, it determines whether to adjust the strategy; when When the network falls below a set threshold, the system re-executes the policy update to adapt to the new network conditions.

[0131] Specifically, the system maintains an experience buffer pool that records past experiences. Round scheduling execution data In some embodiments, this data is used to incrementally train the policy model, employing stochastic gradient descent (SGD) to optimize the parameters:

[0132] ;

[0133] in,

[0134] The learning rate;

[0135] To optimize the objective function, reinforcement learning methods such as PPO and DQN can generally be used for parameter updates.

[0136] Indicates the parameters to be optimized;

[0137] This indicates an update operation, which assigns the value on the right to the value on the left.

[0138] Represents the objective function

[0139] Regarding parameters

[0140] The gradient.

[0141] The system can dynamically adjust its scheduling strategy in different application scenarios. For example, under high load, the system prioritizes low-latency routes, while under low load, it may adopt a balanced allocation strategy to improve resource utilization. Alternatively, multi-task learning can be incorporated to enhance the generalization ability of the scheduling strategy under different application requirements.

[0142] Step S5 further includes the following steps:

[0143] Global resource scheduling is optimized through graph optimization, and global resource allocation and coordination are achieved through local information transmission in a distributed environment.

[0144] Lagrange multipliers are used to optimize local resource allocation so that overall resource scheduling satisfies global constraints.

[0145] Local information includes, but is not limited to:

[0146] Computing resource status: CPU load, GPU load, memory usage, and storage resources;

[0147] Network resource status: bandwidth utilization, link load, packet loss rate, and network latency;

[0148] Task scheduling information: current task queue, task execution priority, and task execution status;

[0149] Energy consumption information: power consumption of a single computing node, thermal management, and battery power status;

[0150] Load balancing information: local node load status, available computing resources, and task migration information;

[0151] Global constraints include, but are not limited to, computing resource constraints, network resource constraints, latency constraints, energy consumption constraints, global optimization of task scheduling, and reliability and fault tolerance constraints.

[0152] Specifically, step S5 mainly undertakes the comprehensive analysis and subsequent optimization of the processing results from S1 to S4, aiming to improve the robustness and accuracy of the overall system. This step has clear designs in terms of data fusion, anomaly detection, and filtering optimization to ensure seamless connection of data flow between modules and achieve the expected technical effect.

[0153] In this embodiment, the implementation method of step S5 is described in detail below. Generally, step S5 includes preprocessing the output data of each module, outlier detection, data fusion, and signal filtering optimization. Alternatively, step S5 first performs outlier detection on the input data, using a mean and standard deviation judgment method, with the following formula:

[0154] ;

[0155] in, Indicates the first The original data of each module; This is data that has been corrected after anomaly detection; The mean of the data set; Standard deviation; A preset constant (e.g., 2 or 3) is used to control the tolerance range; It is the absolute difference between the data points and the mean.

[0156] Specifically, in one possible implementation, the preprocessed data is fused using a weighted average method, as shown in the following formula:

[0157] ;

[0158] in, This is the final output after fusion; Indicates the first Weighting coefficients for module data; The number of modules participating in the fusion processing; It is the sum of all weighted terms; It is the sum of all weights; Number of samples. Weighting coefficients. The calculation can be dynamically performed based on the signal-to-noise ratio or historical error of each module. The calculation formula is as follows:

[0159] ;

[0160] in, Indicates the first Variance of module data; To prevent division by zero of small positive constants.

[0161] In some embodiments, a frequency domain filtering module is added after data fusion to further reduce high-frequency noise interference. Alternatively, the filtering process can be described by the following formula:

[0162] ;

[0163] in,

[0164] This represents the filtered signal in the frequency domain.

[0165]

[0166] The spectrum of the fused signal;

[0167]

[0168] This is the designed filter function. Generally,

[0169] A low-pass filter can be used.

[0170] In one possible implementation, to further improve the system's adaptive performance, step S5 can also introduce a weight adaptive learning mechanism based on error backpropagation. Specifically, the update iterative formula for each weight coefficient is:

[0171] ;

[0172] in, For the first The weight in the th... The value at the next iteration; Indicates the first Module in Weights in the next iteration; The learning rate; For the objective function Weights The partial derivative of represents the degree to which changes in weights affect the loss.

[0173] Generally, each processing step in step S5 works closely with the data output and processing flow described in S1 to S4, ensuring data continuity and enhancing the overall stability and robustness of the system. Alternatively, in this embodiment, each step can be modularly expanded according to actual needs, such as adding data preprocessing algorithms, introducing multiple filter combinations, or employing other adaptive learning algorithms to adapt to the requirements of different application scenarios.

[0174] Step S6 further includes the following steps:

[0175] Self-supervised learning automatically learns network state features through unsupervised pre-training;

[0176] Meta-learning uses a small amount of sample data to quickly adjust model parameters, enabling optimization strategies to adapt to different network environments.

[0177] The sample data includes computing resource status, network status, task scheduling information, energy consumption data, and user demand data;

[0178] Model parameters include computational resource allocation weights, task priority adjustment parameters, reinforcement learning reward functions, and deep learning optimization parameters;

[0179] In step S6, the network state is analyzed in real time using a self-attention mechanism and a Transformer model, specifically as follows:

[0180] By improving policy computation efficiency through self-attention mechanisms, Q-value calculation and policy updates can be accelerated, thereby enhancing the real-time performance of network optimization.

[0181] Specifically, step S6 serves as the final decision-making correction and feedback control. It receives the fused data from step S5 and forms a closed loop with the previous processing results. This step performs final data verification, adjusts system parameters, and ensures overall accuracy and stability.

[0182] In this embodiment, step S6 analyzes the fused output. Generally, step S6 employs a combination of threshold judgment and error correction. Alternatively, step S6 utilizes the fusion result... and preset reference value The decision is based on the deviation between the two factors, and the formula for determining this deviation is as follows:

[0183] ;

[0184] in, Indicates the final decision output; The data after fusion in step S5; This is a reference value; This is the allowable deviation threshold.

[0185] Specifically, in one possible implementation, step S6 employs a dynamic threshold adjustment strategy. In some embodiments, the threshold parameter is updated using the following formula:

[0186] ;

[0187] in, For the first The threshold of time; For the first The threshold of time; The learning rate has a range of 0 < <10.

[0188] In another implementation, step S6 generates a feedback control signal to correct the front-end module. Specifically, the formula for calculating the feedback signal F is:

[0189] ;

[0190] in, For feedback control signals; This is the feedback gain coefficient, set according to system requirements.

[0191] In some embodiments, the least squares method is used to update parameters in order to further correct model errors.

[0192] In this case, the model parameter correction formula is:

[0193] ;

[0194] in, Represents the corrected parameter vector; To design a matrix containing the values ​​of all input features; For designing a matrix The transpose of the matrix; To design the inverse matrix of the matrix obtained by multiplying the transpose of the original matrix by the original matrix; Let be the error vector, which is defined as follows: ,in and Representing the first The actual output and expected output of the sample.

[0195] Generally, the data correction and feedback control in step S6 are closely integrated with the data processing in the aforementioned steps S1 to S5.

[0196] Alternatively, after receiving the fusion result, step S6 can also compare it with the output data of the preprocessing and feature extraction modules to verify the overall consistency of the system. In one possible implementation, fuzzy logic control is also introduced to adaptively adjust the threshold and feedback.

[0197] Specifically, the input to the fuzzy controller is and its rate of change The output is the feedback correction value. .

[0198] In addition, the design of fuzzy rules and membership functions should be determined based on system requirements, empirical values, or statistical analysis.

[0199] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-driven real-time network optimization algorithm, characterized in that, Includes the following steps: S1. Preprocess the network state data using a generative adversarial network, extract optimized network feature data, and generate high-quality augmented data. S2. Based on augmented data, a graph neural network is used to model the network topology, capture the dependencies between nodes, and update node features through a graph convolutional network. S3. Based on node features, a deep reinforcement learning algorithm is used to dynamically optimize and adjust computing, storage, and bandwidth resources, and adaptively allocate resources. S4. Input the resource allocation scheme into the multi-objective optimization method to further minimize latency and packet loss rate, maximize throughput, optimize energy efficiency, and finally output a global optimized resource scheduling strategy. S5. Combining the Lagrange relaxation method and graph optimization techniques, further optimize the global resource scheduling strategy, or the final scheduling scheme; S6. The final scheduling scheme improves the adaptability of the network model in dynamic environments through self-supervised learning and meta-learning, ensuring that the optimization system runs efficiently under different network conditions. In step S4, multi-objective optimization employs multi-task learning, specifically including: By dynamically adjusting weights, the optimization objectives of latency, throughput, packet loss rate, and energy efficiency are balanced under different network conditions. The conflicts between the optimization objectives are addressed by dynamically adjusting the weights using the Pareto optimization method, and by combining reinforcement learning to adaptively adjust the weight coefficients of different optimization objectives. Step S5 further includes the following steps: Global resource scheduling is optimized through graph optimization, and global resource allocation and coordination are achieved through local information transmission in a distributed environment. Lagrange multipliers are used to optimize local resource allocation so that overall resource scheduling satisfies global constraints. The local information includes, but is not limited to: Computing resource status: CPU load, GPU load, memory usage, and storage resources; Network resource status: bandwidth utilization, link load, packet loss rate, and network latency; Task scheduling information: current task queue, task execution priority, and task execution status; Energy consumption information: power consumption of a single computing node, thermal management, and battery power status; Load balancing information: local node load status, available computing resources, and task migration information; Global constraints include, but are not limited to, computing resource constraints, network resource constraints, latency constraints, energy consumption constraints, global optimization of task scheduling, and reliability and fault tolerance constraints.

2. The AI-driven real-time network optimization algorithm according to claim 1, characterized in that, In step S1, the generative adversarial network is trained using a conditional generative adversarial network, specifically including: The generator produces enhanced data based on network status inputs such as network traffic, node load, and bandwidth utilization. The discriminator evaluates the quality of the generated data and provides feedback to optimize the generator, thereby improving the authenticity and validity of the data.

3. The AI-driven real-time network optimization algorithm according to claim 1, characterized in that, In step S2, the graph neural network performs multi-layer feature propagation through a graph convolutional network, updates node features, and fuses neighboring node information to optimize resource allocation strategies. Specifically, this includes: A K-layer graph convolutional network is used to update node features layer by layer, so that the computing power, bandwidth utilization and storage resource information of each node are gradually propagated in the adjacency relationship, thereby enhancing the perception of the global network state. By assigning different weights to adjacent nodes through an attention mechanism, the accuracy of resource allocation decisions is increased. Optimized node embedding vectors are generated and used as input to a reinforcement learning algorithm to improve the decision-making ability for resource allocation.

4. The AI-driven real-time network optimization algorithm according to claim 1, characterized in that, In step S3, the deep reinforcement learning optimization of network resource allocation specifically includes: The reward function is calculated based on network latency, throughput, packet loss rate, and energy efficiency performance indicators, and the weights are dynamically adjusted according to the current network status. Combine deep Q-networks for offline training and leverage proximal policy optimization for online optimization to improve policy adjustment capabilities; The state inputs of the reinforcement learning include node computing power, storage utilization, bandwidth load and topology information, and the action space includes resource allocation strategy adjustment, using features generated by graph neural networks to make policy decisions.

5. The AI-driven real-time network optimization algorithm according to claim 1, characterized in that, Step S6 further includes the following steps: Self-supervised learning automatically learns network state features through unsupervised pre-training; Meta-learning uses a small amount of sample data to quickly adjust model parameters, enabling optimization strategies to adapt to different network environments. The sample data includes computing resource status, network status, task scheduling information, energy consumption data, and user demand data. The model parameters include computational resource allocation weights, task priority adjustment parameters, reinforcement learning reward functions, and deep learning optimization parameters.

6. The AI-driven real-time network optimization algorithm according to claim 1, characterized in that, In step S6, the network state is analyzed in real time using a self-attention mechanism and a Transformer model, specifically as follows: By improving policy computation efficiency through self-attention mechanisms, Q-value calculation and policy updates can be accelerated, thereby enhancing the real-time performance of network optimization.

7. The AI-driven real-time network optimization algorithm according to claim 1, characterized in that, The algorithm has a fault-tolerant mechanism that dynamically adjusts itself in the event of network failure, specifically including: By monitoring bandwidth utilization, response time, and packet loss rate through fault detection, faulty nodes can be identified. Automatically perform load balancing to ensure uninterrupted task execution, and select the optimal node to take over based on resource availability; Dynamic resource reallocation optimizes bandwidth, computing resources, and storage space allocation, reducing the impact of faults on network performance.

8. The AI-driven real-time network optimization algorithm according to claim 7, characterized in that, The fault tolerance mechanism further includes a preventative fault tolerance strategy based on predictive analytics, specifically including: By analyzing historical data through machine learning models, potential failure points are predicted, and the probability of failure is calculated based on long-term trends. Adjust resource allocation in advance within a 5-10 time window before a failure occurs to reduce the impact of future failures; Once the faulty node recovers, resource allocation is adaptively adjusted to optimize global resource utilization, and historical data is used to update the prediction model to improve the accuracy of future predictions.

Images