Agent-based wireless network optimization methods, apparatuses, electronic devices, and media

Abstract

This application relates to the field of communications, providing a method, apparatus, electronic device, and medium for optimizing wireless networks based on intelligent agents. The method includes: acquiring the root cause localization results of a target cell; inputting the root cause localization results into a decision model of an intelligent agent to obtain a parameter adjustment scheme for the target cell's wireless network; optimizing the target cell's wireless network based on the parameter adjustment scheme; the decision model is obtained through reinforcement learning based on samples of root cause localization results from sample cells and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the sample cell's next-level region; the local reward value is calculated based on the dynamic weights of various performance indicators. The method provided in this application can dynamically adapt to complex optimization objectives, effectively avoid local optimization from damaging the overall network performance, and significantly improve the intelligent decision-making level and overall performance of the wireless network.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method, apparatus, electronic device and medium for optimizing wireless networks based on intelligent agents. Background Technology

[0002] With the increasing complexity of mobile communication networks, intelligent agents based on artificial intelligence technologies such as reinforcement learning have been widely used in the parameter tuning and performance optimization of wireless networks. Under existing intelligent agent training and decision-making frameworks, fixed reward functions are typically used to guide agent learning, resulting in poor flexibility and adaptability, making it difficult to adapt to dynamically changing and complex application scenarios. Existing reward calculations are mostly limited to a single-cell perspective or simple global aggregation, with a disconnect between local and global feedback mechanisms, making it difficult to achieve an effective balance between ensuring local user experience and improving overall network performance. Summary of the Invention

[0003] This application provides a wireless network optimization method, apparatus, electronic device, and medium based on intelligent agents to solve the technical problems of existing wireless network optimization methods having poor flexibility and adaptability, making it difficult to adapt to complex application scenarios with dynamic changes, and making it difficult to achieve an effective balance between ensuring local user experience and improving the overall network performance.

[0004] In a first aspect, embodiments of this application provide a wireless network optimization method based on intelligent agents, including: Obtain the root cause location results of the wireless network in the target cell; the root cause location results are determined based on the performance index data of the wireless network in the target cell. The fault root cause localization result is input into the decision model of the intelligent agent to obtain the parameter adjustment scheme of the wireless network of the target cell output by the decision model; Based on the parameter adjustment scheme, the wireless network of the target cell is optimized; The decision model is obtained by the agent through reinforcement learning based on the root cause localization results of the wireless network of the sample cell and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the next-level area of ​​the sample cell; the local reward value is calculated based on the dynamic weights of each performance index.

[0005] In some embodiments, the decision-making model of the agent is trained based on the following steps: Obtain sample fault work order descriptions and fault root cause location results for the wireless network of the sample cell. The fault root cause localization result sample is input into the initial decision model of the agent to obtain the parameter adjustment prediction scheme of the wireless network of the sample cell output by the initial decision model; After the parameter adjustment prediction scheme is executed, the scheme execution result sample of the sample cell is obtained; The fault work order description sample, the fault root cause location result sample, and the solution execution result sample are input into the reward model. The reward model determines the dynamic weight of each performance indicator based on the fault work order description sample, the fault root cause location result sample, and the solution execution result sample. Based on the dynamic weight, the local reward value is calculated, and the regional reward value is calculated. The local reward value and the regional reward value are merged, and the total reward value and / or the status signal of the sample cell are output. Based on the total reward value, the parameters of the initial decision model are iteratively optimized to obtain the decision model.

[0006] In some embodiments, determining the dynamic weights of each performance metric includes: Based on the fault work order description sample, the fault root cause location result sample, and the scheme execution result sample, the network optimization target of the sample cell is determined. User intent is identified based on the network optimization objective, and each performance indicator and its dynamic weight are determined based on the user intent.

[0007] In some embodiments, the fault work order description sample includes data on each of the performance indicators of the sample cell, the scheme execution result sample includes new data on each of the performance indicators of the sample cell, and the calculation of the local reward value based on the dynamic weight includes: Based on the data of each performance indicator and the new data of each performance indicator, calculate the change value of each performance indicator; The local reward value of the sample cell is calculated based on the changes in each of the performance indicators and the dynamic weights.

[0008] In some embodiments, calculating the local reward value of the sample cell based on the changes in each of the performance metrics and the dynamic weights includes: Based on the historical data of each performance indicator of the sample cells, determine the dynamic threshold of each performance indicator; The local reward value of the sample cell is calculated based on the dynamic threshold of each performance indicator, the change value of each performance indicator, and the dynamic weight.

[0009] In some embodiments, calculating the local reward value of the sample cell based on the changes in each of the performance metrics and the dynamic weights includes: The changes in each of the aforementioned performance indicators are normalized to obtain the normalized changes in each of the aforementioned performance indicators. The local reward value of the sample cell is calculated based on the normalized change values ​​of each performance indicator and the dynamic weight.

[0010] In some embodiments, calculating the regional reward value includes: Determine the association relationship between the sample cell and each of its neighboring cells; Using the sample cell and each of the neighboring cells as nodes, and the relationships between the sample cell and each of the neighboring cells as edges, an initial network topology graph of the next-level region of the sample cell is constructed. The first state feature data and semantic embedding vector of the network optimization target of the sample cell are obtained, and the second state feature data of each of the neighboring cells are obtained; the semantic embedding vector is extracted by the reward model based on the fault work order description sample, the fault root cause location result sample and the scheme execution result sample; The first state feature data, the second state feature data, and the semantic embedding vector of the network optimization objective are injected into the initial network topology graph to obtain the network topology graph; The network topology is input into a graph neural network model to obtain the regional reward value corresponding to the next-level region of the sample cell, as output by the graph neural network model.

[0011] In some embodiments, the graph neural network model is trained based on network topology samples of the upper-level region of the sample cell and the regional reward value labels of the upper-level region of the sample cell.

[0012] In some embodiments, fusing the local reward value and the regional reward value includes: The local reward value is normalized to obtain a normalized local reward value, and the regional reward value is normalized to obtain a normalized regional reward value. Determine the local weight of the normalized local reward value and the regional weight of the normalized regional reward value; Based on the local weight and the regional weight, the normalized local reward value and the normalized regional reward value are weighted and summed to obtain the total reward value.

[0013] Secondly, embodiments of this application provide a wireless network optimization device based on an intelligent agent, comprising: The acquisition unit is used to acquire the root cause location result of the wireless network of the target cell; the root cause location result is determined based on the performance index data of the wireless network of the target cell. The decision unit is used to input the fault root cause localization result into the decision model of the agent to obtain the parameter adjustment scheme of the wireless network of the target cell output by the decision model; A network optimization unit is used to optimize the wireless network of the target cell based on the parameter adjustment scheme. The decision model is obtained by the agent through reinforcement learning based on the root cause localization results of the wireless network of the sample cell and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the next-level area of ​​the sample cell; the local reward value is calculated based on the dynamic weights of each performance index.

[0014] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the agent-based wireless network optimization method described in the first or second aspect.

[0015] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the agent-based wireless network optimization method described in the first or second aspect.

[0016] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the agent-based wireless network optimization method described in the first or second aspect.

[0017] The agent-based wireless network optimization method, apparatus, electronic device, and medium provided in this application achieve automated and intelligent optimization of the wireless network by acquiring the root cause localization results of the target cell and outputting a parameter adjustment scheme for the wireless network using an agent decision model trained based on reinforcement learning. During the training phase of the agent decision model, this application introduces a comprehensive feedback mechanism. The total reward value is calculated by fusing the local reward value of the sample cell with the regional reward value of the previous level region, and the local reward value is generated based on the dynamic weights of various performance indicators. This local evaluation method based on dynamic weights can dynamically adapt to complex optimization objectives, overcoming the rigidity problem of traditional methods relying on manually preset fixed weights. Simultaneously, by combining regional reward values, this scheme fully considers the topological linkage effect between local units and the global network, giving the agent a global perspective, thereby effectively avoiding the phenomenon of local optimization damaging the overall network performance. In summary, the method provided in this application offers a more scientific and refined guidance for agent strategy iteration, significantly improving the intelligent decision-making level and overall performance of the wireless network. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the agent-based wireless network optimization method provided in an embodiment of this application. Figure 2 This is a flowchart illustrating the training process of the decision-making model of the intelligent agent provided in the embodiments of this application; Figure 3 This is a schematic diagram of the local reward value provided in the embodiments of this application; Figure 4 This is a schematic diagram of the process for calculating the regional reward value provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of the agent-based wireless network optimization device provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Currently, existing technologies have incorporated artificial intelligence into network resource scheduling or multi-agent cooperative optimization, but they still have significant limitations in terms of reward evaluation mechanisms, specifically as follows: 1) Lack of dynamic adaptation capability for optimization objectives: Existing solutions often rely on pre-defined fixed indicator weights or static reward functions, making it difficult to flexibly address optimization needs in different scenarios. For example, linearly combining indicators such as throughput and latency with fixed weights lacks in-depth analysis of business intent. When network scenarios or optimization goals change, the weight ratio of each indicator in the reward cannot be dynamically adjusted, often requiring significant manual intervention and frequent parameter tuning to rebalance the contributions of each indicator.

[0022] 2) The reward assessment did not fully integrate network topology relationships: Existing reward calculations are mostly limited to single-point or global perspectives, lacking refined modeling of regional node correlations and topological structures. Taking multi-agent cooperative optimization as an example, its reward criteria are mostly based on independent local indicators or simple global summaries, ignoring interference and cooperative effects between neighboring base stations. This fragmented evaluation mechanism results in reward signals that cannot truly reflect regional performance changes.

[0023] 3) The threshold assessment mechanism is rigid and lacks trend awareness: Existing technologies often use fixed thresholds or static benchmarks to determine the quality of metrics, such as setting a fixed lower limit for throughput, lacking the ability to adaptively adjust based on network load or historical trends. In scenarios with gradual changes in traffic or a continuous decline in experience, fixed thresholds can lead to delayed problem detection; while in the face of brief abnormal fluctuations, they are prone to overreacting beyond the threshold. This rigid mechanism, lacking trend awareness, makes it difficult for reward scoring to smoothly and accurately depict the dynamic evolution of network performance.

[0024] 4) Lack of a unified multi-indicator reward normalization mechanism: Existing solutions often lack effective dimensional alignment when handling multi-metric rewards, and directly combining metrics with different dimensions in a linear fashion can easily lead to evaluation imbalance. For example, when the absolute value of throughput is much greater than latency, the total reward will be dominated by the metric with the larger value, causing the optimization direction to deviate significantly from the expected balance. This lack of a normalization mechanism results in drastic fluctuations in reward values ​​under different network scenarios, severely affecting the convergence stability of reinforcement learning algorithms.

[0025] 5) The local and global feedback mechanisms are disconnected and lack effective integration: Existing reward systems typically only apply to a single level, failing to comprehensively consider both local benefits and overall regional / global performance. Focusing solely on improvements in individual cell metrics can lead to inefficient global resource coordination; conversely, rewarding only based on overall network metrics may mask severe degradation issues at local nodes. Due to the lack of a fusion feedback mechanism for local and regional rewards, current technologies struggle to achieve a good balance between ensuring a positive local user experience and improving overall network performance.

[0026] 6) The reward output format is too simplistic and fails to meet the needs of multiple application scenarios: Existing reinforcement learning optimization schemes typically only output single-dimensional numerical rewards for algorithm iteration, lacking a differentiated output mechanism. For example, the numerical scores obtained by the agent cannot provide operators with intuitive network status or fault alarms. This single output mechanism makes it difficult for a single system to simultaneously handle low-level policy training and high-level real-time operation and maintenance management, greatly limiting its engineering practical value in live network deployments.

[0027] In summary, the aforementioned deficiencies severely restrict the adaptive capability and multi-objective collaborative effect of intelligent optimization algorithms for wireless networks in complex scenarios. Therefore, this application provides an agent-based wireless network optimization method, apparatus, electronic device, and medium to address the aforementioned technical problems.

[0028] Figure 1 This is a flowchart illustrating the agent-based wireless network optimization method provided in an embodiment of this application. (Refer to...) Figure 1 This application provides an agent-based wireless network optimization method, which may include: Step 110: Obtain the root cause location results of the wireless network in the target cell; the root cause location results are determined based on the performance index data of the wireless network in the target cell.

[0029] Optionally, the root cause location results of the target cell can be obtained by signaling interaction with the wireless network management system; historical diagnostic logs can be read in batches from local storage devices or cloud databases to obtain the root cause location results of the target cell; or fault work orders of the target cell can be obtained, and the fault work orders can be analyzed to determine the root cause location results.

[0030] The target cell refers to a specific physical base station or logical cell entity in a particular area of ​​the current wireless network that has performance bottlenecks, connection failures, or urgently needs parameter optimization. The root cause analysis results refer to the specific causes derived from in-depth analysis of the network anomalies in the target cell, such as weak coverage, excessive interference from neighboring frequency bands, and unreasonable inter-site spacing settings, as well as the underlying causes of these network problems.

[0031] Step 120: Input the root cause location results into the decision model of the intelligent agent to obtain the parameter adjustment scheme of the target cell's wireless network output by the decision model.

[0032] In this context, an intelligent agent refers to a computational entity with autonomous learning and sequential decision-making capabilities, particularly suitable for automated control in complex wireless network environments with vast state spaces. The decision model is the core algorithmic architecture within the intelligent agent, which can contain multiple layers of neural networks used to map the input network states and target intent features into a specific execution action space.

[0033] After the root cause analysis results are input into the decision model, the model performs forward propagation calculations to generate a set of specific control command vectors or discrete action codes. The parameter adjustment scheme for the target cell's wireless network specifically refers to the implementation strategy for modifying relevant network parameters of the target cell, such as adjusting the maximum transmit power from -10 to -5, or adjusting the antenna electronic downtilt angle, or the handover success rate threshold.

[0034] Step 130: Optimize the wireless network of the target cell based on the parameter adjustment scheme.

[0035] The optimization method can be to directly send the above parameter adjustment scheme to the base station equipment corresponding to the target cell through the network controller and force it to take effect, thereby changing the base station's transmit power, frequency resource allocation, or service connection threshold, thereby improving the service communication quality and user experience of the target cell from the physical layer or network layer.

[0036] The decision model is obtained by the agent through reinforcement learning based on the root cause localization results of the wireless network of the sample cell and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the next-level area of ​​the sample cell; the local reward value is calculated based on the dynamic weights of each performance index.

[0037] It should be noted that reinforcement learning refers to a learning mechanism in which an agent continuously interacts with the network environment and updates its policy parameters based on reward signals from the environment. A sample cell refers to a historical cell used in the offline training or online exploration phase of the decision model, or a virtual cell constructed in a simulation platform for simulated training. The root cause localization result sample is the set of historical fault cause data corresponding to the sample cell.

[0038] The total reward value is a comprehensive feedback score used to guide the decision-making model in updating parameters towards maximizing long-term returns during reinforcement learning training. To balance the communication service quality of individual nodes with the balanced utilization of network resources over a wider area, the total reward value is generated using a fusion approach. This fusion approach can be a linear weighted sum of reward components from multiple spatial levels or a combination of nonlinear strategies.

[0039] The local reward value refers to a fine-grained evaluation score given for the immediate performance improvement of a single cell. The next-level region refers to the local network topology space consisting of the sample cell itself and its neighboring cells within a one- or multi-hop range. The regional reward value is the comprehensive performance evaluation score for the entire local network, used to truly reflect the chain effects of the agent's parameter adjustment actions on neighboring cells and even the overall network topology, such as interference propagation and cooperative gains.

[0040] Performance metrics refer to key business data that measure the communication status of a wireless network, such as throughput, latency, packet loss rate, and wireless disconnection rate. Dynamic weights are weighting coefficients that adaptively change based on current real-time business objectives and network load scenarios. These dynamic weights can be used to weight and sum the absolute values ​​or trend changes of the improvements in various performance metrics to obtain a reasonable feedback score that fits the current scenario, i.e., the local reward value.

[0041] In this embodiment, by acquiring the root cause localization results of the target cell and using an agent decision-making model trained based on reinforcement learning to output a parameter adjustment scheme for the wireless network, closed-loop automated and intelligent optimization of the wireless network is achieved. In the underlying training mechanism of the agent decision-making model, a comprehensive feedback mechanism is introduced. Its total reward value is calculated by fusing the local reward value of the sample cell with the regional reward value of the previous-level region, and the local reward value is generated based on the dynamic weights of various performance indicators. This local evaluation method based on dynamic weights can dynamically adapt to the constantly evolving complex optimization objectives, completely overcoming the rigidity problem caused by the traditional scheme's heavy reliance on manually preset fixed weights. Simultaneously, by combining regional reward values, this scheme fully considers the topological linkage effect between local units and the global network in the reinforcement learning closed loop, giving the agent a global collaborative vision. This effectively avoids the phenomenon of blindly pursuing local optima and damaging the overall network performance, significantly improving the intelligent decision-making level and resource utilization efficiency of the wireless self-optimizing network.

[0042] Figure 2 This is a flowchart illustrating the training process of the decision-making model for an intelligent agent provided in an embodiment of this application. (Refer to...) Figure 2 In some embodiments, the agent's decision-making model is trained based on the following steps: Step 210: Obtain the fault work order description sample and fault root cause location result sample of the wireless network of the sample cell.

[0043] Optionally, fault ticket description samples and fault root cause location result samples can be obtained by batch retrieval from the network operation and maintenance management system or by manual input. Fault ticket description samples typically refer to network optimization goals and business requirements described by network operation and maintenance personnel or users in natural language, such as requesting a reduction in inter-site spacing to address weak coverage issues, or explicitly requiring the guarantee of video stream service quality in specific areas, etc., which are semantic texts with high-level intent.

[0044] Step 220: Input the fault root cause localization result sample into the initial decision model of the intelligent agent to obtain the parameter adjustment prediction scheme of the wireless network of the sample cell output by the initial decision model.

[0045] The initial decision model refers to a decision model whose internal neural network structure has been built but whose parameters have not yet undergone sufficient reinforcement learning iterations, or which is in the process of exploratory training. The parameter adjustment prediction scheme refers to the adjustment actions of the initial decision model in the current training step, based on the current state's exploratory output, to attempt to improve the network environment.

[0046] Step 230: After the parameter adjustment prediction scheme is executed, obtain the scheme execution result sample of the sample cell.

[0047] When a parameter adjustment prediction scheme is applied in a real wireless network or digital twin simulation platform, the operating state of the network environment will change. The scheme execution result sample refers to the actual feedback data generated by various performance indicators of the sample cell within a specific time window after the prediction scheme is implemented. Specifically, it can include the specific values ​​of the performance indicators at each time point before and after parameter adjustment.

[0048] Step 240: Input the fault work order description sample, fault root cause location result sample, and solution execution result sample into the reward model. The reward model determines the dynamic weight of each performance indicator based on the fault work order description sample, fault root cause location result sample, and solution execution result sample. It calculates the local reward value based on the dynamic weight and calculates the regional reward value. It merges the local reward value and the regional reward value and outputs the total reward value and / or the status signal of the sample cell.

[0049] The reward model, within the reinforcement learning closed-loop framework, is a comprehensive algorithm architecture specifically responsible for evaluating the merits of an agent's exploratory actions and generating corresponding feedback signals. The reward model can analyze the business focus contained in the fault work order description samples, and, combined with the current root cause of the fault, automatically assign reasonable weight ratios to different performance indicators such as throughput, latency, and packet loss rate. The total reward value is typically represented as a continuous real-valued scalar, providing smooth gradient feedback containing amplitude information on the quality of the action. The status signal of the sample cell is represented as a discrete, categorical, or hierarchical scoring indicator, such as a binary threshold for determining whether a solution passes or fails, or an indicator signal used to indicate normal, warning, or alarm operational statuses.

[0050] Step 250: Based on the total reward value, iteratively optimize the parameters of the initial decision model to obtain the decision model.

[0051] The agent uses reinforcement learning algorithms, taking the received continuous total reward values ​​as the optimization objective. Through algorithms such as backpropagation, it continuously updates the neural network weight coefficients within the initial decision model, making it inclined to choose actions with higher expected total reward values ​​in future states. After a large amount of environmental interaction and multiple rounds of trial and error training, the model parameters gradually stabilize and converge, ultimately yielding a decision model that can adapt to complex network environments and has excellent decision-making performance.

[0052] In this embodiment, by introducing a reward model that integrates multi-dimensional information to guide the closed-loop training of the agent's decision-making model, high-level semantic intent can be accurately captured and dynamically quantified into evaluation metrics for various performance indicators, making the agent's learning process highly aligned with actual business objectives. Simultaneously, the reward model flexibly outputs continuous total reward values ​​and discrete sample cell state signals. On the one hand, it utilizes the smooth gradient information of the total reward value to ensure the stability and convergence speed of the reinforcement learning algorithm during training in complex state spaces; on the other hand, it uses the state signals to provide a rigorous quality control and security interception mechanism for practical applications. This training framework, which decouples evaluation from learning, not only improves the efficiency of policy iteration but also greatly enhances the reliability and controllability of the optimization model during deployment in the live network.

[0053] In some embodiments, determining the dynamic weights of each performance metric includes: Based on the fault work order description samples, fault root cause location result samples, and solution execution result samples, the network optimization target of the sample cell is determined. User intent is identified based on network optimization objectives, and performance metrics and their dynamic weights are determined based on user intent.

[0054] Network optimization goals refer to the desired improvement or business expectation for addressing specific problems currently existing in the network. Semantic parsing and contextual information extraction can be performed on natural language text samples such as fault work order descriptions and root cause analysis results to obtain network optimization goals. User intent refers to the core business concerns and constraints hidden behind the textual description of network optimization goals.

[0055] Optionally, performance metrics are determined based on user intent, such as those urgently needing improvement, those requiring strict protection, and those that can be sacrificed for overall performance enhancement. Dynamic weights are then assigned to different performance metrics in real time based on the parsed user intent. This dynamic weighting can be determined by incorporating knowledge from communications experts.

[0056] For example, when a fault ticket description sample mentions the need to ensure the quality of video streaming services in a specific area, the reward model can accurately identify that the user's intention has strict requirements for network latency and jitter, thereby significantly increasing the weight of latency-related performance indicators in the final generated dynamic weights, while relatively reducing the weight of other non-critical indicators.

[0057] In this embodiment, by performing deep semantic analysis on the network optimization target and accurately identifying user intent, dynamic weights that are highly consistent with it are generated. This not only significantly reduces the network performance loss and operation and maintenance costs caused by the lag in manual parameter tuning, but also ensures that the optimization exploration direction of the intelligent agent is always highly aligned with the ever-changing actual business objectives. This significantly enhances the versatility and intelligence of the reward evaluation system for complex and ever-changing network scenarios.

[0058] In some embodiments, the fault work order description sample includes data on various performance indicators of the sample cell, and the solution execution result sample includes new data on various performance indicators of the sample cell. The local reward value is calculated based on dynamic weights, including: Based on the data and new data for each performance indicator, calculate the change value of each performance indicator; The local reward value of the sample cell is calculated based on the changes in each performance indicator and its dynamic weight.

[0059] In this step, the data for each performance metric typically refers to the initial baseline values ​​of the sample cell before the parameter adjustment and prediction scheme is implemented, such as throughput, latency, and packet loss rate. The new data for each performance metric refers to the new service operation values ​​exhibited by the sample cell after the scheme is implemented, or even after a specific period of time, such as half an hour later.

[0060] Optionally, the specific differences in performance indicators between different time points can be obtained by performing difference calculations between the new data and the existing data for each performance indicator. Dynamic weights can be used to perform a weighted summation of the changes in the aforementioned performance indicators or the trend rewards derived from these changes to obtain the local reward value for the sample cell.

[0061] In this embodiment, by incorporating the changes in performance metrics before and after execution into the calculation of the local reward value, the evaluation mechanism not only focuses on the static values ​​of network performance but also keenly captures the long-term evolution direction and dynamic trends brought about by optimization actions. This trend-aware evaluation method can provide the agent with more forward-looking reward guidance, enabling the agent to receive positive incentives due to the continuous improvement in performance even when the performance has not yet reached its ideal state. This effectively guides the reinforcement learning algorithm to continuously adjust network parameters in the correct optimization direction.

[0062] In some embodiments, the local reward value of the sample cell is calculated based on the changes in each performance indicator and its dynamic weight, including: Based on historical data of various performance indicators of the sample cells, the dynamic thresholds of each performance indicator are determined. The local reward value of the sample cell is calculated based on the dynamic thresholds of each performance indicator, as well as the changes and dynamic weights of each performance indicator.

[0063] In this step, the dynamic threshold refers to the classification boundary standard generated by the system through a data-driven approach to judge whether the network performance is excellent, good, or poor.

[0064] Optionally, selected statistical or machine learning algorithms can be applied to mine the patterns of natural distribution characteristics of a large amount of historical data. For example, unsupervised clustering algorithms, such as the K-means algorithm, can be used to divide historical data into high, medium, and low performance clusters and determine the points between the cluster center values ​​as boundaries. Alternatively, a mixture Gaussian distribution can be fitted to the historical values ​​of the indicators and boundary values ​​can be determined based on the inflection points of the distribution. The percentile method can also be used to set specific quantiles, such as 90 percent, as the excellent threshold.

[0065] Optionally, these dynamically updated thresholds, which are updated over time as the network environment evolves, can be used to assess the degree of improvement in the changes of various performance indicators and obtain the local reward value for the sample cell. For example, in areas with poor overall performance, a small performance improvement may be sufficient to exceed a lower dynamic threshold and qualify for a reward, while in high-performance areas, a higher dynamic threshold is required to be deemed an effective improvement by the system. Based on this, the final reward score is calculated by combining dynamic weights.

[0066] Figure 3 A schematic diagram of the local reward value provided for an embodiment of this application. (Refer to...) Figure 3 The local reward value diagram provided in this application embodiment intuitively illustrates the dynamic changes of various performance indicators of the sample cell and the comprehensive evaluation results of the local reward value after the agent executes the parameter adjustment scheme.

[0067] exist Figure 3 The top section prominently displays the core quantitative data for evaluating optimization, including the overall degree of improvement, numerical reward signals, and scoring signals used to determine whether a solution is approved. Meanwhile, Figure 3 The document details different evaluation dimensions, including performance metrics where lower values ​​are better, such as wireless call drop rate, flow call drop rate, and connection reconstruction ratio, as well as performance metrics where higher values ​​are better, such as handover success rate, wireless connection success rate, QoS flow establishment success rate, and Evolved Packet System fallback success rate.

[0068] For each specific performance metric, Figure 3The document provides a detailed comparison of specific business metrics before and after optimization, listing both fixed upper and lower limits set by traditional rules, and dynamic thresholds automatically generated by the system based on historical performance data statistical learning—referred to as AI upper or lower limits in the diagram. By comparing the actual data of each performance indicator before and after optimization with the system-generated dynamic thresholds, the improvement or deterioration of each indicator can be assessed in a fine-grained manner and presented with intuitive status indicators.

[0069] Based on the changes in these multi-dimensional performance indicators and the dynamic weights assigned to them by the system, the reward model performs weighted summation and normalization, and finally calculates the overall local reward evaluation result. Figure 3 The numerical reward signal shown is primarily used to provide continuous gradient updates for the reinforcement learning network within the agent, enabling the agent to perceive the quality of its actions. The scoring signal, on the other hand, serves as a strict quality control threshold, used from a global perspective to determine whether the current parameter adjustment scheme is sufficiently safe and practically deployable.

[0070] In this embodiment, a data-driven automatic scale adjustment closed loop is constructed by using historical data for statistical learning to automatically generate dynamic thresholds. This adaptive dynamic threshold mechanism can automatically calibrate the evaluation benchmark according to the actual load conditions of different network regions and time periods. It can avoid overreaction caused by short-term network fluctuations and prevent the problem exposure lag when the network gradually deteriorates, thereby ensuring the fairness and high environmental adaptability of the reward evaluation standard, and significantly reducing the cost of manual maintenance and the risk of human error.

[0071] In some embodiments, the local reward value of the sample cell is calculated based on the changes in each performance indicator and its dynamic weight, including: The changes in each performance index are normalized to obtain the normalized changes in each performance index. The local reward value of the sample cell is calculated based on the normalized change value and dynamic weight of each performance index.

[0072] In this step, normalization refers to scaling the numerical values ​​of performance changes across different dimensions using a specific mathematical transformation function. The changes in each performance indicator or the weighted trend score can be non-linearly mapped using a sigmoid function with monotonic properties, such as the sigmoid function. This normalized change value of each performance indicator is then uniformly compressed and confined to a specific target interval between zero and one.

[0073] In this embodiment, by introducing a unified normalization processing mechanism to standardize the scale of the change values ​​of different performance indicators, the dimensional differences between different network indicators are effectively eliminated, ensuring the fairness and consistency of the joint evaluation results of multiple indicators, and greatly improving the convergence stability and policy robustness of the agent reinforcement learning training process.

[0074] Figure 4 This is a schematic diagram illustrating the process of calculating the regional reward value provided in an embodiment of this application. (Refer to...) Figure 4 In some embodiments, calculating the regional reward value includes: Step 410: Determine the relationship between the sample cell and its neighboring cells.

[0075] Neighboring cells refer to cells that are physically or logically located within one or more hops of the sample cell. Association refers to the objective mutual influence between cells, such as geographical adjacency or frequency interference. This can be determined by reading the neighboring cell list from the wireless network configuration management database or by extracting an interference matrix from a measurement report.

[0076] Step 420: Using the sample cell and each neighboring cell as nodes, and the relationships between the sample cell and each neighboring cell as edges, construct the initial network topology of the next-level region of the sample cell.

[0077] In this context, a node is a basic entity unit in the network topology diagram, corresponding to an independent cell in the wireless network. An edge is a connection line linking the nodes, representing the relationships between cells.

[0078] Step 430: Obtain the first state feature data of the sample cell and the semantic embedding vector of the network optimization target, and obtain the second state feature data of each neighbor cell; the semantic embedding vector is extracted by the reward model based on the fault work order description sample, the fault root cause location result sample and the scheme execution result sample.

[0079] Specifically, the first and second state feature data refer to the numerical values ​​of specific indicators representing the network operating status, such as the throughput, latency, and recent performance changes of the sample cell and its neighboring cells at the current moment. Semantic embedding vectors are vector representations of the meaning of high-level intentions and constraints in the semantic space, generated by the large language model after encoding business requirements in natural language form. Acquisition methods include real-time collection of state data from the network performance monitoring platform and direct reception of the encoded tensors output by the large language model.

[0080] Step 440: Inject the first state feature data, the second state feature data, and the semantic embedding vector of the network optimization objective into the initial network topology graph to obtain the network topology graph.

[0081] Optionally, the first and second state feature data can be assigned to the corresponding cell nodes as local feature vectors, while the semantic embedding vector of the network optimization objective can be incorporated into the graph structure as a global feature. This semantic embedding vector can be concatenated to the feature vector of each graph node, or it can be used as a global appended node connected to all cell nodes. The resulting network topology graph is a complete input data structure that integrates the network physical state and high-level semantic intent.

[0082] Step 450: Input the network topology map into the graph neural network model to obtain the regional reward value corresponding to the next-level region of the sample cell output by the graph neural network model.

[0083] The graph neural network model can be a deep learning architecture specifically designed for processing graph-structured data, such as a graph convolutional network or a graph attention network. The input processing involves each cell node receiving information from its one- or two-hop neighbors in each layer of the graph neural network model's message passing mechanism. This information is then weighted and aggregated based on edge weights or interference levels to update its own hidden representation. Through multi-layer network stacking, the node representations gradually propagate and encompass a wider range of structural influences. Finally, the output layer of the graph neural network model evaluates and extracts data from the entire subgraph, outputting a region reward value representing the degree of overall network performance improvement in that local area.

[0084] In this embodiment, a graph neural network is introduced to model the topology of a wireless cell. Multi-layer message passing within the graph neural network captures the cascading effects and interference propagation across cells, ensuring that the agent can spontaneously balance local and global interests when making decisions. This effectively avoids the problem of drastic performance degradation in neighboring cells due to excessive pursuit of single-cell optimization. Simultaneously, the semantic embedding vectors extracted from a large model are innovatively injected into the topology graph as global features. This allows the same graph neural network to dynamically adjust its focus on various network states based on different high-level service intentions, achieving on-demand customization of regional evaluation standards and significantly improving the accuracy of collaborative optimization of complex networks and the agent's multi-objective adaptability.

[0085] In some embodiments, the graph neural network model is trained based on the network topology graph samples of the upper-level region of the sample cell and the regional reward value labels of the upper-level region of the sample cell.

[0086] Among them, network topology graph samples refer to graph structure input data that includes physical nodes, logical edges, historical state characteristics, and historical parameter change information, generated using historical real network change data or digital twin simulation platforms.

[0087] In the specific training process, a large number of collected network topology graph samples are input into the initial graph neural network to be trained in batches. Forward propagation yields the predicted changes in regional performance or the predicted reward value. Subsequently, a preset loss function is used to calculate the difference and error gradient between the predicted result and the actual regional reward value label. The network weight parameters, such as convolutional layers or attention layers, within the initial graph neural network are iteratively updated and fitted using the backpropagation algorithm. After multiple rounds of iterative training until the model's prediction error converges to a set allowable range, a graph neural network model with accurate prediction capabilities is obtained.

[0088] In addition to using a separate supervised learning approach, the training of this graph neural network model can also be carried out simultaneously with the agent's reinforcement learning strategy optimization process, that is, the model parameters are continuously updated by using new data generated from environmental exploration during runtime.

[0089] In this embodiment, by utilizing historical real-world change data or simulated labeled data to perform fully supervised learning training on the graph neural network model, the model can accurately learn from massive amounts of data and uncover the neighbor interaction patterns and signal interference propagation laws hidden beneath the complex network topology. This data-driven model training method not only reduces manual maintenance costs but also endows the graph neural network with powerful cross-cell cascading effect prediction capabilities. This mechanism can provide reinforcement learning agents with more stable, accurate, and objectively wireless communication physical laws-compliant regional reward evaluation signals, significantly enhancing the overall effectiveness and robustness of multi-cell cooperative network optimization schemes.

[0090] In some embodiments, the merging of local reward values ​​and regional reward values ​​includes: The local reward value is normalized to obtain the normalized local reward value, and the regional reward value is normalized to obtain the normalized regional reward value. Determine the local weight of the normalized local reward value and the regional weight of the normalized regional reward value; Based on local and regional weights, the normalized local reward value and the normalized regional reward value are weighted and summed to obtain the total reward value.

[0091] Normalization refers to using specific mathematical transformations to map feedback values ​​from different computational scales and evaluation dimensions to a unified comparison range based on the same standard. This can be achieved through linear scaling or monotonic nonlinear function mapping, ensuring that both the normalized local reward value and the normalized regional reward value fall within a baseline range of zero to one.

[0092] Optionally, fixed preference values ​​can be pre-set by network operators according to specific control strategies. For example, a higher local weight can be set when pursuing ultimate single-point performance, or a higher regional weight can be set during major communication security activities to ensure network-wide collaborative stability. Furthermore, dynamic adjustments can be made based on the training stage of the agent's decision-making model. For instance, a higher local weight can be assigned in the early stages of reinforcement learning to guide the agent to quickly master basic performance-enhancing actions, while the regional weight can be gradually increased in the later stages of training to guide the agent to learn more complex global collaborative strategies. The calculated total reward value is a single real scalar, which serves as a feedback signal that ultimately integrates multi-dimensional objectives and is directly fed into the reinforcement learning algorithm to guide the gradient update of the policy.

[0093] Alternatively, the total reward value can be calculated as follows: R_total=β R_local+(1-β) R_regional; Where R_local is the normalized local reward value, R_regional is the normalized regional reward value, and β is the local weight. By adjusting the value of β, switching between different optimization strategies can be achieved: β closer to 1 emphasizes individual performance, while β closer to 0 emphasizes overall benefits; this weight can be preset according to the operator's strategy preferences. Alternatively, the agent can be designed to automatically adjust β during the learning process, but the basic principle is that a coordination coefficient is needed to integrate the dual rewards.

[0094] In this embodiment, by uniformly normalizing and flexibly weighting and fusing the local and regional reward components, the system ensures that the agent receives smooth and stable training guidance while pursuing the service quality of individual nodes and maintaining the balance of network resources, thus avoiding drastic fluctuations during the policy learning process. Simultaneously, the flexible and adjustable weight allocation design gives the system a high degree of freedom in policy control, allowing operators to intervene in the agent's decision-making tendencies in real time by fine-tuning parameters without retraining the model. This significantly improves the engineering practicality and multi-objective coordination capabilities of the wireless network self-optimization system in complex live network deployments.

[0095] The following describes the agent-based wireless network optimization apparatus provided in the embodiments of this application. The agent-based wireless network optimization apparatus described below and the agent-based wireless network optimization method described above can be referred to in correspondence.

[0096] Figure 5 A schematic diagram of the structure of the agent-based wireless network optimization device provided in an embodiment of this application. (Refer to...) Figure 5 This application provides an agent-based wireless network optimization device 500, comprising: The acquisition unit 510 is used to acquire the root cause location result of the wireless network of the target cell; the root cause location result is determined based on the performance index data of the wireless network of the target cell. The decision unit 520 is used to input the fault root cause location result into the decision model of the intelligent agent to obtain the parameter adjustment scheme of the target cell's wireless network output by the decision model. The network optimization unit 530 is used to optimize the wireless network of the target cell based on the parameter adjustment scheme; The decision model is obtained by the agent through reinforcement learning based on the root cause localization results of the wireless network of the sample cell and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the next-level area of ​​the sample cell; the local reward value is calculated based on the dynamic weights of each performance index.

[0097] Optionally, the agent's decision-making model is trained based on the following steps: Obtain sample fault work order descriptions and fault root cause location results for the wireless network of the sample cell. The root cause localization results are input into the initial decision model of the agent to obtain the parameter adjustment prediction scheme of the wireless network of the sample cell output by the initial decision model. After the parameter adjustment prediction scheme is executed, obtain the scheme execution result samples of the sample cells; Input the fault work order description sample, fault root cause location result sample, and solution execution result sample into the reward model. The reward model determines the dynamic weight of each performance indicator based on the fault work order description sample, fault root cause location result sample, and solution execution result sample. It calculates the local reward value based on the dynamic weight and calculates the regional reward value. It merges the local reward value and the regional reward value and outputs the total reward value and / or the status signal of the sample cell. Based on the total reward value, the parameters of the initial decision model are iteratively optimized to obtain the decision model.

[0098] Optionally, the dynamic weights of each performance metric are determined, including: Based on the fault work order description samples, fault root cause location result samples, and solution execution result samples, the network optimization target of the sample cell is determined. User intent is identified based on network optimization objectives, and performance metrics and their dynamic weights are determined based on user intent.

[0099] Optionally, the fault work order description sample includes data on various performance indicators of the sample cell, and the solution execution result sample includes new data on various performance indicators of the sample cell. The local reward value is calculated based on dynamic weights, including: Based on the data and new data for each performance indicator, calculate the change value of each performance indicator; The local reward value of the sample cell is calculated based on the changes in each performance indicator and its dynamic weight.

[0100] Optionally, based on the changes in each performance indicator and its dynamic weight, the local reward value of the sample cell is calculated, including: Based on historical data of various performance indicators of the sample cells, the dynamic thresholds of each performance indicator are determined. The local reward value of the sample cell is calculated based on the dynamic thresholds of each performance indicator, as well as the changes and dynamic weights of each performance indicator.

[0101] Optionally, based on the changes in each performance indicator and its dynamic weight, the local reward value of the sample cell is calculated, including: The changes in each performance index are normalized to obtain the normalized changes in each performance index. The local reward value of the sample cell is calculated based on the normalized change value and dynamic weight of each performance index.

[0102] Optionally, the calculation of the regional reward value includes: Determine the relationships between the sample cell and its neighboring cells; Using the sample cell and its neighboring cells as nodes, and the relationships between the sample cell and its neighboring cells as edges, construct the initial network topology of the next-level region of the sample cell. The first state feature data of the sample cell and the semantic embedding vector of the network optimization target are obtained, and the second state feature data of each neighbor cell are obtained. The semantic embedding vector is extracted by the reward model based on the fault work order description sample, the fault root cause location result sample and the scheme execution result sample. The first state feature data, the second state feature data, and the semantic embedding vector of the network optimization objective are injected into the initial network topology graph to obtain the network topology graph. The network topology is input into the graph neural network model to obtain the regional reward value corresponding to the next-level region of the sample cell output by the graph neural network model.

[0103] Optionally, the graph neural network model is trained based on the network topology map samples of the previous level region of the sample cell and the regional reward value labels of the previous level region of the sample cell.

[0104] Optionally, local reward values ​​and regional reward values ​​can be combined, including: The local reward value is normalized to obtain the normalized local reward value, and the regional reward value is normalized to obtain the normalized regional reward value. Determine the local weight of the normalized local reward value and the regional weight of the normalized regional reward value; Based on local and regional weights, the normalized local reward value and the normalized regional reward value are weighted and summed to obtain the total reward value.

[0105] It should be noted that the embodiments of this application provide a wireless network optimization device based on intelligent agents, which can implement all the method steps implemented in the above embodiments of wireless network optimization based on intelligent agents, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail.

[0106] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 6 As shown, the electronic device may include a processor 610, a communication interface 620, a memory 630, and a communication bus 640. The processor 610, communication interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call a computer program in the memory 630 to execute steps of an agent-based wireless network optimization method, such as: obtaining the root cause location results of the target cell's wireless network; the root cause location results are determined based on the performance index data of the target cell's wireless network; inputting the root cause location results into the agent's decision model to obtain a parameter adjustment scheme for the target cell's wireless network output by the decision model; optimizing the target cell's wireless network based on the parameter adjustment scheme; the decision model is obtained through reinforcement learning based on sample cell root cause location result samples and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the sample cell's next-level region; the local reward value is calculated based on the dynamic weights of each performance index.

[0107] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the agent-based wireless network optimization method provided in the above embodiments, such as: obtaining the root cause location result of the wireless network of the target cell; the root cause location result is determined based on the performance index data of the wireless network of the target cell; inputting the root cause location result into the decision model of the agent to obtain the parameter adjustment scheme of the wireless network of the target cell output by the decision model; optimizing the wireless network of the target cell based on the parameter adjustment scheme; the decision model is obtained by reinforcement learning based on the root cause location result samples of sample cells and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the upper-level area of ​​the sample cell; the local reward value is calculated based on the dynamic weight of each performance index.

[0109] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program. The computer program is used to cause a processor to execute the steps of the agent-based wireless network optimization method provided in the above embodiments. For example, it includes: obtaining the root cause location result of a wireless network fault in a target cell; the root cause location result is determined based on the performance index data of the target cell's wireless network; inputting the root cause location result into the agent's decision model to obtain a parameter adjustment scheme for the target cell's wireless network output by the decision model; optimizing the target cell's wireless network based on the parameter adjustment scheme; the decision model is obtained through reinforcement learning based on samples of root cause location results from sample cells and the corresponding total reward value; the total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the sample cell's next-level region; the local reward value is calculated based on the dynamic weights of each performance index.

[0110] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0111] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for optimizing wireless networks based on intelligent agents, characterized in that, include: Obtain the root cause location results of the wireless network fault in the target cell; The fault root cause location result is determined based on the performance index data of the target cell's wireless network; The fault root cause localization result is input into the decision model of the intelligent agent to obtain the parameter adjustment scheme of the wireless network of the target cell output by the decision model; Based on the parameter adjustment scheme, the wireless network of the target cell is optimized; The decision model is obtained by the agent through reinforcement learning based on the root cause localization results of the wireless network in the sample cell and the corresponding total reward value. The total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the next-level region of the sample cell; the local reward value is calculated based on the dynamic weights of each performance indicator.

2. The agent-based wireless network optimization method according to claim 1, characterized in that, The decision-making model of the intelligent agent is trained based on the following steps: Obtain sample fault work order descriptions and fault root cause location results for the sample cell. The fault root cause localization result sample is input into the initial decision model of the agent to obtain the parameter adjustment prediction scheme of the wireless network of the sample cell output by the initial decision model; After the parameter adjustment prediction scheme is executed, the scheme execution result sample of the sample cell is obtained; The fault work order description sample, the fault root cause location result sample, and the solution execution result sample are input into the reward model. The reward model determines the dynamic weight of each performance indicator based on the fault work order description sample, the fault root cause location result sample, and the solution execution result sample. Based on the dynamic weight, the local reward value is calculated, and the regional reward value is calculated. The local reward value and the regional reward value are merged, and the total reward value and / or the status signal of the sample cell are output. Based on the total reward value, the parameters of the initial decision model are iteratively optimized to obtain the decision model.

3. The agent-based wireless network optimization method according to claim 2, characterized in that, The determination of the dynamic weights of each performance index includes: Based on the fault work order description sample, the fault root cause location result sample, and the scheme execution result sample, the network optimization target of the sample cell is determined. User intent is identified based on the network optimization objective, and each performance indicator and its dynamic weight are determined based on the user intent.

4. The agent-based wireless network optimization method according to claim 2, characterized in that, The fault work order description sample includes data for each performance indicator of the sample cell, and the scheme execution result sample includes new data for each performance indicator of the sample cell. The calculation of the local reward value based on the dynamic weight includes: Based on the data of each performance indicator and the new data of each performance indicator, calculate the change value of each performance indicator; The local reward value of the sample cell is calculated based on the changes in each of the performance indicators and the dynamic weights.

5. The agent-based wireless network optimization method according to claim 4, characterized in that, The calculation of the local reward value of the sample cell based on the changes in each of the performance indicators and the dynamic weights includes: Based on the historical data of each performance indicator of the sample cells, determine the dynamic threshold of each performance indicator; The local reward value of the sample cell is calculated based on the dynamic threshold of each performance indicator, the change value of each performance indicator, and the dynamic weight.

6. The agent-based wireless network optimization method according to claim 4, characterized in that, The calculation of the local reward value of the sample cell based on the changes in each of the performance indicators and the dynamic weights includes: The changes in each of the aforementioned performance indicators are normalized to obtain the normalized changes in each of the aforementioned performance indicators. The local reward value of the sample cell is calculated based on the normalized change values ​​of each performance indicator and the dynamic weight.

7. The agent-based wireless network optimization method according to claim 2, characterized in that, The calculation of the regional reward value includes: Determine the association relationship between the sample cell and each of its neighboring cells; Using the sample cell and each of the neighboring cells as nodes, and the relationships between the sample cell and each of the neighboring cells as edges, an initial network topology graph of the next-level region of the sample cell is constructed. The first state feature data and semantic embedding vector of the network optimization target of the sample cell are obtained, and the second state feature data of each of the neighboring cells are obtained; the semantic embedding vector is extracted by the reward model based on the fault work order description sample, the fault root cause location result sample and the scheme execution result sample; The first state feature data, the second state feature data, and the semantic embedding vector of the network optimization objective are injected into the initial network topology graph to obtain the network topology graph; The network topology is input into a graph neural network model to obtain the regional reward value corresponding to the next-level region of the sample cell, as output by the graph neural network model.

8. The agent-based wireless network optimization method according to claim 7, characterized in that, The graph neural network model is trained based on the network topology graph samples of the upper-level regions of the sample cell and the regional reward value labels of the upper-level regions of the sample cell.

9. The agent-based wireless network optimization method according to claim 2, characterized in that, The fusion of the local reward value and the regional reward value includes: The local reward value is normalized to obtain a normalized local reward value, and the regional reward value is normalized to obtain a normalized regional reward value. Determine the local weight of the normalized local reward value and the regional weight of the normalized regional reward value; Based on the local weight and the regional weight, the normalized local reward value and the normalized regional reward value are weighted and summed to obtain the total reward value.

10. A wireless network optimization device based on intelligent agents, characterized in that, include: The acquisition unit is used to acquire the root cause location results of the wireless network fault in the target cell. The fault root cause location result is determined based on the performance index data of the target cell's wireless network; The decision unit is used to input the fault root cause localization result into the decision model of the agent to obtain the parameter adjustment scheme of the wireless network of the target cell output by the decision model; A network optimization unit is used to optimize the wireless network of the target cell based on the parameter adjustment scheme. The decision model is obtained by the agent through reinforcement learning based on the root cause localization results of the wireless network in the sample cell and the corresponding total reward value. The total reward value is obtained by fusing the local reward value of the sample cell and the regional reward value of the next-level region of the sample cell; the local reward value is calculated based on the dynamic weights of each performance indicator.

11. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the agent-based wireless network optimization method according to any one of claims 1 to 9.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the agent-based wireless network optimization method as described in any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the agent-based wireless network optimization method according to any one of claims 1 to 9.

Images