Network coverage optimization method and device based on large and small model collaboration

Abstract

The invention provides a network coverage optimization method and device based on large and small model collaboration. The method comprises the following steps: receiving a network optimization task instruction and current network state information of a network environment; the large model processes a network optimization task instruction and network state information based on a retrieval enhancement generation mechanism of reasoning enhancement, and outputs priori strategy distribution of network parameter adjustment; the small model generates a network parameter adjustment action according to priori strategy distribution and a value function fused with the collaborative alignment degree; executing a network parameter adjustment action, and collecting performance feedback data of the network environment; and updating decision parameters of the small model based on the performance feedback data, and correcting an inference mechanism of the large model until the network performance index reaches a preset target. According to the method, the problems of capability limitation of a single model and lack of an inter-model cooperation mechanism in the prior art can be solved, and global optimization, dynamic self-adaption and autonomous intelligent operation and maintenance of network coverage are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a network coverage optimization method and apparatus based on size model collaboration. Background Technology

[0002] With the explosive growth of mobile communication traffic and the increasing density of network deployments, the complexity of wireless networks has reached unprecedented levels, creating an urgent need for automated operation and maintenance and intelligent management. Network coverage optimization, as a core component of wireless resource management, directly determines user experience quality and network operational efficiency. Its core challenge lies in the need to jointly optimize a series of interdependent parameters, such as base station antenna parameters, transmit power, and carrier frequency, to achieve the optimal balance between coverage, signal quality, and system capacity in a complex electromagnetic environment.

[0003] To overcome the limitations of manual optimization, researchers have introduced machine learning, particularly lightweight models such as reinforcement learning, to automatically adjust network parameters. While these methods have achieved automation to some extent, they still face bottlenecks: significant issues with sample efficiency and convergence, requiring large amounts of online or offline interactive data for training, and high costs and time consumption in real-world network environments; limited generalization ability, where models trained in specific scenarios experience a sharp decline in performance when network topology, user behavior, or interference environments change significantly; and difficulty in designing reward functions, as network optimization objectives are often multi-dimensional, making it challenging to design a reward function that accurately and comprehensively reflects the optimization goal. An unreasonable reward function design can easily lead to agents getting trapped in local optima or exhibiting unexpected policy behaviors.

[0004] In recent years, large-scale models, represented by large language models, have demonstrated enormous potential in various fields due to their superior natural language understanding, logical reasoning, and code generation capabilities. In the networking field, some works have begun to explore the use of large models for network fault diagnosis and resource allocation recommendations. Large models can understand high-level optimization intentions and decompose them into a series of operational steps. However, there are significant obstacles to directly applying general-purpose large models to refined network coverage optimization: the knowledge of large models comes from their training corpora and lacks professional knowledge of the physical and protocol layers in the communications field, which may produce illusory outputs that do not conform to physical laws or protocol specifications when generating specific parameter configurations; their internal knowledge is static and cannot obtain real-time dynamic information such as the latest network topology, device status, and channel characteristics of specific areas; at the same time, the inference process of large models has high computational overhead and high latency, making it difficult to meet the real-time requirements of network control loops.

[0005] Current preliminary explorations of large-scale model collaboration mostly remain at the level of simple task assignment or result integration, lacking an intrinsic, deep collaboration and alignment mechanism. The high-level guidance of large models is often described in vague natural language, while the decision space of small models consists of precise numerical parameters, creating a significant semantic gap. This loose coupling leads to inefficient knowledge transfer and makes it difficult to guarantee collaborative effectiveness.

[0006] In summary, existing technological approaches either rely too heavily on manual intervention, are limited by the generalization capabilities of small models, or suffer from the domain knowledge gap and reliability issues of large models. There is an urgent need for a new paradigm that can organically combine the macro-strategic reasoning capabilities of large models with the micro-tactical execution capabilities of small models, and effectively incorporate domain knowledge, to address the challenges of self-optimization in next-generation wireless networks. Summary of the Invention

[0007] This invention provides a network coverage optimization method and apparatus based on large and small model collaboration, addressing the limitations of single model capabilities and the lack of inter-model collaboration mechanisms in existing technologies. It achieves global optimization, dynamic adaptation, and autonomous intelligent operation and maintenance of network coverage. The technical solution proposed by this invention is as follows: In a first aspect, the present invention provides a network coverage optimization method based on size model collaboration, comprising: Receive network optimization task instructions and current network status information of the network environment; The large model processes the network optimization task instructions and the network state information based on the reasoning-enhanced retrieval generation mechanism, and outputs the prior policy distribution for network parameter adjustment. The small model generates network parameter adjustment actions based on the prior policy distribution and the value function that incorporates the collaborative alignment degree. Perform the network parameter adjustment action and collect network environment performance feedback data; The decision parameters of the small model are updated based on the performance feedback data, and the inference mechanism of the large model is corrected until the network performance indicators reach the preset target.

[0008] Optionally, the large model processes the network optimization task instructions and the network state information based on the reasoning-enhanced retrieval generation mechanism, and outputs the prior policy distribution for network parameter adjustment, including: The network optimization task instructions and the network status information are input into the large model together; The large model, based on the network optimization task instructions and the network state information, decomposes the task instructions into multiple sequentially executed reasoning steps through a thought chain; For each reasoning step, a query is generated by combining the network state information and the reasoning context, and a specified number of knowledge documents most relevant to the query are retrieved from the network knowledge base; The retrieved knowledge documents are fused with the context of the current reasoning step to generate knowledge-enhanced reasoning results; Based on the reasoning results enhanced by all knowledge, a structured distribution of the prior policy is generated.

[0009] Optionally, the value function incorporating collaborative alignment is constructed in the following manner: Calculate the KL divergence between the policy distribution of the small model and the prior policy distribution output by the large model, and use it as the co-alignment degree; The collaborative alignment degree is used as a regularization term and introduced into the original value function of the small model to form the value function that incorporates the collaborative alignment degree.

[0010] Optionally, the network parameter adjustment action is defined as adjusting at least one of the following parameters of the base station: antenna vertical downtilt angle, antenna horizontal azimuth angle, antenna transmit power, and carrier frequency corresponding to the downlink absolute radio frequency channel number.

[0011] Optionally, the performance feedback data is calculated based on the network performance index of the user equipment, which is a weighted sum of the signal-to-noise ratio function and the logarithm of the throughput, wherein the signal-to-noise ratio function approaches a saturation value when the user's signal-to-noise ratio exceeds a preset threshold. The signal-to-noise ratio of the user equipment is determined based on the antenna transmit power, the channel gain from the base station to the user equipment, and the thermal noise power of the user equipment. The channel gain is jointly determined by the antenna gain, path loss, and small-scale fading. The antenna gain is a function of the vertical downtilt deviation and the horizontal azimuth deviation. The vertical downtilt deviation is the deviation between the antenna vertical downtilt angle and the optimal antenna vertical downtilt angle from the base station to the user equipment, and the horizontal azimuth deviation is the deviation between the antenna horizontal azimuth angle and the optimal horizontal azimuth angle from the base station to the user equipment.

[0012] Optionally, the value function that incorporates the cooperative alignment is a KL-regularized Q function, the value of which is determined by the immediate performance reward, the cooperative alignment penalty, and the expected value of the next state. The decision parameters of the small model are updated using a dual-deep Q-network algorithm, the loss function of which is constructed based on the temporal difference error of the Q-function with KL regularization. The correction mechanism for the large model includes minimizing the KL divergence between the prior policy distribution and the policy distribution of the small model obtained from the empirical replay data.

[0013] Secondly, the present invention also provides a network coverage optimization device based on size model collaboration, comprising the following modules: The task receiving module is used to receive network optimization task instructions and current network status information of the network environment; The large model decision module is used to process the network optimization task instructions and the network state information based on the reasoning-enhanced retrieval generation mechanism of the large model, and output the prior policy distribution for network parameter adjustment. The small model decision module is used by the small model to generate network parameter adjustment actions based on the prior policy distribution and the value function that incorporates the collaborative alignment degree. The environment interaction module is used to perform the network parameter adjustment actions and collect network environment performance feedback data; The model update module is used to update the decision parameters of the small model based on the performance feedback data and correct the inference mechanism of the large model until the network performance indicators reach the preset target.

[0014] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the network coverage optimization method based on size model collaboration as described in the first aspect above.

[0015] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the network coverage optimization method based on size model collaboration as described in the first aspect above.

[0016] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the network coverage optimization method based on size model collaboration as described in the first aspect above.

[0017] Based on the above technical solution, the beneficial effects of the present invention compared with the prior art are as follows: The network coverage optimization method and apparatus based on large-scale model collaboration provided by this invention provides high-level guidance to the small model through the prior policy distribution output by the large model. The small model generates network parameter adjustment actions based on the prior policy distribution and a value function incorporating collaboration alignment. This collaborative approach organically combines the high-level strategic reasoning capability of the large model with the micro-tactical execution capability of the small model. Furthermore, by using the value function of collaboration alignment, it narrows the semantic gap between the ambiguous natural language description of the large model and the precise numerical parameters of the small model, improving knowledge transfer efficiency and ensuring collaborative effectiveness. After executing the network parameter adjustment actions, performance feedback data of the network environment is collected, and the decision parameters of the small model are updated based on this data, while the reasoning mechanism of the large model is corrected. This feedback mechanism enables the large and small models to continuously learn and optimize, forming a closed-loop optimization process, which can better meet the challenges of self-optimization in next-generation wireless networks.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention 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 invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the network coverage optimization method based on size model collaboration provided by the present invention.

[0022] Figure 2 This is a schematic diagram of the framework of the network coverage optimization method based on size model collaboration provided by the present invention.

[0023] Figure 3 This is a comparison chart of SNR convergence performance under different schemes.

[0024] Figure 4 This is a comparison chart of coverage rates with different numbers of base stations.

[0025] Figure 5 This is a comparison chart of average throughput under different numbers of user devices.

[0026] Figure 6 This is a schematic diagram of the network coverage optimization device based on size model collaboration provided by the present invention.

[0027] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] Small models based on methods such as reinforcement learning typically have a decision-making perspective limited to the current state and local rewards, lacking a macroscopic understanding of the network's global state and long-term planning capabilities. In complex optimization scenarios with multiple conflicting objectives, they are prone to getting trapped in local optima and cannot fundamentally solve systemic problems such as large-scale contiguous weak coverage.

[0030] While general-purpose large models can perform complex logical reasoning, their reasoning process is disconnected from real-world physical constraints and domain knowledge. In network optimization tasks, this "castle in the air" reasoning leads to strategies lacking feasibility, security, and effectiveness. Direct deployment may cause network performance degradation or even service interruption. Reliability is the biggest obstacle to its application in core network control. Current preliminary explorations of large-scale model collaboration mostly remain at the level of simple task assignment or result integration, lacking an intrinsic, deep collaboration and alignment mechanism. The high-level guidance of large models is often described in vague natural language, while the decision space of small models consists of precise numerical parameters, creating a significant semantic gap. This loose coupling leads to inefficient knowledge transfer and makes it difficult to guarantee collaborative effectiveness.

[0031] To overcome the aforementioned shortcomings, this patent aims to address the core technical problem of how to construct a knowledge-driven, self-evolving, and highly collaborative intelligent network optimization system. This invention aims to design a novel collaborative reasoning paradigm, in which a large model is responsible for global, strategic planning and reasoning, setting optimization directions and strategy boundaries for smaller models; the smaller models are responsible for high-frequency, efficient tactical execution and parameter fine-tuning in real-world environments. Through closed-loop feedback, the execution results of the smaller models can, in turn, correct and optimize the cognition of the large model, forming a symbiotic and evolving intelligent closed loop.

[0032] This method is implemented in next-generation wireless communication network environments (such as 5G-Advanced and 6G), which possess self-organizing network capabilities, and whose base stations and mobile user equipment support corresponding parameter adjustments and status information feedback. Simultaneously, a professional knowledge base containing massive amounts of network domain knowledge is required. This knowledge base covers 3GPP protocol specifications, equipment parameter manuals, historical optimization cases, expert rules, etc., and is capable of efficient knowledge retrieval. Furthermore, a large model and a small model need to be deployed, capable of data interaction and collaborative work. The large model can be a large language model such as Qwen2.5-7B-Instruct, and the small model can be a Double Deep Q-Network (DDQN) agent. Initialization is performed first, including large model initialization, small model initialization, and environment awareness module initialization. Large model initialization refers to loading pre-trained large model parameters to ensure the large model possesses basic network optimization-related knowledge and reasoning capabilities. Small model initialization refers to initializing the online network parameters and target network parameters of the small model, and simultaneously initializing an empty experience replay pool to store experience data during subsequent interactions. Initializing the environment perception module refers to configuring the network environment perception module to ensure that it can accurately collect relevant data from base stations and mobile users to form the current network status information.

[0033] Reference Figure 1 As shown, the network coverage optimization method based on size model collaboration includes the following: S110: Receive network optimization task instructions and current network status information of the network environment.

[0034] The network management system interface receives network optimization task instructions, which specify the optimization objectives, such as improving signal coverage in a specific area, reducing interference levels, or increasing user throughput. The instructions are transmitted in structured data format, containing key information such as task type, target area, and performance requirements.

[0035] By utilizing base station monitoring equipment and user terminal feedback modules, the current network status information of the network environment is collected in real time. This includes base station antenna parameters (such as vertical downtilt angle and horizontal azimuth angle), transmit power, carrier frequency, as well as information such as the location of mobile users, received signal strength, and signal-to-noise ratio.

[0036] S120. The large model uses a reasoning-enhanced retrieval generation mechanism to process the network optimization task instructions and the network state information, and outputs the prior policy distribution for network parameter adjustment.

[0037] The reasoning-enhanced retrieval generation mechanism combines the reasoning capabilities of a large model with retrieval enhancement techniques. The large model first performs logical analysis and judgment on network optimization task instructions and network state information, then retrieves relevant information from external professional knowledge bases to provide more evidence for reasoning, thereby improving the accuracy and reliability of the output. The specific processing flow is as follows: The large model activates its Chain-of-Thought (CoT) capability, decomposing complex network optimization tasks into a series of logically interconnected sub-problems or reasoning steps. For example, if the task is to optimize network coverage in a certain area, the large model might decompose it into steps such as "identifying base stations with weak coverage in the area," "analyzing the key parameters of these base stations," and "determining adjustment strategies, such as downtilt or power." For each reasoning step, a highly focused query is generated by combining the original task and previous reasoning context, retrieving multiple knowledge documents most relevant to the query from a vector database. For example, when analyzing key base station parameters, relevant equipment parameter manuals and historical optimization cases are retrieved. The large model uses the retrieved knowledge documents as additional contextual information, combining them with the current reasoning step to generate knowledge-calibrated and enhanced reasoning results. Finally, based on the entire knowledge-enhanced inference chain, the large model generates a structured, executable distribution of prior policies, indicating which parameter adjustments are more worthwhile to explore in the current network state.

[0038] S130. The small model generates network parameter adjustment actions based on the prior policy distribution and the value function that incorporates the collaborative alignment degree.

[0039] The smaller model receives the prior policy distribution and the current network state from the larger model. During its decision-making process, the smaller model considers the potential environmental rewards of actions based on a value function that incorporates cooperative alignment. This function quantifies the alignment between the smaller model's own policy distribution and the larger model's prior policy distribution by calculating the KL divergence, and incorporates this alignment as a regularization term into the value function. Through this mechanism, when exploring optimal actions, the smaller model tends to select parameter adjustment schemes that improve network performance in the short term while aligning with the larger model's long-term strategic intent, ultimately outputting specific and executable network parameter adjustment actions.

[0040] S140. Perform the network parameter adjustment action and collect network environment performance feedback data.

[0041] The parameter adjustment actions generated by the small model are sent to the corresponding user devices for execution via the network management interface. After the actions are executed and the network status stabilizes, network status information, including throughput and signal-to-noise ratio, is collected again. Network performance metrics are calculated based on a predefined network performance metric calculation system to obtain the immediate performance reward for this adjustment as performance feedback data. Simultaneously, the complete interaction experience is recorded, including the state before adjustment, the actions performed, the reward obtained, the new state after adjustment, and the prior strategies provided by the large model, forming a complete experience tuple for subsequent learning.

[0042] S150. Update the decision parameters of the small model based on the performance feedback data, and correct the inference mechanism of the large model until the network performance indicators reach the preset target.

[0043] A batch of historical experience data is randomly sampled from the accumulated experience replay pool for model updates. For small models, this experience data is used to calculate the gradient of the loss function through a temporal difference learning algorithm, and the parameters of the online network or target network are updated. By continuously updating the parameters, the small model can generate better network parameter adjustment actions in subsequent decisions.

[0044] The process of modifying the large model's inference mechanism involves fine-tuning the large model based on the execution feedback from the small model. For example, by minimizing the KL divergence loss, the prior knowledge of the large model is made closer to the small model's policies, which have proven effective in practice, thus achieving bidirectional adaptation. The relationship between the distribution of the prior policies output by the large model and the actual improvement in network performance is analyzed to identify potential problems in the large model's inference process. Adjustments are then made to the large model's inference algorithm, such as modifying the weights of the knowledge base for retrieval enhancement and adjusting the logical rules of inference.

[0045] Repeat steps S120 to S150 to continuously improve network optimization capabilities through constant interaction, learning, alignment, and fine-tuning. Continuously monitor network performance metrics, and stop iterating when the network performance metrics reach the preset target, thus completing this network coverage optimization task.

[0046] Traditional reinforcement learning models, especially small models, require extensive online or offline interactive data for training, resulting in high trial-and-error costs and time consumption in real-world network environments. In this invention, the large model, based on an inference-enhanced retrieval generation mechanism, leverages existing knowledge and experience to provide a more reasonable prior policy distribution for the small model. The small model then makes decisions based on this, reducing blind trial-and-error, lowering the need for large amounts of training data, improving sample efficiency, and accelerating convergence. When network topology, user behavior, or interference environments change significantly, the performance of traditional small models deteriorates sharply. The large model, with its stronger generalization ability, can understand high-level optimization intentions and transmit this understanding to the small model. This invention, through the collaboration of the large and small models, allows the small model to generate more adaptable network parameter adjustment actions based on the prior policy distribution provided by the large model and combined with new network state information, thereby improving the overall system's adaptability and generalization ability in the face of changing network environments.

[0047] Large models derive their knowledge from training corpora but lack expertise in the physical and protocol layers of communications. This can lead to illusory outputs that contradict physical laws or protocol specifications when generating specific parameter configurations. In this invention, small models focus on micro-tactical execution, handling specific numerical parameter adjustments. Through collaboration with the large model, they can leverage the higher-level guidance of the large model to some extent, while simultaneously incorporating their own domain knowledge to ensure that the generated network parameter adjustments conform to the standards and requirements of the communications field.

[0048] Existing preliminary explorations of large-scale model collaboration mostly remain at the level of simple task assignment or result integration, lacking an inherent, deep collaboration and alignment mechanism. In this invention, the prior policy distribution output by the large model provides high-level guidance to the small model. The small model generates network parameter adjustment actions based on the prior policy distribution and a value function incorporating collaboration alignment. This collaborative approach organically combines the high-level strategic reasoning capabilities of the large model with the micro-tactical execution capabilities of the small model. Furthermore, through the value function of collaboration alignment, it narrows the semantic gap between the ambiguous natural language description of the large model and the precise numerical parameters of the small model, improving knowledge transfer efficiency and ensuring collaborative effectiveness. After executing the network parameter adjustment actions, performance feedback data of the network environment is collected, and the decision parameters of the small model are updated based on this data, while the reasoning mechanism of the large model is corrected. This feedback mechanism enables the large and small models to continuously learn and optimize, forming a closed-loop optimization process, which can better address the challenges of self-optimization in next-generation wireless networks.

[0049] In an optional embodiment, this invention proposes a Reasoning-enhanced Retrieval-Augmented Generation (RRAG) mechanism. This mechanism decomposes complex problems into thought chains and enhances the knowledge of each node in the chain, thereby guiding the reasoning process of a large model and ensuring the professionalism and accuracy of its decisions. The large model described in S120 above processes the network optimization task instructions and the network state information based on the Reasoning-enhanced Retrieval-Augmented Generation mechanism, outputting a priori policy distribution for network parameter adjustment, including: S1201. Input the network optimization task instruction and the network status information into the large model together.

[0050] This invention's technical solution is based on a cellular network system model containing N base stations and M user devices. Network optimization task instructions and current network state information are input into the large model. For example, the network optimization task instruction is "optimize coverage for edge users in area A," and the current network state information includes the set of adjustable network parameters for base station n. The set of adjustable network parameters for base station n is defined as follows: , respectively representing the vertical downtilt angle of its antenna. Antenna horizontal azimuth angle Antenna transmit power (dBm) and the downlink absolute radio frequency channel number (ARFCN) The corresponding carrier frequency and the channel gain from base station n to user equipment m This information provides the large model with the task context and the actual situation of the current network, enabling it to perform targeted inference.

[0051] S1202. The large model, based on the network optimization task instructions and the network state information, decomposes the task instructions into multiple sequentially executed reasoning steps through a thought chain.

[0052] When the large model receives network optimization task instructions and current network state information, it activates its thought process capabilities. In this initial stage, the large model associates or retrieves general quantitative criteria for evaluating network state. These criteria exist in the form of thresholds, suggested ranges, or proportions. For example, it associates criteria with "a suggested configuration range for antenna downtilt angle in a specific scenario" and "a performance threshold defining coverage quality for user channel gain." By systematically comparing specific data in the network state information with these quantitative criteria, the large model identifies significant deviations and anomalies. For example, if a certain proportion of base stations in a certain area have antenna downtilt angle configuration values ​​consistently higher than the upper limit of the suggested range, and simultaneously a certain proportion of users in that area have channel gain measurements consistently lower than the performance threshold, the large model will generate a sequence of inference steps closely related to the current network situation based on this quantitative judgment. . "Identifying base stations involved in overlapping coverage areas" means identifying areas where the proportion of users with channel gain below a performance threshold exceeds a preset threshold, and the set of base stations providing services to those areas. "Analyzing the key parameters of these base stations" means, for the identified set of base stations, analyzing the configuration of their key adjustable parameters and quantifying the degree to which each parameter value deviates from its corresponding performance target range. "Determine the adjustment strategy" means generating a specific network parameter adjustment strategy based on the quantitative analysis results of parameter deviation, clarifying the parameters to be adjusted, the direction of adjustment, and the quantitative adjustment range.

[0053] For example, the task of "optimizing coverage for edge users in region A" might be broken down into the following reasoning steps: Step 1: Identify base stations with weak coverage. Specifically, this involves analyzing the signal coverage of each base station for edge users in area A within the current network state to determine which base stations have insufficient signal strength. This process is achieved through the following operations: First, based on the transmit power p of base station n... n and the channel gain from base station n to user equipment m Calculate the estimated received signal power P of user equipment m. rx =p n +G n,m The calculated P rx Compared with the preset coverage performance threshold P th Perform numerical comparisons and filter out all P values. rx <P th The user equipment is then identified. Next, the serving base station identifiers of these user equipment are matched against the theoretical main serving base station list for the area as defined in the network plan. If the proportion or number of weakly covered user equipment served by a base station exceeds a threshold, then that base station is determined to be a weakly covered base station.

[0054] Step Two: For the weakly covered base stations identified in Step One, analyze whether the current configuration of their adjustable parameters is reasonable. This process is achieved through the following operations: Extracting the set of adjustable network parameters for base station n. Then, the performance target range that the above parameters should meet under the current network scenario is obtained from the network knowledge base. For example, to ensure edge coverage of area A, the antenna vertical downtilt angle... The performance target range is [φ min ,φ max ], transmit power p n The performance target range is [p min ,p max Finally, set the current parameters. The measured values ​​are compared with the corresponding performance target range to determine whether they exceed the performance target range, and the parameter deviation is calculated. The superscript 'i' represents the set of adjustable network parameters. The i-th parameter in.

[0055] Step 3: Generate specific parameter adjustment instructions based on the parameter deviation calculated in Step 2. This process is achieved through the following operations: based on the parameter deviation obtained in Step 2... The sign and magnitude of the value are used to invoke predefined adjustment rules from the knowledge base. These rules are expressed as "If the current value of parameter X is higher / lower than the performance target range, adjust it by Y units towards the center of the range." For example, if the antenna's vertical downtilt angle... The current value is higher than the upper limit of the performance target range φ. max Then generate "will The instruction to "reduce Δ degree" is given, where Δ is selected from a preset set of adjustment step sizes based on the magnitude of the deviation. Finally, the output is a set of specific adjustment actions for each poorly covered base station.

[0056] S1203. For each reasoning step, combine the network state information and the reasoning context to generate a query, and retrieve a specified number of knowledge documents most relevant to the query from the network knowledge base.

[0057] In advance, massive amounts of network domain knowledge, such as 3GPP protocol specifications, equipment parameter manuals, historical optimization cases, expert rules, etc., are structured and divided into semantically coherent knowledge blocks. . Indicates the first Each knowledge block is encoded into a high-dimensional vector using a text embedding model (such as BERT, Sentence-BERT, etc.) and stored in a network database for efficient similarity retrieval.

[0058] For each inference step in the generation of the large model Perform the following operations: combine the "received network optimization task instructions" and the "current inference step". The query is generated by taking the description of the network, relevant network state data and its comparison results with the quantification standard, and historical reasoning conclusions. This query explicitly requests the quantification standard knowledge that matches the current scenario. The query is then used to search the network database and return the top-K most relevant knowledge documents. These documents contain quantitative judgment criteria, equipment specifications, or optimization experience guidelines defined in the form of proportions, thresholds, and ranges.

[0059] The query is generated as follows: The natural language description of the current inference step is analyzed to extract the intended operation and the knowledge domain to be queried. The generated inference context is checked. If the inference context is not empty, the list of identified, confirmed, or locked specific network entity identifiers is directly inherited from the latest conclusion of the inference context; this list is defined as the target entity set for the current inference step. If the inference context is empty (i.e., processing the first inference step), an initial target entity set is determined from the network state information based on the operation and knowledge domain parsed in the first step, using predefined filtering rules. Based on the determined target entity set, all relevant data records (including configuration parameters, performance indicators, measurement reports, etc.) of these entities at the decision-making moment are retrieved and extracted from the complete network state information, forming the fact dataset upon which the current query depends. All historical conclusions stored in the inference context are read sequentially, and confirmed problem diagnoses, verified constraints, or excluded hypotheses are extracted and transformed into a set of structured query background conditions. A query template matching the type of the current inference step is selected. The variable placeholders in the template are instantiated and replaced with specific values ​​from the target entity set, the fact dataset, and background conditions to generate a complete query statement containing explicit facts and historical logical constraints.

[0060] For example: For step one: the generated query could be "how to identify weakly covered base stations based on channel gain", which retrieves relevant expert rules, historical optimization cases, and other knowledge documents from the network knowledge base to obtain methods and experience for judging weakly covered base stations.

[0061] For step two: the query could be "the impact of different antenna parameters on coverage", which can be used to search for documents such as equipment parameter manuals and communication theory to understand the relationship between antenna parameters and coverage.

[0062] Step 3: The query can be "parameter tuning strategies for different coverage problems," retrieving historical optimization cases, best practices, and other documents to obtain practical parameter tuning experience. By retrieving a specified number (e.g., Top-K) of the most relevant knowledge documents, ensure that the information obtained is targeted and effective.

[0063] S1204. The retrieved knowledge documents are fused with the context of the current reasoning step to generate knowledge-enhanced reasoning results.

[0064] The retrieved knowledge documents are fused with the context of the current inference step to generate a knowledge-enhanced inference result. During this fusion process, the fusion of the retrieved knowledge documents and the context of the current inference step is achieved by constructing structured enhanced hints. This process first extracts the context of the current inference step, which consists of the inference conclusions generated by previous inference steps that are directly related to this inference step, along with the entities and conditional constraints they contain. This context is then organized with the network state data directly associated with the current inference step into a basic hint in a fixed format. Subsequently, the retrieved knowledge documents are sorted according to their relevance to the time the query was generated, and the complete texts of the top-ranked documents are sequentially embedded into a dedicated knowledge reference section of the basic hint. Finally, this hint, containing the context of the current inference step and the external professional knowledge sorted by relevance, is input into the large model. The large model performs comprehensive inference based on this hint and generates a text output, which is the knowledge-enhanced inference result. For example, in step one, the retrieved methods and experience for identifying weakly covered base stations are combined with information such as the channel gain of the current network to more accurately identify weakly covered base stations. In step two, based on the knowledge of the relationship between antenna parameters and coverage, and the key parameters of the current base station, a thorough analysis of the factors affecting coverage is conducted. In step three, based on practical experience in parameter adjustment and the current network coverage issues, an adjustment strategy is determined.

[0065] S1205. Based on the reasoning results enhanced by all knowledge, generate a structured distribution of the prior policy.

[0066] The large model receives and synthesizes all the knowledge-enhanced reasoning results it outputs at each reasoning step in the thought chain. These results together form a reasoning chain from problem diagnosis to solution. The large model is based on this inference chain, transforming it into a structured, executable prior policy distribution. , indicating that in a given network state Under these conditions, the output of the large model provides information about the adjustment actions for all possible network parameters. The prior policy probability distribution is as follows. Specifically, the large model comprehensively weighs the specific parameter adjustment suggestion (e.g., "increase the downtilt angle of base station X by 1.5 degrees") pointed to by the final inference step with possible alternative network parameter adjustment actions and their theoretical basis. Based on its internal mechanism, it calculates or evaluates the preference for each network parameter adjustment action in a given network state, and finally outputs a probabilistic prior policy distribution covering all possible network parameter adjustment actions. This prior policy distribution indicates which network parameter adjustment actions are more worthwhile to explore in the current network state. For example, the structured prior policy distribution can be presented in the form of a table or probability distribution, clearly indicating the direction and approximate magnitude of the adjustment of parameters such as antenna vertical downtilt angle, horizontal azimuth angle, and transmit power for each poorly covered base station, as well as the probability of different network parameter adjustment actions.

[0067] This invention decomposes complex tasks into multiple reasoning steps using a thought chain approach and combines this with knowledge retrieval and fusion from a network knowledge base. This enables the large model to gain a more comprehensive and in-depth understanding of the task and network state, thereby generating more accurate and reliable prior policy distributions. For example, when optimizing network coverage, it can more accurately identify coverage problems, analyze influencing factors, and determine reasonable parameter adjustment strategies, avoiding erroneous decisions due to a lack of professional knowledge or an incomplete understanding of the task. The professional knowledge in the network knowledge base provides expert support for the large model's reasoning, making its generated strategies more consistent with the physical laws and protocol specifications of the communication field. Simultaneously, real-time reasoning combined with current network state information allows the strategies to better adapt to different network environments and task requirements. Through retrieval enhancement and knowledge fusion, the large model can utilize the rich information in external knowledge bases to expand its knowledge scope, thereby improving its generalization ability to unseen network scenarios and tasks. For example, when encountering a new network topology or service type, it can retrieve relevant information from the knowledge base, combine it with the current network state to reason, and generate reasonable optimization strategies without relying on large amounts of training data.

[0068] The generated prior policy distribution provides strategic guidance and optimization direction for the small model, enabling it to make more targeted parameter adjustments during real-time decision-making. Simultaneously, the execution results of the small model can be fed back to the large model, promoting its correction and optimization, forming a symbiotic and evolving intelligent closed loop that improves the performance of the entire network optimization system.

[0069] In an optional embodiment, to address the semantic gap between large and small models, this invention aims to establish a quantifiable and optimizable policy alignment mechanism. This invention proposes a collaborative alignment degree concept based on KL divergence in information theory, using it as a mathematical tool to measure the consistency between the policy distribution of the small model and the prior policy distribution of the large model. By introducing this collaborative alignment degree as a regularization term into the optimization objective, it guides the small model to complete specific optimization tasks in a theoretically guaranteed manner, without deviating from the strategic intent of the large model. This ensures that the exploration and decision-making behavior of the small model does not deviate from the strategic direction given by the large model after careful consideration. The value function incorporating the collaborative alignment degree is constructed in the following way: Calculate the KL divergence between the policy distribution of the small model and the prior policy distribution output by the large model, and use it as the co-alignment degree; introduce the co-alignment degree as a regularization term into the original value function of the small model to form the value function that incorporates the co-alignment degree.

[0070] In the formula, For coordinated alignment, Let represent the network state observed at decision time t. In network status The next step is to perform network parameter adjustment actions on the small model. For small models, the policy distribution represents the distribution given the network state. The smaller model selects each possible network parameter adjustment action. The probability of. For the prior policy distribution of the large model, it represents the distribution of policies given the network state. Below, the large model, based on its knowledge-enhanced reasoning, recommends adjusting actions for each possible network parameter. The probability of. This represents the KL divergence. The smaller the KL divergence value, the closer the action selection probability distribution of the smaller model is to the prior policy distribution recommended by the larger model, i.e., the higher the co-alignment.

[0071] This invention incorporates collaborative alignment as an intrinsic reward or penalty into the learning objective of a small model, redefining a KL-regularized Q-function. Specifically, the value function incorporating collaborative alignment is a KL-regularized Q-function, the value of which is jointly determined by the immediate performance reward, the collaborative alignment penalty term, and the expected value of the next network state. In the formula, In network status Perform network parameter adjustment actions Finally, the long-term expected value after the alignment penalty was considered. This is for immediate performance enhancement, i.e., performing network parameter tuning actions. Subsequently, the performance improvement was directly reflected back from the network environment. Alignment weight coefficients. It is a constant greater than 0, used to balance the importance between network performance rewards and policy alignment penalties. This is a penalty for coordinating alignment. This is the discount factor. For information about the next network state The expectation. The expected value of the next network state, i.e., in the next network state. Below, the state value after considering the alignment penalty is taken into account.

[0072] The learning objective of small models is no longer simply to maximize external immediate performance rewards. Instead of simply adjusting network parameters, the smaller model maximizes this comprehensive Q-value. This means that when choosing network parameter adjustment actions, the smaller model weighs two aspects simultaneously. First, can this network parameter adjustment action directly improve network performance? Second, does this network parameter adjustment action align with the more forward-looking strategy recommendations given by the larger model? Hyperparameters are as described above. The importance of controlling alignment was emphasized.

[0073] In an optional embodiment, the network parameter adjustment action is defined as adjusting at least one of the following parameters of the base station: antenna vertical downtilt angle, antenna horizontal azimuth angle, antenna transmit power, and carrier frequency corresponding to the downlink absolute radio frequency channel number.

[0074] Adjusting the antenna's vertical downtilt angle can alter the vertical coverage range of a base station signal. When the vertical downtilt angle increases, the base station signal tilts downwards, effectively covering the area directly below the base station, reducing signal propagation over long distances, and thus enhancing signal strength for nearby users. In areas covered by multiple base stations, signals from different stations may interfere with each other. By adjusting the antenna's vertical downtilt angle, the direction and range of signal propagation can be controlled, reducing overlapping coverage areas between base stations, thereby reducing co-channel and adjacent-channel interference. The antenna's horizontal azimuth angle determines the direction of base station signal propagation in the horizontal plane. By adjusting the horizontal azimuth angle, the base station signal can be more accurately directed to densely populated user areas or areas requiring focused coverage. By adjusting the antenna's transmit power, differentiated services can be provided to different areas or users based on the needs of different services. For example, in areas with high data service demand, appropriately increasing the transmit power improves signal quality and meets users' needs for high-speed data transmission.

[0075] In cellular networks, frequency reuse technology can improve spectrum utilization. By adjusting the carrier frequency corresponding to the downlink absolute radio frequency channel number, frequency reuse can be achieved between different base stations, allowing adjacent base stations to communicate using different frequencies and reducing co-channel interference. When interference occurs in the network, adjusting the carrier frequency can switch affected users to other frequencies, thereby eliminating interference and restoring normal communication for users.

[0076] In an optional embodiment, the performance feedback data is calculated based on the network performance metrics of the user equipment, which are weighted sums of a signal-to-noise ratio function and the logarithm of throughput, wherein the signal-to-noise ratio function approaches a saturation value when the user's signal-to-noise ratio exceeds a preset threshold. The signal-to-noise ratio (SNR) of the user equipment (UE) is determined based on the antenna transmit power, the channel gain from the base station to the UE, and the thermal noise power of the UE. The channel gain is jointly determined by the antenna gain, path loss, and small-scale fading. The antenna gain is a function of the vertical downtilt deviation and the horizontal azimuth deviation; the vertical downtilt deviation is the antenna's vertical downtilt angle. With base station To user equipment Optimal antenna vertical downtilt angle The deviation between them, the horizontal azimuth deviation is the antenna horizontal azimuth angle With base station To user equipment Optimal horizontal azimuth The deviation between them.

[0077] Channel gain from base station n to user equipment m (dB) is a comprehensive physical quantity, and its calculation method is as follows: in: For antenna gain, it is related to the vertical downtilt angle deviation ( ) and azimuth deviation ( The function accurately simulates the effect of antenna pattern on signal strength.

[0078] The path loss model, which follows the standards defined by 3GPP and other standards, is a function of distance and propagation environment (such as urban or suburban areas).

[0079] Multipath propagation effects were modeled for small-scale fading channel coefficients.

[0080] Signal-to-noise ratio received by user equipment m from base station n The throughput that user equipment m can achieve from base station n They are given by the following formulas respectively: ; ; in, It is thermal noise power. It is the channel bandwidth allocated by base station n to user equipment m. It is the transmit power of base station n.

[0081] The core optimization objective of this invention is to maximize a weighted sum of network performance metrics. This metric aims to balance coverage quality and user throughput: in, These are trade-off factors. The first term is the signal-to-noise ratio (SNR) function, used to evaluate coverage quality. The SNR function is a sigmoid function, where... Exceeding the preset threshold When the first term approaches 1, it indicates that the basic coverage requirements are met. The second term is the logarithm of the throughput, representing the data transmission efficiency. This represents the steepness coefficient of the sigmoid function.

[0082] In an optional embodiment, the value function incorporating cooperative alignment is a KL-regularized Q function, the value of which is determined by an immediate performance reward, a cooperative alignment penalty, and the expected value of the next network state. The decision parameters of the small model are updated using a dual-deep Q-network algorithm, the loss function of which is constructed based on the temporal difference error of the Q-function with KL regularization. The correction mechanism for the large model includes minimizing the KL divergence between the prior policy distribution and the policy distribution of the small model obtained from the empirical replay data.

[0083] The Double Deep Q-Network (DQN) consists of two deep neural networks: an online network and a target network. The online network selects the actions to adjust the network parameters, while the target network evaluates the value of those actions. This separate design helps mitigate the overestimation problem in traditional Q-learning.

[0084] Specifically, the complete workflow of this invention is a continuously iterative, self-evolving closed-loop system, as described above. Figure 2 As shown, the specific execution steps are as follows: S210, System Initialization. This includes initializing the parameters of the Large Model (LM). (Typically, a pre-trained model is loaded.) Initialize the online network parameters of the Small Model (SM). and target network parameters Initialize an empty experience replay pool. It is used to store interactive experiences. Figure 2 The memory module in the system is responsible for managing the storage of history and context, as well as storing experience.

[0085] S220, Environmental Awareness and Knowledge-Enhanced Reasoning. At decision time t, multi-dimensional data is collected from the network environment to construct the current network state. Large model reception As input, its RRAG module is activated. The knowledge module enhances reasoning capabilities by using tools such as text embedding, relation extraction, and graph-based Retrieval-Augmented Generation (GraphRAG). It decomposes the problem using thought chains, retrieves knowledge through Retrieval-Augmented Generation (RAG), and finally generates a knowledge-enhanced prior policy distribution. .

[0086] S230, Small Model Decision Making and Environment Interaction. At each time step, the agent (i.e., the reinforcement learning agent) observes the current network state. Online networks depend on network status Based on a preset strategy (such as the ϵ-greedy strategy), and according to the current network state Select a network parameter adjustment action This strategy involves estimating the KL regularized Q-function. The network parameter adjustment action... It is applied to the network environment, for example, to modify base station parameters through the network management interface.

[0087] S240, Environmental Feedback and Experience Storage. After the network environment performs network parameter adjustment actions, it transitions to the next network state. Measure network performance metrics for the next network state, including performance, alarm status, and user feedback, and calculate the immediate performance bonus. Complete empirical tuples Store in the experience replay pool .

[0088] S250, Collaborative Fine-tuning and Update of Large and Small Models. The Large Language Model Agent (LLMAgent) involves the collaborative fine-tuning and updating of both the large and small models.

[0089] The small model update process is as follows: For each experience, the online network parameters are updated using the Dual Deep Q Network (DDQN) algorithm. Its core is to calculate the time-series difference error and perform gradient descent. The target Q-value for TD is... The calculation method is as follows: In the formula, The target Q value. This refers to the immediate performance reward, which is the performance gain directly calculated and fed back by the network environment after executing the network parameter tuning action recorded in the j-th experience sampled from the experience replay pool. In other words, it's the performance reward calculated and fed back immediately after executing the network parameter tuning action. Subsequently, the network environment directly feeds back the network performance indicators mentioned above. The calculated reward. For online network parameters, These are the target network parameters. For collaborative alignment, it represents the network state. Policy distribution of the lower metric small model Prior policy distribution of large models The degree of difference between them. j represents the j-th experience randomly sampled from the experience replay pool.

[0090] The network parameter adjustment action selected for the online network. This indicates that the online network (with parameters...) is being adjusted. Assess the next network state All possible network parameter adjustment actions The Q value is determined, and the optimal network parameter adjustment action is selected based on this. . The future value evaluated for the target network. This represents the value of the target network (with parameters...) To calculate the optimal network parameter adjustment action selected in the previous step. In network status The actual Q value.

[0091] The loss function is The timing difference error is calculated using this loss function. The predicted Q-value for an online network is determined by the online network based on its network state. and network parameter adjustment actions The calculated value of the KL regularization Q function.

[0092] Periodically (e.g., every C steps), copy the online network parameters to the target network: .

[0093] The update process of the large model is as follows: it can be fine-tuned based on the execution feedback from the small model. For example, by minimizing the KL divergence between the prior policy distribution and the effective policy distribution (i.e., the policy distribution statistically obtained from the empirical replay data), the prior knowledge of the large model is made closer to the policies of the small model that have proven effective in practice, achieving bidirectional adaptation. The loss function of the large model... , Co-alignment is defined as the KL divergence between the prior policy distribution of the large model and the policy distribution of the small model. These are the model parameters for a large model.

[0094] S260 and iterative steps S210-S250 continuously improve the network optimization capability through constant interaction, learning, alignment and fine-tuning until the network performance indicators converge or reach the preset operation and maintenance goals.

[0095] This invention separates the selection of network parameter adjustment actions and value evaluation through a dual-deep Q-network, avoiding the overestimation problem that may occur when using the same network to estimate the maximum network parameter adjustment action value in traditional Q-learning. This makes the Q-value estimation more accurate, thereby improving the agent's decision-making quality and enabling it to learn the optimal policy better. The introduction of KL regularization provides additional constraints on the Q-function update, making the update process smoother and more stable. It prevents excessive fluctuations in the Q-function during the update process, avoiding performance degradation caused by overfitting or unstable updates. At the same time, KL regularization can also maintain the similarity between the Q-function and prior knowledge to a certain extent, enhancing the model's robustness and enabling it to better adapt to different environmental changes.

[0096] The prior policy distribution, based on prior knowledge and experience, represents a reasonable expectation of the agent's behavior. By minimizing the KL divergence between the effective policy distribution and the prior policy distribution, large models can be guided to learn policies that better align with prior knowledge, avoiding the learning of unreasonable or suboptimal policies. This helps improve the performance and reliability of large models in practical applications. The policy distribution of small models is statistically derived from replay data of real-world experiences, reflecting the agent's behavioral patterns in real-world environments. By comparing and refining with the prior policy distribution, large models can fully utilize this real-world data to further optimize their policies, making them more closely aligned with actual application scenarios and improving the practicality and effectiveness of the policies. Combining prior knowledge and real-world experience data to refine the reasoning mechanism of large models enables them to better generalize and adapt to new and unseen environments or situations. Prior knowledge provides general guiding principles, while real-world experience data contains specific environmental information; combining the two helps large models generate more general and robust policies, improving the model's generalization ability.

[0097] Compared to existing large-scale model applications, this invention, through the RRAG mechanism, deeply binds the abstract reasoning capabilities of the large model with concrete and verifiable domain knowledge, suppressing the generation of illusions at their source. Simultaneously, through closed-loop interaction with the small model in real-world environments, it provides continuous real-world testing and correction for the large model's decisions, ensuring that its decisions are both forward-looking and grounded in reality. Compared to existing small-scale model methods, this approach, by introducing the macro-planning capabilities of the large model, frees it from the constraints of local optima, enabling it to exhibit unprecedented adaptability and generalization capabilities in more complex and dynamic network environments.

[0098] Unlike the loose and heuristic model combinations in existing technologies, the collaborative alignment mechanism based on KL divergence proposed in this invention provides a mathematically rigorous solution with a solid information theory foundation for the collaboration between large and small models. It transforms the abstract policy alignment into a quantifiable and optimizable objective, making the collaboration between large and small models no longer a black box, but a transparent, efficient, and stable process, which greatly improves the efficiency and final performance of collaborative optimization.

[0099] Most existing AI methods are purely data-driven, and their capabilities are limited by the quantity and quality of training data. This invention, through the RRAG framework, enables intelligent agents to dynamically utilize external knowledge bases. This means that the system can not only learn from historical interaction data, but also proactively apply structured professional knowledge accumulated by humans for reasoning and decision-making. This dual-driven model of knowledge and data enables better handling of novel, unseen zero-shot problems, representing a key step towards achieving higher levels of network intelligence.

[0100] To fully evaluate the performance and advantages of the method of the present invention, detailed experiments were conducted in a high-fidelity network simulation environment.

[0101] The experiment was set up in a dynamic urban macrocellular network simulation environment, where the number of base stations dynamically varied from 5 to 30, and the number of user devices dynamically varied from 10 to 50, to comprehensively test the algorithm's adaptability under different network densities and load conditions. For model implementation, the large model adopted the Qwen2.5-7B-Instruct model, while the small model was built based on the Double Deep Q-Network (DDQN) algorithm. The knowledge base integrates 3GPP physical layer specifications, common antenna model parameters, publicly available papers in the field of coverage optimization, and typical best practice cases, forming a structured source of professional knowledge. The baseline comparisons include: The invention fully implements the Reasoning-enhanced Retrieval-Augmented Generation - Large and Small Model Collaboration (RRAG-LSMC) method, namely... Figure 3 , Figure 4 , Figure 5 The Proposed in; We employ a simplified collaborative approach using a large model to directly generate natural language suggestions based on internal knowledge, without introducing a reasoning-enhanced Retrieval-Augmented Generation (RRAG) mechanism or KL divergence alignment, and a Double Deep Q-Network + Large Language Model (DDQN+LLM). In addition, there are representative independent reinforcement learning benchmarks such as Dual Deep Q Network (DDQN), Proximal Policy Optimization (PPO), and Advantage Actor-Critic (A2C).

[0102] The experimental parameter settings are summarized in Table 1.

[0103] Table 1 This invention compares the SNR convergence performance under different schemes. (Refer to...) Figure 3 As shown, under fixed scenarios, the method of this invention achieves the fastest convergence speed in terms of average SNR (dB) and throughput, and reaches the highest performance level. It is noteworthy that the DDQN+LLM scheme, due to the illusion of LM (e.g., suggesting adjustment when ARFCN should not be adjusted), actually performs worse than the pure DDQN baseline. This highlights the necessity and effectiveness of the RRAG knowledge injection and KL alignment mechanisms in this invention. The method of this invention achieves approximately 12% performance improvement compared to the baseline DDQN. Step represents the number of training steps or iteration steps.

[0104] This invention also performs network scalability analysis, referring to... Figure 4The figure shows a comparison of coverage (%) under different numbers of base stations. As the number of base stations increases, inter-cell interference intensifies, challenging the performance of all methods. However, the method of this invention exhibits the most gradual performance decline in SNR and throughput, consistently maintaining a leading position. Especially in terms of coverage, when the number of base stations exceeds 25, other baseline methods suffer from coverage collapse due to their inability to effectively handle complex interference relationships, while the method of this invention, benefiting from the global planning capabilities of LM and the reference to interference coordination strategies in the knowledge base, still maintains a high coverage rate of over 80%, demonstrating excellent robustness.

[0105] User load adaptability analysis was also performed. (Refer to...) Figure 5 The figure shows a comparison of average throughput (bps) under different numbers of user devices. Average throughput is the average throughput of all user devices. As the number of user devices increases, network resource competition intensifies. Experimental results show that the method of this invention achieves the optimal balance among the three interdependent metrics of SNR, throughput, and coverage. Other methods either sacrifice one aspect for another (e.g., DDQN+LLM may unilaterally pursue throughput at the expense of SNR) or suffer from overall poor performance. This demonstrates that the collaborative decision-making mechanism of the method of this invention can make more comprehensive and informed resource allocation decisions based on the global load situation.

[0106] The following describes the network coverage optimization device based on size model collaboration provided by the present invention. The network coverage optimization device based on size model collaboration described below and the network coverage optimization method based on size model collaboration described above can be referred to in correspondence.

[0107] The network coverage optimization device based on size model collaboration provided by this invention refers to... Figure 6 As shown, it includes: The task receiving module 310 is used to receive network optimization task instructions and current network status information of the network environment. The large model decision module 320 is used to process the network optimization task instructions and the network state information through the large model's reasoning-based enhanced retrieval generation mechanism, and output the prior strategy distribution for network parameter adjustment. Small model decision module 330 is used to generate network parameter adjustment actions based on the prior policy distribution and the value function that incorporates the collaborative alignment degree. The environment interaction module 340 is used to perform the network parameter adjustment action and collect network environment performance feedback data; The model update module 350 is used to update the decision parameters of the small model based on the performance feedback data and correct the inference mechanism of the large model until the network performance indicators reach the preset target.

[0108] Figure 7An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions from the memory 430 to execute a network coverage optimization method based on a size-model collaborative approach.

[0109] Furthermore, the logical instructions in the aforementioned memory 430 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 the present invention, 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 the present invention. 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.

[0110] On the other hand, the present invention 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 is able to execute the network coverage optimization method based on size model collaboration provided by the above methods.

[0111] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the network coverage optimization method based on size model collaboration provided by the above methods.

[0112] 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.

[0113] 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.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A network coverage optimization method based on size model collaboration, characterized in that, include: Receive network optimization task instructions and current network status information of the network environment; The large model processes the network optimization task instructions and the network state information based on the reasoning-enhanced retrieval generation mechanism, and outputs the prior policy distribution for network parameter adjustment. The small model generates network parameter adjustment actions based on the prior policy distribution and the value function that incorporates the collaborative alignment degree. Perform the network parameter adjustment action and collect network environment performance feedback data; The decision parameters of the small model are updated based on the performance feedback data, and the inference mechanism of the large model is corrected until the network performance indicators reach the preset target.

2. The network coverage optimization method based on size model collaboration according to claim 1, characterized in that, The large model, based on an inference-enhanced retrieval generation mechanism, processes the network optimization task instructions and the network state information, and outputs a priori policy distribution for network parameter adjustment, including: The network optimization task instructions and the network status information are input into the large model together; The large model, based on the network optimization task instructions and the network state information, decomposes the task instructions into multiple sequentially executed reasoning steps through a thought chain; For each reasoning step, a query is generated by combining the network state information and the reasoning context, and a specified number of knowledge documents most relevant to the query are retrieved from the network knowledge base; The retrieved knowledge documents are fused with the context of the current reasoning step to generate knowledge-enhanced reasoning results; Based on the reasoning results enhanced by all knowledge, a structured distribution of the prior policy is generated.

3. The network coverage optimization method based on size model collaboration according to claim 1, characterized in that, The value function that incorporates collaborative alignment is constructed in the following manner: Calculate the KL divergence between the policy distribution of the small model and the prior policy distribution output by the large model, and use it as the co-alignment degree; The collaborative alignment degree is used as a regularization term and introduced into the original value function of the small model to form the value function that incorporates the collaborative alignment degree.

4. The network coverage optimization method based on size model collaboration according to claim 1, characterized in that, The network parameter adjustment action is defined as adjusting at least one of the following parameters of the base station: antenna vertical downtilt angle, antenna horizontal azimuth angle, antenna transmit power, and carrier frequency corresponding to the downlink absolute radio frequency channel number.

5. The network coverage optimization method based on size model collaboration according to claim 4, characterized in that, The performance feedback data is calculated based on the network performance indicators of the user equipment. The network performance indicators are a weighted sum of the signal-to-noise ratio function and the logarithm of the throughput, wherein the signal-to-noise ratio function approaches a saturation value when the user's signal-to-noise ratio exceeds a preset threshold. The signal-to-noise ratio of the user equipment is determined based on the antenna transmit power, the channel gain from the base station to the user equipment, and the thermal noise power of the user equipment. The channel gain is jointly determined by the antenna gain, path loss, and small-scale fading. The antenna gain is a function of the vertical downtilt deviation and the horizontal azimuth deviation. The vertical downtilt deviation is the deviation between the antenna vertical downtilt angle and the optimal antenna vertical downtilt angle from the base station to the user equipment, and the horizontal azimuth deviation is the deviation between the antenna horizontal azimuth angle and the optimal horizontal azimuth angle from the base station to the user equipment.

6. The network coverage optimization method based on size model collaboration according to claim 1, characterized in that, The value function that incorporates the collaborative alignment is a KL-regularized Q function, the value of which is determined by the immediate performance reward, the collaborative alignment penalty, and the expected value of the next state. The decision parameters of the small model are updated using a dual-deep Q-network algorithm, the loss function of which is constructed based on the temporal difference error of the Q-function with KL regularization. The correction of the inference mechanism of the large model includes minimizing the KL divergence between the prior policy distribution and the policy distribution of the small model obtained from the empirical replay data.

7. A network coverage optimization device based on size model collaboration, characterized in that, include: The task receiving module is used to receive network optimization task instructions and current network status information of the network environment; The large model decision module is used to process the network optimization task instructions and the network state information based on the reasoning-enhanced retrieval generation mechanism of the large model, and output the prior policy distribution for network parameter adjustment. The small model decision module is used by the small model to generate network parameter adjustment actions based on the prior policy distribution and the value function that incorporates the collaborative alignment degree. The environment interaction module is used to perform the network parameter adjustment actions and collect network environment performance feedback data; The model update module is used to update the decision parameters of the small model based on the performance feedback data and correct the inference mechanism of the large model until the network performance indicators reach the preset target.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the network coverage optimization method based on size model collaboration as described in any one of claims 1 to 6.

9. 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 network coverage optimization method based on size model collaboration as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the network coverage optimization method based on size model collaboration as described in any one of claims 1 to 6.

Images