A communication network GIS design method based on agentic reinforcement learning

Abstract

The application discloses a communication network GIS design method based on Agentic reinforcement learning, comprising the following steps: 1) converting the GIS design operation in the communication network construction process into a structured executable operation instruction; 2) constructing a data set matched with the GIS design operation and the GIS operation instruction of the communication network, so that a large language model learns the mapping relationship between the communication network design operation and the GIS operation instruction; 3) taking the large language model as the core to construct a GIS operation instruction sequence generation model; 4) taking the large model as an intelligent agent to perform closed-loop training in a real GIS environment; 5) model evaluation, complete training; 6) according to the communication network demand text input by a user, using the optimized large model, a complete GIS operation instruction sequence is automatically generated. The application is suitable for various communication network construction scenes, can automatically generate corresponding GIS design schemes according to different regional conditions and planning requirements, and has good universality and engineering application value.

Description

Technical Field

[0001] This invention relates to communication network technology, and more particularly to a communication network GIS design method based on agentic reinforcement learning. Background Technology

[0002] As an indispensable key infrastructure in modern information society, communication networks have comprehensively covered core areas such as daily life, financial transactions, medical treatment, and transportation, becoming a core pillar supporting social operation and economic development. The construction scale of communication networks has shown explosive growth. At the same time, the rapid rise and large-scale application of emerging technologies such as 5G communication, Internet of Things (IoT), Artificial Intelligence (AI), and cloud computing have made the business scenarios of communication networks increasingly rich. The network architecture is evolving from the traditional centralized structure to a distributed and heterogeneous direction. The number of network nodes and the amount of data transmission are growing exponentially, and the overall complexity of communication networks is increasing dramatically.

[0003] In the early design phase of communication network construction, Geographic Information System (GIS) platforms are currently a widely used core tool in the industry. The core advantage of this platform lies in its ability to integrate geospatial data, including multi-dimensional information such as topographic data, building distribution data, transportation network data, and underground pipeline data. Through spatial analysis algorithms, it can quickly select the optimal fiber optic cable laying path, effectively avoiding complex terrains such as mountains and rivers, as well as densely populated urban areas. This reduces the technical difficulty of construction, shortens the construction period, and controls construction costs. Simultaneously, the GIS platform has powerful annotation capabilities, which can visually annotate key network nodes such as base station locations, fiber optic junction box locations, optical splitter deployment points, and core equipment rooms, forming intuitive network design drawings. This facilitates accurate identification of target locations by construction personnel during on-site operations, improving the precision of construction operations.

[0004] However, despite the important role GIS platforms play in communication network design, the current GIS-based communication network design process still suffers from many prominent problems, severely restricting design efficiency and quality: First, the design process is cumbersome and inefficient. Existing GIS design models require engineers to manually complete a series of operations, including data collection, layer loading, path planning, parameter setting, and drawing, resulting in numerous steps and repetitive tasks. Second, design quality is heavily influenced by human factors, making consistency difficult to guarantee. Significant differences in the professional backgrounds, technical expertise, and practical experience of different engineers lead to substantial variations in the final design schemes regarding path rationality, resource utilization, and cost control. Some designs even contain potential technical vulnerabilities, affecting the stability of the network after its construction. Third, the network's adaptability to expansion and upgrades is poor. Under traditional design models, each network expansion or upgrade requires engineers to conduct extensive manual work, including site surveys, data updates, and path replanning. This not only increases labor and time costs but may also cause network upgrades to lag behind business needs due to excessively long design cycles, impacting user experience.

[0005] The aforementioned problems mean that traditional GIS-based communication network design methods are no longer suitable for the trend of large-scale, complex, and rapidly iterating communication networks. There is an urgent need for an efficient, intelligent, and automated design method to break through the existing bottlenecks. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of existing technologies by providing a communication network GIS design method based on agentic reinforcement learning. By leveraging the powerful data analysis, logical reasoning, and autonomous decision-making capabilities of large models, this method enables the automatic generation of communication network GIS maps, intelligent path planning, and optimized deployment of key nodes, thereby simplifying the design process and improving design efficiency.

[0007] The technical solution adopted by this invention to solve its technical problem is: a communication network GIS design method based on agentic reinforcement learning, comprising the following steps: 1) Transform the GIS design operations during the construction of communication networks into structured, executable operation instructions; 2) Construct a dataset that matches communication network GIS design operations with GIS operation commands, so that the large language model can learn the mapping relationship between communication network design operations and GIS operation commands; 3) Construct a GIS operation command sequence generation model with a large language model as the core. The model input includes network coverage, service user scale, equipment model and specifications, geographical constraints, and engineering constraints. The model output is a sequence of GIS operation commands that can be executed directly. 4) Use the large model as an intelligent agent and perform closed-loop training in a real GIS environment: 5) Set the model termination training conditions, and complete the model parameter convergence and model evaluation; 6) Based on the natural language text of the communication network requirements input by the user, obtain the communication network planning requirements, geospatial constraints and engineering rules input by the model, and use the optimized large model to automatically generate a complete sequence of GIS operation instructions.

[0008] According to the above scheme, in step 1), the GIS design operations in the communication network construction process are abstracted and modeled, and the manual GIS operations, including optical cable path drawing, node connection, and topology construction, are uniformly represented as standardized and structured executable operation instructions.

[0009] According to the above scheme, step 1) specifically includes the following: The transformation of GIS design operations into structured executable instructions is divided into two categories: core processes and auxiliary features. The core processes are used to ensure the integrity of the planning chain, while the auxiliary features are used to improve the accuracy of instruction coding. 1.1) Two-level classification of GIS design operation instructions; GIS design operation instructions include two categories: core processes and auxiliary features. Core process instructions are used to ensure the integrity of the planning process, while auxiliary feature instructions are used to improve the accuracy of instruction coding. 1.1.1) The core process is divided into three main processes according to the construction sequence of the communication network: preliminary analysis, mid-term deployment, and post-construction verification and optimization; The preliminary analysis includes: geospatial analysis, terrain adaptation analysis, and business density analysis; Mid-term deployment includes: base station deployment, fiber optic cable routing, antenna configuration, and network slicing deployment; Post-implementation verification includes: constraint verification, engineering feasibility review, and dynamic interference optimization; 1.1.2) The auxiliary feature design operation includes two dimensions: the characteristics of the operation object and the characteristics of the execution logic; in, Characteristics of the objects being operated on: geographic space, physical equipment, logical resources, and business requirements. Execution logic characteristics: divided into analysis and decision-making, execution configuration, verification and auditing, and optimization and iteration; 1.2) Standardize and normalize the instruction parameters; 1.2.1) Parameter classification and normalization; For continuous parameters: the linear mapping is normalized to the interval [0, 1]; For discrete parameters: one-hot encoding is used; For semi-continuous parameters: normalize first and then retain gradient features.

[0010] 1.2.2) Fixed-length sequence alignment; The longest valid instruction is used as the benchmark to standardize the fixed sequence length; Invalid bits are padded with the terminator, and empty parameter bits are padded with the default value of -3 to adapt to large models with fixed input lengths.

[0011] 1.3) Construct a requirement instruction mapping rule to map the natural language text of the operation requirement into a fixed-length standard instruction sequence; The natural language requirements used in this step include core elements such as coverage area, terrain, equipment configuration, business requirements, and constraints. Define mapping rules: strictly output the structured sequence (instruction name, [parameter set]) corresponding to the natural language requirements according to the early, middle, and late time sequence; Establish a mapping to obtain a standard sequence of fixed length.

[0012] According to the above scheme, in step 4), the GIS system is used to execute the operation instructions generated by the model and return the execution results. The execution results include structured state information and map screenshots. The execution results are automatically identified and judged by a multimodal model. The success of the instruction execution, the spatial rationality of the design results, and the compliance of engineering specifications are transformed into reward signals and fed back to the reinforcement learning framework to drive the model to optimize the instruction generation strategy.

[0013] According to the above scheme, step 4) specifically includes the following: 4.1) Construction of reinforcement learning paradigm: The large model is embedded in a real GIS environment as an intelligent agent with autonomous decision-making ability to construct a closed-loop training paradigm of generating instructions, executing instructions, obtaining feedback and optimizing strategies. 4.2) Multi-source feedback acquisition: After the GIS system executes the command, it synchronously returns two types of feedback information: structured status data and map screenshots; The structured state data is the quantified topology and resource attribute data parsed and output by the GIS system, including node coordinates, line attributes, and topological connection relationships; The map screenshot reflects the spatial layout results, including spatial layout, route direction, and environmental constraint information; 4.3) Multi-level reward evaluation: A multi-level reward algorithm is adopted to generate a comprehensive reward signal, including: format reward for verifying the legality of instruction structure and the completeness of fields, accuracy reward for evaluating the semantic consistency between the instruction and the standard instruction, and actual execution reward for judging the rationality of design and engineering compliance.

[0014] 4.4) Model optimization: Based on the cumulative reward of the complete interaction trajectory, the agent iteratively updates the model parameters through reinforcement learning algorithm, continuously optimizes the instruction generation strategy, and avoids invalid instructions and non-compliant designs.

[0015] The beneficial effects of this invention are: 1. This invention transforms communication network planning and design into the problem of generating GIS operation command sequences, enabling automatic generation of commands from the model and automatic execution by GIS, significantly improving design efficiency and reducing reliance on designers.

[0016] 2. This invention combines Agentic reinforcement learning with a multimodal feedback mechanism, enabling the model to continuously optimize its instruction generation strategy in a real GIS environment, ensuring that the automated design results are engineering-executable and compliant with standards.

[0017] 3. This invention is applicable to various communication network construction scenarios and can automatically generate corresponding GIS design schemes according to different regional conditions and planning requirements, thus possessing good versatility and engineering application value. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] like Figure 1 As shown, a GIS design method for communication networks based on agentic reinforcement learning includes the following steps: 1) Transform the GIS design operations during the construction of communication networks into structured, executable operation instructions; Abstracting and modeling the GIS design operations in the process of communication network construction, and unifying manual GIS operations, including optical cable path drawing, node connection, and topology construction, into standardized and structured executable operation instructions; The transformation of GIS design operations into structured executable instructions is divided into two categories: core processes and auxiliary features. The core processes are used to ensure the integrity of the planning chain, while the auxiliary features are used to improve the accuracy of instruction coding. 1.1) GIS design operation instructions for two-layer classification modeling The GIS design operation includes two categories: core processes and auxiliary features. The core processes are used to ensure the integrity of the planning chain, while the auxiliary features are used to improve the accuracy of instruction coding. 1.1.1) The core process is divided into three main processes according to the construction sequence of the communication network: preliminary analysis, mid-term deployment, and post-construction verification and optimization; The preliminary analysis includes: geospatial analysis, terrain adaptation analysis, and business density analysis; Mid-term deployment includes: base station deployment, fiber optic cable routing, antenna configuration, and network slicing deployment; Post-implementation verification includes: constraint verification, engineering feasibility review, and dynamic interference optimization; The initial analysis commands are positioned as data support for planning and decision-making. Their core function is to collect and analyze basic information about the target area, including geography, business, and environment, to provide quantitative basis for resource deployment. This includes three sub-categories of commands: Geospatial Analysis Commands: Core parameters are latitude and longitude boundaries (e.g., 112.3°-112.6°E, 36.8°-37.1°N, continuous parameters) and minimum received power threshold (-86dBm, semi-continuous parameter), used to define coverage area and signal quality benchmarks; Terrain Adaptation Analysis Commands: Core parameters are terrain type codes (…). Mountainous / plain / water areas correspond to [1, 2, 3], discrete parameters), altitude range (continuous parameters), terrain occlusion coefficient (0.3-0.8, semi-continuous parameters), quantifying the impact of terrain on signal propagation; business density analysis instructions: core parameters are user density threshold (300 people / square kilometer, continuous parameters), business type weight (voice / data / video correspond to [1, 2, 3], discrete parameters), time period distribution coefficient (peak / off-peak / valley correspond to 1.2 / 1.0 / 0.7, semi-continuous parameters), linking geospatial data with business requirements.

[0021] Mid-term deployment commands are positioned as the core implementation stage of planning, with the core function of realizing the spatial configuration and parameter setting of physical equipment and logical resources. They include four sub-categories of commands: Base station deployment commands: core parameters are the equipment's latitude and longitude coordinates (accurate to 6 decimal places, corresponding to 11 cm resolution, continuous parameter), equipment model (5G enhanced macro base station / micro base station Pro corresponds to [1, 2], discrete parameter), and transmit power (45dBm, equipment limited range 35dBm-53dBm, semi-continuous parameter); Optical cable routing instructions: core parameters are the route start / end latitude and longitude (continuous parameter), optical cable core count (216 cores, discrete parameter), and route embedment... Depth (1.2 meters, continuous parameter), obstacle avoidance markers (crossing rivers / railways corresponding to [1, 0], discrete parameter); Antenna configuration instructions: core parameters are antenna model code (Massive MIMO corresponding to 3, discrete parameter), downtilt angle (7.5°, continuous parameter), azimuth angle (120°, continuous parameter), beamwidth (65°, semi-continuous parameter); Network slice deployment instructions: core parameters are slice type (eMBB / uRLLC / mMTC corresponding to [1, 2, 3], discrete parameter), bandwidth allocation (200MHz, continuous parameter), latency threshold (1ms, continuous parameter), connection density (continuous parameter).

[0022] Post-implementation verification and optimization commands are positioned as a closed-loop guarantee for planning quality. Their core function is to verify the compliance and feasibility of deployment plans and dynamically optimize performance. They include three subcategories: Constraint Verification Commands: Core parameters are the verification object ID (base station / optical cable route index, discrete parameter), constraint threshold (minimum distance between base stations 580 meters / restricted area distance 1.3 kilometers, continuous parameter), and terrain loss correction coefficient (7.8dB increase in hilly areas compared to plains, semi-continuous parameter); Engineering Feasibility Review Commands: Core parameters are the construction difficulty coefficient (1.5 / 1.0 for urban / suburban areas, semi-continuous parameter), equipment power supply radius (800 meters, continuous parameter), and maintenance accessibility score (0-10 points, continuous parameter); Dynamic Interference Optimization Commands: Core parameters are the interference source device ID (discrete parameter), interference threshold (-103dBm, semi-continuous parameter), power adjustment step size (2dB, continuous parameter), and number of optimization iterations (5 times, discrete parameter).

[0023] 1.1.2) The auxiliary feature design operation includes two dimensions: the characteristics of the operation object and the characteristics of the execution logic; in, Characteristics of the objects being operated on: geographic space, physical equipment, logical resources, and business requirements. Execution logic characteristics: divided into analysis and decision-making, execution configuration, verification and auditing, and optimization and iteration; Geospatial category: Includes geospatial analysis, terrain adaptation analysis, and fiber optic cable routing instructions. Core parameters focus on spatial attributes such as latitude and longitude, altitude, and terrain. Physical equipment category: includes base station deployment and antenna configuration instructions, with core parameters focusing on physical attributes such as equipment model, power, and hardware specifications; Logical resource category: Includes network slice deployment instructions, with core parameters focusing on virtual resource attributes such as slice type, bandwidth, and latency; Business requirement category: Includes business density analysis commands, with core parameters focusing on demand attributes such as user density, business weight, and service quality; Analysis and decision-making commands: including commands for geospatial analysis, terrain adaptation analysis, and business density analysis; Execution configuration commands: including commands for base station deployment, fiber optic cable routing, antenna configuration, and network slicing deployment; Verification and auditing instructions: including constraint verification and project feasibility audit instructions; Optimization Iteration Class: Includes dynamic interference optimization instructions.

[0024] 1.2) Standardize and normalize the instruction parameters; 1.2.1) Parameter classification and normalization; For continuous parameters (such as latitude and longitude, power, bandwidth, distance): linear mapping normalizes to the [0, 1] interval; For discrete parameters (such as terrain coding, equipment model, and tile type): one-hot coding is used; For semi-continuous parameters (such as occlusion coefficient, tilt angle, loss coefficient): normalize first and then retain gradient features.

[0025] 1.2.2) Fixed-length sequence alignment; The longest valid instruction is used as the benchmark to standardize the fixed sequence length; Invalid bits are padded with the terminator, and empty parameter bits are padded with the default value of -3 to adapt to large models with fixed input lengths.

[0026] 1.2.3) High-precision parameter quantization; The normalized [0, 1] interval parameters are quantized into 224 levels (0~223) at equal intervals; the geographic coordinate quantization error is controlled within 10.5 meters, accurately locking critical constraints such as base station spacing and restricted construction zone distance.

[0027] 1.3) Construct a requirement instruction mapping rule to map the natural language text of the operation requirement into a fixed-length standard instruction sequence; Natural language requirement input specifications: Natural language requirements must include core elements such as coverage area, terrain, equipment configuration, business requirements, and constraints; Define mapping rules: strictly output the structured sequence (instruction name, [parameter set]) corresponding to the natural language requirements according to the early, middle and late time sequence, and add at the end; Establish a mapping to obtain a fixed-length standard sequence. An example of a mapping result is: "(CoverageAnalysis, [112.3, 36.8, 112.6, 37.1, -86, ...]), (TerrainAdaptation, [2, 150, 0.5, ...]), (TrafficDensityAnalysis, [300, 2, 1.2, ...]), (BaseStationPlacement, [112.35, 36.85, 1, 45, ...]), (FiberRouteLaying, [112.4, 36.9, 112.5, 37.0, 216, 1.2, 1, ...]), (ConstraintCheck, [0, 1, 580, ...]), (DynamicInterferenceOptimization, [0, 3, -103, 2, 5, ...]), ( )"; 2) Construct a GIS operation instruction dataset to enable the large language model to learn the mapping relationship between communication network design operations and GIS operation instructions; 3) Construct a GIS operation command sequence generation model with a large language model as the core. The model input includes network coverage, service user scale, equipment model and specifications, geographical constraints, and engineering constraints. The model output is a complete and directly executable GIS operation command sequence. Model architecture selection: Based on the Transformer architecture, the Qwen3-14B large language model is selected as the core foundation, relying on its powerful semantic understanding and long sequence generation capabilities to adapt to multi-dimensional complex requirement parsing and full-process instruction generation tasks. The input is tokenized and encoded by a tokenizer, and then combined with position and type encoding before being incorporated into the model; the output is a structured sequence of GIS operation instructions that can be directly connected to a GIS system. Supervised fine-tuning training: AdamW optimizer was used, with a learning rate of 2×10⁻⁶. -5 With a batch sample size of 32, the model was subjected to 10 epochs of supervised fine-tuning using the LoRA parameter with a rank of 16, enabling the model to quickly learn expert-level GIS operation logic and acquire basic command generation capabilities.

[0028] Based on the analyzed requirements and learned operational logic, the model automatically generates a logically coherent and complete sequence of GIS operation instructions, covering the entire process of geographic analysis, resource deployment, and constraint verification. Command issuance and execution: The generated command sequence can be sent directly to the GIS system without manual intervention, through the sandbox execution program, driving the GIS system to automatically complete operations such as spatial modeling, topology construction, and resource placement of the communication network.

[0029] 4) Use the large model as an agent and conduct closed-loop training in a real GIS environment: The GIS system executes the operation instructions generated by the model and returns the structured state information and map screenshots after execution; the multimodal model automatically identifies and judges the execution results, and transforms the success of instruction execution, the spatial rationality of design results and compliance with engineering specifications into reward signals, which are fed back to the reinforcement learning framework to drive the model to optimize the instruction generation strategy. 4.1) Construction of reinforcement learning paradigm: The large model is embedded in the real GIS environment as an intelligent agent with autonomous decision-making ability to construct a closed-loop training paradigm of "generating instructions → executing instructions → obtaining feedback → policy optimization". 4.2) Multi-source feedback acquisition: After the GIS system executes the command, it synchronously returns two types of feedback information: structured status data (node ​​coordinates, line attributes, topological connection relationships, etc.) and map screenshots (including spatial layout, line direction, environmental constraints, and other visual information). 4.3) Multi-level reward evaluation: The multi-level reward algorithm proposed in this invention is used to generate a comprehensive reward signal: Format validity rewards and field completeness rewards. Regarding format constraints, GIS operation instructions are typically described using a structured DSL (such as GUAL), including operation type, parameter fields, geographic coordinates, and topological objects. If an instruction contains unclosed parentheses, mismatched parameter numbers, or missing required fields, GIS execution will directly fail; therefore, these must be explicitly modeled in the rewards. Format rewards first check structural validity, mathematically expressed as follows: a positive reward is given when the instruction can be correctly parsed; otherwise, a negative reward is given. This can be represented as: (1) To assess whether the instruction provides the necessary parameters, a normalized evaluation of field completeness is performed. If the actual number of fields is... n hit The standard template requires a certain number of fields. n req The completeness reward is then defined as: (2) Accuracy bonus (assessing semantic consistency between the instruction and the standard instruction through regular expression matching); The inputs for this step include three aspects: the overall format reward, the generated GIS operation instructions, and the standard operation instructions. The overall format reward is the reward signal output from step 2, serving as the basis for instruction quality. The generated GIS operation instructions are the same as those in step 1, i.e., the model at the time step... t The generated operation instructions are Standard operating instructions The predefined standard operation instructions in the dataset serve as a reference for matching with the instructions generated by the model.

[0030] The output of this step is the instruction matching degree. And accuracy rewards .

[0031] The model at time step t The generated operation instructions are The standard instructions for the corresponding tasks in the training set are By constructing a set of regular expression templates covering GIS design operations (such as DrawFiberRoute, CreateJointBox, BindConduit, UpdateAttribute, etc.), it is possible to calculate... and The matching degree is obtained with A structural consistency measure. Define a matching score. ∈[0,1], its expression is as follows: (3) In the formula K For the set of key fields contained in standard commands (such as command type, geographic object, coordinate parameters, topological relationship fields, etc.), 1( `)` is an indicator function that takes a value of 1 when the model output successfully matches a field of the standard instruction under its respective regular expression template, and 0 otherwise. Therefore, This reflects the overall consistency of the generated instructions in terms of structure, semantic placeholders, and parameter format.

[0032] To suppress noise from low-matching instructions while encouraging the model to rapidly improve quality within the parsable instruction region, this invention employs a threshold-corrected exponential reward function. When When the value is less than 0.3, the instruction is considered to lack executable meaning and receives zero reward. Once the value exceeds the threshold, the reward increases rapidly in an exponential manner to reflect the value of high-quality instructions in practical engineering design. The reward function is defined as follows: (4) In the formula, α Control the baseline scale of rewards, β Controlling the growth rate of rewards. Through this exponential incentive structure, the model gains stronger motivation to optimize parameter generation, structure filling, and operational semantic expression after meeting basic format requirements, thereby gradually approaching expert-level GIS design instructions during reinforcement learning.

[0033] Actual implementation reward (based on GIS implementation results and Qwen3-VL multimodal visual assessment to determine the rationality of the design and the compliance of the project); The inputs for this step are: accuracy bonus and generated GIS operation instructions. Accuracy Bonus The reward from step 3 serves as feedback on the semantic accuracy of the instructions. The generated GIS operation instructions are the same as those from step 1, meaning the model is at the same time step. t The generated operation instructions are .

[0034] The output of this step: 1) Rewards based on the execution results of the GIS system. If the instruction is executed successfully, a positive reward is returned; if the execution fails, a negative reward is returned.

[0035] 2) Visual large-scale model evaluation results: Evaluation results based on multimodal visual models (such as Qwen3-VL) provide feedback on the design rationality and engineering effectiveness after instruction execution. Higher rewards are given if the design meets the expected engineering requirements; lower or negative rewards are given if problems exist.

[0036] Specifically, during reinforcement learning training, the policy model generates GIS operation instructions based on the current network planning state. These instructions include, but are not limited to, drawing fiber optic cable paths, connecting nodes, establishing topology relationships, and configuring attributes. The generated GIS operation instructions are sent to the GIS system in the sandbox environment for actual execution to obtain real execution results.

[0037] When the GIS system successfully executes the operation command, it first awards a positive reward to the policy model based on the command's executability. If the command cannot be parsed or fails to execute in the GIS system, a negative reward is directly awarded, and the subsequent evaluation of the current decision step is terminated. In this way, the actual execution results of the GIS system are mapped to the basic reward signal for reinforcement learning, effectively constraining the policy model to output unexecutable or semantically incomplete commands, thus avoiding wasted computational resources caused by invalid commands during training.

[0038] Model optimization: Based on the cumulative rewards of the complete interaction trajectory, the agent iteratively updates the model parameters through reinforcement learning algorithms, continuously optimizes the instruction generation strategy, and avoids invalid instructions and non-compliant designs.

[0039] 5) Model evaluation, terminate training; The current training round terminates when the training process reaches the preset maximum number of decision steps, the policy model generates an end command, or the multimodal evaluation module determines that the planning objective has been achieved. Subsequently, the policy model uses a reinforcement learning algorithm to iteratively update the model parameters based on the cumulative reward of the entire interaction trajectory to maximize long-term rewards.

[0040] 6) Based on the input communication network planning requirements, geospatial constraints, and engineering rules, the optimized large model is used to automatically generate a complete sequence of GIS operation instructions. Once trained, the model can directly receive user planning requirements and automatically complete the entire process of requirement analysis, instruction generation, instruction execution, and result optimization without human intervention, thus achieving full automation of the communication network GIS design process and adapting to various communication network construction scenarios.

[0041] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A GIS design method for communication networks based on agentic reinforcement learning, characterized in that, Includes the following steps: 1) Transform the GIS design operations during the construction of communication networks into structured, executable operation instructions; 2) Construct a dataset that matches communication network GIS design operations with GIS operation commands, so that the large language model can learn the mapping relationship between communication network design operations and GIS operation commands; 3) Construct a GIS operation command sequence generation model with a large language model as the core. The model input includes network coverage, service user scale, equipment model and specifications, geographical constraints, and engineering constraints. The model output is a sequence of GIS operation commands that can be executed directly. 4) Use the large model as an intelligent agent and perform closed-loop training in a real GIS environment: 5) Set the model termination training conditions, and complete the model parameter convergence and model evaluation; 6) Based on the natural language text of the communication network requirements input by the user, obtain the communication network planning requirements, geospatial constraints and engineering rules input by the model, and use the optimized large model to automatically generate a complete sequence of GIS operation instructions.

2. The GIS design method for communication networks based on agentic reinforcement learning according to claim 1, characterized in that, In step 1), the GIS design operations in the communication network construction process are abstracted and modeled, and the manual GIS operations, including optical cable path drawing, node connection, and topology construction, are uniformly represented as standardized and structured executable operation instructions.

3. The GIS design method for communication networks based on agentic reinforcement learning according to claim 1, characterized in that, In step 1), the specific details are as follows: 1.1) GIS design operation instructions for two-layer classification modeling; The GIS design operations are converted into structured executable operation instructions, which are divided into two categories: core process instructions and auxiliary feature instructions. Core process instructions are used to ensure the integrity of the planning chain, while auxiliary feature instructions are used to improve the accuracy of instruction coding. 1.1.1) The core process is divided into three main processes according to the construction sequence of the communication network: preliminary analysis, mid-term deployment, and post-construction verification and optimization; The preliminary analysis includes: geospatial analysis, terrain adaptation analysis, and business density analysis; Mid-term deployment includes: base station deployment, fiber optic cable routing, antenna configuration, and network slicing deployment; Post-implementation verification includes: constraint verification, engineering feasibility review, and dynamic interference optimization; 1.1.2) The auxiliary feature design operation includes two dimensions: the characteristics of the operation object and the characteristics of the execution logic; in, Characteristics of the objects being operated on: geographic space, physical equipment, logical resources, and business requirements. Execution logic characteristics: divided into analysis and decision-making, execution configuration, verification and auditing, and optimization and iteration; 1.2) Standardize and normalize the instruction parameters; 1.2.1) Parameter classification and normalization; 1.2.2) Fixed-length sequence alignment; 1.3) Construct a requirement instruction mapping rule to map the natural language text of the operation requirement into a fixed-length standard instruction sequence; Natural language requirements include core elements such as coverage area, terrain, equipment configuration, business requirements, and constraints. Define mapping rules: strictly follow the early, middle, and late time sequence to output a structured sequence of natural language requirements in the form of (instruction name, [parameter set]); Establish a mapping to obtain a standard sequence of fixed length.

4. The GIS design method for communication networks based on agentic reinforcement learning according to claim 3, characterized in that, In step 1.2), the parameter classification and normalization are performed as follows: For continuous parameters: the linear mapping is normalized to the interval [0, 1]; For discrete parameters: one-hot encoding is used; For semi-continuous parameters: normalize first and then retain gradient features.

5. The GIS design method for communication networks based on agentic reinforcement learning according to claim 1, characterized in that, In step 3), the Transformer architecture large language model is selected as the base, and LoRA parameters are used for efficient fine-tuning. The GIS operation command sequence generated by the model is directly sent to the GIS system through the sandbox program to automatically complete spatial modeling, node placement and topology construction.

6. The GIS design method for communication networks based on agentic reinforcement learning according to claim 1, characterized in that, In step 4), the GIS system is used to execute the operation instructions generated by the model and return the execution results, which include structured status information and map screenshots. By using a multimodal model to automatically identify and judge the execution results, the success of instruction execution, the spatial rationality of the design results, and the compliance with engineering specifications are transformed into reward signals, which are fed back to the reinforcement learning framework to drive the model to optimize the instruction generation strategy.

7. The GIS design method for communication networks based on agentic reinforcement learning according to claim 6, characterized in that, In step 4), the specific details are as follows: 4.1) Construction of reinforcement learning paradigm: The large model is embedded in a real GIS environment as an intelligent agent with autonomous decision-making ability to construct a closed-loop training paradigm of generating instructions, executing instructions, obtaining feedback and optimizing strategies. 4.2) Multi-source feedback acquisition: After the GIS system executes the command, it synchronously returns two types of feedback information: structured status data and map screenshots; The structured state data is the quantified topology and resource attribute data parsed and output by the GIS system, including node coordinates, line attributes, and topological connection relationships; The map screenshot reflects the spatial layout results, including spatial layout, route direction, and environmental constraint information; 4.3) Multi-level reward evaluation: A multi-level reward algorithm is adopted to generate a comprehensive reward signal, including: format reward for verifying the legality of the instruction structure and the completeness of the fields, accuracy reward for evaluating the semantic consistency between the instruction and the standard instruction, and actual execution reward for judging the rationality of the design and the compliance of the project. 4.4) Model optimization: Based on the cumulative reward of the complete interaction trajectory, the agent iteratively updates the model parameters through reinforcement learning algorithm, continuously optimizes the instruction generation strategy, and avoids invalid instructions and non-compliant designs.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images