Communication resource arrangement method and system based on big language model collaboration

Abstract

The invention discloses a communication resource arrangement method and system based on large language model cooperation. The method comprises the following steps: constructing a multi-layer wireless intelligent communication network architecture; based on a multilayer wireless intelligent communication network architecture, expert knowledge in the communication field is obtained, a plurality of large language models are embedded, and a large language model cooperation mechanism is constructed; generating a traditional wireless communication network data processing algorithm and an intelligent data processing algorithm, and constructing an algorithm resource library; based on a multi-layer wireless intelligent communication network architecture, task initiation, analysis and generation are carried out in combination with an algorithm resource library, and a communication network resource arrangement scheme is obtained; and issuing the communication network resource arrangement scheme to a near real-time control layer for task execution, thereby realizing communication resource arrangement. According to the invention, the rationality of resource allocation of the wireless communication system can be enhanced, and the system efficiency of the wireless communication system is improved. The communication resource arrangement method and system based on large language model collaboration can be widely applied to the technical field of mobile communication.

Description

Technical Field

[0001] This invention relates to the field of mobile communication technology, and in particular to a communication resource orchestration method and system based on large language model collaboration. Background Technology

[0002] With the approach of the 6G era, achieving ubiquitous intelligence through the deep integration of communication and artificial intelligence has become a core development direction. Artificial intelligence technology is gradually permeating all levels of the network ecosystem. Currently, the field of native AI networks has achieved two key advancements: First, communication networks have significantly improved service quality / experience by deploying a large number of well-trained AI models to implement resource orchestration schemes. Second, communication networks themselves are evolving into a core component of AI systems. This paradigm shift from "AI empowering communication networks" to "communication networks supporting AI" is reshaping the technological landscape. Notably, the breakthrough progress of large language models, represented by ChatGPT, with its unique natural language interaction and multimodal data processing capabilities, provides new opportunities to improve the performance and interactivity of 6G networks. Existing technical approaches can be divided into two categories: one is training task-oriented AI for specific functions, and the other is utilizing large language models to handle the full-scenario needs of 6G. However, both methods have significant drawbacks in scenarios with limited computing and storage resources: the former suffers from insufficient generalization ability and faces high retraining costs for new tasks, while the latter is difficult to deploy effectively at the edge due to its resource-intensive nature. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a communication resource orchestration method and system based on large language model collaboration, which can enhance the rationality of resource allocation in wireless communication systems and improve the system efficiency of wireless communication systems.

[0004] The first technical solution adopted in this invention is: a communication resource orchestration method based on large language model collaboration, comprising the following steps:

[0005] Determine the functions of nodes in a wireless network and their computing and storage capabilities, and construct a multi-layered wireless intelligent communication network architecture by combining open-source wireless access network protocol standards and wireless access network intelligent controllers.

[0006] Based on a multi-layer wireless intelligent communication network architecture, knowledge from experts in the field of communication is acquired, several large language models are embedded, and a large language model collaboration mechanism is constructed.

[0007] Based on the large language model collaboration mechanism, traditional wireless communication network data processing algorithms and intelligent data processing algorithms are generated, and an algorithm resource library is constructed.

[0008] Based on a multi-layered wireless intelligent communication network architecture, and combined with an algorithm resource library, a communication network resource orchestration scheme is obtained by initiating, parsing, and generating tasks.

[0009] The communication network resource orchestration scheme is distributed to the near real-time control layer for task execution, thereby realizing communication resource orchestration.

[0010] Furthermore, the multi-layered wireless intelligent communication network architecture specifically includes service management and orchestration, a near real-time control layer, a wireless access network, and terminals, wherein:

[0011] The service management and orchestration are deployed on a central server with massive computing and storage resources. It is used to retrieve the required communication knowledge from the expert knowledge base and extract the corresponding indicators of the algorithm from the algorithm resource library through the collaboration of multiple language models, and finally generate a resource orchestration scheme.

[0012] The near real-time control layer is configured with an algorithm resource library for business execution;

[0013] The wireless access network includes several wireless access points;

[0014] The terminal consists of several user terminals, each with computing and storage resources. Each user terminal can transmit its own needs to the wireless access network via the air interface or obtain the results of service execution from the wireless access network through semantic encoding or semantic decoding.

[0015] Furthermore, the step of acquiring expert knowledge in the communication field, embedding several large language models, and constructing a large language model collaboration mechanism based on a multi-layer wireless intelligent communication network architecture specifically includes:

[0016] Based on the multi-layer wireless intelligent communication network architecture, several large language models are selected and the task types that each language model is built for are determined as the functional bias of the large language model.

[0017] Based on the evaluation of several selected large language models, one large language model was determined as a general agent, and the remaining large language models were determined as special agents.

[0018] Acquire knowledge from experts in the communications field and build an expert knowledge base;

[0019] The general agent is used to retrieve the expert knowledge base. Based on the functional bias of each language model, the user terminal's needs are divided into several sub-tasks. These sub-tasks are assigned to dedicated agents, which process the corresponding sub-tasks and return the corresponding sub-task processing solutions, thus building a collaborative mechanism for large language models.

[0020] Furthermore, the step of generating traditional wireless communication network data processing algorithms and intelligent data processing algorithms based on the large language model collaboration mechanism, and constructing an algorithm resource library, specifically includes:

[0021] Based on the code generation capability of the large language model collaboration mechanism, generate data processing algorithms for traditional wireless communication networks.

[0022] Based on the large language model collaboration mechanism, an AI network model is constructed, training and validation datasets are generated for iterative optimization, and intelligent data processing algorithms are generated.

[0023] An algorithm resource library is constructed by combining traditional wireless communication network data processing algorithms with intelligent data processing algorithms.

[0024] Furthermore, the step of generating traditional wireless communication network data processing algorithms based on the code generation capability of the large language model collaboration mechanism specifically includes:

[0025] The principle description of traditional wireless communication network data processing algorithms and the programming language type and format requirements are input into the large language model through the application programming interface of the large language model, and the preliminary traditional wireless communication network data processing algorithm is output.

[0026] Acquire test data and input it into the large language model. Copy the preliminary traditional wireless communication network data processing algorithm into the corresponding compilation software, compile and execute it, and obtain the compilation and execution results.

[0027] If the compilation and execution results in an error, the specific error information is extracted and input into the large language model through the application programming interface of the large language model. The model is then recompiled and executed until the compilation and execution results are displayed normally, thus generating a traditional wireless communication network data processing algorithm.

[0028] Furthermore, the step of constructing an AI network model based on a large language model collaboration mechanism, generating training and validation datasets for iterative optimization, and generating an intelligent data processing algorithm specifically includes:

[0029] Obtain the AI ​​network model code from the intelligent data processing algorithm, download and import the code into the code compilation software, and build the AI ​​network model at the software level;

[0030] Configure parameters for the AI ​​network model to generate training and testing datasets;

[0031] The AI ​​network model is trained using a training dataset to obtain the training output results;

[0032] Set the loss function for training the AI ​​network model, take the training output and the labels in the training dataset as inputs, input them into the loss function for calculation, and backfeed the calculation result to train the AI ​​network model to obtain the trained AI network model.

[0033] The trained AI network model is tested using a test dataset to obtain the test output results;

[0034] Set the test accuracy function for the trained AI network model, take the test output and the labels in the test dataset as inputs, and input them into the test accuracy function for calculation. Repeat the training and testing steps of the AI ​​network model until the calculation result of the test accuracy function meets the preset requirements, and generate an intelligent data processing algorithm.

[0035] Furthermore, the step of initiating, parsing, and generating a communication network resource orchestration scheme based on a multi-layered wireless intelligent communication network architecture and in conjunction with an algorithm resource library specifically includes:

[0036] Based on a multi-layer wireless intelligent communication network architecture, a general agent running in the non-real-time control layer of the central server collects service requests from user terminals and status information of near-real-time control layer nodes.

[0037] The business request and status information are sent to a dedicated agent. Based on the expert knowledge base, the corresponding professional knowledge is matched through semantic matching to generate several extended prompt words with similar semantic information.

[0038] Several extended prompt words with similar semantic information are sent back to the general agent. Through a multi-dimensional semantic matching mechanism, relevant knowledge is extracted from the expert knowledge base to generate corresponding communication services.

[0039] Based on the large language model collaboration mechanism, and through the hierarchical knowledge filtering mechanism, the communication service is divided into several sub-tasks and assigned to a dedicated agent for parsing to obtain the sub-task processing solution.

[0040] The subtask processing scheme is evaluated. If the subtask processing scheme is not accepted, it is sent back to the dedicated agent for multi-dimensional semantic matching again until the subtask processing scheme is accepted. Based on the order of the subtasks, a directed acyclic graph is generated to represent the corresponding communication network resource orchestration scheme.

[0041] Furthermore, the step of distributing the communication network resource orchestration scheme to the near real-time control layer for task execution, thereby realizing communication resource orchestration, specifically includes:

[0042] The non-real-time control layer distributes the communication network resource orchestration scheme to the near real-time control layer for task execution and obtains the processing results of the original services;

[0043] Feedback and evaluation of the processing results of the original business are provided and sent to the large language model for optimization, and directions for improvement of several sub-tasks are obtained;

[0044] Several sub-tasks are improved by sending directions to a dedicated agent for multi-dimensional semantic matching until the user terminal no longer provides feedback on the final result of the received business processing, thus achieving communication resource orchestration.

[0045] The second technical solution adopted in this invention is: a communication resource orchestration system based on large language model collaboration, comprising:

[0046] The first module is used to determine the functions of nodes in the wireless network and their computing and storage capabilities, and to construct a multi-layer wireless intelligent communication network architecture by combining open-source wireless access network protocol standards and wireless access network intelligent controllers.

[0047] The second module is used to acquire expert knowledge in the field of communication based on a multi-layer wireless intelligent communication network architecture, embed several large language models, and build a large language model collaboration mechanism.

[0048] The third module is used to generate traditional wireless communication network data processing algorithms and intelligent data processing algorithms based on the large language model collaboration mechanism, and to build an algorithm resource library.

[0049] The fourth module is used to initiate, parse, and generate tasks based on a multi-layer wireless intelligent communication network architecture and in conjunction with an algorithm resource library, so as to obtain a communication network resource orchestration scheme.

[0050] The fifth module is used to distribute the communication network resource orchestration scheme to the near real-time control layer for task execution, thereby realizing communication resource orchestration.

[0051] The beneficial effects of the method and system of this invention are as follows: By determining the functions of nodes in a wireless network and their computing and storage capabilities, and combining open-source wireless access network protocol standards and wireless access network intelligent controllers, this invention constructs a multi-layered wireless intelligent communication network architecture. Based on the availability of computing and storage resources at each level, it can deploy corresponding decision-making units more efficiently and flexibly, and determine service execution nodes more rationally. It can adapt to various communication environment requirements and has strong portability. Furthermore, based on a multi-layer wireless intelligent communication network architecture, it acquires expert knowledge in the communication field, embeds several large language models, and constructs a large language model collaboration mechanism. It can construct the target service type based on the large language model, and purposefully exert its potential in specific fields, thereby reducing the service carrying pressure of a single large language model and more fully stimulating the reasoning potential of the large language model. Based on the large language model collaboration mechanism, it generates traditional wireless communication network data processing algorithms and intelligent data processing algorithms, and constructs an algorithm resource library. Finally, based on the multi-layer wireless intelligent communication network architecture and combined with the algorithm resource library, it performs task initiation, parsing, and generation to obtain a communication network resource orchestration scheme, realizes the intelligent orchestration strategy of the communication network, significantly improves the intelligence level of the wireless communication network, enhances the rationality of wireless communication system resource allocation, and improves the system efficiency of the wireless communication system. Attached Figure Description

[0052] Figure 1 This is a flowchart of the steps of a communication resource orchestration method based on large language model collaboration according to the present invention;

[0053] Figure 2 This is a structural block diagram of a communication resource orchestration system based on large language model collaboration according to the present invention;

[0054] Figure 3 This is a schematic diagram illustrating the system average and rate improvement obtained by the communication network resource orchestration method using large language model collaboration, provided in a specific embodiment of the present invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0056] Reference Figure 1 This invention provides a communication resource orchestration method based on large language model collaboration, the method comprising the following steps:

[0057] S100. Determine the functions of nodes in the wireless network and their computing and storage capabilities, and construct a multi-layer wireless intelligent communication network architecture by combining open-source wireless access network protocol standards and wireless access network intelligent controllers.

[0058] It should be noted that the multi-layered wireless intelligent communication network architecture specifically includes service management and orchestration, a near real-time control layer, a wireless access network, and terminals. The service management and orchestration is deployed in a central server with massive computing and storage resources. It is used to retrieve the required communication knowledge from an expert knowledge base and extract the corresponding indicator from an algorithm resource library through collaboration of multiple language models, ultimately generating a resource orchestration scheme. The near real-time control layer is configured with an algorithm resource library for service execution. The wireless access network includes several wireless access points. The terminals consist of several user terminals, each with computing and storage resources. They can upload their own needs to the wireless access network via the air interface or obtain the results of service execution from the wireless access network through semantic encoding or semantic decoding.

[0059] In this embodiment, the functions of nodes in the wireless network and their computing and storage capabilities are determined. Based on the open-source Radio Access Network (O-RAN) protocol standard and the Radio Access Network Intelligent Controller (RIC), a multi-layered wireless intelligent communication network architecture is constructed. This multi-layered wireless intelligent communication network architecture includes four communication layers: Service Management and Orchestration (SMO), Near-Real-Time RIC (near-RT RIC), RAN, and terminals. The SMO is implemented in a central server with massive computing and storage resources. It is used to retrieve the necessary communication knowledge from an expert knowledge base and extract the corresponding indicator from the algorithm resource library through multi-language model collaboration, ultimately generating a resource orchestration scheme. The aforementioned multi-language model collaboration is implemented in the non-real-time RIC (non-RT RIC) module within the SMO. The near-RT RIC is equipped with an algorithm resource library and has service execution capabilities. The RAN includes K = 5 radio access points (APs). The terminals consist of U = 3 user terminals, each with limited computing and storage resources. They can transmit their needs to the RAN via the air interface or obtain the service execution results from the RAN through semantic encoding or semantic decoding. Under this framework, the AP can upload terminal requests to the near-RT RIC via the standard E2 interface in ORAN and to the non-RT RIC via the standard O1 interface. The near-RT RIC can receive resource orchestration schemes issued by the SMO from the A1 interface, extract the corresponding algorithms from the algorithm resource library according to the algorithm indicator to implement service, or send the corresponding algorithms to the AP or UE for local execution.

[0060] Among them, the knowledge of experts in the field of communications includes communication signal data collected from real communication environments, synthetic data generated using the nrCDLChannel function of MATLAB software based on the CDL channel specified in the 3GPP38.901 protocol, communication protocol rules formulated by the internationally recognized international communications organization 3GPP, and industry white papers formulated by internationally renowned companies.

[0061] S200, based on a multi-layer wireless intelligent communication network architecture, acquires expert knowledge in the field of communication, embeds several large language models, and constructs a large language model collaboration mechanism.

[0062] Specifically, based on a multi-layered wireless intelligent communication network architecture, several large language models are selected, and the task types that each language model is designed for are determined, serving as the functional bias of that large language model. The selected large language models are evaluated, and one large language model is selected as a general agent, while the remaining large language models serve as dedicated agents. Expert knowledge in the communication field is acquired to construct an expert knowledge base. The general agent is used to retrieve information from the expert knowledge base. Based on the functional biases of each language model, the user terminal's needs are divided into several sub-tasks, which are then assigned to dedicated agents. The dedicated agents process the corresponding sub-tasks and return the corresponding sub-task processing solutions, thus constructing a large language model collaboration mechanism.

[0063] In this embodiment, existing 5GSpecAnalyzerAI, GPT-4-based, GPT-3-based, Gemini-based, and Llama3-8B-based algorithms are used as core modules for generating communication network decisions. By leveraging collaboration among multiple large language models, the reasoning ability of each model is maximized, ultimately generating a communication resource orchestration scheme. Simultaneously, traditional wireless communication network data processing algorithms, including minimum mean square error estimation algorithms, zero-forcing algorithms, and the latest intelligent data processing algorithms from academia and industry, including deep learning-based signal angle of arrival estimation algorithms and deep learning-based digital and analog joint beamforming algorithms, are used as communication resources in the multi-layered wireless communication network architecture, providing a usable algorithm resource library for the execution of the final communication network resource orchestration scheme. The collaboration mechanism among multiple large language models and the construction method of the algorithm resource library are as follows: the collaboration mechanism among multiple large language models consists of general agents and dedicated agents. To construct this collaborative mechanism, several large language models were first selected from existing large language models as collaborative decision-making modules in the multi-language model collaboration mechanism, including 5GSpecAnalyzerAI, GPT-4-based, GPT-3-based, Gemini-based, and Llama3-8B-based. Simultaneously, the task types targeted by each major language model were determined, serving as the functional bias of the major language model. Specifically, GPT-4-based was used for communication signal angle of arrival estimation, Gemini-based for generating and solving optimization equations, Llama3-8B-based for handling linear algebra and precoding implementation issues, and GPT-3-based for enhancing prompt words. Next, an evaluation was conducted within the selected major language models, comprehensively considering language processing capabilities and cross-domain knowledge mastery. 5GSpecAnalyzerAI was selected as the general agent, while the remaining major language models were designated as specialized agents. Then, the functions of the general and specialized agents were determined. Specifically, the general agent needed to retrieve user requirements from the expert knowledge base, categorizing them into signal angle of arrival estimation, channel estimation, precoding, and beamforming. Based on the functional bias descriptions of the major language models, subtasks were assigned to GPT-4-based, Gemini-based, and Llama3-8B-based, respectively. The specialized agents processed the corresponding subtasks and returned corresponding subtask processing solutions. These solutions included the subtask processing flow and the specific algorithms used at each step. The general agent 5GSpecAnalyzerAI finally summarizes the task processing schemes generated by the dedicated agent to generate the final communication network resource orchestration scheme.

[0064] The aforementioned expert knowledge in the communications field includes communication signal data collected from real-world communication environments, synthetic data generated using methods such as digital twin technology, ray tracing technology, and generative artificial intelligence technology, communication protocol rules established by recognized international communications organizations, and industry white papers developed by internationally renowned companies. This expert knowledge in the communications field can be obtained from open-source platforms on the Internet, vectorized using general database management software, and ultimately stored in a vector database, serving as an important basis for subsequent multi-language model collaborative decision-making resource orchestration schemes.

[0065] S300, based on the large language model collaboration mechanism, generates traditional wireless communication network data processing algorithms and intelligent data processing algorithms, and builds an algorithm resource library;

[0066] First, it's important to clarify that the algorithm resource library provides algorithm sources for each subtask in the communication network resource orchestration scheme generated by the large language model. To construct this library, the algorithm resources are first divided into traditional wireless communication network data processing algorithms and the latest intelligent data processing algorithms researched in academia and industry. This division is based on whether artificial intelligence is used as the core processing unit of the algorithm. Second, for traditional wireless communication network data processing algorithms, corresponding implementation schemes and code can be found in existing communication system simulation software and stored in the algorithm resource library. For intelligent data processing algorithms, corresponding implementation schemes and code can be found in open-source code repositories on the internet and stored in the algorithm resource library. Then, each algorithm stored in the resource library is indexed and numbered, and a functional description of the algorithm is generated based on its target task and the data type it processes. Finally, these functional descriptions are pre-input into each dedicated agent as prompts and background knowledge during the subtask processing scheme generation process.

[0067] Based on whether artificial intelligence (AI) is used as the core processing unit of the algorithm, traditional wireless communication network data processing algorithms are divided into those researched in academia and industry and those that are the latest intelligent data processing algorithms. The specific algorithms to be implemented for each are then determined. Traditional wireless communication network data processing algorithms include minimum mean square error estimation algorithms, while intelligent data processing algorithms include deep learning-based signal angle of arrival estimation algorithms and deep learning-based digital and analog joint beamforming algorithms. S310 utilizes code generation capabilities based on a large language model collaborative mechanism to generate traditional wireless communication network data processing algorithms.

[0068] Specifically, the principle description of the traditional wireless communication network data processing algorithm, the programming language type and its format requirements are input into the large language model through the application programming interface of the large language model, and a preliminary traditional wireless communication network data processing algorithm is output. Test data is obtained and input into the large language model. The preliminary traditional wireless communication network data processing algorithm is copied to the corresponding compilation software for compilation and execution, and the compilation and execution results are obtained. If there are errors in the compilation and execution results, the specific error information is extracted and input into the large language model through the application programming interface of the large language model for recompilation and execution until the compilation and execution results are displayed normally, thus generating the traditional wireless communication network data processing algorithm.

[0069] In this embodiment, the code generation capabilities of the Gemini-based large language model are utilized to generate the required traditional wireless communication network data processing algorithm. This includes inputting a description of the principle of the required traditional wireless communication network data processing algorithm into the large language model via its API. This description states, "With the core objective of finding a matrix G such that the received data GY is as close as possible to the original data x, and combining fundamental knowledge of linear algebra and matrix operations, generate code for the Minimum Mean Square Error Estimation (MMSE) algorithm." The required programming language description and its format requirements are also provided, namely, "The code language is Python and must conform to Python code format requirements." Furthermore, test data is generated using industry- and academically recognized algorithm platforms and data generation modules. The algorithm platform used is the MATLAB simulation platform, and the data generation module is the nrCDLChannel built-in library function. The parameters of this function are set to the parameters in step S100, and the other parameters are kept at default settings. Running the function code on the MATLAB platform generates test data. The generated test data is then used as input, and the traditional wireless communication network data processing algorithm code generated by the large language model is copied to the corresponding compiler software for compilation and execution. Finally, if the implemented traditional wireless communication network data processing algorithm runs normally and produces the corresponding output results, the algorithm is considered successfully generated; otherwise, the specific error information during code execution is extracted and used as input, fed into the large language model via the large language model API interface, and the model is required to regenerate the algorithm code. This process is repeated until the generated algorithm code runs normally and outputs results. Since the MMSE algorithm implemented in this study is relatively general, the implemented algorithm can pass the test and does not need to be regenerated.

[0070] S320: Based on the large language model collaboration mechanism, construct an AI network model, generate training and validation datasets for iterative optimization, and generate intelligent data processing algorithms.

[0071] Specifically, the process involves: acquiring the AI ​​network model code from the intelligent data processing algorithm; downloading and importing the code into code compilation software to build the AI ​​network model at the software level; configuring the parameters of the AI ​​network model to generate training and testing datasets; training the AI ​​network model using the training dataset to obtain the training output results; setting the loss function for training the AI ​​network model, taking the training output results and labels from the training dataset as inputs to the loss function for calculation, and then feeding the calculation results back to train the AI ​​network model to obtain the trained AI network model; testing the trained AI network model using the testing dataset to obtain the test output results; setting the test accuracy function for the trained AI network model, taking the test output results and labels from the test dataset as inputs to the test accuracy function for calculation, and repeating the training and testing steps of the AI ​​network model until the calculation result of the test accuracy function meets the preset requirements, thus generating the intelligent data processing algorithm.

[0072] In this embodiment, the process of finding and building an AI network model, generating training and validation datasets, and customizing the required intelligent data processing algorithm includes the following steps:

[0073] S321. Find the latest AI network model code implemented in intelligent data processing algorithms from open-source platforms or open-source algorithm libraries on the Internet, download or import the code into the corresponding code compilation software, and then build the AI ​​network model at the software level. Specifically, for the signal angle of arrival estimation algorithm based on deep learning, the Visual Transformer (ViT) model from the torchvision timm library in Python is used as the AI ​​network model, with its input "image_size" set to 64, "patch_size" set to 8, "channel" set to 2, and the output "num_classes" of the last layer set to 3. Similarly, the digital and analog joint beamforming algorithm based on deep learning is implemented using the same ViT network model, requiring its input "image_size" to be set to 64, "patch_size" to be set to 8, "channel" to be set to 2, and the output "num_classes" of the last layer set to 64.

[0074] S322. Utilizing industry- and academically recognized algorithm platforms and data generation modules, create the corresponding simulation environment at the software level and set the corresponding parameters. The algorithm platform used is the MATLAB simulation platform, and the data generation module is the nrCDLChannel built-in library function. The parameters of this function are set to the parameters in step S100, while the remaining parameters remain at their default settings. After configuration, run the software code to generate 100,000 training data points and 20,000 test data points, constructing training and validation datasets. Each data point in these two datasets includes the original generated data and its corresponding output label. The original generated data is the function output, and the output label is the signal angle of arrival and the corresponding signal beam of the function output.

[0075] S323. Customize the training of the constructed AI network model using the training dataset. Specifically, input each original generated data in the training dataset into the constructed AI network model and generate the corresponding output.

[0076] S324. Set the loss function required for training the AI ​​network model. This loss function is defined as the mean squared error (MSE) between the output of the ViT network model and the output label. Then, take the corresponding output generated in S323 and the corresponding output label in the training dataset as the input of this loss function, and use the calculated result of the function as the loss for training the AI ​​network model. Train the AI ​​network through the backpropagation mechanism.

[0077] S325. Customize the test of the built AI network model using the test dataset. Specifically, input each original generated data in the test dataset into the built AI network model and generate the corresponding output.

[0078] S326. Set the test accuracy function required for AI network model validation. This test accuracy function is also defined as the MSE between the output and output label of the ViT network model. The smaller the MSE value, the higher the accuracy. Then, use the corresponding output generated in S325 and the corresponding output label in the test dataset as input to this test accuracy function, and use the calculation result of the function as the test accuracy of the AI ​​network model training.

[0079] S327. Repeat steps S323 to S326 for 500 rounds, and select the AI ​​network model corresponding to the result with the highest test accuracy in the verification results as the final intelligent data processing algorithm model, thereby obtaining the required intelligent data processing algorithm.

[0080] S330 combines traditional wireless communication network data processing algorithms with intelligent data processing algorithms to construct an algorithm resource library.

[0081] S400, based on a multi-layer wireless intelligent communication network architecture, combined with an algorithm resource library, performs task initiation, parsing, and generation to obtain a communication network resource orchestration scheme;

[0082] Specifically, based on a multi-layered wireless intelligent communication network architecture, a general agent running in the non-real-time control layer of the central server collects service requests from user terminals and status information from near-real-time control layer nodes. The service requests and status information are sent to a dedicated agent, which, based on an expert knowledge base, matches corresponding professional knowledge through semantic matching to generate several extended prompt words with similar semantic information. These extended prompt words are then sent back to the general agent, where a multi-dimensional semantic matching mechanism extracts relevant knowledge from the expert knowledge base to generate corresponding communication services. Based on a large language model collaboration mechanism, a hierarchical knowledge filtering mechanism divides the communication services into several sub-tasks and assigns them to dedicated agents for parsing to obtain sub-task processing solutions. The sub-task processing solutions are evaluated; if an unacceptable solution is not found, it is sent back to the dedicated agent for re-multi-dimensional semantic matching until an acceptable solution is found. Based on the order of the sub-tasks, a directed acyclic graph is generated to represent the corresponding communication network resource orchestration scheme.

[0083] In this embodiment, during the service initiation phase, the general agent 5GSpecAnalyzerAI running on the non-RT RIC of the central server collects service requests from user terminals and status information from near-RT RIC nodes. This includes the terminal device using semantic encoding technology to vectorize user requests for higher throughput to watch 4K ultra-high-definition video, extracting corresponding semantic features through a neural network model to generate a structured task description. This description needs to include the service type—improving user throughput, a description of the user's required service quality (i.e., high QoS), and the user's required service experience (i.e., high QoE). This task description, along with the real-time status information of the near-RT RIC nodes, is input to the general agent as prompt words. Real-time status information for RIC nodes needs to include, but is not limited to, the remaining available computing and storage resources and the current working status of the nodes. The general agent 5GSpecAnalyzerAI sends the received task description and the real-time status information of the near-RTRIC nodes to the dedicated agent GPT-3-based that processes the prompt words. Based on these vectorized prompt words, the agent matches the corresponding professional knowledge from the vector database storing expert knowledge through semantic matching and generates multiple extended prompt words with similar semantic information. These extended prompt words and the original prompt words together constitute the context of knowledge retrieval and are returned to the general agent 5GSpecAnalyzerAI. The general agent extracts relevant knowledge from the expert knowledge vector database through a multi-dimensional semantic matching mechanism and generates the corresponding communication services.

[0084] In the service analysis phase, this invention, based on the collaboration of multiple language models and employing a hierarchical knowledge filtering mechanism, divides the task into signal angle of arrival estimation, channel estimation, precoding, and beamforming, and assigns them sequentially to GPT-4-based, Gemini-based, and Llama3-8B-based systems. After the general agent 5GSpecAnalyzerAI generates the corresponding communication service, it further divides the communication service into continuous sub-tasks of signal angle of arrival estimation, channel estimation, precoding, and beamforming based on relevant knowledge extracted from the expert knowledge vector database, and distributes them sequentially to GPT-4-based, Gemini-based, and Llama3-8B-based systems. Each dedicated agent receives a vectorized representation of the task content and performance requirements from the general agent. It then performs a multi-dimensional semantic matching mechanism in the expert knowledge vector database to obtain relevant knowledge as background information for decision-making. Based on this knowledge, it retrieves the corresponding execution algorithms from the algorithm resource library: a deep learning-based signal angle of arrival estimation algorithm, a minimum mean square error estimation algorithm for channel estimation, and a deep learning-based digital and analog joint beamforming algorithm for precoding and beamforming. Each dedicated agent returns a subtask processing plan to the general agent, presented as a tuple including a description of the task content and an index of the selected execution algorithm.

[0085] Task generation phase: After receiving the sub-task processing schemes generated by each dedicated agent, the general agent 5GSpecAnalyzerAI selects the sub-task processing schemes provided and generates a directed acyclic graph to represent the corresponding final generated communication network resource orchestration scheme.

[0086] S500 distributes the communication network resource orchestration scheme to the near real-time control layer for task execution, thereby realizing communication resource orchestration.

[0087] Specifically, the non-real-time control layer distributes the communication network resource orchestration scheme to the near-real-time control layer for task execution and obtains the processing results of the original services; it then provides feedback and evaluation on the processing results of the original services and distributes them to the large language model for optimization, obtaining directions for improvement of several sub-tasks; finally, it distributes the directions for improvement of several sub-tasks to the dedicated agent for multi-dimensional semantic matching, until the user terminal no longer provides feedback on the final results of the received service processing, thus realizing communication resource orchestration.

[0088] In this embodiment, after generating the directed acyclic graph (DAG) representing the service, the general agent distributes the task to the near-RT RIC node via a standardized AI interface for subsequent task execution. The process is as follows: First, after receiving the DAG representing the service, the near-RT RIC node reads the descriptions of the signal angle of arrival estimation, channel estimation, precoding, and beamforming subtasks and their corresponding algorithm index values ​​in the order specified by the DAG. Then, based on the algorithm index values, it extracts the corresponding algorithms from the algorithm database and executes the corresponding algorithms in the near-RT RIC. After the near-RT RIC node has processed all the subtasks in sequence, it distributes the results to the RAN via the E2 interface. The RAN then distributes the final processing result of the original service to the user terminal via the air interface. After receiving the final result of the service processing, the user terminal can evaluate and provide feedback on the result according to its own needs. In this example, the user does not provide feedback on the final result of the service processing, indicating that the communication network resource orchestration scheme has been finally accepted and the service execution has been completed.

[0089] Define the system average sum rate as:

[0090]

[0091] where R sum The unit is bits per second per hertz. To perform the averaging operation, Let U be the set of user terminals, with a total of U = 3 user terminals B. u,k =1 represents the bandwidth allocated by the k-th AP to the u-th user, SINR u This represents the ratio of the received signal energy to the interference and noise energy for the u-th user. (Refer to...) Figure 3 The resource orchestration method provided by this invention, with the collaboration of K=5 AP nodes and multiple large language models, achieves a significant improvement in sum rate compared to traditional AI-free schemes combining minimum mean square error estimation and zero-forcing algorithms, as well as schemes based on a single large language model, using system average sum rate as a metric. Furthermore, it more closely approximates the ideal situation. This demonstrates that the communication network resource orchestration method based on large language model collaboration proposed in this invention has a more reasonable resource orchestration scheme and can achieve higher system efficiency.

[0092] Therefore, the embodiments of the present invention have the following advantages compared with the prior art:

[0093] 1) The wireless communication system adopts a multi-level wireless communication architecture consisting of SMO, near-RT RIC, RAN, and terminals. This architecture allows for more efficient and flexible deployment of corresponding decision-making units and more rational determination of service execution nodes based on the available computing and storage resources at each level. It can adapt to various communication environment requirements and has strong portability.

[0094] 2) In response to the working mode and potential of large language models, this invention constructs a collaborative architecture for large language models, which can be based on the business type to be addressed and can give full play to its potential in specific fields, thereby reducing the business burden of individual large language models and more fully stimulating the reasoning potential of large language models.

[0095] 3) By adopting a combination of large language model collaboration mechanism and multi-level wireless communication architecture, an intelligent orchestration strategy for communication networks is realized, which greatly improves the intelligence level of wireless communication networks, enhances the rationality of wireless communication system resource allocation, and improves the system efficiency of wireless communication systems.

[0096] Reference Figure 2 A communication resource orchestration system based on large language model collaboration, comprising:

[0097] The first module 201 is used to determine the functions of nodes in the wireless network and their computing and storage capabilities, and to construct a multi-layer wireless intelligent communication network architecture by combining open source wireless access network protocol standards and wireless access network intelligent controllers.

[0098] The second module 202 is used to acquire expert knowledge in the field of communication based on a multi-layer wireless intelligent communication network architecture, embed several large language models, and build a large language model collaboration mechanism.

[0099] The third module 203 is used to generate traditional wireless communication network data processing algorithms and intelligent data processing algorithms based on the large language model collaboration mechanism, and to build an algorithm resource library.

[0100] The fourth module 204 is used to initiate, parse, and generate tasks based on a multi-layer wireless intelligent communication network architecture and in conjunction with an algorithm resource library, so as to obtain a communication network resource orchestration scheme.

[0101] The fifth module 205 is used to distribute the communication network resource orchestration scheme to the near real-time control layer for task execution, thereby realizing communication resource orchestration.

[0102] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0103] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A communication resource orchestration method based on large language model collaboration, characterized in that, Includes the following steps: Determine the functions of nodes in a wireless network and their computing and storage capabilities, and construct a multi-layered wireless intelligent communication network architecture by combining open-source wireless access network protocol standards and wireless access network intelligent controllers. Based on a multi-layer wireless intelligent communication network architecture, knowledge from experts in the field of communication is acquired, several large language models are embedded, and a large language model collaboration mechanism is constructed. Based on the large language model collaboration mechanism, traditional wireless communication network data processing algorithms and intelligent data processing algorithms are generated, and an algorithm resource library is constructed. Based on a multi-layered wireless intelligent communication network architecture, and combined with an algorithm resource library, a communication network resource orchestration scheme is obtained by initiating, parsing, and generating tasks. The communication network resource orchestration scheme is distributed to the near real-time control layer for task execution, thereby realizing communication resource orchestration.

2. The communication resource orchestration method based on large language model collaboration according to claim 1, characterized in that, The multi-layered wireless intelligent communication network architecture specifically includes service management and orchestration, a near real-time control layer, a wireless access network, and terminals, wherein: The service management and orchestration are deployed on a central server with massive computing and storage resources. It is used to retrieve the required communication knowledge from the expert knowledge base and extract the corresponding indicators of the algorithm from the algorithm resource library through the collaboration of multiple language models, and finally generate a resource orchestration scheme. The near real-time control layer is configured with an algorithm resource library for business execution; The wireless access network includes several wireless access points; The terminal consists of several user terminals, each with computing and storage resources. Each user terminal can transmit its own needs to the wireless access network via the air interface or obtain the results of service execution from the wireless access network through semantic encoding or semantic decoding.

3. The communication resource orchestration method based on large language model collaboration according to claim 2, characterized in that, The step of acquiring expert knowledge in the communication field, embedding several large language models, and constructing a large language model collaboration mechanism based on a multi-layer wireless intelligent communication network architecture specifically includes: Based on the multi-layer wireless intelligent communication network architecture, several large language models are selected and the task types that each language model is built for are determined as the functional bias of the large language model. Based on the evaluation of several selected large language models, one large language model was determined as a general agent, and the remaining large language models were determined as special agents. Acquire knowledge from experts in the communications field and build an expert knowledge base; The general agent is used to retrieve the expert knowledge base. Based on the functional bias of each language model, the user terminal's needs are divided into several sub-tasks. These sub-tasks are assigned to dedicated agents, which process the corresponding sub-tasks and return the corresponding sub-task processing solutions, thus building a collaborative mechanism for large language models.

4. The communication resource orchestration method based on large language model collaboration according to claim 3, characterized in that, The step of generating traditional wireless communication network data processing algorithms and intelligent data processing algorithms based on the large language model collaboration mechanism, and constructing an algorithm resource library, specifically includes: Based on the code generation capability of the large language model collaboration mechanism, generate data processing algorithms for traditional wireless communication networks. Based on the large language model collaboration mechanism, an AI network model is constructed, training and validation datasets are generated for iterative optimization, and intelligent data processing algorithms are generated. An algorithm resource library is constructed by combining traditional wireless communication network data processing algorithms with intelligent data processing algorithms.

5. The communication resource orchestration method based on large language model collaboration according to claim 4, characterized in that, The step of generating traditional wireless communication network data processing algorithms based on the code generation capability of the large language model collaboration mechanism specifically includes: The principle description of traditional wireless communication network data processing algorithms and the programming language type and format requirements are input into the large language model through the application programming interface of the large language model, and the preliminary traditional wireless communication network data processing algorithm is output. Acquire test data and input it into the large language model. Copy the preliminary traditional wireless communication network data processing algorithm into the corresponding compilation software, compile and execute it, and obtain the compilation and execution results. If the compilation and execution results in an error, the specific error information is extracted and input into the large language model through the application programming interface of the large language model. The model is then recompiled and executed until the compilation and execution results are displayed normally, thus generating a traditional wireless communication network data processing algorithm.

6. The communication resource orchestration method based on large language model collaboration according to claim 5, characterized in that, The step of constructing an AI network model based on a large language model collaboration mechanism, generating training and validation datasets for iterative optimization, and generating an intelligent data processing algorithm specifically includes: Obtain the AI ​​network model code from the intelligent data processing algorithm, download and import the code into the code compilation software, and build the AI ​​network model at the software level; Configure parameters for the AI ​​network model to generate training and testing datasets; The AI ​​network model is trained using a training dataset to obtain the training output results; Set the loss function for training the AI ​​network model, take the training output and the labels in the training dataset as inputs, input them into the loss function for calculation, and backfeed the calculation result to train the AI ​​network model to obtain the trained AI network model. The trained AI network model is tested using a test dataset to obtain the test output results; Set the test accuracy function for the trained AI network model, take the test output and the labels in the test dataset as inputs, and input them into the test accuracy function for calculation. Repeat the training and testing steps of the AI ​​network model until the calculation result of the test accuracy function meets the preset requirements, and generate an intelligent data processing algorithm.

7. The communication resource orchestration method based on large language model collaboration according to claim 6, characterized in that, The step of initiating, parsing, and generating a communication network resource orchestration scheme based on a multi-layer wireless intelligent communication network architecture and in conjunction with an algorithm resource library specifically includes: Based on a multi-layer wireless intelligent communication network architecture, a general agent running in the non-real-time control layer of the central server collects service requests from user terminals and status information of near-real-time control layer nodes. The business request and status information are sent to a dedicated agent. Based on the expert knowledge base, the corresponding professional knowledge is matched through semantic matching to generate several extended prompt words with similar semantic information. Several extended prompt words with similar semantic information are sent back to the general agent. Through a multi-dimensional semantic matching mechanism, relevant knowledge is extracted from the expert knowledge base to generate corresponding communication services. Based on the large language model collaboration mechanism, and through the hierarchical knowledge filtering mechanism, the communication service is divided into several sub-tasks and assigned to a dedicated agent for parsing to obtain the sub-task processing solution. The subtask processing scheme is evaluated. If the subtask processing scheme is not accepted, it is sent back to the dedicated agent for multi-dimensional semantic matching again until the subtask processing scheme is accepted. Based on the order of the subtasks, a directed acyclic graph is generated to represent the corresponding communication network resource orchestration scheme.

8. The communication resource orchestration method based on large language model collaboration according to claim 7, characterized in that, The step of distributing the communication network resource orchestration scheme to the near real-time control layer for task execution, thereby realizing the communication resource orchestration, specifically includes: The non-real-time control layer distributes the communication network resource orchestration scheme to the near real-time control layer for task execution and obtains the processing results of the original services; Feedback and evaluation of the processing results of the original business are provided and sent to the large language model for optimization, and directions for improvement of several sub-tasks are obtained; Several sub-tasks are improved by sending directions to a dedicated agent for multi-dimensional semantic matching until the user terminal no longer provides feedback on the final result of the received business processing, thus achieving communication resource orchestration.

9. A communication resource orchestration system based on large language model collaboration, characterized in that, Includes the following modules: The first module is used to determine the functions of nodes in the wireless network and their computing and storage capabilities, and to construct a multi-layer wireless intelligent communication network architecture by combining open-source wireless access network protocol standards and wireless access network intelligent controllers. The second module is used to acquire expert knowledge in the field of communication based on a multi-layer wireless intelligent communication network architecture, embed several large language models, and build a large language model collaboration mechanism. The third module is used to generate traditional wireless communication network data processing algorithms and intelligent data processing algorithms based on the large language model collaboration mechanism, and to build an algorithm resource library. The fourth module is used to initiate, parse, and generate tasks based on a multi-layer wireless intelligent communication network architecture and in conjunction with an algorithm resource library, so as to obtain a communication network resource orchestration scheme. The fifth module is used to distribute the communication network resource orchestration scheme to the near real-time control layer for task execution, thereby realizing communication resource orchestration.

Images