A core network control method and a core network system

Abstract

The application discloses a core network control method and a core network system, relates to the technical field of communication, and is applied to a core network system with an intelligent agent structure and comprises the following steps: receiving requirements from the outside about network control, performance optimization or management; performing task type identification on the requirements by a task management unit; judging whether there is an intelligent agent-based strategy corresponding to the task, verified and in a frozen state according to the identification result; if not, triggering an intelligent decision process to generate a new intelligent agent-based strategy, and setting the verified strategy to a frozen release state; performing deterministic inference and control operation based on the strategy in the frozen release state to realize real-time closed-loop control of the core network; wherein the method realizes decoupling of the intelligent decision process and real-time control operation through the intelligent agent structure. In this way, the core network control can be ensured to be low in time delay, predictable and high in reliability, strategy unified management multiplexing is realized, and network stability is improved.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a core network control method and core network system. Background Technology

[0002] Traditional network control and management typically rely on manual configuration or rule-driven automation, resulting in weak intelligence and autonomy. This makes it difficult to cope with the challenges posed by network expansion, diverse service models, and highly dynamic operating environments. Looking towards future evolution, networks need lower latency, higher reliability, and stronger determinism, along with highly intelligent adaptive capabilities to meet the application demands of new intelligent services, integrated air-space-ground systems, and complex scenarios. Traditional networks are no longer adequate in terms of decision-making flexibility, cross-domain collaboration, and dynamic optimization capabilities. While multi-agent network architectures can improve network autonomy and collaboration through intelligent cooperation, and some solutions introduce large language models to enhance the reasoning and decision-making capabilities of agents, these models, based on probabilistic generation mechanisms, suffer from uncertain results, computational complexity, and high latency. Directly applying them to network control poses operational risks, highlighting significant shortcomings in real-time performance and determinism in existing architectures. These limitations fail to meet the future network requirements for high reliability, low latency, and stable operation. Summary of the Invention

[0003] The purpose of this invention is to provide a core network control method and a core network system that can ensure low latency, predictability, and high reliability of core network control, achieve unified policy management and reuse, and improve network stability.

[0004] To address the aforementioned technical problems, this invention provides a core network control method. The method is applied to a core network system with an intelligent agent structure, and includes: Receive requests from external sources regarding network control, performance optimization, or management; The task management unit identifies the task type of the requirements. Based on the task type identification result, determine whether there is an agent-based strategy that corresponds to the task, has been verified, and is in a frozen state. If it does not exist, the intelligent decision-making process is triggered to generate a new agent-based policy, the new agent-based policy is verified, and the verified policy is set to the frozen release state. Based on the strategy in a frozen release state, deterministic inference and control operations are performed to achieve real-time closed-loop control of the core network; The method decouples the intelligent decision-making process from the real-time control operation through the intelligent agent structure.

[0005] To address the aforementioned technical problems, the present invention also provides a core network system implementing the above-described core network control method, comprising: The requirement input layer is used to receive external requirements regarding network control, performance optimization, or management. The task management unit is used to identify the task type of the requirements. The decision unit is used to determine, based on the task type identification result, whether there is an agent-based strategy that corresponds to the task, has been verified, and is in a frozen state; if not, it triggers an intelligent decision-making process to generate a new agent-based strategy, verifies the new agent-based strategy, and sets the verified strategy to a frozen release state; based on the strategy in the frozen release state, it performs deterministic inference and control operations to achieve real-time closed-loop control of the core network. The core network system decouples the intelligent decision-making process from real-time control operations through a hierarchical intelligent agent structure.

[0006] The beneficial effects of this invention are as follows: The core network control method provided by this invention receives external network control, performance optimization, or management requests in a core network system with an intelligent agent structure. The task management unit identifies the task type and determines whether there is a corresponding verified and frozen intelligent agent policy. When no available frozen policy is available, intelligent decision-making is triggered to generate a new policy, which is then verified and set to a frozen release state. Based on the frozen policy, deterministic inference and control are executed to achieve real-time closed-loop control of the core network. Furthermore, by decoupling the intelligent decision-making process from the real-time control operation through the intelligent agent structure, complex inference models can be avoided from directly entering the real-time control link, ensuring low latency, predictability, and repeatability of network control. This enables unified policy management and reuse, allowing the network to support both rapid modeling of new tasks and rapid invocation of known tasks. It balances intelligent decision-making capabilities with deterministic real-time control requirements, effectively reduces the computational complexity of real-time decision-making, avoids the uncertainty risks brought by probabilistic models, and significantly improves the stability, reliability, autonomy, and engineering deployability of the core network in complex dynamic environments. It better meets the comprehensive requirements of the network for real-time performance, stability, and intelligent collaboration, and has high engineering practicality and application value.

[0007] In addition, the present invention also provides a corresponding core network system for the core network control method, which has the same or corresponding technical features as the core network control method mentioned above, and has the same effect. Attached Figure Description

[0008] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart of the core network control method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the core network system provided in an embodiment of the present invention. Detailed Implementation

[0010] Fifth-generation mobile communication systems and their enhanced versions (5G-Advanced, 5GA) have been widely deployed. The 5GA core network, based on service-oriented architecture, network function virtualization, and software-defined networking technologies, achieves network function decoupling and flexible deployment, supporting access and service requirements for various service scenarios. However, the control and management mechanisms of traditional 5GA core networks are still mainly based on manual configuration or rule-based automation, with limited intelligence and autonomy, making it difficult to cope with the challenges brought by network expansion, diversified service forms, and highly dynamic operating environments. For sixth-generation mobile communication systems (6G), the core network needs to support communication services with lower latency, higher reliability, and stronger determinism, and possess high intelligence and adaptability to meet the needs of native AI services, integrated air-space-ground networks, and ultra-complex service scenarios. Traditional 5GA core networks are no longer sufficient to meet the requirements of 6G development in terms of decision-making flexibility, cross-domain collaboration capabilities, and dynamic optimization capabilities. To enhance the autonomy and collaboration capabilities of the core network, a multi-agent core network architecture can abstract different network functions or control tasks in the core network into multiple agents, achieving network control and management through collaboration between these agents. However, some solutions attempt to introduce Large Language Models (LLMs) to enhance the reasoning and decision-making capabilities of agents. Since LLMs are based on probabilistic generation mechanisms, their decision results are non-deterministic, and the reasoning process is computationally complex and has high latency, making it difficult to meet the stringent real-time and deterministic requirements of core network control. Directly applying them to core network control carries significant risks. Therefore, current LLM-based multi-agent core network architectures have significant shortcomings in terms of real-time performance and determinism, and cannot yet meet the requirements of 6G core networks for high reliability, low latency, and predictable operation. To address these issues, this invention provides a core network control method that can improve the intelligence level of the core network and ensure the real-time and deterministic operation of the core network.

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0012] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0013] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] An embodiment of the present invention provides a core network control method. Figure 1 This is a flowchart of a core network control method provided in an embodiment of the present invention. This method is applied to a core network system with an intelligent agent structure, such as... Figure 1 As shown, the method includes: S101, Receive requests from external sources regarding network control, performance optimization, or management.

[0015] It should be noted that the executing entity of this invention is a core network system with an intelligent agent structure. In this invention, the core network system is a system that adopts an intelligent agent architecture to decouple the intelligent decision-making process from the implementation of control operations, and performs deterministic execution of the verified and frozen intelligent agent strategy to complete the network closed-loop control.

[0016] Figure 2 This is a schematic diagram of the core network system provided in an embodiment of the present invention. Figure 2 As shown, the demand input layer 1 can receive demands from external sources regarding network control, performance optimization, or management. Here, "external" refers to the demand side, which can include UE terminals, application service providers, developers, etc.

[0017] S102. The task type is identified through the task management unit.

[0018] In implementation, such as Figure 2 As shown, the task management unit (also known as the task management center) 2 can be used to identify the task type of the requirements received by the requirement input layer 1, obtain the task type identification result, and then schedule the corresponding functional layer to process the task and provide feedback on the task execution progress to the requester.

[0019] S103. Based on the task type identification result, determine whether there is an agent-based strategy that corresponds to the task, has been verified, and is in a frozen state.

[0020] In implementation, this invention, based on task type identification results, determines whether there are agent-based strategies that match the current task, have been verified, and are in a frozen state. This enables deterministic and rapid reuse of known tasks and intelligent decision-making for unknown tasks. On the one hand, calling only frozen and verified strategies ensures that the control logic is predictable, repeatable, and has low latency, directly avoiding the uncertainty and computational overhead introduced by unverified strategies and ensuring the stability and reliability of the real-time control link. On the other hand, by accurately matching task types and frozen strategies, unified management and efficient retrieval of strategies can be achieved.

[0021] S104. If it does not exist, trigger the intelligent decision-making process to generate a new agent-based policy, verify the new agent-based policy, and set the verified policy to the frozen release state.

[0022] In practice, when there is no matching freeze strategy, this invention generates a new agent-based strategy by triggering an intelligent decision-making process, verifies it, and sets the verified strategy to a frozen release state. This provides intelligent strategy generation capability for unknown or new tasks, ensuring the flexibility and scalability of the core network in response to dynamic needs. Furthermore, through the forced verification and freeze release mechanism, it ensures that the new strategy is safe, reliable, meets performance standards, and has a deterministic execution logic, thus preventing untested strategies from entering the real-time control link from the source.

[0023] S105. Based on the strategy in the frozen release state, perform deterministic inference and control operations to achieve real-time closed-loop control of the core network; wherein, the core network control method decouples the intelligent decision-making process from the real-time control operation through the agent structure.

[0024] In implementation, this invention relies on the strategy of freezing the release state to execute deterministic inference and control operations. From the execution level, it ensures that the core network control logic has no random fluctuations, no additional computational overhead, and predictable latency, thereby achieving highly reliable, low-latency real-time closed-loop control. At the same time, the intelligent agent structure decouples the intelligent decision-making process from the real-time control operation, allowing non-real-time processes such as complex reasoning, strategy generation and verification to run independently. Only mature and determined frozen strategies are used for field control, which not only preserves the intelligent iteration and optimization capabilities of the core network, but also completely isolates the impact of decision uncertainty on real-time control, thus ensuring the stability, determinism and real-time performance of the core network control from the architecture.

[0025] In the core network control method provided by the embodiments of the present invention, external network control, performance optimization, or management requests are received in a core network system with an intelligent agent structure. The task management unit identifies the task type and determines whether there is a corresponding verified and frozen intelligent agent policy. When no available frozen policy is available, intelligent decision-making is triggered to generate a new policy, which is then verified and set to a frozen release state. Deterministic inference and control are then performed based on the frozen policy to achieve real-time closed-loop control of the core network. Furthermore, by decoupling the intelligent decision-making process from the real-time control operation through the intelligent agent structure, complex inference models can be avoided from directly entering the real-time control link, ensuring low latency, predictability, and repeatability of network control. This enables unified management and reuse of policies, allowing the network to support both rapid modeling of new tasks and rapid invocation of known tasks. It balances intelligent decision-making capabilities with deterministic real-time control requirements, effectively reduces the computational complexity of real-time decision-making, avoids the uncertainty risks brought by probabilistic models, and significantly improves the stability, reliability, autonomy, and engineering deployability of the core network in complex dynamic environments. This better meets the network's comprehensive requirements for real-time performance, stability, and intelligent collaboration, and has high engineering practicality and application value.

[0026] Furthermore, in specific implementations, in the core network control method provided in the embodiments of the present invention, such as... Figure 2 As shown, the intelligent agent structure may include: Cognitive Layer 3, Shadow Execution Layer 4, and Execution Layer 5; Cognitive Layer 3 is deployed in the cloud, while Shadow Execution Layer 4 and Execution Layer 5 are deployed at the edge. Execution Layer 5 runs in parallel with the main path of Shadow Execution Layer 4.

[0027] In the above steps, the intelligent decision-making process and real-time control operation are decoupled through the intelligent agent structure. Specifically, this may include: using the cognitive layer 3 to generate a basic model structure or policy template; using the shadow execution layer 4 to optimize candidate policies based on the basic model structure or policy template, verifying the candidate policies, and setting the verified policies to a frozen release state; using the execution layer 5 to load the verified and frozen policy versions, performing deterministic inference based on the network running state, and generating control actions so that the underlying tool layer can execute corresponding network control operations according to the control actions.

[0028] It should be noted that, as Figure 2As shown, the core network system architecture of this invention can be divided into three main layers from top to bottom: a demand input layer 1, a two-layer intelligent agent decision layer, and a bottom-level tool layer 6. The two-layer intelligent agent decision layer includes a task management unit 2, a model repository, a cognitive layer 3, a shadow execution layer 4, and an execution layer 5. The demand input layer 1 can submit control, optimization, or management requests to the network, such as performance optimization goals, policy adjustment requests, or anomaly handling requests. After receiving external demands, the task management unit 2 can identify the task type (such as a new type of task or an existing type of task), decide whether to schedule the cognitive layer 3 for modeling or directly call the execution layer 5 to freeze the policy, and provide feedback on task progress and results, realizing a unified task entry point and scheduling logic for multiple services and scenarios. When a new type of task modeling request is received from the task management unit 2, the cognitive layer 3 generates a basic model structure and stores it in the model repository. The shadow execution layer 4 optimizes the policy parameters of the basic model structure, generates candidate policies, and submits them to the model repository. When the execution layer 5 receives a strategy issuance command or strategy invocation request from the task management unit 2, it loads the frozen strategy from the model repository, performs deterministic inference based on the network operating status, and generates control actions. The underlying tool layer 6 receives the control actions issued by the execution layer 5 and executes the corresponding network control operations. The model repository can store basic models, fine-tuning models, task templates, and frozen / verified strategy versions. "Frozen" means the strategy version is fixed, the parameters are locked, and it cannot be modified arbitrarily, but it can be invoked by the execution layer 5. "Verified" means the strategy version has passed verification and meets the usage requirements. The model repository supports strategy template reuse and cross-task migration, reducing the cost of repetitive modeling; it supports frozen release of strategy versions, making the inference path of the execution layer 5 stable and controllable; and it supports strategy consistency verification, scope management, and rollback, making strategy iteration engineering controllable.

[0029] This invention employs a cloud-edge collaboration: Cognitive Layer 3 is deployed in the cloud, suitable for handling complex reasoning and model generation; Shadow Execution Layer 4 and Execution Layer 5 are both deployed at the edge to meet low-latency control requirements. Cognitive Layer 3 consists of multiple Language Model-based Agents (LLM-based Agents). Shadow Execution Layer 4 and Execution Layer 5 each consist of multiple Reinforcement Learning-based Agents (RL-based Agents). Execution Layer 5 primarily includes the Primary RL-based Agent. Cognitive Layer 3 is responsible for "cognitive decision-making," including task understanding, policy generation, initial model design, and policy updates. When encountering a new type of task for the first time, Cognitive Layer 3 performs reasoning, generates a basic model for task solving, and stores the results in a structured form in a model library. It can be considered the "brain" of the network but does not directly participate in real-time control, only generating policy templates and basic model structures. In addition, Cognitive Layer 3 also performs incremental optimization or transfer learning on the basic model based on execution logs, environmental feedback, and Key Performance Indicators (KPIs) from Execution Layer 5, achieving "collective experience accumulation." Shadow Execution Layer 4 does not directly control the network, but based on the same state input and the policy template model provided by Cognition Layer 3, it can use online reinforcement learning or rapid fine-tuning to optimize the policy and generate candidate policies for subsequent iterations. Simulation verification is performed on the candidate policies. If the optimized policy consistently outperforms the current version over a long period, a policy freeze and version release are issued after meeting the conditions. Simultaneously, a safety boundary configuration for the fine-tunable parameters of Execution Layer 5 is defined. In other words, Shadow Execution Layer 4 can be responsible for online optimization and, after verification, enter the freeze and release process. Execution Layer 5 can be responsible for "real-time adaptive decision-making," including traffic scheduling, resource allocation, and QoS guarantees. It directly controls the network, loads "frozen / verified" policy versions, performs deterministic inference, and outputs control actions. During execution, it does not explore policies, but the main execution agent can make deterministic parameter adjustments within the preset safety boundary, based on input parameters and environmental feedback, without changing the policy structure or introducing randomness. Furthermore, Execution Layer 5 also needs to perform performance monitoring and assurance, mainly including collecting performance KPIs, monitoring environmental feedback and system performance. If performance decline occurs, a policy version rollback mechanism is immediately activated. Finally, the execution logs, environmental feedback, and performance KPIs of all main executing agents need to be fed back to the cognitive layer 3 for continuous optimization.

[0030] This invention confines the LLM-based Agent with probabilistic generation characteristics to the non-real-time cognitive layer 3, which only undertakes offline / non-real-time tasks such as policy generation, model modeling, and policy updates. The real-time closed loop is executed by the PrimaryRL-based Agent deployed at the edge, which executes the frozen policy version and performs deterministic inference. This avoids high-complexity inference models from directly entering the real-time control path, ensuring that the network control link has low latency, predictability, and repeatability.

[0031] Furthermore, in specific implementation, in the core network control method provided in the embodiments of the present invention, step S102 identifies the task type of the requirement through the task management unit, which may specifically include: after identifying the requirement as a new type of task, sending a new type of task modeling request to the cognitive layer to trigger the process of generating a basic model structure or strategy template; after identifying the requirement as an existing type of task, sending a strategy invocation request to the execution layer to trigger the process of loading the strategy version that has passed verification and is in a frozen state.

[0032] It should be noted that Task Management Unit 2 may include a Task Parsing Unit and a Task Scheduling Unit. The Task Parsing Unit can parse various task-related information in the requirements, generate task identifiers and task descriptions, combine them with task constraint information to generate task constraint descriptions, associate and store the task constraint descriptions with the task identifiers and task descriptions, and return task response messages to the outside. The Task Scheduling Unit can, upon identifying a new requirement as a new type of task, send a new type of task modeling request to the Cognition Layer 3 to trigger the process of generating a basic model structure or strategy template; this new type of task modeling request includes a task identifier ID, task type description, task objective, and constraints; upon identifying a requirement as an existing type of task, it sends a strategy invocation request to the Execution Layer 5 to trigger the process of loading a validated and frozen strategy version.

[0033] In implementation, external requests for network control, performance optimization, or management are sent to the demand input layer 1. These external requests need to interact with the task management unit 2. The external request sends a task request message to the task management unit 2 through the demand input layer 1. This task request message can include at least the requester's identifier, task type identifier, task objective description (e.g., performance optimization, resource adjustment, anomaly handling), task scope (e.g., specific slice, user group, region), task priority information, and task constraint information. The task constraint information includes performance constraints (e.g., latency cap, packet loss rate threshold), resource constraints (e.g., bandwidth, computing resource limits), and strategy preference information (e.g., stability priority or performance priority). Upon receiving the task request message, the task parsing unit within the task management unit 2 parses the task based on the relevant task information, generating a task identifier and task description. Simultaneously, it generates a task constraint description based on the constraint information and stores it in association with the task identifier and task description information. Subsequently, the task management unit 2 returns a task response message to the external request. This task response message can include the task acceptance result, task identifier ID, and a description of the subsequent processing flow.

[0034] After identifying a task as a new type of task, the task scheduling unit within task management unit 2 can send a new type of task modeling request to cognitive layer 3 to trigger the process of generating a basic model structure or strategy template. It can also receive a basic model generation completion message returned by cognitive layer 3; this message includes the model version number and a callable identifier. Additionally, after receiving a strategy freeze notification message from the model repository, the task scheduling unit can send a strategy issuance instruction to execution layer 5; this instruction may include the new frozen strategy version number, task identifier ID, task type, and current network status parameters. Furthermore, after identifying a task as an existing type of task, the task scheduling unit can send a strategy invocation request to execution layer 5 to trigger the process of loading a validated and frozen strategy version.

[0035] Furthermore, during implementation, external entities may need to adjust task objectives or cancel tasks during execution, requiring control interaction with the task management unit 2. The requirement input layer 1 can also receive task adjustment or cancellation request messages from external entities and send such messages to the task scheduling unit if external entities adjust task objectives or cancel tasks during execution. This task adjustment or cancellation request message may include: task identifier ID, adjustment type description (e.g., parameter adjustment, target change, task cancellation), and updated task objective or parameter information. The task parsing unit can also receive task adjustment or cancellation request information, determine the feasibility of adjustment or cancellation based on the current task status, perform the corresponding operation, and return the operation execution result and updated task information to the external entity. This task control response message must at least include the operation execution result and updated task information.

[0036] When an external entity needs to understand the progress or stage status of task processing, the task management unit 2 needs to interact with the external entity to return the task execution progress. During task processing, the task management unit 2 sends a task status update message to the external entity. This task status update message includes at least the task identifier ID and the current processing stage (e.g., model generation in progress, strategy evaluation in progress, or already entered execution layer 5).

[0037] Furthermore, in specific implementation, in the core network control method provided in the embodiments of the present invention, the generation of basic model structure or strategy template using cognitive layer 3 may specifically include: for new task types, using cognitive layer 3 to parse the modeling request of the new type of task, based on the parsed task semantics, combined with the intelligent agent of cognitive layer 3, generating the corresponding basic model structure or strategy template through language model and writing it into the model repository; after receiving the model storage confirmation message returned by the model repository, returning the basic model generation completion message to the task management unit 2, and the task management unit 2 sending the strategy optimization start message to the shadow execution layer 4.

[0038] It should be noted that if a new type of task, previously unprocessed, is proposed to the network for the first time from an external source, Task Management Unit 2 needs to interact with Cognitive Layer 3. Cognitive Layer 3 may include a Policy Generation Unit and a Model Structuring Unit. The Policy Generation Unit, after parsing the new task modeling request sent by Task Management Unit 2, generates corresponding policy templates and basic models based on the parsed task semantics and the intelligent agents available in Cognitive Layer 3, using a language model (such as LLM), and sends the generated results to the Model Structuring Unit. The Model Structuring Unit converts the basic model into a structured representation and sends a model storage request message to the model repository. This message may include: task type identifier, basic model structure, initial parameter set, and version number information. After receiving a model storage confirmation message from the model repository, it returns a basic model generation completion message to Task Management Unit 2. Subsequently, Task Management Unit 2 sends a policy optimization start message to Shadow Execution Layer 4.

[0039] Accordingly, the candidate strategy is optimized and verified using the shadow execution layer 4. Specifically, the shadow execution layer 4 generates candidate strategies based on the basic model structure or strategy template in the model repository, performs consistency verification on the candidate strategies, and sends a candidate strategy submission message to the model repository after the verification is passed, so that the model repository marks the candidate strategy as frozen release status.

[0040] It should be noted that Shadow Execution Layer 4 may include a Shadow Training Unit and a Simulation Verification Unit. The Model Repository may include a Policy Verification Unit. The Shadow Training Unit can be used to obtain the basic model structure or policy template from the model repository for new types of tasks, obtain the corresponding initial policy parameters, generate candidate policy parameters through reinforcement learning or fine-tuning, and send them to the Simulation Verification Unit; or, for existing types of tasks, obtain the currently frozen policy from the model repository, obtain the corresponding initial policy parameters, generate candidate policy parameters through reinforcement learning or fine-tuning, and send them to the Simulation Verification Unit. The Simulation Verification Unit can be used to verify the candidate policy parameters based on the simulation environment or historical running data; when the verification result meets the set conditions, a candidate policy submission message is sent to the model repository so that the Policy Verification Unit can perform consistency verification on the candidate policy parameters. After the verification passes, the policy state is marked as frozen.

[0041] In implementation, to ensure the stability and effectiveness of the model used by execution layer 5, shadow execution layer 4 needs to perform further reinforcement learning training and parameter verification on the basic model. The process is as follows: Shadow execution layer 4 first obtains the basic model for this task type from the model repository and sends the initial policy parameters to its internal shadow training unit. The shadow training unit performs online reinforcement learning or rapid fine-tuning based on the same state input as execution layer 5 and generates candidate policy parameters. The candidate policy parameters are sent to the simulation verification unit of shadow execution layer 4, which verifies the candidate policy parameters based on the simulation environment or historical running data. The verification task includes at least: verification data: historical running data playback / synthetic scene / digital twin simulation; verification indicators: average gain, latency, jitter, backoff trigger rate, action change rate, etc.; pass threshold: gain improvement ≥ A%, stability indicators do not deteriorate, action change rate ≤ B; when the verification results meet the performance improvement and stability conditions, shadow execution layer 4 sends a candidate policy submission message to the model repository. This candidate policy submission message may include: candidate policy parameters, verification indicators, suggested freeze flag, and associated task identifier ID. After the policy verification unit inside the model repository performs an eventual consistency check on the candidate policies, it marks the policy status as "freezeable".

[0042] Furthermore, in specific implementation, in the core network control method provided in the embodiments of the present invention, the verification of candidate strategies by the shadow execution layer 4 can be achieved through at least one of a simulation verification environment, a digital twin environment, a historical playback environment, an A / B test sandbox, or an offline evaluation platform.

[0043] After verifying the consistency of candidate policies, the process may further include: sending a policy freeze notification message to task management unit 2 to trigger the execution layer to load the frozen policy version. This policy freeze notification message may include the new policy version number and the applicable task scope. Task management unit 2 then sends a policy issuance instruction to execution layer 5.

[0044] It should be noted that when Shadow Execution Layer 4 needs to optimize policies without affecting network operation, it needs to interact with Cognition Layer 3 and the model repository. Shadow Execution Layer 4 can first obtain the current frozen policy from the model repository and send the initial policy parameters to its internal shadow training unit. The shadow training unit performs online reinforcement learning or rapid fine-tuning based on the same state input as Execution Layer 5 and generates candidate policy parameters. The candidate policy parameters are sent to Shadow Execution Layer 4, which verifies the candidate policies based on the simulation environment or historical running data. When the verification results meet the performance improvement and stability conditions, Shadow Execution Layer 4 sends a candidate policy submission message to the model repository, including the candidate policy parameters, verification metrics, and suggested freeze flag.

[0045] If the candidate strategy submitted by Shadow Execution Layer 4 meets the release conditions, the model repository needs to interact with Task Management Unit 2 and Execution Layer 5. After performing an eventual consistency check on the candidate strategy, the strategy verification unit within the model repository marks the strategy status as "freezeable." The model repository sends a strategy freeze notification message to Task Management Unit 2, including the new strategy version number and the applicable task scope. Task Management Unit 2 can then send a strategy switching instruction message to Execution Layer 5, including the new frozen strategy version number. Execution Layer 5 completes the strategy switch and returns a switch completion confirmation message.

[0046] It should be noted that this invention achieves long-term strategy evolution through experience accumulation in the cognitive layer 3 and continuous online optimization in the shadow execution layer 4. However, the evolution process is isolated from the existing network control, which can gradually enhance the network's autonomy and avoid introducing high-risk intelligent control all at once. The shadow execution layer 4 performs online reinforcement learning or rapid fine-tuning in a path parallel to the execution layer 5. All candidate strategies must undergo simulation verification and consistency check before entering the "freezeable / releaseable" state, thereby isolating exploratory and trial-and-error optimization from the real-time control of the existing network and achieving "continuous optimization without risks entering the main link." Furthermore, the shadow execution layer 4 runs in parallel with the execution layer 5, inputting homomorphic states but not controlling the existing network. It obtains candidate strategies through online reinforcement learning / rapid fine-tuning and verifies them in the simulation verification unit based on historical playback / digital twins. Only after meeting the conditions does it enter the release process, forming a safe closed loop from "exploration and verification to frozen release."

[0047] Furthermore, in specific implementation, the core network control method provided in the embodiments of the present invention may further include: for existing types of tasks, the task management unit 2 determines the freeze policy version number matching the requirements based on the task identifier, and sends the freeze policy version number and the current network state parameters to the execution layer 5; the policy loading unit of the execution layer 5 loads the corresponding freeze policy version from the model repository based on the freeze policy version number; the main execution agent of the execution layer 5 performs deterministic inference based on the current network state parameters and the loaded freeze policy version to generate control actions; the action constraint unit of the execution layer 5 verifies and corrects the control actions based on preset security boundaries, and sends the verified and corrected control actions to the underlying tool layer 6; the preset security boundaries are expressed in at least one of the following ways: interval threshold, rate of change limit, whitelist rule, feasible region constraint, or hard constraint optimization.

[0048] It should be noted that Execution Layer 5 may include a policy loading unit, a main execution agent, and an action constraint unit. The policy loading unit receives policy issuance instructions or policy invocation requests from Task Management Unit 2, reads the corresponding frozen policy from the model repository based on the frozen policy version number in the instruction or request, and sends the policy loading result to the main execution agent. The main execution agent performs deterministic inference based on the input network state parameters and the frozen policy, generates control actions, and sends them to the action constraint unit. The action constraint unit verifies and corrects the control actions based on preset safety boundary parameters and sends the verified and corrected control actions to the underlying tool layer 6. After completing policy issuance, Execution Layer 5 sends a policy issuance completion confirmation message to Task Management Unit 2.

[0049] Furthermore, if an external entity proposes a previously processed task type to the network, considering that the task model already has a corresponding frozen policy version in the model repository, the task management unit 2 needs to interact with the execution layer 5. The task management unit 2 sends a policy invocation request message to the execution layer 5, which includes: task identifier ID, task type, frozen policy version number, and current network state parameters. The policy loading unit within the execution layer 5 reads the corresponding policy from the model repository based on the frozen policy version number and sends the policy loading result to the main execution agent. The main execution agent performs deterministic inference based on the input network state parameters and the frozen policy, generates deterministic actions, and sends the control actions to the action constraint unit of the execution layer 5. The action constraint unit verifies and corrects the control actions based on preset safety boundary parameters before sending the final control actions to the underlying tool layer 6 for execution, achieving hard constraint guarantees. The execution layer 5's output actions are verified and corrected in the action constraint unit based on safety boundaries, avoiding out-of-bounds configurations and overly aggressive scheduling, ensuring that control actions meet hard constraints and safety policies, reducing network oscillation risks, and enhancing operational stability.

[0050] It is important to note that Execution Layer 5 allows the main executing agent to make deterministic parameter adjustments within a safe boundary without changing the policy structure or introducing randomness, thus achieving "controllable adaptation." In other words, Execution Layer 5 uses a frozen policy version for deterministic inference and avoids randomness caused by online exploration; when adaptation is needed, deterministic parameter adjustments are only allowed within a preset safe boundary, and are forcibly pruned by the action constraint unit to ensure the predictability and repeatability of control actions.

[0051] Furthermore, in specific implementation, the core network control method provided in the above embodiments of the present invention may further include: using the execution layer 5 to collect performance indicators; when the collected performance indicators are lower than a preset performance threshold, sending a policy rollback request message to the model repository so that the model repository returns historical stable policy version information; and switching to the historical stable policy version according to the historical stable policy version information returned by the model repository to restore core network control.

[0052] It should be noted that, considering the potential performance degradation of Execution Layer 5 during real-time operation, Execution Layer 5 needs to perform internal performance monitoring and interact with the model repository. Execution Layer 5 can also be used for safety verification and correction of generated control actions, as well as performance monitoring, anomaly detection, and policy rollback, feeding operational data back to Cognition Layer 3. Cognition Layer 3 can also be used to optimize and update the basic model structure based on operational data. Based on this, Execution Layer 5 can also include a performance acquisition unit, a performance evaluation unit, and a policy rollback control unit. The performance acquisition unit can collect performance metrics (KPIs) and send them to the performance evaluation unit. The performance evaluation unit can send a performance anomaly notification message to the policy rollback control unit when the collected performance metrics are lower than a preset performance threshold; this notification message can include the anomaly type and anomaly metric. The policy rollback control unit can send a policy rollback request message to the model repository based on the performance anomaly notification message, so that the model repository returns historical stable policy version information; this request message can include the current policy version number and task identifier. Execution Layer 5 can also switch to a historical version and restore network control based on the historical stable policy version information. This enables the execution layer 5 to integrate performance acquisition, evaluation, and policy rollback mechanisms. When performance degradation or anomalies are detected, it can immediately switch to a historically stable policy version, forming a controllable closed-loop guarantee capability. Simultaneously, through version records in the model repository and policy status marking, the policy lifecycle is traceable, manageable, and auditable. Through threshold evaluation and anomaly triggering mechanisms, it can quickly roll back to a stable version when performance declines, ensuring network service continuity and reliability and preventing widespread service degradation caused by policy failure.

[0053] Furthermore, in specific implementation, the core network control method provided in the embodiments of the present invention may further include: collecting execution logs, environmental feedback and performance indicators of the main executing agent using the execution layer, performing pattern analysis on the collected information, incrementally optimizing or transferring learning the basic model structure or policy template based on the analysis results, and writing the updated model into the model repository.

[0054] It should be noted that, considering the need for long-term knowledge accumulation and model optimization in the cognitive layer 3, the execution layer 5 needs to interact with the cognitive layer 3. The execution layer 5 may also include a log aggregation unit and a model optimization unit. The cognitive layer 3 may also include an experience analysis unit. The log aggregation unit can collect the execution logs, environmental feedback, and performance metrics of the main executing agent and send execution feedback messages to the experience analysis unit. The experience analysis unit can perform pattern analysis based on the execution feedback information and send the analysis results to the model optimization unit. The model optimization unit can perform incremental optimization or transfer learning on the basic model structure based on the analysis results, write the updated model into the model repository, and form a new version of the basic model.

[0055] Furthermore, in a specific implementation, when the execution layer of the core network control method provided in the embodiments of the present invention includes multiple parallel-running main execution agents, it may also include: distributing subtasks to multiple main execution agents according to the task load of the execution layer, collecting the control actions generated by each main execution agent, performing conflict detection and consistency solving on the collected control actions, and performing action constraint verification on the aggregation result after completing the action aggregation.

[0056] It should be noted that, considering that both Execution Layer 5 and Shadow Execution Layer 4 consist of multiple RL-based Agents, collaborative scheduling is required within each layer. Execution Layer 5 may also include an agent scheduling unit and an action aggregation unit. The agent scheduling unit can distribute subtasks to multiple main executing agents based on task load. The action aggregation unit can collect control actions from each main executing agent, perform conflict detection and consistency resolution on the collected control actions, and then send the aggregated control actions to the underlying tool layer 6 after action constraint verification.

[0057] When multiple master agents are responsible for different subdomains (such as the wireless side, bearer side, core network UPF side, etc.), and may output conflicting actions, the agent scheduling unit within execution layer 5 distributes the subtasks to multiple master agents and collects the control actions and confidence / priority flags output by each agent. The action aggregation unit performs conflict detection (such as duplicate allocation of the same resource, concurrent execution of mutually exclusive actions, etc.) and performs consistency resolution according to preset rules: priority arbitration (such as task priority / slice level), constraint satisfaction (such as hard constraint priority), and the principle of minimum modification (such as keeping the actions as close as possible to the original actions of each agent). The aggregated actions are then sent to the action constraint unit for safety verification before being executed. In this way, after multiple master agents infer in parallel within execution layer 5, the action aggregation unit coordinates them uniformly to resolve action conflicts and resource contention issues, enhance cross-domain collaboration capabilities, ensure global consistency and interpretable arbitration results during multi-agent parallel control, and improve the ability to handle complex business scenarios.

[0058] Furthermore, in specific implementation, in the core network control method provided in the embodiments of the present invention, the underlying tool layer 6 can be used to receive control actions issued by the action constraint unit and execute network control operations including traffic scheduling, resource allocation, and quality of service assurance. The underlying tool layer 6 may include a monitoring module or a probe module; the monitoring module or probe module can be used to monitor the network operating status, and when an unexpected situation is detected (such as sudden congestion in a local area), it sends a status input message to the execution layer 5 to trigger the abnormal handling process of the execution layer 5; the status input message may include information such as area identifier, congestion level, critical link or network element load, and current key performance indicators; it can also be used to feed back various status data and execution results during network operation to the execution layer 5, providing basic data for the performance monitoring, policy adjustment, and rollback of the execution layer 5, and ensuring that upper-layer decisions can be made based on the real network status.

[0059] In implementation, when large-scale events cause sudden congestion in local areas, requiring real-time traffic scheduling, the monitoring / probe module of the underlying tool layer 6 can send "congestion alarm status input" to the execution layer 5, including: area identifier, congestion level, critical link / network element load, and current KPIs (such as latency / packet loss / retransmission). The execution layer 5 triggers "congestion handling strategy invocation": the strategy loading unit loads the frozen strategy version; the main execution agent deterministically infers and generates control actions (such as routing / diversion, queue scheduling parameters, rate limiting / shaping, UPF routing, switching threshold adjustment, etc.). The shadow execution layer 4 performs parallel online optimization: under homomorphic state input, it performs rapid fine-tuning / online RL training to obtain candidate strategies; the simulation verification unit verifies the stability and benefits of candidate strategies based on historical playback / digital twin environment; after meeting the conditions, it submits to the model repository to enter the freezeable process. In this way, the real-time closed loop is completely completed by edge deterministic strategies; the shadow layer continuously produces better strategies without interfering with the live network.

[0060] It should be noted that this invention decouples the intelligent decision-making process in the network into layers, introducing different types of intelligent agents to undertake non-real-time policy generation tasks and real-time decision execution tasks, respectively. Specifically, agents with strong reasoning capabilities are used to complete policy modeling, model generation, or decision rule construction in offline or non-real-time scenarios; agents with rapid response capabilities are used to perform network control and resource scheduling based on the policy model during real-time operation, thereby avoiding the direct participation of highly complex intelligent models in the real-time network control closed loop.

[0061] Through the above scheme, this invention enhances network intelligence and autonomy by introducing a multi-agent mechanism, while effectively reducing the computational complexity in real-time decision-making paths and avoiding the uncertainty risks brought about by probabilistic models directly controlling the network. This ensures low latency, high reliability, and strong determinism in network operation. This invention can be applied to complex scenarios involving multiple slices, multiple regions, and multiple services, achieving cross-domain collaborative control and unified consistency assurance.

[0062] It should be noted that the hierarchical structure and module composition of this invention can be implemented in various equivalent variations. Task management unit 2 can be transformed from a centralized module into a distributed scheduler, deployed in a distributed manner according to regions, slices, or network elements to reduce single-point bottlenecks. This variation is still an equivalent extension of the task identification and scheduling concept of this invention. The model repository can adopt a multi-level form combining a cloud main repository and an edge cache repository. Execution layer 5 prioritizes using frozen policy versions cached at the edge, while the cloud is responsible for unified version management and synchronization, thereby improving edge autonomy and availability during network outages. Cognitive layer 3 can replace the language model with a domain fine-tuning model, retrieval enhancement model, or symbolic reasoning enhancement model without entering the real-time control closed loop to complete policy template generation and rule construction, all of which fall within the protection scope of this invention. Shadow execution layer 4 can generate candidate policies using online reinforcement learning, offline reinforcement learning, imitation learning, policy distillation, or rapid fine-tuning methods. As long as the core principles of not directly controlling the existing network and requiring verification before freezing and release are met, they are considered equivalent replacements of this invention.

[0063] The freeze release and verification mechanism of this invention also supports multiple equivalent variations. The triggering condition for freeze release can be modified from long-term stability better than the current version to judgment rules that satisfy multiple consecutive cycles of improved benefits without degradation of stability and meet multi-scenario coverage requirements, which are equivalent forms of the same release principle. The security boundary can be expressed in different ways, such as interval thresholds, rate of change limits, whitelist rules, feasible region constraints, or hard constraint optimization, but its essence is to constrain, verify, and prune control actions, which are still within the protection scope of this invention.

[0064] Based on the same inventive concept, embodiments of the present invention also provide a core network system for implementing the above-described core network control method. The system includes: The requirement input layer is used to receive external requirements regarding network control, performance optimization, or management. The task management unit is used to identify the task type of the requirements. The decision-making unit is used to determine whether there is an agent-based policy that corresponds to the task, has been verified, and is in a frozen state based on the task type identification result. If it does not exist, the intelligent decision-making process is triggered to generate a new agent-based policy, the new agent-based policy is verified, and the verified policy is set to a frozen release state. Based on the policy in the frozen release state, deterministic inference and control operations are performed to achieve real-time closed-loop control of the core network. The core network system decouples the intelligent decision-making process from real-time control operations through a hierarchical intelligent agent structure.

[0065] In the core network system provided in the embodiments of the present invention, complex inference models can be avoided from directly entering the real-time control link, ensuring low latency, predictability, and repeatability of network control, realizing unified management and reuse of strategies, enabling the network to support rapid modeling of new tasks and rapid invocation of known tasks, taking into account both intelligent decision-making capabilities and deterministic real-time control requirements, effectively reducing the computational complexity of real-time decision-making, avoiding the uncertainty risks brought by probabilistic models, significantly improving the stability, reliability, autonomy, and engineering deployability of the core network in complex dynamic environments, better meeting the network's comprehensive requirements for real-time performance, stability, and intelligent collaboration, and possessing high engineering practicality and application value.

[0066] Since the embodiments of the core network system correspond to the embodiments of the core network control method described above, the descriptions of the features in the core network system embodiments can be found in the relevant descriptions of the core network control method embodiments described above, and will not be repeated here. Furthermore, it has the same beneficial effects as the core network control method mentioned above.

[0067] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0068] The core network control method and core network system provided by this invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of this invention.

Claims

1. A core network control method, characterized in that, The method is applied to a core network system with an intelligent agent structure, and the method includes: Receive requests from external sources regarding network control, performance optimization, or management; The task management unit identifies the task type of the requirements. Based on the task type identification result, determine whether there is an agent-based strategy that corresponds to the task, has been verified, and is in a frozen state. If it does not exist, the intelligent decision-making process is triggered to generate a new agent-based policy, the new agent-based policy is verified, and the verified policy is set to the frozen release state. Based on the strategy in a frozen release state, deterministic inference and control operations are performed to achieve real-time closed-loop control of the core network; The method decouples the intelligent decision-making process from the real-time control operation through the intelligent agent structure.

2. The core network control method according to claim 1, characterized in that, The intelligent agent structure includes: a cognitive layer, a shadow execution layer, and an execution layer; the cognitive layer is deployed in the cloud, and the shadow execution layer and the execution layer are deployed at the edge. The decoupling of intelligent decision-making processes from real-time control operations through the aforementioned agent structure includes: The cognitive layer is used to generate a basic model structure or strategy template; Based on the aforementioned basic model structure or strategy template, the shadow execution layer is used to optimize candidate strategies, verify the candidate strategies, and set the verified strategies to a frozen release state. The execution layer loads and verifies the policy version, which is in a frozen state. Based on the network operating status, deterministic inference is performed to generate control actions, so that the underlying tool layer can perform corresponding network control operations according to the control actions.

3. The core network control method according to claim 1 or 2, characterized in that, The task management unit identifies the task type of the requirements, including: After identifying the requirement as a new type of task, a new type of task modeling request is sent to the cognitive layer to trigger the process of generating a basic model structure or strategy template. After identifying the requirement as an existing type of task, a policy invocation request is sent to the execution layer to trigger the process of loading the policy version that has passed verification and is in a frozen state.

4. The core network control method according to claim 3, characterized in that, Generate a basic model structure or policy template using the cognitive layer, including: For new task types, the cognitive layer is used to parse the modeling request of the new task type. Based on the parsed task semantics, combined with the intelligent agent of the cognitive layer, the corresponding basic model structure or strategy template is generated through the language model and written into the model repository. After receiving the model storage confirmation message returned by the model repository, a basic model generation completion message is returned to the task management unit, and the task management unit sends a strategy optimization start message to the shadow execution layer. The candidate strategy is optimized using the shadow execution layer, and the candidate strategy is verified, including: The shadow execution layer generates candidate strategies based on the basic model structure or strategy template in the model repository, performs consistency verification on the candidate strategies, and sends a candidate strategy submission message to the model repository after the verification passes, so that the model repository marks the candidate strategies as frozen release status.

5. The core network control method according to claim 4, characterized in that, The verification of candidate strategies by the shadow execution layer is achieved through at least one of the following: simulation verification environment, digital twin environment, historical replay environment, A / B testing sandbox, or offline evaluation platform. After performing consistency checks on the candidate strategies, the following is also included: A policy freeze notification message is sent to the task management unit to trigger the execution layer to load the freeze policy version.

6. The core network control method according to claim 3, characterized in that, Also includes: For existing types of tasks, the task management unit determines the freeze policy version number that matches the requirement based on the task identifier, and sends the freeze policy version number and the current network status parameters to the execution layer. The execution layer's strategy loading unit loads the corresponding freeze strategy version from the model repository based on the freeze strategy version number; The main execution agent of the execution layer performs deterministic inference and generates control actions based on the current network state parameters and the loaded freeze policy version. The action constraint unit of the execution layer verifies and corrects the control action according to the preset safety boundary, and sends the verified and corrected control action to the underlying tool layer; the preset safety boundary is expressed in at least one of the following ways: interval threshold, rate of change limit, whitelist rule, feasible region constraint or hard constraint optimization.

7. The core network control method according to claim 6, characterized in that, Also includes: The execution layer collects performance metrics. When the collected performance metrics are lower than a preset performance threshold, a strategy rollback request message is sent to the model repository so that the model repository returns historical stable strategy version information. The system then switches to the historical stable policy version based on the historical stable policy version information returned by the model repository in order to restore core network control.

8. The core network control method according to claim 6, characterized in that, Also includes: The execution layer collects the execution logs, environmental feedback, and performance metrics of the main executing agent, performs pattern analysis on the collected information, and incrementally optimizes or performs transfer learning on the basic model structure or policy template based on the analysis results, and writes the updated model into the model repository.

9. The core network control method according to claim 6, characterized in that, When the execution layer includes multiple main execution agents running in parallel, it also includes: The execution layer distributes subtasks to multiple main execution agents based on task load, collects control actions generated by each main execution agent, performs conflict detection and consistency resolution on the collected control actions, and performs action constraint verification on the aggregation result after completing action aggregation.

10. A core network system implementing the core network control method as described in any one of claims 1 to 9, characterized in that, include: The requirement input layer is used to receive external requirements regarding network control, performance optimization, or management. The task management unit is used to identify the task type of the requirements. The decision unit is used to determine, based on the task type identification result, whether there is an agent-based strategy that corresponds to the task, has been verified, and is in a frozen state. If it does not exist, the intelligent decision-making process is triggered to generate a new agent-based policy, the new agent-based policy is verified, and the verified policy is set to the frozen release state; based on the policy in the frozen release state, deterministic inference and control operations are performed to achieve real-time closed-loop control of the core network. The core network system decouples the intelligent decision-making process from real-time control operations through a hierarchical intelligent agent structure.

Images