Communication method, apparatus, storage medium, and computer program product

By introducing intelligent agents and evaluator modules generated by generative networks, the problem of insufficient intelligent decision-making in the 5G core network is solved, and more efficient intelligent agent behavior optimization and decision quality improvement are achieved.

WO2026238105A1PCT designated stage Publication Date: 2026-11-19HUAWEI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2026/097285
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-16
Filing Date
2026-05-15
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

The service-oriented architecture of the 5G core network lacks a differentiated update mechanism, making it difficult to provide deeper intelligent decision-making capabilities.

Method used

Generative networks are introduced to generate agents for overall evaluation. Through generative pre-trained models and evaluator modules, global policies are adaptively estimated to distinguish the degree of contribution of agents and optimize their behavior.

Benefits of technology

It enhances the adaptability and intelligent decision-making capabilities of the next-generation core network, and strengthens the robustness and quality of decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2026097285_19112026_PF_FP_ABST
    Figure CN2026097285_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Embodiments of the present disclosure provide a communication method, an apparatus, a storage medium, and a computer program product. In the method, an agent for evaluation acquires a user request for executing a task. The agent for evaluation receives a plurality of decisions from a plurality of agents, and individual evaluation information corresponding to each of the plurality of decisions. The agent for evaluation determines, on the basis of the user request, the plurality of decisions, and the individual evaluation information, overall evaluation information for instructing the plurality of agents to generate a generative network (GN), wherein the GN is generated on the basis of the plurality of decisions. In this way, in the embodiments of the present disclosure, the agent for evaluation can adaptively estimate a global policy, thereby improving the adaptability of an agent-based next-generation core network A-Core.
Need to check novelty before this filing date? Find Prior Art

Description

Communication methods, devices, storage media and computer program products

[0001] This application claims priority to Chinese Patent Application No. 202510643772.5, filed on May 16, 2025, entitled "Communication Method, Apparatus, Storage Medium and Computer Program Product", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This disclosure generally relates to the field of communications, and more specifically to a communication method, apparatus, computer-readable storage medium, and computer program product. Background Technology

[0003] The 5G core network adopts a service-based architecture (SBA), the core idea of ​​which is to modularize large network functions and break them down into multiple independent service units. These service units collaborate with each other through standardized interfaces, thus providing a more flexible, scalable, and easily deployed and managed network solution. However, this network architecture based on standardized interfaces lacks differentiated update mechanisms to some extent, making it difficult to provide deeper intelligent decision-making capabilities. Summary of the Invention

[0004] The embodiments of this disclosure provide a communication technology solution, specifically providing a method and apparatus for performing an overall evaluation of agents participating in the generation of a Generative Network (GN), which can provide a deeper level of intelligent decision-making capabilities.

[0005] Firstly, a communication method is provided. The executing entity of the method provided in the first aspect can be an intelligent agent for evaluation. Unless otherwise specified, the intelligent agent for evaluation in this disclosure can refer to the intelligent agent itself (e.g., in some examples, it may involve core network devices), a module within the intelligent agent for evaluation (e.g., a processor, chip, or chip system), or a logical module or software capable of implementing all or part of the functions of the intelligent agent for evaluation. In this method, the intelligent agent for evaluation obtains a user request for performing a task. The intelligent agent for evaluation receives multiple decisions from multiple intelligent agents, and individual evaluation information corresponding to each of the multiple decisions. Based on the user request, the multiple decisions, and the individual evaluation information, the intelligent agent for evaluation determines overall evaluation information for instructing the multiple intelligent agents to generate a generative network GN, where the GN is generated based on the multiple decisions. In this way, the intelligent agent for evaluation can adaptively estimate the global policy, thereby improving the adaptability of the agent-based next-generation core network A-Core.

[0006] In some exemplary embodiments, the agent used for evaluation obtains first information based on a user request, wherein the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests. In some exemplary embodiments, during the process of the agent used for evaluation obtaining the first information, the agent sends a first request and receives first information in response to the first request, wherein the first request includes the user request. Thus, the agent used for evaluation, by leveraging semi-structured retrieval information, can fully utilize historical reasoning trajectory information to perform in-depth learning on the semi-structured retrieval information generated during the interaction process.

[0007] In some exemplary embodiments, during the process of determining the overall evaluation information by the agent used for evaluation, the overall evaluation information is determined based on user requests, multiple decisions, individual evaluation information, and first information. For example, the agent used for evaluation uses a generative pre-trained model to determine the overall evaluation information based on user requests, multiple decisions, individual evaluation information, and first information. In this way, during the adaptive estimation of the global policy, the agent used for evaluation can make full use of historical inference trajectory information, thus further improving the adaptability of the agent-based next-generation core network A-Core.

[0008] In some exemplary embodiments, the agent used for evaluation also receives information indicating the generation result of the GN. In some exemplary embodiments, the agent used for evaluation also determines the generation error information of the GN based on the information of the generation result and the overall evaluation information. In this way, if the GN is generated correctly in function, the agent can receive a reward, while if the GN fails to be generated correctly in function, the agent will face a penalty.

[0009] In some exemplary embodiments, the agent used for evaluation also sends generation error information and multiple contributions from multiple agents to the generation of the GN. In some exemplary embodiments, the first contribution of the first agent among the multiple agents to the generation of the GN is determined based on information about the generation result and individual evaluation information corresponding to the first agent's first decision. In this way, different agents can receive differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their level of contribution in the task, thereby optimizing their behavior in a targeted manner.

[0010] In some exemplary embodiments, the agent used for evaluation also updates the generative pre-trained model based on the generated error information. In this way, the agent used for evaluation can also receive differentiated reward feedback based on its actual performance in the task, thereby optimizing its behavior.

[0011] In some exemplary embodiments, the overall evaluation information includes the probability that the GN is correctly generated. In some exemplary embodiments, individual evaluation information is used to determine the probability that the agent's decision results in the correct generation of the GN. Thus, the agent's estimation of the global policy for evaluation can be quantified based on probabilistic information.

[0012] In some exemplary embodiments, user requests are based on natural language representations. In this case, since the agent used for evaluation is applied to the interaction scenario of a network agent (NetAgent) in the field of communication, the interaction and value assessment mode of the agent used for evaluation based on structured data is integrated into the natural language-based NetAgent interaction system, thus enabling it to match the interaction mechanism with NetAgent.

[0013] Secondly, a communication method is provided. The executing entity of the method provided in this second aspect can be a first intelligent agent. Unless otherwise specified, the first intelligent agent in this embodiment can refer to the first intelligent agent itself (e.g., in some examples, it may involve core network equipment), a module within the first intelligent agent (e.g., a processor, chip, or chip system), or a logical module or software capable of implementing all or part of the functions of the first intelligent agent. In this method, the first intelligent agent receives generation error information from a generative network GN and a first contribution of the first intelligent agent among multiple intelligent agents to the generation of GN, wherein GN is generated based on multiple decisions of multiple intelligent agents. The first intelligent agent updates its generative pre-trained model based on the generation error information and the first contribution.

[0014] In some exemplary embodiments, during the process of the first agent updating the generative pre-trained model, the first agent determines reward information based on generation error information and a first contribution, and updates the generative pre-trained model based on the reward information. In this way, different agents can receive differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their own contribution level in the task, thereby optimizing their behavior in a targeted manner.

[0015] Thirdly, a communication method is provided. The executing entity of the method provided in this third aspect can be a first intelligent agent. Unless otherwise specified, the first intelligent agent in this embodiment can refer to the first intelligent agent itself (e.g., in some examples, it may involve core network equipment), a module within the first intelligent agent (e.g., a processor, chip, or chip system), or a logic module or software capable of implementing all or part of the functions of the first intelligent agent. In this method, the first intelligent agent receives input information. Based on the input information, the first intelligent agent determines a first decision among multiple intelligent agents. The first intelligent agent determines its individual evaluation information for generating a generative network (GN). The first intelligent agent sends the first decision to a second intelligent agent among multiple intelligent agents.

[0016] In some exemplary embodiments, the first agent also sends first decision and individual evaluation information to the agent used for evaluation.

[0017] In some exemplary embodiments, during the process of the first agent determining a first decision, the first agent sends a second request, receives second information in response to the second request, and uses a generative pre-trained model of the first agent to determine the first decision based on the second information and input information. The second request includes input information, and the second information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information.

[0018] In some exemplary embodiments, during the process of the first agent determining individual evaluation information, the first agent sends input information and a first decision to the evaluation module in the first agent, and receives individual evaluation information from the evaluation module, wherein the individual evaluation information is determined by the evaluation module using a generative pre-trained model based on third information, and the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information.

[0019] In some exemplary embodiments, the input information includes a user request for performing a task, the user request being based on a natural language representation, and the first agent also sends the input information to an agent for evaluation. Alternatively, the input information includes a third decision by a third agent among a plurality of agents.

[0020] Fourthly, a communication method is provided. The execution entity of the method provided in this fourth aspect can be an evaluation module in a first intelligent agent. Unless otherwise specified, the evaluation module in this embodiment can refer to the evaluation module itself (e.g., in some examples, it may involve core network devices), a module within the evaluation module (e.g., a processor, chip, or chip system), or a logic module or software capable of implementing all or part of the evaluation module's functions. In this method, the evaluation module receives input information and a first decision of the first intelligent agent. The evaluation module sends a third request to a fifth network functional entity, wherein the third request includes the input information and the first decision, and the fifth network functional entity is used to store data. The evaluation module receives third information from the fifth network functional entity in response to the third request, wherein the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. The evaluation module determines individual evaluation information based on the third information. In this way, the intelligent agent can perform self-evaluation when making decisions, thereby introducing the dimension of decision quality. When faced with complex tasks, the agent can attempt multiple outputs based on its self-assessment and select the highest-quality output. This process effectively enhances the robustness of the decision-making process and improves its quality.

[0021] In some exemplary embodiments, the evaluation module also sends individual evaluation information. In some exemplary embodiments, the input information includes a user request for performing a task, the user request being based on a natural language representation. Alternatively, the input information includes a third decision by a third agent among multiple agents.

[0022] Fifthly, a communication method is provided. The execution entity of the method provided in the fifth aspect can be a common memory. Unless otherwise specified, the common memory in this disclosure can refer to the common memory itself (e.g., in some examples, it can involve core network devices), a module within the common memory (e.g., a processor, chip, or chip system), or a logical module or software capable of implementing all or part of the common memory's functions. In this method, the common memory receives a first request, wherein the first request includes a user request for the generative network (GN) to perform a task, and the GN is generated based on multiple decisions made by multiple agents. The common memory determines first information based on the user request, wherein the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests. The common memory sends the first information.

[0023] In some exemplary embodiments, the common memory also receives overall evaluation information for instructing multiple agents to generate GN.

[0024] In some exemplary embodiments, the common memory also receives a second request, determines second information based on the input information, and sends the second information. The second request includes the input information, and the second information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information.

[0025] In some exemplary embodiments, the common memory further receives a third request, determines third information based on the input information and the first decision, and sends the third information. The third request includes the input information and the first decision of the first agent among a plurality of agents, and the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information.

[0026] In some exemplary embodiments, the input information includes a user request based on a natural language representation; or the input information includes a third decision of a third agent among a plurality of agents.

[0027] A sixth aspect provides a communication apparatus, including an acquisition unit, a receiving unit, and a determination unit. In some embodiments, the acquisition unit may be configured to acquire a user request for performing a task, the receiving unit may be configured to receive multiple decisions from multiple agents and individual evaluation information corresponding to each of the multiple decisions, and the determination unit may be configured to determine overall evaluation information for instructing the multiple agents to generate a generative network GN based on the user request, the multiple decisions, and the individual evaluation information, wherein the GN is generated based on the multiple decisions.

[0028] In a seventh aspect, a communication apparatus is provided, including a receiving unit and an updating unit. In some embodiments, the receiving unit may be configured to receive generation error information of a generative network GN and a first contribution of a first agent among a plurality of agents to the generation of GN, wherein GN is generated based on multiple decisions of the plurality of agents, and the updating unit may be configured to update the generative pre-trained model of the first agent based on the generation error information and the first contribution.

[0029] Eighthly, a communication apparatus is provided, including a receiving unit, a first determining unit, a second determining unit, and a transmitting unit. In some embodiments, the receiving unit may be configured to receive input information, the first determining unit may be configured to determine a first decision of a first agent among a plurality of agents based on the input information, the second determining unit may be configured to determine individual evaluation information of the first agent among the plurality of agents for generating a GN, and the transmitting unit may be configured to transmit the first decision to a second agent among the plurality of agents.

[0030] A ninth aspect provides a communication apparatus, including a first receiving unit, a transmitting unit, a second receiving unit, and a determining unit. In some embodiments, the first receiving unit may be configured to receive input information and a first decision of a first agent; the transmitting unit may be configured to send a third request to a fifth network functional entity, wherein the third request includes the input information and the first decision, and the fifth network functional entity is used to store data; the second receiving unit may be configured to receive third information from the fifth network functional entity in response to the third request, wherein the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information; and the determining unit may be configured to determine individual evaluation information based on the third information.

[0031] A tenth aspect provides a communication apparatus, including a receiving unit, a determining unit, and a transmitting unit. In some embodiments, the receiving unit may be configured to receive a first request, wherein the first request includes a user request for a generative network (GN) to perform a task, and the GN is generated based on multiple decisions of multiple agents; the determining unit may be configured to determine first information based on the user request, wherein the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests; and the transmitting unit may be configured to transmit the first information.

[0032] The aforementioned communication device has the function of implementing the behavior in the method examples of the first or second aspect. The function can be implemented in hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the aforementioned function. The descriptions of the intelligent agent for evaluation in the first aspect, the first intelligent agent in the second aspect, the first intelligent agent in the third aspect, the evaluation module in the first intelligent agent in the fourth aspect, and the common memory in the fifth aspect also apply to the communication device in the sixth, seventh, eighth, ninth, and tenth aspects. That is, the communication device in the sixth, seventh, eighth, ninth, and tenth aspects can refer to the communication device itself, or to components within the communication device (e.g., processor, chip, or chip system), or to logic modules or software capable of implementing all or part of the functions of the communication device. In one possible design, the communication device includes a unit that executes the methods of the first, second, third, fourth, and fifth aspects or their implementations. The beneficial effects achieved by the communication devices provided in the sixth, seventh, eighth, ninth, and tenth aspects can be found in the descriptions of the first, second, third, fourth, and fifth aspects, and will not be repeated here.

[0033] Eleventhly, an apparatus is provided, including a processor and a memory storing a computer program or instructions, which, when executed by the processor, cause the apparatus to perform any method according to the first aspect, the second aspect, the third aspect, the fourth aspect, the fifth aspect, and implementations thereof.

[0034] In a twelfth aspect, a computer-readable storage medium is provided, which stores a computer program or instructions that, when executed by an electronic device, cause the electronic device to perform the methods described in the preceding aspects.

[0035] In a thirteenth aspect, a computer program or computer program product is provided, the computer program or computer program product including computer program instructions, which, when executed by an electronic device, cause the electronic device to perform the methods described in the preceding aspects.

[0036] In a fourteenth aspect, embodiments of this disclosure provide a chip system including a processor for implementing the functions of the apparatus in the methods of the foregoing aspects. In one possible design, the chip system further includes a memory for storing computer programs or instructions and / or data. The chip system may be composed of chips or may include chips and other discrete devices.

[0037] In a fifteenth aspect, embodiments of this disclosure also provide a communication system including an intelligent agent for performing an evaluation in the first aspect, a first intelligent agent for performing a second aspect, a first intelligent agent for performing a third aspect, an evaluation module in the first intelligent agent for performing a fourth aspect, and a common memory for performing a fifth aspect. Attached Figure Description

[0038] Figure 1 illustrates a possible application framework schematic diagram that can be used to implement some embodiments of the communication system according to this application.

[0039] Figure 2 illustrates a system architecture of an agent-based next-generation core network (A-Core) according to some embodiments of the present disclosure.

[0040] Figure 3 illustrates the internal components of an intelligent agent according to some embodiments of the present disclosure.

[0041] Figure 4 illustrates a communication flow diagram according to some embodiments of the present disclosure.

[0042] Figure 5 illustrates a communication flow diagram according to some embodiments of the present disclosure.

[0043] Figure 6 illustrates a communication flow diagram according to some embodiments of the present disclosure.

[0044] Figure 7 illustrates a communication flow diagram according to some embodiments of the present disclosure.

[0045] Figure 8 illustrates a communication flow diagram according to some embodiments of the present disclosure.

[0046] Figure 9 illustrates a flowchart of an overall evaluation process according to some embodiments of the present disclosure.

[0047] Figure 10 illustrates a schematic diagram of an evaluation agent performing an overall evaluation according to some embodiments of the present disclosure.

[0048] Figure 11 illustrates a schematic flowchart of feedback allocation according to some embodiments of the present disclosure.

[0049] Figure 12 shows a schematic flowchart of a preliminary evaluation according to some embodiments of the present disclosure.

[0050] Figure 13 shows a schematic flowchart of implementation at the agent for evaluation according to some embodiments of the present disclosure.

[0051] Figure 14 shows a schematic flowchart implemented at a first intelligent agent according to some embodiments of the present disclosure.

[0052] Figure 15 shows a schematic flowchart implemented at a first intelligent agent according to some embodiments of the present disclosure.

[0053] Figure 16 shows a schematic flowchart of the implementation at the evaluation module in the first intelligent agent according to some embodiments of the present disclosure.

[0054] Figure 17 shows a schematic flowchart of implementation at a common memory according to some embodiments of the present disclosure.

[0055] Figure 18 is a block diagram of a device that can be used to implement some embodiments of this application.

[0056] Figure 19 is a schematic diagram of the structure of an apparatus according to some embodiments of this application.

[0057] Figure 20 is a schematic diagram of the structure of an apparatus according to some embodiments of this application.

[0058] Figure 21 is a schematic diagram of the structure of an apparatus according to some embodiments of this application.

[0059] Figure 22 is a schematic diagram of the structure of an apparatus according to some embodiments of this application.

[0060] Figure 23 is a schematic diagram of the structure of an apparatus according to some embodiments of this application. Detailed Implementation

[0061] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that embodiments of this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0062] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0063] In recent years, the rapid development of Artificial Intelligence (AI) technology has driven continuous progress in self-learning and self-optimization algorithms. Generative AI, as an important branch of AI, breaks through the limitations of conventional AI algorithms, enabling the generation of more creative and flexible content based on input data. Among them, Large Language Models (LLMs), as a typical representative of generative AI, have demonstrated outstanding text understanding and generation capabilities in the field of natural language processing, providing strong support for the application of AI technology in language-related fields.

[0064] Furthermore, the concept of an intelligent agent, as an emerging one, is widely applied in the field of autonomous intelligent systems. An intelligent agent is defined as an entity capable of perceiving its surrounding environment and making autonomous decisions accordingly. By embedding a large language model (LLM) into a dynamic environment, a network intelligent agent (NetAgent) based on the large language model for the communication domain is formed. AI systems can achieve more natural interactions and possess the ability to handle complex tasks. NetAgents possess core capabilities such as memory, tool use, planning, and action, enabling them to autonomously complete tasks and collaborate with other intelligent agents in complex environments. Integrating NetAgents with future communication technologies holds the promise of bringing significant changes to future communication networks. For example, after a large model is pre-trained based on communication data to obtain NetAgents, through interactions between numerous NetAgents, the network can dynamically generate, execute, and retrieve customized services based on user intent. This integration not only enhances the intelligence level of communication networks but also makes it possible to achieve highly personalized and flexible services, pointing to a new direction for the development of future communication networks.

[0065] In a NetAgent-based collaborative communication architecture, a key issue is how to optimize agent decision-making and support its continuous evolution through network feedback. However, after multiple agents collaborate to generate customized services, the system can only provide a single overall feedback signal. The core challenge is how to rationally distribute this single feedback signal to each agent based on their contribution.

[0066] Figure 1 illustrates a schematic diagram of a possible application framework 100 that can be used to implement some embodiments of this application in a communication system. As shown in Figure 1, network elements in the communication system are connected via interfaces (e.g., NG, Xn) or air interfaces. These network element nodes, such as core network device 130, access network node (RAN node) 120, terminal 110, OAM 140, or cloud server (such as an OTT (Over-The-Top) server, not shown in Figure 1), are equipped with one or more AI modules (only one is shown in Figure 1 for clarity). The aforementioned access network node can be a single RAN node or can include multiple RAN nodes, for example, including CU 122-1, 122-2 and DU 124-1, 124-2. CU 122-1 (or CU 122-2) and DU 124-1 (or DU 124-2) can also be equipped with one or more AI modules. Optionally, the CU can also be split into CU-CP and CU-UP. One or more AI modules are configured in CU-CP and / or CU-UP.

[0067] In multi-agent systems, the policies of individual agents differ in quality, but the behavior of each agent impacts the final performance of the entire system. QMIX is an algorithmic framework for multi-agent reinforcement learning, aiming to optimize the overall performance of multi-agent systems by combining centralized training with distributed execution, while maintaining the independent decision-making capabilities of each agent. However, the QMIX algorithm is built within a reinforcement learning framework, and its information interaction and value evaluation rely on specific, pre-defined structured data such as state and action spaces. Therefore, the QMIX algorithm is incompatible with the interaction mechanism of NetAgent. Furthermore, the QMIX algorithm lacks a differentiated update mechanism.

[0068] In view of this, embodiments of this disclosure propose a communication technology solution, such as a communication technology solution based on the architecture of a next-generation core network (A-Core) based on agents. This solution is based on a novel A-Core architecture and introduces an evaluator module within the agents, enabling semantic interaction with other components. This evaluator module uses its own large model to evaluate and score the output decisions. Furthermore, an evaluation agent (Critic agent) is introduced in the A-Core, responsible for quantifying the contribution of each agent through natural language interaction. During network operation, based on the execution results of the generative network (GN), the evaluation agent generates optimization update information for different agents and transmits this feedback to the agents participating in the GN generation process to guide their performance optimization. This communication technology solution distinguishes the contribution of each agent through natural language information exchanged between participating NetAgents and adaptively optimizes the performance of each agent in a differentiated manner based on feedback from a single communication network. Specific embodiments of this disclosure will be further described in the embodiments shown in Figures 2 to 23.

[0069] A typical system architecture of this disclosure is A-Core, a next-generation core network based on intelligent agents with different roles. The intelligent agents use an LLM (Limited Linear Model) as their intelligent brain, pre-trained using knowledge from the communication domain to become communication expert models. Figure 2 illustrates the system architecture of A-Core according to some embodiments of this disclosure, and Figure 3 illustrates the internal components of the intelligent agents according to some embodiments of this disclosure. As shown in Figure 2, A-Core consists of multiple intelligent agents with different functions and common components. The internal components of each intelligent agent are shown in Figure 3, which integrate relevant network functions, capabilities, and information. This disclosure is applicable to the next-generation core network architecture shown in Figure 2, and the functions of each intelligent agent in Figure 2 are described below.

[0070] Network Planning Agent 212: Understands the complex requirements of user input and decomposes the user's intent into a series of simple and easy-to-execute sub-tasks. Network Assembly Agent 214: Based on the input task or sub-task description, selects the necessary functionalities or external application programming interfaces (APIs) from the toolbox 226. Network Connection Agent 216: Responsible for intelligent connection management, including configuring terminal and base station functions, and establishing and updating end-to-end network topology and connections according to business requirements. Network Execution Agent 218: Responsible for computing resource management, including the deployment, updating, and deletion of functional instances, and dynamic scheduling of computing resources.

[0071] Network Generative Pre-trained Transformer (NetGPT) 220: A large-scale generative pre-trained model for mobile networks that infers expected results based on input prompts, etc. Agent-based Interface (ABI) 222: Represents a declarative interface between agents and public components. It understands the purpose of a message based on NetGPT and determines how to forward the message to the appropriate agent or public component. Unlike imperative interfaces that specify service operations, ABI declarative interfaces specify goals. Consumers only need to provide a goal description when invoking agents or public components; the ABI understands the purpose of the message and forwards it to the appropriate agent or public component for processing based on the goal. Public Memory 224: Used to collect and store network data, knowledge, etc., for agents to query. Toolbox 226: Stores a large amount of network function and application function information, which agents can flexibly arrange and assemble to generate personalized application networks, etc. Sandbox 228: Used to verify the reasoning results of intelligent agents, simulating the operation and performance of the application network before deployment.

[0072] Evaluation Agent 230: Used to evaluate the overall performance of all agents participating in the GN 219 generation process, and to generate optimization information for each agent based on the GN execution results. The agents and components shown in Figure 2 are only examples; the future core network may not be limited to the agents and components described above, and may also include other functions.

[0073] The aforementioned network elements are agents and common components defined in A-Core, used to help to more clearly illustrate the technical solutions of the embodiments of this disclosure. The technical solutions of the embodiments of this disclosure are not limited thereto. The functional descriptions of each component of agent 310 in Figure 3 (such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 shown in Figure 2) are as follows.

[0074] AgentGPT 312: A large-scale agent model for reasoning and decision-making based on input. AgentGPT is a large model pre-trained on knowledge of communication networks. Received information (e.g., user information, decision-making and evaluation processes of individual agents, relevant historical information returned from memory) can be used as part of a prompt, input into the large model, which then performs reasoning and decision-making based on the prompt. The reasoning and decision-making process involves segmenting and encoding the input prompt before inputting it into the large model (e.g., the large model may include a Transformer-based neural network). This neural network generates a probability distribution for the next token each time and selects the next generated token based on the probability distribution, where a token can be a character, a word, or a phrase. Agents can distinguish between reasoning and decision-making content by designing output formats, such as <reasoning>…<\reasoning> and <decision>…<\decision> format fields.

[0075] Tool 314: A tool library of network and application functions inherent in the intelligent agent itself. Private Memory 316: The intelligent agent's private memory, recording work logs not sent to the ABI. Actor Module 318: Interfaces for different platforms, connecting the physical and digital worlds, responsible for semantic translation.

[0076] The Evaluator module 320 is a large-model-based AI module that evaluates the merits of the agent's decisions based on stored memories and knowledge of the core network environment, and outputs an assessment of the probability that the agent's current decision can correctly generate the GN. Correctly generating the GN means that the tool modules invoked by the agent can be correctly deployed within the GN and complete end-to-end operation, logically fulfilling the user's intent. If any functions deployed within the GN are missing, or the workflow between functional modules cannot be correctly established (e.g., the output of one function does not match the input of the next), thus preventing the establishment of end-to-end service functions, it means that the GN has failed to execute its task.

[0077] Figure 4 illustrates a schematic diagram of a communication flow 400 according to some embodiments of the present disclosure. As shown in Figure 4, flow 400 involves a first intelligent agent 402, a second intelligent agent 404, an intelligent agent 410 for evaluation, and a common memory 412. Unless otherwise specified, the term "device" in this application can refer to the device itself, a module within the device (e.g., a processor, chip, or chip system), or a logic module or software capable of implementing all or part of the device's functions. Generative networks (GNs) are generated based on multiple decisions made by multiple intelligent agents.

[0078] An example of the first agent 402 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the second agent 404 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the agent 410 used for evaluation may refer to evaluation agent 230 as shown in Figure 2. An example of the common memory 412 may refer to common memory 224 as shown in Figure 2.

[0079] In process 400, at 422, the first agent 402 receives input information 420. In some exemplary embodiments, the input information includes a user request for performing a task. Optionally, the user request is based on a natural language representation; for example, the user request may be unstructured data containing semantic elements. Optionally, the user request may also be based on structured or semi-structured data representations; for example, the user request may contain relevant tags used to segment semantic elements. Optionally, the first agent also sends the input information to an agent for evaluation. Accordingly, at 423, the agent for evaluation 410 obtains the user request for performing the task.

[0080] Alternatively, in some exemplary embodiments, the input information includes a third decision of a third agent among a plurality of agents (e.g., the plurality of agents include a first agent 402, a second agent 404, a third agent, etc.), the third agent being an upstream agent that outputs the third decision to the first agent, and the first agent determining a first decision based on the third decision.

[0081] At 424, based on input information 420, the first agent 402 determines a first decision 430. In some exemplary embodiments, during the determination of the first decision, the first agent sends a second request to a fifth network function entity (the fifth network function entity may involve the public memory 412 shown in FIG. 4, such as the public memory 224 shown in FIG. 2, or the private memory 316 of the first agent 310 shown in FIG. 3), the second request including input information, and the fifth network function entity is used to store data. Throughout this disclosure, a network function entity may be equivalent to a network function instance; for example, the fifth network function entity may be considered as a network function instance.

[0082] Accordingly, in some exemplary embodiments, the fifth network functional entity receives a second request from the first agent, determines second information based on the input information included in the second request, and sends the second information to the first agent. Exemplarily, the second information indicates and / or is used to determine historical reasoning trajectory information. For example, the second information includes at least one of the following: historical input information related to the input information (e.g., historical input information similar to the input information), and multiple historical decisions (i.e., historical orchestration information) and historical individual evaluation information corresponding to the historical input information. Accordingly, in some exemplary embodiments, the first agent receives the second information in response to the second request from the fifth network functional entity, and uses the generative pre-trained model of the first agent to determine a first decision based on the second information and the input information.

[0083] At 426, the first agent 402 determines the first agent evaluation information 442 for generating the GN. The individual evaluation information is the information the agent uses to self-evaluate its own decision to correctly generate the GN during decision-making. When encountering complex tasks, the agent can make multiple output attempts based on the self-evaluation information and select a high-quality output solution from these attempts. This process effectively enhances the robustness of the decision and improves the quality of the decision. In some exemplary embodiments, during the determination of the first agent evaluation information, the first agent 402 sends input information and the first decision to the evaluation module within the first agent. For example, the agent uses its own large model, AgentGPT, to make decisions and performs self-evaluation.

[0084] In some exemplary embodiments, a fifth network functional entity receives a third request from an evaluation module, including input information and a first decision; determines third information based on the input information and the first decision; and sends the third information to the evaluation module. Exemplarily, the third information indicates and / or is used to determine historical reasoning trajectory information. For example, the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. Accordingly, in some exemplary embodiments, a first agent receives first agent evaluation information from the evaluation module. The first agent evaluation information is determined by the evaluation module using a generative pre-trained model based on the third information.

[0085] Subsequently, at 428, the first agent 402 sends a first decision 430 to the second agent 404 among a plurality of agents. Accordingly, in some exemplary embodiments, at 432, the second agent 404 receives the first decision 430 from the first agent 402. Subsequently, based on the first decision 430, the second agent 404 determines a second decision and determines second individual evaluation information for generating the GN.

[0086] Subsequently, in some exemplary embodiments, at 440, the first agent 402 further sends a first decision 430 and first agent evaluation information 442 to the agent 410 used for evaluation. Optionally, in some exemplary embodiments, the second agent further sends a second decision and second agent evaluation information to the agent used for evaluation.

[0087] Accordingly, at 460, the agent 410 for evaluation receives multiple decisions from multiple agents, and individual evaluation information corresponding to each of the multiple decisions (e.g., a first decision and first individual evaluation information from a first agent, and a second decision and second individual evaluation information from a second agent). In some exemplary embodiments, based on a user request, at 462, the agent 410 for evaluation sends a first request 464 including the user request to a common memory 412.

[0088] Accordingly, at 466, the common memory 412 receives a first request 464 from the agent 410 for evaluation. At 468, the common memory 412 determines first information 472 based on the user request. Exemplarily, the first information indicates and / or is used to determine historical reasoning trajectory information. For example, the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests. Thereafter, at 470, the common memory 412 sends the first information 472 to the agent 410 for evaluation. Accordingly, in some exemplary embodiments, at 474, the agent 410 for evaluation receives first information 468 from the common memory 412 in response to the first request 464 to obtain the first information 472. Thus, the agent for evaluation, by means of semi-structured retrieval information, can fully utilize historical reasoning trajectory information to perform in-depth learning on the semi-structured retrieval information generated during the interaction.

[0089] Subsequently, at 480, the agent 410 for evaluation, based on user requests, multiple decisions, and individual evaluation information, determines overall evaluation information (e.g., an overall evaluation of the entire cooperative orchestration process, eva_total, and this overall evaluation can be viewed as an end-to-end evaluation) to instruct multiple agents to generate the GN. In some exemplary embodiments, the common memory also receives the overall evaluation information from the agent 410 for evaluation. In this way, the agent for evaluation can adaptively estimate the global policy, thereby improving the adaptability of the agent-based next-generation core network A-Core. Furthermore, when the user request is based on natural language representation, the agent for evaluation is applied to the interaction scenario of a network agent (NetAgent) in the communications domain. The interaction and value evaluation mode of the agent for evaluation based on structured data is integrated into the natural language-based NetAgent interaction system, thus enabling it to match the interaction mechanism with NetAgent.

[0090] In some exemplary embodiments, the overall evaluation information is determined based on user requests, multiple decisions, individual evaluation information, and first information. For example, the agent used for evaluation uses a generative pre-trained model to determine the overall evaluation information based on user requests, multiple decisions, individual evaluation information, and first information. In this way, the adaptability of the agent-based next-generation core network A-Core is further improved during the adaptive estimation of the global policy because the agent used for evaluation can make full use of historical inference trajectory information.

[0091] In some exemplary embodiments, the agent used for evaluation also receives information from the GN agent indicating the generation result of the GN. In some exemplary embodiments, the agent used for evaluation also determines the generation error information of the GN (e.g., overall error information) based on the information of the generation result and the overall evaluation information. In this way, the agent can be rewarded if the GN is generated correctly in function, and will be penalized if the GN fails to be generated correctly in function.

[0092] In some exemplary embodiments, the agent used for evaluation also sends generation error information (e.g., overall error information) and multiple contributions (e.g., contribution allocation) of the multiple agents to the generation of the GN to multiple agents. In some exemplary embodiments, the first contribution of the first agent among the multiple agents to the generation of the GN is determined based on information about the generation result and individual evaluation information corresponding to the first agent's first decision. In this way, different agents can receive differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their degree of contribution in the task, thereby optimizing their behavior in a targeted manner.

[0093] In some exemplary embodiments, the agent used for evaluation also updates the generative pre-trained model based on generation error information (e.g., global error information). This allows the agent to receive differentiated reward feedback based on its actual performance in the task, thereby optimizing its behavior. In some exemplary embodiments, the global evaluation information includes the probability that GN is correctly generated, and individual evaluation information is used to determine the probability that the agent's decisions result in the correct generation of GN. Thus, the agent's estimation of the global policy can be quantified based on probabilistic information.

[0094] In one example, the agent used for evaluation generates update information based on the success or failure result of the GN task (e.g., a 1 for success and a 0 for failure). This update information is used to update its own large model, AgentGPT, and different update information is generated for different agents, thus providing feedback to each agent to guide its performance optimization. The GN agent feeds back the GN result information to the agent used for evaluation. If the GN is successfully generated and functionally executable, the feedback result information GN_feedback = 1. Otherwise, the feedback result information GN_feedback = 0.

[0095] The agent used for evaluation generates updated information through the following calculations:

[0096] The overall error information is diff = GN_feedback - eva_total. For each agent, the contribution allocation (weights leading to success / failure) is calculated: the contribution allocation for each agent is contrb_ass = |1 - GN_feedback - eva|, where eva is the agent's evaluation result.

[0097] Subsequently, in this example, the agent used for evaluation uses the calculated overall error information diff as its reward feedback to update its large model AgentGPT.

[0098] Figure 5 illustrates a schematic diagram of a communication flow 500 according to some embodiments of the present disclosure. As shown in Figure 5, the flow 500 involves a first agent 402, a second agent 404, a third agent 406, an agent 410 for evaluation, and a common memory 412. Unless otherwise specified, the device in this application can refer to the device itself, a module within the device (e.g., a processor, a chip, or a chip system), or a logic module or software capable of implementing all or part of the device's functions. The generative network GN is generated based on multiple decisions made by multiple agents. For example, in Figure 5, GN is generated based on the decisions of at least three agents (first agent 402, second agent 404, and third agent 406), wherein the second agent 404 determines a second decision 536 based on a first decision 530 made by the first agent 402, the third agent 406 determines a third decision based on the second decision 536 made by the second agent 404, and the GN agent determines GN (not shown in Figure 5) based on the third decision made by the third agent 406.

[0099] An example of the first agent 402 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the second agent 404 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the third agent 406 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the agent 410 used for evaluation may refer to the evaluation agent 230 as shown in Figure 2. An example of the common memory 412 may refer to the common memory 224 as shown in Figure 2.

[0100] In process 500, at 522, the first agent 402 receives input information 520. In some exemplary embodiments, the input information includes a user request for performing a task. Optionally, the user request is based on natural language representation; for example, the user request may be unstructured data containing semantic elements. Optionally, the user request may also be based on structured or semi-structured data representation; for example, the user request may contain relevant tags used to segment semantic elements. Optionally, the first agent also sends the input information to the agent for evaluation. Accordingly, at 523, the agent for evaluation 410 obtains the user request for performing the task.

[0101] Alternatively, in some exemplary embodiments, the input information includes a fourth decision of a fourth agent among a plurality of agents (e.g., the plurality of agents include a first agent 402, a second agent 404, a third agent 406, a fourth agent, etc.), the fourth agent being an upstream agent that outputs the fourth decision to the first agent, and the first agent determining a first decision based on the fourth decision.

[0102] At 524, based on the input information 520, the first agent 402 determines a first decision 530. In some exemplary embodiments, during the process of determining the first decision, the first agent sends a second request to a fifth network functional entity (which may involve the public memory 412 shown in FIG. 5, such as the public memory 224 shown in FIG. 2, or the private memory 316 of the first agent 310 shown in FIG. 3), the second request including input information, and the fifth network functional entity is used to store data.

[0103] Accordingly, in some exemplary embodiments, the fifth network functional entity receives a second request from the first agent, determines second information based on the input information included in the second request, and sends the second information to the first agent. Exemplarily, the second information indicates and / or is used to determine historical reasoning trajectory information. For example, the second information includes at least one of the following: historical input information related to the input information (e.g., historical input information similar to the input information), and multiple historical decisions (i.e., historical orchestration information) and historical individual evaluation information corresponding to the historical input information. Accordingly, in some exemplary embodiments, the first agent receives the second information in response to the second request from the fifth network functional entity, and uses the generative pre-trained model of the first agent to determine a first decision based on the second information and the input information.

[0104] At 526, the first agent 402 determines the first volume evaluation information 542 for generating the GN. In some exemplary embodiments, during the determination of the first volume evaluation information, the first agent 402 sends input information and a first decision to the evaluation module within the first agent. For example, the agent makes a decision using its own large model AgentGPT and performs a self-evaluation.

[0105] In some exemplary embodiments, a fifth network functional entity receives input information and a first decision third request from an evaluation module, determines third information based on the input information and the first decision, and sends the third information to the evaluation module. Exemplarily, the third information indicates and / or is used to determine historical reasoning trajectory information. For example, the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. Accordingly, in some exemplary embodiments, a first agent receives first agent evaluation information from the evaluation module. The first agent evaluation information is determined by the evaluation module using a generative pre-trained model based on the third information.

[0106] Subsequently, at 528, the first agent 402 sends a first decision 530 to the second agent 404 among a plurality of agents. Accordingly, in some exemplary embodiments, at 532, the second agent 404 receives the first decision 530 from the first agent 402. Subsequently, based on the first decision 530, the second agent 404 determines a second decision and determines second individual evaluation information for generating the GN (not shown in FIG5).

[0107] Subsequently, at 534, the second agent 404 sends a second decision 536 to the third agent 406 among the plurality of agents. Accordingly, in some exemplary embodiments, at 538, the third agent 406 receives the second decision 536 from the second agent 404. Subsequently, based on the second decision 536, the third agent 406 determines its third decision and determines the third agent 406's third agent evaluation information for generating the GN (not shown in FIG5).

[0108] Subsequently, in some exemplary embodiments, at 540, the first agent 402 further sends a first decision 530 and first agent evaluation information 542 to the agent 410 for evaluation. In some exemplary embodiments, at 544, the second agent 404 further sends a second decision 536 and second agent evaluation information 546 to the agent 410 for evaluation. In some exemplary embodiments, at 548, the third agent 406 further sends a third decision and third agent evaluation information 550 to the agent 410 for evaluation.

[0109] Accordingly, at 560, the agent 410 for evaluation receives multiple decisions from multiple agents, and individual evaluation information corresponding to each of the multiple decisions (e.g., a first decision and first individual evaluation information from a first agent, a second decision and second individual evaluation information from a second agent, and a third decision and third individual evaluation information from a third agent). In some exemplary embodiments, based on a user request, at 562, the agent 410 for evaluation sends a first request 564 including the user request to a common memory 412.

[0110] Accordingly, at 566, the common memory 412 receives a first request 564 from the agent 410 for evaluation. At 568, the common memory 412 determines first information 572 based on the user request. Exemplarily, the first information indicates and / or is used to determine historical reasoning trajectory information. For example, the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests. Thereafter, at 570, the common memory 412 sends the first information 572 to the agent 410 for evaluation. Accordingly, in some exemplary embodiments, at 574, the agent 410 for evaluation receives the first information 572 in response to the first request 564 from the common memory 412 to obtain the first information 572. Thus, the agent for evaluation, by means of semi-structured retrieval information, can fully utilize historical reasoning trajectory information to perform in-depth learning on the semi-structured retrieval information generated during the interaction.

[0111] Subsequently, at 580, the agent 410 for evaluation determines overall evaluation information (e.g., overall evaluation eva_total) to instruct multiple agents to generate the GN based on user requests, multiple decisions, and individual evaluation information. In some exemplary embodiments, the common memory also receives the overall evaluation information from the agent 410 for evaluation. In this way, the agent for evaluation can adaptively estimate the global policy, thereby improving the adaptability of the agent-based next-generation core network A-Core. Furthermore, when the user request is based on natural language representation, the agent for evaluation is applied to the NetAgent interaction scenario. The interaction and value evaluation patterns of the agent for evaluation based on structured data are integrated into the natural language-based NetAgent interaction system, thus enabling it to match the interaction mechanism with NetAgent.

[0112] In some exemplary embodiments, during the process of determining the overall evaluation information by the agent used for evaluation, the overall evaluation information is determined based on user requests, multiple decisions, individual evaluation information, and first information. For example, the agent used for evaluation uses a generative pre-trained model to determine the overall evaluation information based on user requests, multiple decisions, individual evaluation information, and first information. In this way, during the adaptive estimation of the global policy, the agent used for evaluation can make full use of historical inference trajectory information, thus further improving the adaptability of the agent-based next-generation core network A-Core.

[0113] In some exemplary embodiments, the agent used for evaluation also receives information from the GN agent indicating the generation result of the GN. In some exemplary embodiments, the agent used for evaluation also determines the generation error information of the GN (e.g., overall error information) based on the information of the generation result and the overall evaluation information. In this way, the agent can be rewarded if the GN is generated correctly in function, and will be penalized if the GN fails to be generated correctly in function.

[0114] In some exemplary embodiments, the agent used for evaluation also sends generation error information (e.g., overall error information) and multiple contributions (e.g., contribution allocation) of the multiple agents to the generation of the GN to multiple agents. In some exemplary embodiments, the first contribution of the first agent among the multiple agents to the generation of the GN is determined based on information about the generation result and individual evaluation information corresponding to the first agent's first decision. In this way, different agents can receive differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their degree of contribution in the task, thereby optimizing their behavior in a targeted manner.

[0115] In some exemplary embodiments, the agent used for evaluation also updates the generative pre-trained model based on generation error information (e.g., overall error information). In some exemplary embodiments, the overall evaluation information includes the probability that GN is correctly generated, and individual evaluation information is used to determine the probability that the agent's decisions result in the correct generation of GN. Thus, the agent used for evaluation can also receive differentiated reward feedback based on its actual performance in the task, thereby optimizing its behavior. In some exemplary embodiments, the overall evaluation information includes the probability that GN is correctly generated, and individual evaluation information is used to determine the probability that the agent's decisions result in the correct generation of GN. Thus, the agent used for evaluation can quantify its estimation of the global policy based on probabilistic information.

[0116] In one example, the agent used for evaluation generates update information based on the success or failure result of the GN task (e.g., a 1 for success and a 0 for failure). This update information is used to update its own large model, AgentGPT, and different update information is generated for different agents, thus providing feedback to each agent to guide its performance optimization. The GN agent feeds back the GN result information to the agent used for evaluation. If the GN is successfully generated and functionally executable, the feedback result information GN_feedback = 1. Otherwise, the feedback result information GN_feedback = 0.

[0117] The agent used for evaluation generates updated information through the following calculations:

[0118] The overall error information is diff = GN_feedback - eva_total. For each agent, the contribution allocation (weights leading to success / failure) is calculated: the contribution allocation for each agent is contrb_ass = |1 - GN_feedback - eva|, where eva is the agent's evaluation result.

[0119] Subsequently, in this example, the agent used for evaluation uses the calculated overall error information diff as its reward feedback to update its large model AgentGPT.

[0120] Figure 6 illustrates a schematic diagram of a communication flow 600 according to some embodiments of the present disclosure. As shown in Figure 6, flow 600 involves a first agent 402, a second agent 404, a third agent 406, an agent 410 for evaluation, and a common memory 412. Unless otherwise specified, the device in this application can refer to the device itself, a module within the device (e.g., a processor, a chip, or a chip system), or a logic module or software capable of implementing all or part of the device's functions. The generative network GN is generated based on multiple decisions made by multiple agents. For example, in Figure 6, GN is generated based on the decisions of at least three agents (first agent 402, second agent 404, and third agent 406), wherein the second agent 404 determines a second decision based on a first decision 630 of the first agent 402, the third agent 406 determines a third decision based on the first decision 630 of the first agent 402, and the GN agent determines GN (not shown in Figure 6) based on the second decision of the second agent 404 and the third decision of the third agent 406.

[0121] An example of the first agent 402 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the second agent 404 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the third agent 406 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the agent 410 used for evaluation may refer to the evaluation agent 230 as shown in Figure 2. An example of the common memory 412 may refer to the common memory 224 as shown in Figure 2.

[0122] In process 600, at 622, the first agent 402 receives input information 620. In some exemplary embodiments, the input information includes a user request for performing a task. Optionally, the user request is based on a natural language representation; for example, the user request may be unstructured data containing semantic elements. Optionally, the user request may also be based on structured or semi-structured data representations; for example, the user request may contain relevant tags used to segment semantic elements. Optionally, the first agent also sends the input information to an agent for evaluation. Accordingly, at 623, the agent for evaluation 410 obtains the user request for performing the task.

[0123] Alternatively, in some exemplary embodiments, the input information includes a fourth decision of a fourth agent among a plurality of agents (e.g., the plurality of agents include a first agent 402, a second agent 404, a third agent 406, a fourth agent, etc.), the fourth agent being an upstream agent that outputs the fourth decision to the first agent, and the first agent determining a first decision based on the fourth decision.

[0124] At 624, based on the input information 620, the first agent 402 determines a first decision 630. In some exemplary embodiments, during the determination of the first decision, the first agent sends a second request to a fifth network functional entity (which may involve the public memory 412 shown in FIG. 6, such as the public memory 224 shown in FIG. 2, or the private memory 316 of the first agent 310 shown in FIG. 3), the second request including input information, and the fifth network functional entity is used to store data.

[0125] Accordingly, in some exemplary embodiments, the fifth network functional entity receives a second request from the first agent, determines second information based on the input information included in the second request, and sends the second information to the first agent. Exemplarily, the second information indicates and / or is used to determine historical reasoning trajectory information. For example, the second information includes at least one of the following: historical input information related to the input information (e.g., historical input information similar to the input information), and multiple historical decisions (i.e., historical orchestration information) and historical individual evaluation information corresponding to the historical input information. Accordingly, in some exemplary embodiments, the first agent receives the second information in response to the second request from the fifth network functional entity, and uses the generative pre-trained model of the first agent to determine a first decision based on the second information and the input information.

[0126] At 626, the first agent 402 determines the first volume evaluation information 642 for generating the GN. In some exemplary embodiments, during the determination of the first volume evaluation information, the first agent 402 sends input information and a first decision to the evaluation module within the first agent. For example, the agent makes a decision using its own large model AgentGPT and performs a self-evaluation.

[0127] In some exemplary embodiments, a fifth network functional entity receives a third request from an evaluation module, including input information and a first decision; determines third information based on the input information and the first decision; and sends the third information to the evaluation module. Exemplarily, the third information indicates and / or is used to determine historical reasoning trajectory information. For example, the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. Accordingly, in some exemplary embodiments, a first agent receives first agent evaluation information from the evaluation module. The first agent evaluation information is determined by the evaluation module using a generative pre-trained model based on the third information.

[0128] Subsequently, at 628, the first agent 402 sends a first decision 630 to the second agent 404 among a plurality of agents. Accordingly, in some exemplary embodiments, at 632, the second agent 404 receives the first decision 630 from the first agent 402. Subsequently, based on the first decision 630, the second agent 404 determines a second decision and determines second individual evaluation information for generating the GN (not shown in FIG6).

[0129] Furthermore, at 634, the first agent 402 sends a first decision 630 to the third agent 406 among a plurality of agents. Accordingly, in some exemplary embodiments, at 638, the third agent 406 receives the first decision 630 from the first agent 402. Thereafter, based on the first decision 630, the third agent 406 determines its third decision and determines third agent evaluation information (not shown in FIG6) for generating the GN.

[0130] Subsequently, in some exemplary embodiments, at 640, the first agent 402 further sends a first decision 630 and first agent evaluation information 642 to the agent 410 for evaluation. In some exemplary embodiments, at 644, the second agent 404 further sends a second decision 636 and second agent evaluation information 646 to the agent 410 for evaluation. In some exemplary embodiments, at 648, the third agent 406 further sends a third decision and third agent evaluation information 650 to the agent 410 for evaluation.

[0131] Accordingly, at 660, the agent 410 for evaluation receives multiple decisions from multiple agents, and individual evaluation information corresponding to each of the multiple decisions (e.g., a first decision and first individual evaluation information from a first agent, a second decision and second individual evaluation information from a second agent, and a third decision and third individual evaluation information from a third agent). In some exemplary embodiments, based on a user request, at 662, the agent 410 for evaluation sends a first request 664 including the user request to a common memory 412.

[0132] Accordingly, at 666, the common memory 412 receives a first request 664 from the agent 410 for evaluation. At 668, the common memory 412 determines first information 672 based on the user request. Exemplarily, the first information indicates and / or is used to determine historical reasoning trajectory information. For example, the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests. Thereafter, at 670, the common memory 412 sends the first information 672 to the agent 410 for evaluation. Accordingly, in some exemplary embodiments, at 674, the agent 410 for evaluation receives the first information 672 in response to the first request 664 from the common memory 412 to obtain the first information 672. Thus, the agent for evaluation, by means of semi-structured retrieval information, can fully utilize historical reasoning trajectory information to perform in-depth learning on the semi-structured retrieval information generated during the interaction.

[0133] Subsequently, at 680, the agent 410 for evaluation determines overall evaluation information (e.g., overall evaluation eva_total) to instruct multiple agents to generate the GN based on user requests, multiple decisions, and individual evaluation information. In some exemplary embodiments, the common memory also receives the overall evaluation information from the agent 410 for evaluation. In this way, the agent for evaluation can adaptively estimate the global policy, thereby improving the adaptability of the agent-based next-generation core network A-Core. Furthermore, when the user request is based on natural language representation, the agent for evaluation is applied to the NetAgent interaction scenario. The interaction and value evaluation patterns of the agent for evaluation based on structured data are integrated into the natural language-based NetAgent interaction system, thus enabling it to match the interaction mechanism with NetAgent.

[0134] In some exemplary embodiments, during the process of determining the overall evaluation information by the agent used for evaluation, the overall evaluation information is determined based on user requests, multiple decisions, individual evaluation information, and first information. For example, the agent used for evaluation uses a generative pre-trained model to determine the overall evaluation information based on user requests, multiple decisions, individual evaluation information, and first information. In this way, during the adaptive estimation of the global policy, the agent used for evaluation can make full use of historical inference trajectory information, thus further improving the adaptability of the agent-based next-generation core network A-Core.

[0135] In some exemplary embodiments, the agent used for evaluation also receives information from the GN agent indicating the generation result of the GN. In some exemplary embodiments, the agent used for evaluation also determines the generation error information of the GN (e.g., overall error information) based on the information of the generation result and the overall evaluation information. In this way, the agent can be rewarded if the GN is generated correctly in function, and will be penalized if the GN fails to be generated correctly in function.

[0136] In some exemplary embodiments, the agent used for evaluation also sends generation error information (e.g., overall error information) and multiple contributions (e.g., contribution allocation) of the multiple agents to the generation of the GN to multiple agents. In some exemplary embodiments, the first contribution of the first agent among the multiple agents to the generation of the GN is determined based on information about the generation result and individual evaluation information corresponding to the first agent's first decision. In this way, different agents can receive differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their degree of contribution in the task, thereby optimizing their behavior in a targeted manner.

[0137] In some exemplary embodiments, the agent used for evaluation also updates the generative pre-trained model based on generation error information (e.g., overall error information). In some exemplary embodiments, the overall evaluation information includes the probability that GN is correctly generated, and individual evaluation information is used to determine the probability that the agent's decisions result in the correct generation of GN. Thus, the agent used for evaluation can also receive differentiated reward feedback based on its actual performance in the task, thereby optimizing its behavior. In some exemplary embodiments, the overall evaluation information includes the probability that GN is correctly generated, and individual evaluation information is used to determine the probability that the agent's decisions result in the correct generation of GN. Thus, the agent used for evaluation can quantify its estimation of the global policy based on probabilistic information.

[0138] In one example, the agent used for evaluation generates update information based on the success or failure result of the GN task (e.g., a 1 for success and a 0 for failure). This update information is used to update its own large model, AgentGPT, and different update information is generated for different agents, thus providing feedback to each agent to guide its performance optimization. The GN agent feeds back the GN result information to the agent used for evaluation. If the GN is successfully generated and functionally executable, the feedback result information GN_feedback = 1. Otherwise, the feedback result information GN_feedback = 0.

[0139] The agent used for evaluation generates updated information through the following calculations:

[0140] The overall error information is diff = GN_feedback - eva_total. For each agent, the contribution allocation (weights leading to success / failure) is calculated: the contribution allocation for each agent is contrb_ass = |1 - GN_feedback - eva|, where eva is the agent's evaluation result.

[0141] Subsequently, in this example, the agent used for evaluation uses the calculated overall error information diff as its reward feedback to update its large model AgentGPT.

[0142] The embodiments shown in Figures 4 to 6 construct the interaction pattern between the agent used for evaluation and the agents participating in the GN generation process. This embodiment is based on key information such as user input requests, the decision outputs of each agent, and the evaluation results. It also fully utilizes relevant historical information stored in a shared memory (e.g., historical reasoning trajectory information) and leverages the evaluation agent's own large model, AgentGPT, to perform deep learning on the semi-structured natural language generated during the interaction. Through deep learning, the evaluation agent can flexibly learn the communication network environment, facilitating the adaptive adjustment of each agent's contribution.

[0143] In the field of multi-agent systems, multi-agent reinforcement learning techniques, represented by the QMIX algorithm, typically rely on pre-defined state and action spaces to estimate the value of the global policy in the system. This approach is limited by a predefined framework, resulting in relatively insufficient flexibility and adaptability. In the embodiments shown in Figures 4 to 6, the evaluation agent utilizes semi-structured natural language generated during the interaction process and contextual information from the communication network's operation stored in a shared memory (e.g., relevant historical information such as historical reasoning trajectories) to adaptively estimate the global policy, thus improving the adaptability of A-Core.

[0144] Figure 7 illustrates a schematic diagram of a communication flow 700 according to some embodiments of the present disclosure, which relates to a feedback allocation mechanism applicable to A-Core. As shown in Figure 7, flow 700 involves a first agent 402 and an agent 410 for evaluation. Unless otherwise specified, the term "device" in this application can refer to the device itself, a module within the device (e.g., a processor, chip, or chip system), or a logic module or software capable of implementing all or part of the device's functions. The generative network GN is generated based on multiple decisions made by multiple agents.

[0145] An example of the first agent 402 may involve agent 310 as shown in FIG3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in FIG2. An example of the agent 410 for evaluation may involve evaluation agent 230 as shown in FIG2. In some exemplary embodiments, the agent for evaluation updates the generative pre-trained model based on the generation error information (e.g., overall error information) of the GN. In some exemplary embodiments, the agent for evaluation sends generation error information (e.g., overall error information) and multiple contributions of the multiple agents to the generation of the GN to multiple agents. In some exemplary embodiments, the first contribution of the first agent among the multiple agents to the generation of the GN is determined based on information of the generation result and individual evaluation information corresponding to the first decision of the first agent.

[0146] In process 700, at 720, the agent 410 for evaluation sends generation error information (e.g., overall error information) and a first contribution 722 of the first agent to the generation of GN to the first agent 402. Accordingly, at 724, the first agent 402 receives the generation error information (e.g., overall error information) and the first contribution 722 from the agent 410 for evaluation. Subsequently, at 730, the first agent 402 updates its generative pre-trained model based on the generation error information (e.g., overall error information) and the first contribution 722. In some exemplary embodiments, during the process of the first agent 402 updating the generative pre-trained model, the first agent 402 determines reward information based on the generation error information (e.g., overall error information) and the first contribution 722, and updates the generative pre-trained model based on the reward information. In this way, different agents can obtain differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their contribution to the task, thereby optimizing their behavior accordingly.

[0147] In one example, the agent used for evaluation generates update information based on the success or failure result of the GN task (e.g., a 1 for success and a 0 for failure). This update information is used to update its own large model, AgentGPT, and different update information is generated for different agents, thus providing feedback to each agent to guide its performance optimization. The GN agent feeds back the GN result information to the agent used for evaluation. If the GN is successfully generated and functionally executable, the feedback result information GN_feedback = 1. Otherwise, the feedback result information GN_feedback = 0.

[0148] Successfully generating a GN (in other words, successfully executing the GN task) means that the tool modules invoked by the agent can be correctly deployed within the GN and complete end-to-end operation, logically fulfilling the user's intent. If any functionality deployed within the GN is missing, or the workflow between various functional modules cannot be correctly established (e.g., the output of one function does not match the input of the next), thus preventing the establishment of end-to-end service functionality, it means that the GN task execution has failed.

[0149] The agent used for evaluation generates updated information through the following calculations:

[0150] The overall error information is diff = GN_feedback - eva_total. For each agent, the contribution allocation (weights leading to success / failure) is calculated: the contribution allocation for each agent is contrb_ass = |1 - GN_feedback - eva|, where eva is the agent's evaluation result.

[0151] Subsequently, in this example, the agent used for evaluation updates its large model, AgentGPT, using the calculated overall error information (diff) as its reward feedback. The agent uses the calculated overall error information (diff), its contribution allocation (contrb_ass), and the result information (GN_feedback) as feedback to itself. Furthermore, the agent multiplies diff by contrb_ass as its reward feedback and updates its large model, AgentGPT.

[0152] The embodiment in Figure 7 constructs an interaction pattern between the agent used for evaluation and the agents participating in the GN generation process. This embodiment is based on key information such as user input requests, the decision outputs of each agent, and evaluation results. It also fully utilizes relevant historical information stored in a shared memory (e.g., historical reasoning trajectory information) and leverages the large model AgentGPT of the agent used for evaluation to perform deep learning on the semi-structured natural language generated during the interaction. Through deep learning, the agent used for evaluation can flexibly learn the communication network environment, facilitating the adaptive adjustment of each agent's contribution.

[0153] In the field of multi-agent systems, multi-agent reinforcement learning techniques, represented by the QMIX algorithm, typically rely on pre-defined state and action spaces to estimate the value of the global policy in the system. This approach is limited by a predefined framework, resulting in relatively insufficient flexibility and adaptability. In the embodiment shown in Figure 7, the agent used for evaluation utilizes semi-structured natural language generated during the interaction process and contextual information from the communication network's operation stored in a shared memory (e.g., relevant historical information such as historical reasoning trajectories) to adaptively estimate the global policy, thus improving the adaptability of A-Core.

[0154] Figure 8 illustrates a schematic diagram of a communication flow 800 according to some embodiments of the present disclosure. As shown in Figure 8, flow 800 involves a first intelligent agent 402, an evaluation module 403 in the first intelligent agent 402, and a common memory 412. Unless otherwise specified, the term "device" in this application may refer to the device itself, a module within the device (e.g., a processor, chip, or chip system), or a logic module or software capable of implementing all or part of the device's functions.

[0155] An example of the first agent 402 may refer to agent 310 as shown in Figure 3, such as network planning agent 212, network assembly agent 214, network connection agent 216, and network operation agent 218 as shown in Figure 2. An example of the evaluation module 403 in the first agent 402 may refer to evaluation module 320 as shown in Figure 3. An example of the common memory 412 may refer to common memory 224 as shown in Figure 2.

[0156] In process 800, at 822, the first agent 402 receives input information 820. In some exemplary embodiments, the input information includes a user request for performing a task. Optionally, the user request is based on natural language representation; for example, the user request may be unstructured data containing semantic elements. Optionally, the user request may also be based on structured or semi-structured data representation; for example, the user request may contain relevant tags used to segment semantic elements. Alternatively, in some exemplary embodiments, the input information includes a third decision of a third agent among multiple agents (e.g., multiple agents including the first agent 402 and a third agent, etc.), the third agent being an upstream agent that outputs the third decision to the first agent, and the first agent determining a first decision based on the third decision. At 824, based on the input information 820, the first agent 402 determines its first decision 830. At 826, the first agent 402 sends the input information 820 and the first decision 830 to the evaluation module 403. Accordingly, at 828, the evaluation module 403 receives input information 820 and a first decision 830 from the first agent 402.

[0157] Subsequently, at 840, the evaluation module 403 sends a third request 842 to the fifth network functional entity (the fifth network functional entity may involve the public memory 412 shown in FIG8, such as the public memory 224 shown in FIG2, or the private memory 316 of the first agent 310 shown in FIG3). The third request 842 includes input information 820 and a first decision 830, and the fifth network functional entity is used to store data.

[0158] Accordingly, in cases where the memory involves the common memory 412 shown in FIG. 8 (e.g., the common memory 224 shown in FIG. 2), at 844, the common memory 412 receives a third request 842 from the evaluation module 403. Subsequently, at 846, the common memory 412 determines third information 850 based on the input information 820 and the first decision 830. Exemplarily, the third information indicates and / or is used to determine historical reasoning trajectory information. For example, the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. At 848, the common memory 412 sends the third information 850 to the evaluation module 403.

[0159] Accordingly, at 852, the evaluation module 403 receives third information 850 in response to the third request 842 from the fifth network functional entity (which may involve the public memory 412 shown in FIG. 8, such as the public memory 224 shown in FIG. 2, or the private memory 316 of the first agent 310 shown in FIG. 3). Subsequently, at 860, the evaluation module 403 determines first agent evaluation information based on the third information 850. For example, the evaluation module 403 uses a generative pre-trained model to determine the first agent evaluation information based on the third information 850. In some exemplary embodiments, the evaluation module 403 also sends the first agent evaluation information to the first agent 402. In some exemplary embodiments, the first agent evaluation information is used to determine the probability that a first decision results in the correct generation of the generative network GN.

[0160] The embodiment in Figure 8 designs a mechanism in A-Core where NetAgent utilizes memory to output value assessment, realizing value assessment based on semi-structured natural language information. This mechanism leverages relevant historical information stored in memory (e.g., historical reasoning trajectory information) and an external knowledge base, fully utilizing the autonomous learning capabilities of the large model to enable it to autonomously learn assessment patterns and value assessment rules. Unlike conventional methods that rely on specific, pre-defined structured data such as states and action spaces, this mechanism eliminates dependence on structured data, improving the adaptability of the communication system to complex language tasks. Furthermore, when encountering complex tasks, the agent can output multiple times for the input, selecting the decision output with the highest evaluation result as the final decision.

[0161] Conventional reinforcement learning methods, such as the QMIX algorithm, typically assume that the agent's action and state spaces are well-defined and definite, making them ill-suited to the complexity and diversity of language tasks. In the next-generation agent-based core network A-Core architecture, conventional reinforcement learning algorithms are less adaptable because the output of language tasks is mostly semi-structured natural language. The embodiment shown in Figure 8 implements value assessment based on semi-structured natural language information, independent of specific, pre-defined structured data such as state and action spaces. By utilizing relevant historical information stored in memory (e.g., historical reasoning trajectory information) and an external knowledge base, large models can autonomously learn assessment patterns and value assessment rules, enabling them to better adapt to complex communication networks.

[0162] In this way, the agent can perform self-evaluation when making decisions, thus introducing the dimension of decision quality. When faced with complex tasks, the agent can attempt multiple outputs based on the results of its self-evaluation and select the highest-quality output. This process effectively enhances the robustness of the decision and improves its quality.

[0163] Figure 9 illustrates a schematic diagram of a process 900 for overall evaluation according to some embodiments of the present disclosure, and Figure 10 illustrates a schematic diagram of an evaluation agent performing an overall evaluation according to some embodiments of the present disclosure. As shown in Figure 9, the evaluation agent 910 and the various agents 902 and 904 involved in the generation of GN 908 utilize ABI to achieve semantic interaction. During the interaction, the evaluation agent 910 fully utilizes the relevant historical information stored in the common memory 912, and based on this, summarizes the overall orchestration information of the A-Core and generates an overall evaluation result.

[0164] After agents collaborate to generate a GN, all agents involved in the GN generation process send their decision outputs and evaluation results to the evaluation agent. The evaluation agent mixes user requirements, relevant historical information, and the decision outputs and evaluation results of all agents to obtain an evaluation result for the entire GN generation process. Figure 9 illustrates the overall evaluation process 900 of the evaluation agent 910, using two agents 902 and 904 involved in GN 908 generation as an example. In Figure 10, the evaluation agent utilizes its large model AgentGPT (i.e., the hybrid model within the evaluation agent) to output the orchestration information of the entire GN generation process and the end-to-end evaluation results based on the input information. AgentGPT is a large model pre-trained on communication network knowledge. Received information (e.g., user information, the decision-making and evaluation processes of each agent, and relevant historical information returned from memory) can be used as part of a prompt and input into the large model, which then performs reasoning and decision-making based on the prompt. The reasoning and decision-making process involves segmenting and encoding the input prompts before feeding them into a large model (e.g., a Transformer-based neural network). This neural network generates a probability distribution for the next token each time and selects the next generated token based on this distribution. A token can be a character, a word, or a phrase. The agent can differentiate between reasoning and decision-making content by designing output formats, such as <reasoning>…<\reasoning> and <decision>…<\decision> format fields.

[0165] At 920, agent 1 902 receives a request for user input.

[0166] In 922, Agent 1 902 uses its own large model AgentGPT to make decisions and performs self-evaluation.

[0167] At 924, agent 1 (902) sends its decision output to agent 2 (904).

[0168] In 926, Agent 2 904 makes a decision based on the decision output received from Agent 1 902, using its own large model AgentGPT, and performs a self-evaluation.

[0169] In 928, agent 2 904 sends its decision output to GN 908. Using the internal GN agents, GN 908 generates specific GN instances and provides customized services to users.

[0170] At 930, agent 1 902 sends the request of the user input it received to the evaluation agent 910.

[0171] At 940, agent 1 902 sends its decision output and evaluation results to the evaluation agent 910.

[0172] At 942, agent 2 904 sends its decision output and evaluation results to the evaluation agent 910.

[0173] At 944, the evaluation agent 910 sends the received user input request, the decision outputs of agent 1 902 and agent 2 904 participating in the GN generation process, and the evaluation results to the common memory 912 via the ABI to request retrieval. Based on the retrieval request message, the common memory 912 retrieves relevant historical information stored therein (e.g., historical reasoning trajectory information).

[0174] At 946, the common memory 912 returns relevant historical information (e.g., historical requests similar to user input requests, as well as historical orchestration information, historical evaluation results, and historical reward feedback information corresponding to the historical requests) to the evaluation agent 910.

[0175] In 950, the evaluation agent 910 utilizes its own large model AgentGPT to form orchestration information orchestration and an overall evaluation of the entire cooperative orchestration process eva_total (i.e., end-to-end evaluation) based on the received user input requests, the decision outputs and evaluation results of each agent, and the relevant historical information returned by the common memory 912 (e.g., historical requests similar to the user input requests in 946, as well as historical orchestration information, historical evaluation results, and historical reward feedback information corresponding to the historical requests).

[0176] At 952, the evaluation agent 910 can output orchestration information and overall evaluation eva_total. The evaluation agent 910 sends the output orchestration information and overall evaluation eva_total to the common memory 912 for storage.

[0177] The embodiment in Figure 9 constructs an interaction pattern between the evaluation agent and the agents participating in the GN generation process. This embodiment is based on key information such as user input requests, the decision outputs of each agent, and the evaluation results. It also fully utilizes relevant historical information stored in a shared memory (e.g., historical reasoning trajectory information) and leverages the evaluation agent's own large model, AgentGPT, to perform deep learning on the semi-structured natural language generated during the interaction. Through deep learning, the evaluation agent can flexibly learn the communication network environment, facilitating the adaptive adjustment of each agent's contribution.

[0178] In the field of multi-agent systems, multi-agent reinforcement learning techniques, represented by the QMIX algorithm, typically rely on pre-defined state and action spaces to estimate the value of the global policy in the system. This approach is limited by a predefined framework, resulting in relatively insufficient flexibility and adaptability. In the embodiment shown in Figure 9, the evaluation agent utilizes semi-structured natural language generated during the interaction process and contextual information from the communication network's operation stored in a shared memory (e.g., relevant historical information such as historical reasoning trajectories) to adaptively estimate the global policy, thus improving the adaptability of A-Core.

[0179] Figure 11 illustrates a schematic diagram of a feedback allocation process 1100 according to some embodiments of the present disclosure, which relates to a feedback allocation mechanism applicable to A-Core. As shown in Figure 11, in the A-Core architecture, the evaluation agent 1110 can determine the contribution of different agents in the task execution process by combining the feedback information of the task execution performed by GN 1108 with the semantic interaction content between all agents 1102 and 1104 participating in the GN 1108 generation process and utilizing relevant historical information stored in memory.

[0180] The evaluation agent generates update information based on the success or failure results of the GN task (e.g., a success result of 1 and a failure result of 0). It then uses this update information to update its own large model, AgentGPT, and generates different update information for different agents, providing feedback to each agent to guide its performance optimization. Figure 11 illustrates the overall evaluation process 1100 of the evaluation agent 1110, using agents 1102 and 1104 participating in the GN 1108 generation process as examples.

[0181] The GN agent generates GN 1108 based on the orchestration results and collected tools, and at 1120, feeds back the GN's results to the evaluation agent 1110. If the GN is successfully generated and functionally executable, the feedback result information GN_feedback = 1. Otherwise, the feedback result information GN_feedback = 0.

[0182] At 1122, the evaluating agent 1110 generates updated information through the following calculations:

[0183] The overall error information is diff = GN_feedback - eva_total.

[0184] Calculate the contribution allocation (weights that lead to success / failure) for each agent:

[0185] The contribution allocation for each agent is contrb_ass = |1-GN_feedback-eva|, where eva is the agent's evaluation result.

[0186] At 1124, the evaluation agent 1110 uses the overall error information diff calculated in 1122 as its reward feedback to update its large model AgentGPT.

[0187] At 1126, the evaluation agent 1110 feeds back the overall error information diff calculated in 1122, the contribution allocation contrb_ass1 of agent 1 1102, and the result information GN_feedback to agent 1 1102.

[0188] At 1128, the evaluation agent 1110 feeds back the overall error information diff calculated in 1122, the contribution allocation contrb_ass2 of agent 2 1104, and the result information GN_feedback to agent 2 1104.

[0189] At 1130, Agent 1 at 1102 multiplies diff by contrb_ass1 as its reward feedback and updates its large model AgentGPT.

[0190] At 1140, Agent 2 at 1104 multiplies diff by contrb_ass2 as its reward feedback and updates its own large model AgentGPT.

[0191] In the embodiment shown in Figure 11, the calculation method employed by 1122 ensures that if the GN is generated correctly, the evaluating agent 1110 and all agents participating in the GN generation process will receive positive reward feedback. Furthermore, agents with higher evaluation results receive relatively more positive reward feedback, while agents with lower evaluation results receive relatively less. Conversely, if the GN fails to be generated correctly, the evaluating agent 1110 and all agents participating in the GN generation process will face negative reward feedback (i.e., penalty). In this case, agents with higher evaluation results will receive relatively lighter penalties, while agents with lower evaluation results will receive relatively heavier penalties. In this way, different agents can receive differentiated reward feedback based on their actual performance in the task. This helps agents with different roles to recognize their contribution to the task and thus optimize their behavior accordingly.

[0192] The QMIX algorithm feeds back the same signal to each agent, failing to enable each agent to recognize its own contribution level and perform differentiated updates. The embodiment in Figure 11 generates differentiated update signals for agents with different contribution levels based on the GN generation, allowing each agent to optimize its behavior in a targeted manner, thereby improving the overall performance of the A-Core system.

[0193] Figure 12 illustrates a schematic diagram of a preliminary evaluation process 1200 according to some embodiments of the present disclosure. As shown in Figure 12, the agent 1200 (e.g., a network planning agent) uses its internal evaluator module 1203 to perform a self-evaluation (i.e., a preliminary evaluation) by interacting with other modules at the natural language level and by comprehensively utilizing relevant historical information stored in the memory 1210, on the probability that its decision output can correctly generate GN.

[0194] After the agent makes a decision, the evaluator module uses its own large model, AgentGPT, to perform a preliminary evaluation of the agent's decision based on relevant historical information in memory (e.g., historical reasoning trajectory information). For example, a specific evaluation metric could be the probability of correctly generating the final GN, so the output self-evaluation result is between 0 and 1.

[0195] At 1220, agent 1202 receives input information, which can be a request input by a user or a decision output from another agent. For example, a network planning agent can receive user-input request information.

[0196] At 1222, agent 1202 sends an input message to memory 1210. If retrieval from the public memory is required, a message is sent to the public memory via the ABI. If only retrieval from the internal private memory is required, a message is sent directly to the private memory to invoke it.

[0197] In 1224, the memory 1210 retrieves relevant historical information stored according to the message (e.g., historical reasoning trajectory information), and can search for relevant expert knowledge and useful document information, etc.

[0198] At 1226, after retrieving relevant historical information from memory 1210, semi-structured natural language information is returned to agent 1202, which includes relevant historical information (e.g., historical reasoning trajectory information, each historical reasoning trajectory information includes historical requests similar to the input information, as well as historical arrangement information, historical evaluation results, and historical reward feedback information corresponding to the historical requests), and URLs of useful document information or web page databases, etc.

[0199] In step 1228, agent 1202 combines retrieved relevant historical information with the input information to form enhanced prompts, and uses its large model AgentGPT to perform corresponding inference tasks to generate its own decision output. For example, a network planning agent can output an understanding of the user's intent and a list of sub-tasks.

[0200] At 1230, agent 1202 sends the input information and its own decision output to the internal evaluator module 1203.

[0201] In step 1232, the evaluator module 1203 sends the input information and the decision output of the agent 1202 to the memory 1210 to request a retrieval. If a retrieval from the public memory is required, a message is sent to the public memory via the ABI. If only the internal private memory is needed for retrieval, a message is sent directly to the private memory to invoke it.

[0202] In 1234, the memory 1210 retrieves relevant historical information stored according to the message (e.g., historical reasoning trajectory information), and can search for relevant expert knowledge and useful document information, etc.

[0203] In step 1236, after retrieving relevant historical information from memory 1210, semi-structured natural language information is returned to evaluator module 1203, which includes relevant historical information (e.g., historical reasoning trajectory information), as well as relevant expert knowledge and useful document information.

[0204] At 1240, the evaluator module 1203 uses a large model GPT (which can be its own large model AgentGPT or an external large model GPT) to learn the network environment and feedback information based on the semi-structured natural language information returned by the memory 1210, and performs a self-evaluation on the probability that the decision made by the agent 1202 based on the current input information can ultimately correctly generate GN.

[0205] At 1242, the evaluator module 1203 returns a self-evaluation result to the agent 1202. For example, the evaluator module returns a self-evaluation result of 0.95 based on the decision output generated by the network planning agent in response to the user's input requirements. This indicates that the evaluator module determines that the decision output results in a 95% probability of correctly generating the GN.

[0206] The embodiment shown in Figure 12 designs a mechanism in A-Core where NetAgent utilizes memory to output value assessment, realizing value assessment based on semi-structured natural language information. This mechanism leverages relevant historical information stored in memory (e.g., historical reasoning trajectory information) and an external knowledge base, fully utilizing the autonomous learning capabilities of the large model to enable it to autonomously learn assessment patterns and value assessment rules. Unlike conventional methods that rely on specific, pre-defined structured data such as state and action spaces, this mechanism eliminates the dependence on structured data, improving the adaptability of the communication system to complex language tasks. Furthermore, when encountering complex tasks, the agent can output multiple times for the input, selecting the decision output with the highest evaluation result as the final decision.

[0207] Conventional reinforcement learning methods, such as the QMIX algorithm, typically assume that the agent's action and state spaces are well-defined and definite, making them ill-suited to the complexity and diversity of language tasks. In the next-generation agent-based core network A-Core architecture, conventional reinforcement learning algorithms are less adaptable because the output of language tasks is mostly semi-structured natural language. The embodiment shown in Figure 12 implements value assessment based on semi-structured natural language information, independent of specific, pre-defined structured data such as state and action spaces. By utilizing relevant historical information stored in memory (e.g., historical reasoning trajectory information) and an external knowledge base, large models can autonomously learn assessment patterns and value assessment rules, enabling them to better adapt to complex communication networks.

[0208] In this way, the agent can perform self-evaluation when making decisions, thus introducing the dimension of decision quality. When faced with complex tasks, the agent can attempt multiple outputs based on the results of its self-evaluation and select the highest-quality output. This process effectively enhances the robustness of the decision and improves its quality.

[0209] Figure 13 illustrates a schematic flowchart implemented at an agent for evaluation according to some embodiments of the present disclosure. As shown in Figure 13, process 1300 can be executed by an agent for evaluation, such as a chip, module, or unit in evaluation agent 230. At block 1310, the agent for evaluation receives a user request for performing a task. At block 1320, the agent for evaluation receives multiple decisions from multiple agents, and individual evaluation information corresponding to each of the multiple decisions. At block 1330, the agent for evaluation determines overall evaluation information for instructing the multiple agents to generate a generative network GN based on the user request, the multiple decisions, and the individual evaluation information, wherein the GN is generated based on the multiple decisions. In some embodiments, process 1300 may also include other operations performed at agent 410 for evaluation as described in conjunction with Figures 4 to 7 in embodiments of the present disclosure.

[0210] Figure 14 shows a schematic flowchart implemented at an agent according to some embodiments of the present disclosure. As shown in Figure 14, process 1400 can be executed by an agent, such as a chip, module, or unit in network planning agent 212, network assembly agent 214, network connection agent 216, or network operation agent 218. In block 1410, the agent receives generation error information of a generative network GN and a first contribution of a first agent among a plurality of agents to the generation of GN, wherein GN is generated based on multiple decisions of the plurality of agents. In block 1420, the agent updates the generative pre-trained model of the first agent based on the generation error information and the first contribution. In some embodiments, process 1400 may also include other operations performed at the first agent 402 as described in conjunction with Figures 4 to 8 in embodiments of the present disclosure.

[0211] Figure 15 shows a schematic flowchart implemented at an agent according to some embodiments of the present disclosure. As shown in Figure 15, process 1500 can be executed by an agent, such as a chip, module, or unit in network planning agent 212, network assembly agent 214, network connection agent 216, or network operation agent 218. In block 1510, the agent receives input information. In block 1520, the agent determines a first decision of a first agent among a plurality of agents based on the input information. In block 1530, the agent determines individual evaluation information of the first agent for generating a generative network GN. In block 1540, the agent sends the first decision to a second agent among a plurality of agents. In some embodiments, process 1500 may also include other operations performed at the first agent 402 as described in conjunction with Figures 4 to 8 in embodiments of the present disclosure.

[0212] Figure 16 illustrates a schematic flowchart of an evaluation module implemented in an agent according to some embodiments of the present disclosure. As shown in Figure 16, process 1600 can be executed by an evaluation module in an agent, such as a chip, module, or assembly in evaluation module 320. In block 1610, the evaluation module receives input information and a first decision of a first agent. In block 1620, the evaluation module sends a third request to a fifth network functional entity, wherein the third request includes the input information and the first decision, and the fifth network functional entity is used to store data. In block 1630, the evaluation module receives third information from the fifth network functional entity in response to the third request, wherein the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. In block 1640, the evaluation module determines individual evaluation information based on the third information. In some embodiments, process 1600 may also include other operations performed at evaluation module 403 in conjunction with Figure 8 in embodiments of the present disclosure.

[0213] Figure 17 illustrates a schematic flowchart of implementation at a common memory according to some embodiments of the present disclosure. As shown in Figure 17, process 1700 can be executed by a common memory, such as a chip, module, or assembly in common memory 224. In block 1710, the common memory receives a first request, wherein the first request includes a user request for the generative network GN to perform a task, and the GN is generated based on multiple decisions of multiple agents. In block 1720, the common memory determines first information based on the user request, wherein the first information includes at least one of the following: a historical user request associated with the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user request. In block 1730, the common memory sends the first information. In some embodiments, process 1700 may also include other operations performed at common memory 412 as described in conjunction with Figures 4 to 6 and Figure 8 in embodiments of the present disclosure.

[0214] Figure 18 is a block diagram of a device 1800 that can be used to implement some embodiments of the present application. In some embodiments, device 1800 may be an element of a communication network infrastructure, such as a base station (e.g., NodeB, evolved NodeB (eNodeB or eNB), next-generation NodeB (sometimes called next-generation NodeB, gNodeB or gNB), home subscriber server (HSS), gateway (GW), such as packet gateway (PGW) or serving gateway (SGW), or various other nodes or functions within a core network (CN) or Public Land Mobility Network (PLMN). In other embodiments, device 1800 may be a device connected to the network infrastructure via a wireless interface, such as a mobile phone, smartphone, or other such device that can be classified as user equipment (UE). In some embodiments, device 1800 may be machine-type communication (M2C). Communications (MTC) devices (also known as machine-to-machine (M2M) devices), or another type of device that, although not providing direct service to a user, can be classified as a UE. In some embodiments, device 1800 may be a roadside unit (RSU), a vehicle UE (V-UE), a pedestrian UE (P-UE), or an infrastructure UE (I-UE). In some scenarios, device 1800 may also be referred to as a mobile device, a term intended to reflect a device connected to a mobile network, regardless of whether the device itself is designed for or capable of being mobile. A particular device may utilize all or only a subset of the components shown, and the level of integration may vary depending on the device. Furthermore, device 1800 may contain multiple instances of components, such as multiple processors, memory, transmitters, receivers, etc.

[0215] In cases where device 1800 is a node or function within a core network or public terrestrial mobile network, device 1800 typically includes a processor 1802, such as a central processing unit (CPU), and may further include a dedicated processor, such as a graphics processing unit (GPU) or other such processor, memory 1804, a network interface 1806, and a bus 1808 for connecting the components of device 1800.

[0216] Memory 1804 may include any type of non-transitory system memory readable by processor 1802, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), read-only memory (ROM), or combinations thereof. In one embodiment, memory 1804 may include more than one type of memory, such as ROM used at boot time and DRAM used for program and data storage during program execution. Bus 1808 may be one or more of a plurality of bus architectures of any type, including a memory bus or memory controller, a peripheral bus, or a video bus. In some examples, memory 1804 and processor 1802 may be separate devices. In other examples, memory 1804 may be integrated with processor 1802 as a single device.

[0217] Device 1800 may also include one or more network interfaces 1806, which may include at least one of wired network interfaces and wireless network interfaces. As shown in FIG18, network interface 1806 may include a wired network interface for connecting to network 1822, and may also include a wireless access network interface 1820 for connecting to other devices via a wireless link. When device 1800 is a network infrastructure element, the wireless access network interface 1820 may be omitted for nodes or functions that are elements of a PLMN rather than elements at the wireless edge. When device 1800 is infrastructure at the wireless edge of a network, it may include both wired and wireless network interfaces. Network interface 1806 allows device 1800 to communicate with remote entities such as those connected to network 1822.

[0218] Mass storage 1810 may include any type of non-transitory storage device configured to store data, programs, and other information and make the data, programs, and other information accessible via bus 1808. Mass storage 1810 may include, for example, one or more of a solid-state drive, hard disk drive, disk drive, or optical disk drive. In some embodiments, mass storage 1810 may be located remotely from device 1800 and may be accessed using a network interface such as interface 1806. In the illustrated embodiment, mass storage 1810 is distinct from the memory 1804 that includes it, and mass storage 1810 typically performs storage tasks compatible with higher latency but typically provides low or no fluctuation. In some embodiments, mass storage 1810 may be integrated with heterogeneous memory 1804.

[0219] Figure 19 is a schematic diagram of the structure of a device 1900 according to some embodiments of this application. As shown in Figure 19, the device 1900 includes an acquisition unit 1902, a receiving unit 1904, and a determination unit 1906. The device 1900 can be applied to the communication scenario shown in Figure 1 and can implement the methods provided in the preceding embodiments, such as method 1300. Optionally, the physical manifestation of the device 1900 can be a communication device, such as an intelligent agent for evaluation. Alternatively, the device 1900 can be other devices capable of implementing the functions of a communication device, such as a processor or chip inside a communication device. Specifically, the device 1900 can be a programmable chip, such as a field-programmable gate array (FPGA), a complex programmable logic device (CPLD), an application-specific integrated circuit (ASIC), or a system on a chip (SOC).

[0220] In some embodiments, the acquisition unit 1902 may be configured to acquire a user request for performing a task. The receiving unit 1904 may be configured to receive multiple decisions from multiple agents, and individual evaluation information corresponding to each of the multiple decisions. The determining unit 1906 may be configured to determine overall evaluation information for instructing the multiple agents to generate a generative network GN based on the user request, the multiple decisions, and the individual evaluation information, wherein the GN is generated based on the multiple decisions.

[0221] In some other embodiments, apparatus 1900 may include various other units or modules that can be configured to perform the various operations or functions described with respect to the foregoing method embodiments. Specific details can be obtained by referring to the detailed description of the foregoing method embodiments, and will not be repeated here.

[0222] Figure 20 is a schematic diagram of the structure of a device 2000 according to some other embodiments of this application. As shown in Figure 20, the device 2000 includes a receiving unit 2002 and an updating unit 2004. The device 2000 can be applied to the communication scenario shown in Figure 1 and can implement the methods provided in the preceding embodiments, such as method 1400. Optionally, the physical manifestation of the device 2000 can be a communication device, such as an intelligent agent. Alternatively, the device 2000 can be other devices capable of implementing the functions of a communication device, such as a processor or chip inside a communication device. Specifically, the device 2000 can be a programmable chip, such as a field-programmable gate array (FPGA), a complex programmable logic device (CPLD), an application-specific integrated circuit (ASIC), or a system on a chip (SOC).

[0223] In some embodiments, the receiving unit 2002 may be configured to receive generation error information of the generative network GN and a first contribution of a first agent among multiple agents to the generation of GN, wherein GN is generated based on multiple decisions of multiple agents. The updating unit 2004 may be configured to update the generative pre-trained model of the first agent based on the generation error information and the first contribution.

[0224] In some embodiments, the apparatus 2000 may include various other units or modules that can be configured to perform the various operations or functions described with respect to the foregoing method embodiments. Specific details can be obtained by referring to the detailed description of the foregoing method embodiments, and will not be repeated here.

[0225] Figure 21 is a schematic diagram of the structure of a device 2100 according to some other embodiments of this application. As shown in Figure 21, the device 2100 includes a receiving unit 2102, a first determining unit 2104, a second determining unit 2106, and a transmitting unit 2108. The device 2100 can be applied to the communication scenario shown in Figure 1 and can implement the methods provided in the preceding embodiments, such as method 1500. Optionally, the physical manifestation of the device 2100 can be a communication device, such as an intelligent agent. Alternatively, the device 2100 can be other devices capable of implementing the functions of a communication device, such as a processor or chip inside a communication device. Specifically, the device 2100 can be a programmable chip, such as a field-programmable gate array (FPGA), a complex programmable logic device (CPLD), an application-specific integrated circuit (ASIC), or a system on a chip (SOC).

[0226] In some embodiments, the receiving unit 2102 may be configured to receive input information. The first determining unit 2104 may be configured to determine a first decision of a first agent among a plurality of agents based on the input information. The second determining unit 2106 may be configured to determine individual evaluation information of the first agent among a plurality of agents for generating the GN. The sending unit 2108 may be configured to send the first decision to a second agent among a plurality of agents.

[0227] In some embodiments, the apparatus 2100 may include various other units or modules that can be configured to perform the various operations or functions described with respect to the foregoing method embodiments. Specific details can be obtained by referring to the detailed description of the foregoing method embodiments, and will not be repeated here.

[0228] Figure 22 is a schematic diagram of the structure of a device 2200 according to some other embodiments of this application. As shown in Figure 22, the device 2200 includes a first receiving unit 2202, a transmitting unit 2204, a second receiving unit 2206, and a determining unit 2208. The device 2200 can be applied to the communication scenario shown in Figure 1 and can implement the methods provided in the preceding embodiments, such as method 1600. Optionally, the physical manifestation of the device 2200 can be a communication device, such as an evaluation module in an intelligent agent. Alternatively, the device 2200 can be other devices capable of implementing the functions of a communication device, such as a processor or chip inside a communication device. Specifically, the device 2200 can be a programmable chip, such as a field-programmable gate array (FPGA), a complex programmable logic device (CPLD), an application-specific integrated circuit (ASIC), or a system on a chip (SOC).

[0229] In some embodiments, the first receiving unit 2202 may be configured to receive input information and a first decision of a first agent. The sending unit 2204 may be configured to send a third request to a fifth network functional entity, wherein the third request includes the input information and the first decision, and the fifth network functional entity is used to store data. The second receiving unit 2206 may be configured to receive third information from the fifth network functional entity in response to the third request, wherein the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information. The determining unit 2208 may be configured to determine individual evaluation information based on the third information.

[0230] In some embodiments, the apparatus 2200 may include various other units or modules that may be configured to perform the various operations or functions described with respect to the foregoing method embodiments. Specific details can be obtained by referring to the detailed description of the foregoing method embodiments, and will not be repeated here.

[0231] Figure 23 is a schematic diagram of the structure of device 2300 according to some other embodiments of this application. As shown in Figure 23, device 2300 includes a receiving unit 2302, a determining unit 2304, and a transmitting unit 2306. Device 2300 can be applied to the communication scenario shown in Figure 1 and can implement the methods provided in the foregoing embodiments, such as method 1700. Optionally, the physical manifestation of device 2300 can be a communication device, such as a shared memory. Alternatively, device 2300 can be other devices capable of implementing the functions of a communication device, such as a processor or chip inside a communication device. Specifically, device 1700 can be a programmable chip, such as a field-programmable gate array (FPGA), a complex programmable logic device (CPLD), an application-specific integrated circuit (ASIC), or a system on a chip (SOC).

[0232] In some embodiments, the receiving unit 2302 may be configured to receive a first request, wherein the first request includes a user request for the generative network (GN) to perform a task, and the GN is generated based on multiple decisions made by multiple agents. The determining unit 2304 may be configured to determine first information based on the user request, wherein the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests. The sending unit 2306 may be configured to send the first information.

[0233] In some embodiments, the apparatus 2300 may include various other units or modules that may be configured to perform the various operations or functions described with respect to the foregoing method embodiments. Specific details can be obtained by referring to the detailed description of the foregoing method embodiments, and will not be repeated here.

[0234] It should be noted that the module division in the above embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, exist as separate physical units, or be integrated into one unit by two or more units. The integrated units described above can be implemented in hardware or as software functional units.

[0235] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or all or part of it, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0236] Based on the above embodiments, this application also provides a computer program that, when run on a computer, causes the computer to execute any of the methods provided in the above embodiments.

[0237] Based on the above embodiments, this application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a computer, it causes the computer to perform any of the methods provided in the above embodiments. The storage medium can be any available medium that can be accessed by a computer. By way of example, but not limited to, a computer-readable medium may include RAM, ROM, electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code having the form of instructions or data structures and that can be accessed by a computer.

[0238] Based on the above embodiments, this application also provides a chip for reading a computer program stored in a memory and implementing any of the methods provided in the above embodiments.

[0239] Based on the above embodiments, this application provides a chip system including a processor for supporting a computer device in implementing the functions involved in the communication devices in the above embodiments. In one possible design, the chip system further includes a memory for storing necessary programs and data of the computer device. This chip system may be composed of chips or may include chips and other discrete components.

[0240] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0241] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0242] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0243] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

Claims

1. A communication method, comprising: Obtain the user request used to perform the task; Receive multiple decisions from multiple agents, and individual evaluation information corresponding to each of the multiple decisions; as well as Based on the user request, the multiple decisions, and the individual evaluation information, overall evaluation information is determined to instruct the multiple agents to generate a generative network GN, wherein the GN is generated based on the multiple decisions.

2. The method according to claim 1, further comprising: First information is obtained based on the user request, wherein the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests.

3. The method according to claim 2, wherein obtaining the first information includes: Send a first request, wherein the first request includes the user request; as well as Receive the first information in response to the first request.

4. The method according to claim 2, wherein determining the overall evaluation information includes: The overall evaluation information is determined based on the user request, the multiple decisions, the individual evaluation information, and the first information.

5. The method according to claim 1, further comprising: Receive information indicating the generation result of the GN.

6. The method according to claim 5, further comprising: Based on the information from the generated results and the overall evaluation information, the generation error information of the GN is determined.

7. The method according to claim 6, further comprising: The generation error information and the multiple contributions of the multiple agents to the generation of the GN are sent.

8. The method of claim 7, wherein the first agent among the plurality of agents' first contribution to generating the GN is determined based on the following: The information of the generated result, and the individual evaluation information corresponding to the first decision of the first agent.

9. The method according to claim 6, further comprising: The generative pre-trained model is updated based on the generated error information.

10. The method according to any one of claims 1 to 9, wherein the overall evaluation information includes the probability that the GN is correctly generated.

11. The method according to any one of claims 1 to 9, wherein the individual evaluation information is used to determine the probability that the agent's decision results in the correct generation of the GN.

12. The method according to any one of claims 1 to 11, wherein the user request is based on natural language representation.

13. A communication method, comprising: Receive the generation error information of the generative network GN, and the first contribution of the first agent among multiple agents to the generation of the GN, wherein the GN is generated based on multiple decisions of the multiple agents; as well as Based on the generation error information and the first contribution, the generative pre-trained model of the first agent is updated.

14. The method of claim 13, wherein updating the generative pre-trained model comprises: Based on the generated error information and the first contribution amount, the reward information is determined; as well as The generative pre-trained model is updated based on the reward information.

15. A communication method, comprising: Receive input information; Based on the input information, determine the first decision of the first agent among multiple agents; Determine the individual evaluation information of the first agent for the generative network GN; as well as The first decision is sent to the second agent among the plurality of agents.

16. The method of claim 15, further comprising: The first decision and the individual evaluation information are sent to the agent used for evaluation.

17. The method of claim 15, wherein determining the first decision comprises: Send a second request, wherein the second request includes the input information; Receive second information in response to the second request, wherein the second information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information; as well as The first decision is determined using the generative pre-trained model of the first agent, based on the second information and the input information.

18. The method of claim 15, wherein determining the individual assessment information comprises: Send the input information and the first decision to the evaluation module in the first intelligent agent; as well as The individual evaluation information is received from the evaluation module, wherein the individual evaluation information is determined by the evaluation module using a generative pre-trained model based on third information, and the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information.

19. The method according to any one of claims 15 to 18, wherein: The input information includes a user request for performing a task, and the method further includes sending the input information to an agent for evaluation; or The input information includes the third decision of the third agent among the plurality of agents.

20. The method of claim 19, wherein the user request is based on natural language representation.

21. A communication method, comprising: Receive input information and the first decision of the first intelligent agent; A third request is sent to a fifth network function entity, wherein the third request includes the input information and the first decision, and the fifth network function entity is used to store data; The fifth network functional entity receives third information in response to the third request, wherein the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information; and The individual assessment information is determined based on the third information.

22. The method of claim 21, further comprising: Send the individual assessment information.

23. The method according to claim 21, wherein: The input information includes a user request for performing the task; or The input information includes the third decision of the third agent among the plurality of agents.

24. The method of claim 23, wherein the user request is based on natural language representation.

25. A communication method, comprising: Receive a first request, wherein the first request includes a user request for the generative network GN to perform a task, and the GN is generated based on multiple decisions of multiple agents; Based on the user request, first information is determined, wherein the first information includes at least one of the following: historical user requests related to the user request, and multiple historical decisions and historical individual evaluation information corresponding to the historical user requests; and Send the first message.

26. The method of claim 25, further comprising: Receive overall evaluation information for instructing the plurality of agents to generate the GN.

27. The method of claim 25, further comprising: Receive a second request, wherein the second request includes input information; Based on the input information, second information is determined, wherein the second information includes at least one of the following: historical input information related to the input information, and multiple historical decision and historical individual evaluation information corresponding to the historical input information; and Send the second message.

28. The method of claim 25, further comprising: Receive a third request, wherein the third request includes the input information and a first decision of the first agent among the plurality of agents; Based on the input information and the first decision, third information is determined, wherein the third information includes at least one of the following: historical input information related to the input information, and multiple historical decisions and historical individual evaluation information corresponding to the historical input information; and Send the third message.

29. The method according to claim 27 or 28, wherein: The input information includes user requests; or The input information includes the third decision of the third agent among the plurality of agents.

30. The method of claim 29, wherein the user request is based on natural language representation.

31. A communication device, comprising: Units or modules for performing the method according to any one of claims 1 to 12, or for performing the method according to any one of claims 13 to 14, or for performing the method according to any one of claims 15 to 20, or for performing the method according to any one of claims 21 to 24, or for performing the method according to any one of claims 25 to 30.

32. A communication device, comprising: The processor is configured to perform the method according to any one of claims 1 to 12, or to perform the method according to any one of claims 13 to 14, or to perform the method according to any one of claims 15 to 20, or to perform the method according to any one of claims 21 to 24, or to perform the method according to any one of claims 25 to 30.

33. A computer-readable storage medium storing instructions that, when executed by a device, cause the device to perform the method according to any one of claims 1 to 12, or be configured to perform the method according to any one of claims 13 to 14, or be configured to perform the method according to any one of claims 15 to 20, or be configured to perform the method according to any one of claims 21 to 24, or be configured to perform the method according to any one of claims 25 to 30.

34. A computer program product comprising instructions, which, when executed by a device, cause the device to perform the method according to any one of claims 1 to 12, or be configured to perform the method according to any one of claims 13 to 14, or be configured to perform the method according to any one of claims 15 to 20, or be configured to perform the method according to any one of claims 21 to 24, or be configured to perform the method according to any one of claims 25 to 30.