A method and system for integrated sensing and computing for agent networks

Abstract

This invention provides a method and system for integrated sensing, computing, and intelligence processing for agent networks. The method includes: responding to user-input natural language text by invoking a Large Language Model (LLM); converting the user intent contained in the natural language text into multiple sub-tasks; performing multi-dimensional matching and evaluation of the requirements of each sub-task with the capabilities of each agent in the agent network to generate a task allocation scheme; extracting global dependency information based on the sub-tasks and the dependencies between them and distributing it to each agent in the agent network; obtaining a resource scheduling decision that converges to a preset shared value function through multiple iterations; allocating the sub-tasks to the corresponding agents according to the task allocation scheme; and scheduling resources for integrated sensing, computing, and intelligence processing according to the resource scheduling decision.

Description

Technical Field

[0001] This invention relates to the field of integrated intelligent agent network technology, and in particular to an integrated method and system for sensing, computing and intelligence in intelligent agent networks. Background Technology

[0002] In recent years, thanks to breakthroughs in large-scale pre-trained models, intelligent agents based on these models have demonstrated strong application potential in multiple fields. These intelligent agents possess multiple capabilities, including autonomous perception, learning, decision-making, and execution, enabling them to interact and adjust in complex environments. From personalized services and industrial automation to smart city management, intelligent agents are gradually transforming from independent task executors into key roles capable of participating in systematic and social collaboration. They autonomously form collaborative organizations in dynamically changing open environments, completing complex tasks that a single intelligent agent cannot handle independently. Multiple intelligent agents interact through task allocation, knowledge sharing, and joint decision-making to achieve a higher level of autonomy and intelligence.

[0003] Figure 2 This is a diagram of the existing intelligent agent network architecture. The Internet of Agents (IoA), as an emerging infrastructure paradigm, uses intelligent agents as the basic interaction unit. By building a cross-domain collaborative network foundation, it achieves unified access and intelligent collaborative support for large-scale heterogeneous intelligent agents. In terms of interaction methods, IoA drives the evolution of interaction interfaces from graphical user interfaces to semantically driven, goal-oriented autonomous negotiation between intelligent agents. It can directly understand and exchange task information based on natural language, thereby achieving real-time response and collaborative reasoning for dynamic tasks. For example... Figure 1 As shown, the architecture of IoA aims to provide real-time communication, discovery, and coordination among heterogeneous agents, supporting the orchestration of complex tasks in dynamic environments. In recent years, large-scale agent networks have attracted significant attention in academia, with extensive research conducted on infrastructure, protocol standards, and agent collaboration. At the infrastructure level, researchers are dedicated to building novel network systems capable of supporting the interconnection and collaboration of massive numbers of agents, exploring key technologies such as agent autonomous addressing, capability identification, and semantic communication. Regarding protocol standards, researchers aim to establish a universal language compatible with different types of agents, focusing on semantic interoperability frameworks for communication protocols, standardization of interactions with external data sources and tools, and decentralized communication protocols and interaction rules. In the crucial area of ​​agent collaboration, efforts are focused on solving collaborative decision-making problems among agents in complex scenarios such as goal conflict, incomplete information, and resource constraints. These efforts include both centralized coordinator-based task allocation and global optimization methods, as well as distributed negotiation and swarm intelligence emergence strategies emphasizing decentralized autonomy. These studies collectively drive the development of IoA from proof-of-concept to a more systematic and precise approach.

[0004] Although significant progress has been made in the existing research on agent networks, they still face severe challenges in terms of resource scheduling and collaborative efficiency in practical deployment. The new characteristics demonstrated by agent collaboration pose new requirements for the network architecture: on the one hand, each agent is both an executor of tasks and an information node and decision-making entity in the network. It is necessary to form an efficient self-organizing system through continuous environmental perception, information exchange, and behavior negotiation; on the other hand, the task coupling between agents is extremely high, and resource bottlenecks or delays in any link may lead to collaborative failure, posing unprecedented requirements for the real-time, reliability, and elasticity of underlying communication, computing, sensing, and intelligent resource allocation. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a communication, sensing, computing, and intelligence integration method and system for agent networks to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides a communication, sensing, computing, and intelligence integration method for agent networks. The method includes the following steps: calling a large language model (LLM) in response to a natural language text input by a user, converting the user intent contained in the natural language text into multiple subtasks, performing multi-dimensional matching evaluation on the requirements of each subtask and the capabilities of each agent in the agent network, and generating a task allocation plan; extracting global dependency information based on the subtasks and the dependency relationships between the subtasks and distributing it to each agent in the agent network, and obtaining a resource scheduling decision that converges to a preset shared value function through multiple rounds of iteration; allocating the subtasks to the corresponding agents according to the task allocation plan, and each agent schedules the resources for communication, sensing, computing, and intelligence integration processing according to the resource scheduling decision.

[0007] In some embodiments of the present invention, for the resource scheduling decision, the method further includes: optimizing the resource scheduling decision using a genetic algorithm; wherein the genetic algorithm uses all possible resource scheduling decisions as the search space, uses the resource scheduling decision as the chromosome, evaluates the chromosome using a pre-constructed multi-dimensional decision-making efficiency evaluation system, and generates a new chromosome based on the evaluation result feedback.

[0008] In some embodiments of the present invention, the dimensions of the multi-dimensional decision-making efficiency evaluation system include some or all of the dimensions of task completion rate, response delay, resource utilization rate, energy consumption efficiency, and fault tolerance performance.

[0009] In some embodiments of the present invention, prior to the step of invoking a large language model (LLM) in response to user-input natural language text, the method further includes: performing syntactic analysis and key information extraction on the user-input natural language text, identifying semantic elements including core verbs, objects of operation, and constraints, eliminating ambiguity, and adjusting it into a structured representation.

[0010] In some embodiments of the present invention, the subtasks and the dependencies between subtasks are recorded in the form of a directed graph, wherein the directed graph has subtasks as nodes and the dependencies between subtasks as edges; the step of extracting global dependency information based on the dependencies between subtasks includes: using a graph neural network (GNN) to aggregate the information of each node's neighboring nodes in the directed graph to obtain global dependency information.

[0011] In some embodiments of the present invention, the step of obtaining a resource scheduling decision that converges to a preset shared value function through multiple iterations employs a multi-agent reinforcement learning strategy. Each round includes: each agent generating a resource scheduling decision based on its local observation information and the global dependency information; performing joint actions based on the resource scheduling decisions of each agent; evaluating the value of the joint actions of all agents using a globally shared value function; updating the network parameters of the centralized collaborative strategy network using the value evaluation as a feedback reward signal; updating the parameters of the individual strategy networks of each agent based on the updated centralized collaborative strategy network; the updated centralized collaborative strategy network and individual strategy networks are used for the next round of iteration; and ending the iteration when the centralized collaborative strategy network reaches convergence. Each agent maintains an individual strategy network for selecting the optimal resource scheduling strategy for its assigned subtasks, and the agent networks jointly maintain a centralized collaborative strategy network for conflict resolution and collaborative optimization.

[0012] In some embodiments of the present invention, the preset shared value function includes a global resource utilization reward, a task completion efficiency reward, a collaboration cost penalty, and a constraint violation penalty, and the global sharing of the preset shared value function is implemented using a global Critic network.

[0013] In some embodiments of the present invention, for the resources used for integrated processing of communication, sensing, computing and intelligence, the method further includes: continuously collecting real-time operating status data of intelligent agent resources, including various communication facilities, sensing devices, computing units and data resources, as the physical basis for supporting the collaborative operation of intelligent agents; pre-constructing and maintaining a global resource map that supports real-time querying, and using the global resource map to maintain real-time operating status data of various intelligent agent resources.

[0014] Corresponding to the above method, the present invention also provides an integrated sensory-computing-intelligence system for agent networks. The system includes: a session management layer, used to invoke a Large Language Model (LLM) in response to user-input natural language text, converting the user intent contained in the natural language text into multiple sub-tasks, performing multi-dimensional matching and evaluation of the requirements of each sub-task with the capabilities of each agent in the agent network, and generating a task allocation scheme; a resource scheduling layer, used to extract global dependency information based on the sub-tasks and the dependencies between sub-tasks and distribute it to each agent in the agent network, obtaining a resource scheduling decision that converges to a preset shared value function through multiple iterations; and an agent collaboration layer, used to allocate sub-tasks to corresponding agents according to the task allocation scheme, and each agent scheduling resources for integrated sensory-computing-intelligence processing according to the resource scheduling decision.

[0015] In some embodiments of the present invention, the system further includes: a resource awareness layer, used to pre-build and maintain a global resource map that supports real-time querying, and use the global resource map to maintain real-time operating status data of various intelligent agent resources; and a basic resource layer, used to continuously collect real-time operating status data of intelligent agent resources, including various communication facilities, sensing devices, computing units and data resources, as a physical basis to support the collaborative operation of intelligent agents.

[0016] This invention proposes an integrated sensory-computing-intelligence method for agent networks. It utilizes LLM (Limited Learning Model) to transform and break down user input into multiple sub-tasks, assigning these sub-tasks to each agent. Through multi-agent hierarchical reinforcement learning, it selects resource scheduling decisions, constructing a two-layer intelligent decision-making architecture that combines individual autonomy with global collaboration to achieve optimal performance under set constraints. This enables the construction of an end-to-end solution model from natural language instructions to automated resource scheduling. LLM accurately parses and decomposes fuzzy and heterogeneous user needs into structured, executable task trees. Combined with a hierarchical reinforcement learning framework, it dynamically allocates tasks and collaboration strategies among multiple agents, achieving dual-drive from autonomous decision-making by individual agents to global optimization by a central coordinator. This facilitates efficient response to higher-order user intents, automatically balancing local efficiency with overall system goals, thereby significantly improving the adaptability, interpretability, and overall performance of the resource scheduling system under strong constraints.

[0017] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0018] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a flowchart of an integrated sensing, computing, and intelligence method for intelligent agent networks according to an embodiment of the present invention.

[0020] Figure 2 This is a diagram of the existing intelligent agent network architecture.

[0021] Figure 3 This is a diagram of an integrated sensing, computing, and intelligence system architecture for intelligent agent networks according to an embodiment of the present invention.

[0022] Figure 4 This is a flowchart of the integrated sensing, computing, and intelligence system for intelligent agent networks according to one embodiment of the present invention.

[0023] Figure 5 This is a flowchart of an agent task orchestration method based on LLM in one embodiment of the present invention.

[0024] Figure 6 This is a schematic diagram of a multi-agent hierarchical reinforcement learning scheduling algorithm based on task dependency constraints in one embodiment of the present invention.

[0025] Figure 7 This is a flowchart of a strategy optimization and system resilience assurance method based on genetic algorithms in one embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0027] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0028] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0029] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0030] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0031] With the evolution of networks, the integration of sensing, computing, and intelligence is gradually becoming a core trend in future network architecture. Specifically, this refers to networks that simultaneously possess environmental awareness, ubiquitous communication, elastic computing, and distributed intelligence. Through a unified resource management platform and collaborative scheduling mechanism, they achieve deep integration and dynamic allocation of multi-dimensional resources, thereby enabling the network to possess new closed-loop information flow intelligent Q-interaction and processing capabilities, as well as wide-area intelligent collaboration capabilities. Through the integration of sensing, computing, and intelligence, intelligent agent networks can maintain high response speed and resource utilization efficiency even in high-concurrency, dynamically changing collaborative scenarios.

[0032] To overcome the problems of existing technologies, this invention proposes a multi-intelligent collaborative network based on the concept of integrating intelligent agent networks and synesthetic computing resources. This network is task-oriented, covers cross-scenario, cross-modal, and cross-level synesthetic computing integration, and uses a unified intent-driven, task-centric orchestration approach. It deeply integrates perception and connectivity, intelligence and computing power into a common foundation, forming standardized and reusable capability components and collaboration specifications. This weaves dispersed intelligent agent capabilities into replicable and scalable swarm intelligence, enabling stable and efficient collaboration in heterogeneous resource and multi-domain environments, achieving optimization and continuous improvement in efficiency, cost, and risk. Through this design, this invention not only provides an intelligent collaboration framework and standard interaction paradigm for intelligent agent networks but also introduces dynamic orchestration capabilities for cross-domain multi-dimensional resources, providing key methodological support for future large-scale intelligent agent collaboration scenarios.

[0033] Figure 1 This is a flowchart of a sensor-computer interface integrated method for intelligent agent networks according to an embodiment of the present invention. The method includes the following steps: Step S110: In response to the natural language text input by the user, call the Large Language Model (LLM) to transform the user intent contained in the natural language text into multiple sub-tasks. The requirements of each sub-task are matched and evaluated in multiple dimensions with the capabilities of each agent in the agent network to generate a task allocation scheme.

[0034] The role of the Large Language Model (LLM) includes text preprocessing of natural language text, extraction of key information / key semantics, and extraction of user intent. The task allocation scheme mentioned above includes at least a formatted natural language description of elements such as a subtask list, the allocation relationship between subtasks and agents, and the execution sequence between subtasks. The multi-dimensional matching evaluation mentioned above can include three dimensions: indicator fit, resource utilization, and task coupling, and can also be adjusted according to specific circumstances.

[0035] The steps described above for modeling the requirements of each subtask with the capabilities of each agent in the agent network and performing multi-dimensional matching and evaluation to generate a task allocation scheme include: after obtaining the multi-dimensional matching and evaluation results, allocating subtasks according to the evaluation results, binding highly coupled subtasks to the same agent to avoid adding new execution subjects; outputting a load-balanced task allocation scheme, and simultaneously specifying the execution sequence and feedback nodes.

[0036] Step S120: Extract global dependency information based on subtasks and the dependencies between subtasks and distribute it to each agent in the agent network. Through multiple rounds of iteration, obtain a resource scheduling decision that converges to the preset shared value function.

[0037] Among these, the value function sharing mechanism enables the coordinated consistency of distributed decision-making through centralized value assessment. The value function measures the quality of a state or state-action pair, and is typically divided into state value functions and action value functions.

[0038] Step S130: Assign subtasks to the corresponding intelligent agents according to the task allocation scheme, and each intelligent agent schedules resources for integrated processing of sensing, computing and intelligence according to the resource scheduling decision.

[0039] The process of all intelligent agents scheduling resources for the integrated processing of sensing, computing, and intelligence according to the resource scheduling decision includes: each agent selecting the optimal resource scheduling strategy for its assigned subtasks, accessing the resource awareness layer and the basic resource layer, and the user-input natural language text (task) entering the physical execution phase. After the scheduling strategy is determined, the resource awareness layer maps the logical resource allocation in the strategy to specific physical or virtual resource entities at the underlying level, completing the dynamic binding of resources. Subsequently, precise configuration instructions and task start commands are generated and sent to the basic resource layer through a standard interface. Various intelligent agents, computing nodes, network devices, and sensors in the basic resource layer receive the instructions, are activated, and are ready to execute specific operations.

[0040] This invention employs a sensory-computing-intelligence integrated approach for agent networks, utilizing LLM to transform and break down user input into multiple sub-tasks, assigning these sub-tasks to each agent. Through multi-agent hierarchical reinforcement learning, resource scheduling decisions are selected, constructing a two-layer intelligent decision-making architecture that combines individual autonomy with global collaboration to achieve optimal performance under set constraints. This approach builds an end-to-end solution model from natural language instructions to automated resource scheduling. LLM accurately parses and decomposes fuzzy and heterogeneous user needs into structured, executable task trees. Combined with a hierarchical reinforcement learning framework, tasks and collaboration strategies are dynamically allocated among multiple agents. This enables dual-drive from autonomous decision-making by individual agents to global optimization by a central coordinator, facilitating efficient responses to higher-order user intents and automatically balancing local efficiency with overall system goals. Consequently, it significantly improves the adaptability, interpretability, and overall performance of the resource scheduling system under strong constraints.

[0041] In some embodiments of the present invention, for the resource scheduling decision, the method further includes: optimizing the resource scheduling decision using a genetic algorithm; wherein the genetic algorithm uses all possible resource scheduling decisions as the search space, uses the resource scheduling decisions as chromosomes, evaluates the chromosomes using a pre-constructed multi-dimensional decision performance evaluation system, and generates new chromosomes based on the evaluation results.

[0042] By employing this embodiment of the invention, the resource scheduling decision is optimized using a genetic algorithm. This approach introduces the powerful global optimization and adaptive search capabilities of the genetic algorithm into the field of resource scheduling. Through iterative evolution of scheduling schemes (chromosomes) through selection, crossover, and mutation, the algorithm efficiently approaches the global optimal solution for objectives such as resource utilization, task completion time, or overall cost from a vast number of possible solutions. This helps overcome the limitations of traditional rule-based or heuristic methods, which are prone to getting stuck in local optima and are difficult to adapt to dynamic and complex scenarios, thereby achieving self-optimization and continuous improvement of the scheduling strategy.

[0043] In some embodiments of the present invention, the dimensions of the multi-dimensional decision-making performance evaluation system include some or all of the dimensions of task completion rate, response latency, resource utilization rate, energy efficiency, and fault tolerance performance.

[0044] In other embodiments of the present invention, various resource scheduling decisions are abstracted into coded structured units. Each unit contains adjustable operating parameters and logical components. These decision structured units constitute a continuously accumulating strategy knowledge base. The units with better current performance are selected from the strategy knowledge base as the basis, and a new generation of strategy variants are generated through evolutionary operations such as structural reorganization and parameter perturbation.

[0045] By employing this embodiment of the invention, resource scheduling decisions are managed and updated using a strategy knowledge base. The effectiveness of resource scheduling decisions is evaluated from multiple dimensions, enabling the construction of a dynamic and intelligent resource scheduling decision-making system. Through the continuous accumulation and updating of the strategy knowledge base, the decision-making basis is transformed from static rules into evolvable experience, thereby improving the system's adaptability to complex and ever-changing business scenarios. This helps to form a closed loop of "decision-execution-evaluation-optimization," ultimately systematically improving the global optimality and long-term robustness of resource utilization.

[0046] In some embodiments of the present invention, prior to the step of invoking the Large Language Model (LLM) in response to user-input natural language text, the method further includes: performing syntactic analysis and key information extraction on the user-input natural language text, identifying semantic elements including core verbs, operation objects, and constraints, eliminating ambiguity, and adjusting it into a structured representation.

[0047] By employing this embodiment of the invention, a pre-processing step is set before calling the Large Language Model (LLM) to standardize the heterogeneous input natural language text. This can transform complex and heterogeneous natural language into standardized prompts with clear structure, explicit intent, and standardized format. This helps to significantly reduce the randomness and ambiguity of the input text and improve the accuracy of the large language model's understanding of core instructions and the consistency of task execution.

[0048] In some embodiments of the present invention, the subtasks and the dependencies between subtasks are recorded in the form of a directed graph, wherein the directed graph has subtasks as nodes and the dependencies between subtasks as edges.

[0049] Accordingly, in some embodiments of the present invention, the step of extracting global dependency information based on subtasks and the dependencies between subtasks includes: using a graph neural network (GNN) to aggregate the information of each node's neighboring nodes in the directed graph to obtain global dependency information.

[0050] This invention employs a directed graph to record subtasks and their dependencies, and utilizes a graph neural network (GNN) to extract global dependency information from the directed graph. This allows complex subtask dependencies to be modeled as structured graph data. By leveraging the message passing mechanism of the GNN, the state information of neighboring nodes is aggregated, thereby accurately and efficiently capturing the multi-layered dependency features of local and global dependencies in the task chain. This significantly improves the resource scheduling system's ability to understand complex workflows and provides a better global scheduling strategy for dependency-sensitive scenarios.

[0051] In some embodiments of the present invention, the step of obtaining a resource scheduling decision that converges to a preset shared value function through multiple rounds of iteration adopts a multi-agent reinforcement learning strategy, which includes the following in each round: (1) Each agent generates a resource scheduling decision based on its local observation information and the global dependency information; (2) Performs joint actions based on the resource scheduling decisions of each agent, evaluates the value of the joint actions of all agents using a globally shared value function, updates the network parameters of the centralized collaborative strategy network with the value evaluation as a feedback reward signal, updates the parameters of the individual strategy networks of each agent based on the updated centralized collaborative strategy network, and uses the updated centralized collaborative strategy network and individual strategy networks for the next round of iteration; (3) When the centralized collaborative strategy network reaches convergence, the iteration ends; wherein, each agent maintains an individual strategy network for selecting the optimal resource scheduling strategy for the subtasks assigned to it, and the agent network jointly maintains a centralized collaborative strategy network for conflict resolution and collaborative optimization.

[0052] This invention employs an individual policy network and a centralized collaborative policy network to coordinate the balance between individuals and the global situation. A multi-agent reinforcement learning strategy is introduced to iteratively update both the individual and centralized collaborative policy networks to achieve optimal performance under set constraints. This constructs a two-layer intelligent decision-making architecture that balances individual autonomy and global collaboration. Through the training mechanism of multi-agent reinforcement learning, the individual policy network learns independent optimal actions in complex environments. Simultaneously, the centralized collaborative policy network integrates global information and coordinates individual conflicts to achieve overall goals. This helps to effectively solve the "collaborative game" problem between individual optimization and system optimization in dynamic, multi-constrained resource scheduling scenarios, achieving a leap from decentralized decision-making to collective intelligence.

[0053] In some embodiments of the present invention, the preset shared value function includes a global resource utilization reward, a task completion efficiency reward, a collaboration cost penalty, and a constraint violation penalty, and the global sharing of the preset shared value function is implemented using a global Critic network.

[0054] This invention employs multiple dimensions of a preset shared value function and provides a global sharing implementation method for the preset shared value function. This facilitates the leap in decision-making capabilities of multi-agent systems from local perception to global collaboration in complex dependent environments. It also helps to preserve excellent conflict resolution and collaborative optimization capabilities among agents while ensuring the individual policy network resource selection and environmental adaptability.

[0055] In some embodiments of the present invention, for the resources used for integrated processing of communication, sensing, computing and intelligence, the method further includes: (1) continuously collecting real-time operating status data of intelligent agent resources, including various communication facilities, sensing devices, computing units and data resources, as the physical basis for supporting the collaborative operation of intelligent agents; (2) pre-constructing and maintaining a global resource map that supports real-time query, and using the global resource map to maintain real-time operating status data of various intelligent agent resources.

[0056] By adopting this embodiment of the invention, a resource awareness layer and a basic resource layer are introduced to jointly maintain the resources used for integrated processing of communication, sensing, computing and intelligence. The resource awareness layer dynamically monitors and abstracts the real-time status and performance of various heterogeneous resources, and the basic resource layer realizes unified resource pooling, scheduling and lifecycle management. This helps to break down the resource barriers between the four functional domains of communication, sensing, computing and intelligence, and helps to improve the overall performance of complex tasks in terms of processing efficiency, resource utilization and system adaptability.

[0057] On the other hand, this invention also proposes an integrated sensing, computing, and intelligence system for intelligent agent networks, which includes the following structure: (1) Session management layer, which is used to call the large language model LLM in response to the natural language text input by the user, transform the user intent contained in the natural language text into multiple sub-tasks, and perform multi-dimensional matching and evaluation of the requirements of each sub-task with the capabilities of each agent in the agent network to generate a task allocation scheme.

[0058] (2) Resource scheduling layer, which is used to extract global dependency information based on subtasks and the dependencies between subtasks and distribute it to each agent in the agent network. Through multiple rounds of iteration, it obtains resource scheduling decisions that converge to the preset shared value function.

[0059] (3) Intelligent agent collaboration layer, used to allocate sub-tasks to the corresponding intelligent agents according to the task allocation scheme, and each intelligent agent schedules the resources used for integrated processing of sensing, computing and intelligence according to the resource scheduling decision.

[0060] In some embodiments of the present invention, the system further includes: (4) a resource perception layer, used to pre-build and maintain a global resource map that supports real-time query, and use the global resource map to maintain real-time operating status data of various intelligent agent resources; (5) a basic resource layer, used to continuously collect real-time operating status data of intelligent agent resources, including various communication facilities, sensing devices, computing units and data resources, as the physical basis for supporting the collaborative operation of intelligent agents.

[0061] Figure 3 This is a diagram of an integrated sensing, computing, and intelligence system architecture for intelligent agent networks according to an embodiment of the present invention. Figure 3As shown, this invention constructs a multi-layered intelligent agent network architecture. From top to bottom, this architecture sequentially encompasses a session management layer for understanding user intent, an intelligent agent collaboration layer responsible for organizing inter-agent coordination, a resource scheduling layer for resource allocation, a resource awareness layer for maintaining a global resource view, and a basic resource layer encapsulating heterogeneous physical resources. Each layer interacts through standard interfaces and a closed-loop feedback mechanism, jointly completing the entire process from user intent parsing to cross-domain resource scheduling, ultimately achieving efficient and reliable collaboration among intelligent agents in complex and dynamic environments.

[0062] Specifically, the functions of each layer in the intelligent agent network architecture are as follows: (1) Session Management Layer: As the top-level interface for system-user interaction, this layer focuses on understanding user intent and undertakes the core function of transforming user intent into executable tasks. This layer performs deep semantic analysis on user input through an intent parsing engine, identifies task elements and constraints, and transforms them into structured task instances. During the task execution phase, this layer centrally tracks and manages the entire task lifecycle, dynamically updates the progress based on feedback from lower layers, and coordinates task replanning and indicator adjustments when anomalies occur, ensuring that user goals can be accurately and reliably transmitted to the lower-level execution system.

[0063] (2) Agent Collaboration Layer: As the core hub for task execution, this layer is responsible for transforming abstract business requirements into actionable plans and for organizing collaboration among agents. This layer first structurally decomposes complex tasks, selects suitable agents from a candidate pool to form a conversation team based on sub-task requirements, and assigns specific roles and contexts. During task execution, it maintains collaborative work among agents by establishing dedicated communication channels, managing data exchange and state synchronization to ensure the process proceeds as planned. Simultaneously, this layer also assumes governance responsibilities for the collaboration process, using a rule engine to detect various collaboration anomalies such as resource contention and behavioral conflicts in real time, and promptly resolves conflicts through pre-set arbitration mechanisms or negotiation interventions, ensuring the efficiency and consistency of the entire collaboration system.

[0064] (3) Resource scheduling layer: This is the core execution engine for the system's capability resource utilization. It is responsible for transforming the action plans planned by the upper layer into the scheduling and allocation of heterogeneous resources such as communication, sensing, and computing in the network environment. Based on the dependencies between tasks and the global resource status, this layer generates the optimal resource combination through scheduling algorithms and makes dynamic adjustments to ensure that tasks meet performance requirements while maximizing the efficiency of global resource utilization.

[0065] (4) Resource Awareness Layer: This is the foundational layer for the interaction between the system and physical and virtualized resources. It constructs a real-time, dynamic digital understanding of intelligent agents and heterogeneous resources across the entire domain, providing accurate and timely data support for upper-level decision-making and maintaining a global resource view. This layer continuously collects status information from a wide range of computing resources, communication links, and sensing devices, establishes functional dependencies and performance correlations between resources, and ultimately constructs a global resource map that supports real-time queries. It also models the capability attributes and interactive behaviors of intelligent agents, forming dynamic capability profiles, and providing accurate and comprehensive situational awareness for upper-level scheduling and collaboration.

[0066] (5) Basic Resource Support Layer: As the core of the system architecture, this layer is responsible for providing resources to all functions in the upper layers and is used to encapsulate heterogeneous physical resources. This layer continuously collects real-time operational status data of various communication facilities, sensing devices, computing units, and data resources to build the physical foundation supporting the collaborative operation of intelligent agents. Communication resources include a full range of network facilities from 5G / 6G base stations and core network equipment to satellite communication terminals, which together build a reliable connection channel to ensure real-time interaction between intelligent agents; sensing resources cover a multimodal sensing system, including visual acquisition devices and environmental monitoring sensors, which provide real environmental data input for intelligent agent decision-making; computing and storage resources integrate heterogeneous computing power covering cloud computing centers, edge nodes, and terminal devices, providing elastically schedulable computing resources. In addition, different types of intelligent agents are also included in a unified resource management framework, including dedicated intelligent agents with environmental perception and autonomous execution capabilities, as well as general intelligent agent platforms built on large language models. These platforms provide the system with powerful cognitive and reasoning capabilities through standardized interaction interfaces.

[0067] By adopting the above-mentioned integrated system architecture of sensing, computing and intelligence for intelligent agent networks, and jointly scheduling intelligent agent networks and sensing, computing and intelligence resources, a new network paradigm can be formed with intelligent agent collaboration as the core and sensing, computing and intelligence integration as the capability foundation. It is possible to construct the functional division and inter-layer interaction mechanism of the conversation management layer, intelligent agent collaboration layer, resource scheduling layer, resource perception layer and basic resource layer in the architecture.

[0068] Figure 4 This is a flowchart illustrating the workflow of an integrated sensing, computing, and intelligence system for intelligent agent networks according to one embodiment of the present invention. Figure 4 As shown, in this invention, each layer works in close collaboration, transforming high-level user intents into a unified, interconnected action encompassing communication, sensing, computing power, and intelligent resources through layer-by-layer parsing, orchestration, and scheduling. Continuous optimization is achieved through closed-loop feedback during execution. This process can be systematically divided into the following key steps: (1) User intent understanding and task instantiation The session management layer receives business objectives, constraints, and key expectations from users via natural language, structured forms, or visual interfaces. Subsequently, an intent parsing engine performs deep semantic analysis and contextual inference on the user's ambiguous intent, identifying core tasks, key entities, and inherent logical connections. Qualitative requirements such as "fast" and "accurate" are then transformed into quantifiable service level objectives (SLOs). Finally, the parsed task objectives and SLOs are encapsulated as task instances, added to a dynamic task list, and entered into the lifecycle management process.

[0069] (2) Task decomposition and conversation team creation The agent collaboration layer first decomposes complex macro-level tasks into a series of logically related, sequentially or in parallel atomic subtasks, forming a clear execution path diagram. Based on the requirements of the subtasks, this layer initiates a selection process from the global agent candidate pool, matching suitable agent execution units for each subtask according to capability profiles, real-time status, and historical performance, and formally creating a unified session context environment.

[0070] (3) Initialization of agent collaboration context and establishment of internal communication The agent collaboration layer performs unified context initialization for the entire conversational team, injecting key information such as task objectives, global parameters, collaboration protocols, and success criteria into all members to ensure they have a shared and accurate understanding of the task. Simultaneously, it establishes and maintains a dedicated and reliable communication channel within the team, supporting various interaction modes such as event-driven and publish-subscribe, and guaranteeing deterministic message delivery based on task requirements. Through this channel, agents can send task requests, return execution results, and notify of abnormal situations according to a pre-defined collaboration process.

[0071] (4) Resource demand mapping and global scheduling strategy generation The resource scheduling layer receives the decomposed collaborative execution plan from the upper layer, performs business requirement analysis, and formally extracts the subtasks' resource requirements for communication, sensing, and computing, as well as performance indicators such as latency, bandwidth, and computing power. Subsequently, combined with the real-time global resource status provided by the resource sensing layer, it uses optimization algorithms and intelligent decision-making technology to generate one or more globally optimized resource scheduling plans, specifying in detail the communication links to be used by each subtask, the sensing devices to be invoked, and the computing units to be allocated.

[0072] (5) Dynamic binding of resources and issuance of task instructions Once the scheduling strategy is determined, the resource awareness layer maps the logical resource allocation in the strategy to specific physical or virtual resource entities at the underlying level, completing the dynamic binding of resources. Subsequently, precise configuration instructions and task start commands are generated and sent to the basic resource layer through a standard interface. Various intelligent agents, computing nodes, network devices, and sensors in the basic resource layer receive the instructions, are activated, and are ready to execute specific operations.

[0073] (6) Distributed execution and real-time state synchronization The task enters the physical execution phase. Each agent in the session team executes its subtasks in parallel or sequentially according to assignments, and performs necessary data exchange, intermediate result transmission, and execution status synchronization through the communication channels established during the initialization phase. The resource awareness layer continuously monitors various indicators from physical resources to agent states, providing the upper layers with real-time status quo information about the system's operation. Meanwhile, the agent collaboration layer, as the core of the entire system's collaborative control, is responsible for monitoring the progress of the entire collaborative session.

[0074] (7) Implement governance and optimize the closed loop The agent collaboration layer and resource scheduling layer continuously monitor and evaluate the execution process. Once the system detects performance deviations, resource conflicts, agent anomalies, or external environmental disturbances, it immediately triggers a governance process. Based on preset strategies, it automatically resolves conflicts, partially retries or reassigns tasks, dynamically adjusts resource allocation, and, if necessary, interrupts the current session and requests the session management layer to initiate task replanning. After execution, the results are fed back to the user, and all process data is recorded for analysis and optimization.

[0075] The above steps (1)-(7) constitute the workflow of the integrated sensory computing and intelligence system for intelligent agent networks. It can realize end-to-end automated execution and management from intention to result by using intelligent agents as collaborative units and sensory computing and intelligence resources as a unified foundation.

[0076] Adopting the aforementioned workflow of the integrated sensory-computing-intelligence system for agent networks, the integrated sensory-computing-intelligence architecture faces multiple challenges in resource scheduling, collaborative decision-making, and system resilience, given the complex requirements of agent networks in dynamic task environments and heterogeneous resource collaboration. To achieve efficient end-to-end collaboration and continuous optimization from user intent to resource execution, it is urgently necessary to introduce a key methodology system with global perception, intelligent decision-making, and adaptive evolution capabilities. Based on this, this invention proposes three key components supporting the integrated sensory-computing-intelligence architecture for agent networks, focusing on the entire process from task decomposition to underlying resource scheduling within the architecture.

[0077] This invention proposes a multi-dimensional resource-guaranteed intelligent agent collaborative workflow, which can form a closed-loop execution system driven by task objectives, centered on task orchestration, and guaranteed by resource collaboration. Using this workflow, a working method can be obtained that includes: "user intent understanding and task instantiation – task decomposition and session team creation – intelligent agent collaboration context initialization and internal communication establishment – ​​resource requirement mapping and global scheduling strategy generation – dynamic resource binding and task instruction issuance – distributed execution and real-time state synchronization – execution governance and optimization closed loop."

[0078] The first key part: an agent task orchestration method based on LLM (corresponding to step S110).

[0079] In the integrated architecture of sensory computing and intelligence for agent networks, the quality of task decomposition and allocation decisions directly determines the efficiency of system collaboration. Traditional task decomposition and allocation methods based on rules or optimization algorithms are difficult to adapt to complex and ever-changing semantic environments, and face challenges of high modeling complexity and insufficient flexibility when dealing with multi-task modeling and matching the capabilities of heterogeneous agents. To address this, this invention proposes an end-to-end task orchestration method based on Large Language Models (LLM). By leveraging the deep semantic understanding, logical processing, and context awareness capabilities of LLM, it achieves intelligent orchestration from macro-level task intent to specific task decomposition and assignment.

[0080] Large Language Models (LLMs) are a class of artificial intelligence models built on a Transformer architecture and pre-trained on ultra-large-scale corpora, possessing powerful natural language understanding and generation capabilities. In the collaborative architecture of agent networks, introducing an LLM-based task orchestration method enables intelligent generation and optimization from task decomposition to task assignment schemes. This method leverages the semantic understanding capabilities of LLMs to parse complex task requirements, utilizes their logical reasoning abilities to construct a hierarchical task decomposition tree, and generates the optimal agent assignment scheme based on the real-time system state. Compared to traditional methods, LLM-driven task orchestration can not only handle explicit constraints but also understand implicit business rules through context learning, achieving more intelligent and adaptive workflow construction.

[0081] Figure 5 This is a flowchart of an agent task orchestration method based on LLM in one embodiment of the present invention. Figure 5 As shown, the LLM-based agent task orchestration method can be divided into three modules, which are described in detail below: 1) Input Preprocessing and Prompt Construction Module. This module, serving as the pre-processing layer of LLM, plays a crucial role in transforming heterogeneous input data into standardized prompts. First, it performs syntactic analysis and key information extraction on the user-input natural language text, identifying semantic elements such as core verbs, operational objects, and constraints, eliminating ambiguity, and adjusting it into a structured representation. Then, it transforms machine-readable data from the agent's capability model, such as functional types, performance parameters, and real-time states, into natural language text easily understood by LLM through template conversion. Finally, based on a pre-defined multi-turn dialogue framework, it combines elements such as task objectives, agent capabilities, and output format requirements according to Chain of Thought (CoT) reasoning logic to construct complete prompts containing task context, decomposition rules, and allocation constraints, providing high-quality reasoning guidance for LLM.

[0082] 2) LLM Core Inference Module. This module is based on the core of a large language model using the Transformer architecture, achieving end-to-end task orchestration and inference through a self-attention mechanism. It utilizes a multi-head attention layer to deeply encode the task semantics and agent capability descriptions in the prompts, establishing semantic associations and dependencies between words. The logical inference component, guided by thought chain prompts, progressively executes task decomposition planning through a layered decoding strategy: first, it identifies the parallelizable units and temporal dependencies of the macro-task; then, it performs multi-dimensional matching and evaluation of sub-task requirements and agent capabilities; finally, it generates an allocation scheme by comprehensively considering load balancing and other optimization goals—that is, a formatted natural language expression containing elements such as a sub-task list, allocation relationships, and execution sequence.

[0083] 3) Solution Verification and Feedback Module. This module performs multi-dimensional verification and evaluation of the task allocation scheme generated by LLM. The verification engine first performs a logical consistency check, verifying the acyclicity of task dependencies through a graph traversal algorithm to ensure that there are no circular waits between subtasks; simultaneously, it performs a capability matching audit, verifying whether the requirements of each subtask and the capability specifications of the assigned agent meet preset thresholds. The conflict detection component identifies resource allocation conflicts and agent overload risks, and evaluates the feasibility of the scheme. The verification results generate a detailed evaluation report through a feedback mechanism, marking potential problems and improvement suggestions in the scheme. When a serious defect is detected, the module triggers a reprocessing process, feeding back the error type and correction guidelines to the input preprocessing module, guiding the system to build more accurate prompts for a new round of reasoning, forming a continuous optimization quality loop.

[0084] Based on the collaborative working mechanism of the three modules, the key process of this method can be summarized into the following six steps: 1) The input preprocessing module first performs deep semantic analysis and key information extraction on the natural language text, identifies the core objectives, key parameters, execution constraints and other elements of the task, and transforms them into a structured description of the overall task objective.

[0085] 2) Obtain a real-time capability profile of the global intelligent agent from the resource perception layer, including its functional type, performance specifications and current status.

[0086] 3) Based on the output of the first two steps, the prompting module intelligently combines the structured task objectives, agent capability descriptions, and output format requirements according to the reasoning logic of the thought chain, to construct a structured prompt word that is complete in information and highly guiding.

[0087] 4) LLM breaks down macro-level tasks into a set of sub-tasks with clear logical relationships and clarifies the dependencies between them, ultimately forming a hierarchical task decomposition tree or task list.

[0088] 5) Taking into account the agent's functional expertise, real-time load status, and historical performance, select and assign the most suitable agent to each subtask to complete the initial role allocation.

[0089] 6) Perform rigorous automated verification of the solution, including logical consistency checks to ensure that task dependencies are acyclic, as well as capability matching audits and resource conflict detection, to assess the actual executability of the solution.

[0090] The key points of the first crucial part are: ① By combining LLM with prompt word construction and constraint reasoning, the user's natural language text task intent is automatically transformed into a structured, executable task decomposition and allocation scheme. ② A three-level processing flow: input preprocessing and prompt construction → core reasoning of the large language model → scheme verification and feedback optimization, achieving end-to-end intelligent transformation from task description to executable allocation.

[0091] The first key protection points are: ① Methods for semantic analysis and unified parsing of task inputs in natural language and structured forms using large language models, including specific implementation mechanisms for task element identification, constraint extraction, and service level indicator quantification. ② Task decomposition and task allocation methods based on the reasoning capabilities of large language models, including methods for sub-task dependency analysis, agent capability profile matching, and dynamic team building strategies. ③ Methods for multi-dimensional verification of the solution, including logical consistency checks, capability matching audits, resource conflict identification, and prompt word optimization and re-reasoning mechanisms based on verification results.

[0092] The first key component, an LLM-based agent task orchestration method, achieves end-to-end intelligent transformation from natural language task description to executable assignment by constructing a three-stage pipeline architecture of "preprocessing-inference-verification". This method fully leverages the advantages of large language models in semantic understanding, logical reasoning, and context awareness. Through cue word engineering, it transforms the complex task decomposition and assignment problem into a controllable text generation task, effectively addressing the shortcomings of traditional methods in semantic adaptability and flexibility. A verification feedback mechanism ensures the reliability and feasibility of the output scheme, forming a continuously optimized closed-loop system. This method provides an efficient and intelligent task collaboration solution for integrated sensory-computing intelligence systems oriented towards agent networks.

[0093] The second key part: a multi-agent hierarchical reinforcement learning scheduling algorithm based on task dependency constraints (corresponding to step S120).

[0094] In agent network collaborative environments, agent collaborative tasks often exhibit complex dependency characteristics, including not only temporal constraints but also various forms of association such as data transmission, resource sharing, and result coupling. This dependency leads to strong correlation in resource requirements: sensing devices need to coordinate acquisition timing to ensure data consistency, communication bandwidth needs to meet the timeliness of data transmission between dependent tasks, and computing resources need to ensure the processing synchronization of dependent links. Traditional scheduling methods struggle to effectively characterize and satisfy the resource collaboration needs under such dependency constraints, resulting in problems such as resource allocation conflicts, waiting delays for dependent tasks, and low overall execution efficiency. To address this, this invention proposes a multi-agent hierarchical reinforcement learning scheduling algorithm based on task dependency constraints. By explicitly modeling task dependencies, designing a dependency-aware reward mechanism and a hierarchical decision architecture, it achieves optimized resource collaboration scheduling under dependency constraints.

[0095] Graph Neural Networks (GNNs) are a class of deep learning models specifically designed for processing graph-structured data. They utilize message passing mechanisms to propagate information and learn features between nodes in a graph, effectively handling complex relational data in non-Euclidean spaces. In agent network environments, task dependencies form a directed graph structure—nodes represent subtasks, and edges represent various dependencies between subtasks—providing deep structural awareness support for subsequent resource scheduling decisions.

[0096] Multi-Agent Reinforcement Learning (MARL) is an extension of reinforcement learning in multi-agent environments. It studies how multiple agents learn optimal decision-making strategies through interactive learning (cooperative learning or competitive learning) in a shared environment. Unlike traditional single-agent reinforcement learning, MARL faces unique challenges such as environmental non-stationarity, credit allocation, and inter-agent coordination. In agent network resource scheduling scenarios, each agent (such as a drone, robot, or edge server) acts as an independent MARL agent, aiming to select the optimal resource scheduling strategy for its assigned subtasks. Each agent makes resource allocation decisions based on local observations while simultaneously learning collaborative strategies through interaction with other agents. This architecture maintains the agents' autonomous decision-making capabilities while achieving system-level optimization through MARL's collaborative mechanisms.

[0097] Figure 6 This is a schematic diagram of a multi-agent hierarchical reinforcement learning scheduling algorithm based on task dependency constraints in one embodiment of the present invention. Figure 6 As shown, the multi-agent hierarchical reinforcement learning scheduling algorithm based on task dependency constraints can be divided into a task graph dependency modeling module and a distributed multi-agent cooperative scheduling module, which are described in detail below: 1) Task Dependency Graph Modeling Module The task dependency graph modeling module is a core component of this algorithm, responsible for transforming the complex dependencies in agent-cooperative tasks into a structured representation that computers can understand and process. Based on GNN technology, this module uses deep learning and feature extraction to analyze inter-task relationships, providing crucial constraint information for subsequent resource scheduling decisions.

[0098] The task decomposition results received by the module contain two core components: a set of subtasks and a dependency description. Each task in the subtask set needs to be represented using features, including task type identifiers (e.g., perception, computation, communication), resource requirements (bandwidth, computing power, storage space), and execution constraints (latest completion time, priority weight), among other quantitative features. The dependency description needs to be parsed into explicit association pairs, including metadata such as dependency type (time-sequence dependency, data dependency, resource dependency), dependency strength (strong dependency, weak dependency), and dependency direction.

[0099] The module first constructs an initial directed graph based on the resolved dependencies. During this process, the system performs loop detection and elimination, and uses a topological sorting algorithm to ensure the generated graph structure satisfies the properties of a directed acyclic graph (DAG). For each subtask node, a corresponding feature vector is generated, containing task attributes and constraints. Edges are assigned different weights and attributes based on the dependency type; for example, temporal dependency edges may contain minimum time interval constraints, while data dependency edges may contain data transmission volume information.

[0100] The constructed DAG is used as input to the GNN for processing. The GNN performs feature aggregation and transformation at each node through multi-layer graph convolution operations. The first convolution layer mainly aggregates information from directly adjacent nodes to learn local dependency patterns; deeper convolutions can capture dependencies at greater distances and global topological features. Through an attention mechanism, the GNN can identify key dependency paths and important nodes, assigning differentiated importance weights to different dependencies.

[0101] After processing by the GNN, the module outputs standardized graph structure data, comprising two main components: a node feature matrix and an adjacency matrix. Each row vector in the node feature matrix represents the embedding representation of the corresponding task learned by the GNN; these vectors encode the task's own attributes and its structural information within the dependency network. The adjacency matrix quantifies the strength of dependencies between tasks, where each element reflects the importance and constraint strength of the corresponding dependency.

[0102] 2) Distributed multi-agent cooperative scheduling module The distributed collaborative scheduling module based on GNN-MARL fusion is a key execution component of this algorithm, responsible for transforming the structured information output by the task dependency graph modeling module into actual resource scheduling decisions. This module employs a multi-agent reinforcement learning framework, distributing the global dependency information extracted by the GNN to each agent to achieve decentralized collaborative decision-making. This maintains the autonomy of individual decisions while ensuring the overall collaborative efficiency of the system.

[0103] The module first constructs an agent cooperative network based on the parsed dependency graph information. During this process, the system optimizes the information distribution strategy, using an attention allocation algorithm to ensure each agent receives the most relevant global information. For each agent node, a customized information packet is generated, containing a global graph embedding summary, relevant dependency context, and critical path hints. Connection weights are dynamically adjusted based on the degree of cooperation between agents; for example, highly dependent agent pairs need to share more state information, while competing agent pairs need to coordinate resource allocation strategies.

[0104] The constructed agent cooperative network is processed as input to MARL. MARL utilizes a multi-agent cooperative mechanism, enabling policy learning and decision optimization at each agent level. The individual policy network primarily learns individual decisions based on local observations, mastering resource selection and environmental adaptation capabilities; the centralized cooperative policy network learns multi-agent cooperative patterns, mastering conflict resolution and cooperative optimization capabilities. Through a value function sharing mechanism, MARL can coordinate decision conflicts among agents, providing differentiated policy guidance for different cooperative scenarios.

[0105] Specifically, the individual policy network, serving as the foundational decision engine for each agent, learns resource selection patterns and environmental adaptation strategies, enabling agents to quickly respond to changes in the local environment and make effective independent decisions. The network input includes the agent's local task requirements, the state of available resources, and limited neighborhood environmental information; the output consists of specific resource selection decisions and execution parameter configurations.

[0106] The centralized collaborative policy network serves as the core of the system's collaborative optimization. By analyzing the dependencies, resource competition, and task coupling among agents, it acquires advanced capabilities for conflict detection and resolution. The network receives policy intent predictions from other agents, a summary of the system's global state, and dependency strength information, and outputs collaborative policy adjustment instructions and conflict resolution solutions.

[0107] The value function sharing mechanism achieves coordinated consensus in distributed decision-making through centralized value assessment. This mechanism employs a global Critic network to assess the overall system state and the joint actions of all agents, generating coordination signals to guide individual policy optimization. Through the shared value function, agents can understand the impact of their decisions on the overall system and achieve a balance between individual interests and collective goals.

[0108] In a specific embodiment of the present invention, to achieve the goal of agent cooperation under dependency constraints, a closed-loop process from dependency modeling to multi-agent reinforcement learning is constructed for the second key part, which can be summarized as the following six key steps: ① Construct a Directed Acyclic Graph (DAG) with subtasks and dependencies. ② The DAG is input into the GNN, aggregating the information of its neighboring nodes at each task node. ③ Global dependency information is distributed to each agent in the MARL. ④ Each agent makes cooperative decisions based on local observations and received global information. ⑤ The environment (neighborhood environment information) provides reward signals based on the joint actions of the agents. ⑥ The centralized cooperative policy network updates network parameters and guides the individual policy networks to update. ⑦ Return to step ④, and start the next iteration based on the updated policy network in the new state. The loop continues until the policy network converges.

[0109] The key points of the second crucial part include: ① Automatically transforming complex task dependencies into structured, executable resource scheduling strategies through graph neural networks combined with a multi-agent reinforcement learning framework. ② A two-layer processing architecture: task dependency graph modeling → distributed multi-agent collaborative scheduling, enabling intelligent decision-making from dependency identification to resource collaborative optimization.

[0110] The second key protection points include: ① Methods for modeling and extracting features from task dependencies using graph neural networks, including specific implementation mechanisms for subtask feature representation, dependency parsing, and graph structure construction. ② Distributed cooperative scheduling methods based on multi-agent reinforcement learning, including complete methods such as individual policy network design, centralized cooperative policy network construction, and value function sharing mechanisms.

[0111] Graph Neural Networks (GNNs) are deep learning-based models specifically designed for processing graph-structured data. They capture the relationships between nodes and the characteristics of the global graph by leveraging node features and graph structural information.

[0112] The second key component, a multi-agent hierarchical reinforcement learning scheduling algorithm based on task dependency constraints, fully leverages the advantages of graph neural networks in structured relationship modeling and multi-agent reinforcement learning in distributed collaborative decision-making. Through dependency-aware reward design and hierarchical optimization mechanisms, it transforms the complex resource scheduling problem into a learnable collaborative decision-making task, effectively addressing the shortcomings of traditional methods in handling dependency constraints and improving collaborative efficiency. This provides an efficient and collaborative resource scheduling solution for intelligent agent network-based integrated sensing and computing systems.

[0113] The third key part: a strategy optimization and system resilience assurance method based on genetic algorithms.

[0114] In integrated networks of sensing, computing, and intelligence, static policy configurations and fixed collaboration patterns are insufficient to cope with the continuous evolution of environmental dynamism and task diversity during long-term operation. The core of the strategy optimization and system resilience assurance method based on genetic algorithms lies in establishing a feedback optimization mechanism from system operating state to policy parameters, enabling the network architecture to have the ability to gradually adjust and continuously improve, thereby exhibiting stronger adaptability and stability in the face of resource fluctuations, task mode changes, or external disturbances.

[0115] Genetic algorithms, as a global optimization method that simulates the natural evolutionary process, can effectively explore complex policy spaces and find optimal or near-optimal solutions through operations such as selection, crossover, and mutation. In agent network environments, this method encodes various resource scheduling decisions into a chromosome population, continuously improves policy performance through multiple generations of evolution, and ultimately obtains high-quality policy solutions that can adapt to complex environments.

[0116] Figure 7 This is a flowchart of a strategy optimization and system resilience assurance method based on a genetic algorithm in one embodiment of the present invention. Figure 7 As shown, the specific process of this method can be summarized in the following steps: (1) Establish a multi-dimensional decision-making effectiveness evaluation system, covering multiple key dimensions such as task completion rate, response latency, resource utilization rate, energy efficiency, and fault tolerance performance. In actual operation, it is necessary to continuously collect various performance data, calculate the specific values ​​of each indicator in real time, and compare and analyze them with historical baseline data or preset expected targets to form an accurate judgment on the current strategy execution effect.

[0117] (2) Various resource scheduling decisions are abstracted into coded structured units. These units contain adjustable operating parameters and logical components. All these decision-making units together constitute a continuously accumulating strategy knowledge base. The evolutionary algorithm engine, as the core driving mechanism of the entire process, will be automatically triggered according to a preset cycle or when performance degradation is detected. It will select the currently performing units from the strategy base as the basis and generate a new generation of strategy variants through evolutionary operations such as structural reorganization and parameter perturbation. / / This is equivalent to initialization (3) For newly generated strategy variants, a comprehensive evaluation of their overall performance and robustness under different conditions is required. Only strategy variants that demonstrate significant performance improvements in simulation tests will be deployed to the actual operating environment. New strategies will be deployed to the actual operating environment in a gradual and controllable manner.

[0118] (4) The operational data of the new strategy in the actual environment will be fed back to the evaluation system in real time, thereby completing a complete optimization cycle. By continuously comparing the actual performance of different strategy versions, the system can retain effective strategy units, eliminate ineffective or poorly performing parts, and identify and strengthen those strategy features that help maintain system stability under abnormal conditions, thereby enhancing the resilience of the entire network.

[0119] The key points of the third crucial part include: ① Transforming the system's operating state and performance indicators into an evolutionary strategy optimization scheme through a genetic algorithm, thereby achieving continuous self-improvement of the strategy. ② Adopting a closed-loop architecture of "evaluation-optimization-deployment-feedback" to form an intelligent decision-making cycle for strategy performance monitoring and dynamic adjustment.

[0120] The third key protection point includes: ① a closed-loop control method for policy optimization based on genetic algorithms, including policy encoding rules, fitness function design, and specific implementation mechanisms of genetic operations. ② a multi-dimensional evaluation system for decision-making efficiency, covering quantitative evaluation methods for key indicators such as task completion rate, response latency, and resource utilization.

[0121] The third key component, a strategy optimization and system resilience assurance method based on genetic algorithms, introduces biological evolution mechanisms into the network strategy management process. This provides the integrated sensory-computing-intelligence network with an inherent ability for continuous self-improvement, enabling the system to continuously adapt to complex and ever-changing environments and naturally form reliable resilience assurances in the long-term evolution.

[0122] Corresponding to the above method, the present invention also provides an electronic device, which includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the electronic device performs the steps of the method as described above.

[0123] Corresponding to the methods described above, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the above embodiments. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0124] Corresponding to the above methods, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method as described in any of the above embodiments.

[0125] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0126] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0127] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for integrating sensing, computing, and intelligence in intelligent agent networks, characterized in that, include: In response to the natural language text input by the user, the large language model LLM is invoked to transform the user intent contained in the natural language text into multiple sub-tasks. The requirements of each sub-task are matched and evaluated with the capabilities of each agent in the agent network in a multi-dimensional way to generate a task allocation scheme. Global dependency information is extracted based on subtasks and the dependencies between subtasks and distributed to each agent in the agent network. Through multiple rounds of iteration, a resource scheduling decision that converges to the preset shared value function is obtained. Subtasks are assigned to the corresponding intelligent agents according to the task allocation scheme, and each intelligent agent schedules resources for integrated processing of sensing, computing and intelligence according to the resource scheduling decision.

2. The method according to claim 1, characterized in that, For the resource scheduling decision, the method further includes: optimizing the resource scheduling decision using a genetic algorithm; wherein the genetic algorithm uses all possible resource scheduling decisions as the search space, uses the resource scheduling decisions as chromosomes, evaluates the chromosomes using a pre-constructed multi-dimensional decision performance evaluation system, and generates new chromosomes based on the evaluation results.

3. The method according to claim 2, characterized in that, The dimensions of the multi-dimensional decision-making effectiveness evaluation system include some or all of the dimensions such as task completion rate, response latency, resource utilization rate, energy efficiency, and fault tolerance performance.

4. The method according to claim 1, characterized in that, Prior to the step of invoking the Large Language Model (LLM) in response to user-input natural language text, the method further includes: performing syntactic analysis and key information extraction on the user-input natural language text, identifying semantic elements including core verbs, objects of operation, and constraints, eliminating ambiguity, and adjusting it into a structured representation.

5. The method according to claim 1, characterized in that, The subtasks and their dependencies are recorded in the form of a directed graph, where subtasks are nodes and the dependencies between subtasks are edges. The step of extracting global dependency information based on subtasks and the dependencies between subtasks includes: using a graph neural network (GNN) to aggregate the information of each node's neighboring nodes in the directed graph to obtain global dependency information.

6. The method according to claim 1, characterized in that, The step of obtaining a resource scheduling decision that converges to a preset shared value function through multiple iterations employs a multi-agent reinforcement learning strategy, and each round includes: Each intelligent agent generates resource scheduling decisions based on its local observation information and the global dependency information; Based on the resource scheduling decisions of each agent, joint actions are carried out. For the joint actions of all agents, a globally shared value function is used for value evaluation. The value evaluation is used as a feedback reward signal to update the network parameters of the centralized collaborative strategy network. Based on the updated centralized collaborative strategy network, the parameters of the individual strategy networks of each agent are updated. The updated centralized collaborative strategy network and individual strategy networks are used for the next round of iteration. The iteration ends when the centralized collaborative policy network reaches convergence. Each agent maintains an individual policy network to select the optimal resource scheduling policy for the subtasks assigned to it, and the agent network jointly maintains a centralized collaborative policy network to achieve conflict resolution and collaborative optimization.

7. The method according to claim 1, characterized in that, The preset shared value function includes global resource utilization reward, task completion efficiency reward, collaboration cost penalty and constraint violation penalty. The global sharing of the preset shared value function is implemented using a global Critic network.

8. The method according to claim 1, characterized in that, For the resources used in the integrated processing of sensing, computing, and intelligence, the method further includes: Continuously collect real-time operational status data of intelligent agent resources, including various communication facilities, sensing devices, computing units and data resources, as the physical basis to support the collaborative operation of intelligent agents; Pre-build and maintain a global resource map that supports real-time queries, and use the global resource map to maintain real-time operational status data of various intelligent agent resources.

9. A sensor-computer intelligence integrated system for intelligent agent networks, characterized in that, The system includes: The conversation management layer is used to respond to user input natural language text by calling the Large Language Model (LLM). It transforms the user intent contained in the natural language text into multiple sub-tasks, performs multi-dimensional matching and evaluation of the requirements of each sub-task with the capabilities of each agent in the agent network, and generates a task allocation scheme. The resource scheduling layer is used to extract global dependency information based on subtasks and the dependencies between subtasks and distribute it to each agent in the agent network. Through multiple rounds of iteration, it obtains a resource scheduling decision that converges to the preset shared value function. The agent collaboration layer is used to allocate subtasks to the corresponding agents according to the task allocation scheme, and each agent schedules resources for integrated processing of sensing, computing and intelligence according to the resource scheduling decision.

10. The integrated sensing and computing system according to claim 9, characterized in that, The system also includes: The resource awareness layer is used to pre-build and maintain a global resource map that supports real-time queries, and to maintain real-time operational status data of various intelligent agent resources using the global resource map. The basic resource layer is used to continuously collect real-time operational status data of intelligent agent resources, including various communication facilities, sensing devices, computing units, and data resources, as the physical foundation supporting the collaborative operation of intelligent agents.

Images