Intelligent management scheduling system and method based on multi-agent and retrieval enhanced generation

Abstract

The application discloses an intelligent management scheduling system and method based on multi-agent and retrieval enhanced generation, and belongs to the field of artificial intelligence and intelligent management technology. The system comprises a task agent layer, a coordination agent layer, a knowledge agent layer and a communication module. The task agent generates an initial scheduling scheme based on a reinforcement learning model; the coordination agent models and predicts potential conflicts between tasks through a graph neural network; the knowledge agent integrates a retrieval enhanced generation module, and in response to conflict prediction, retrieves relevant information from a dynamic knowledge base and generates an optimized strategy using a large language model. The application integrates perception, prediction, knowledge seeking, decision making and execution through a "conflict driven" mechanism, realizes dynamic knowledge injection and closed loop optimization, effectively solves the problems of static passivity, knowledge update lag and weak collaboration ability of existing systems, and significantly improves the adaptability, decision quality and overall collaboration efficiency of the system in a complex dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent management technology, and in particular relates to an intelligent management and scheduling system and method based on multi-agent and retrieval augmentation generation (RAG) technology, which is applicable to task scheduling and resource coordination in complex and dynamic scenarios such as smart campuses, smart warehouses, and smart retail. Background Technology

[0002] With the rapid development of artificial intelligence and the Internet of Things (IoT) technologies, intelligent management and scheduling systems are widely used in complex and dynamic scenarios such as smart campuses, smart warehousing, and smart retail. These scenarios typically involve multi-task concurrency, heterogeneous resources, and real-time changes in environmental parameters, placing extremely high demands on the autonomy, adaptability, and collaborative efficiency of the scheduling system. Realizing a system capable of efficient, intelligent decision-making and collaborative control in such highly dynamic and uncertain environments has become an important research direction at the intersection of artificial intelligence and intelligent management.

[0003] Currently, existing technical solutions in this field mainly suffer from the following three core defects: First, the systems mostly rely on preset static rules or single optimization models, and their decision-making logic is rigid, unable to adapt to real-time changes in the environment, resulting in delayed response and suboptimal decisions; Second, the system's knowledge base is usually static and isolated, making it difficult to integrate and utilize constantly evolving external domain knowledge (such as new regulations, market dynamics, and historical cases) in real time, leading to a decline in decision quality as the environment changes and compliance risks; Third, multiple executing entities in the system (such as robots or subsystems with different functions) often present an "information silo" state, lacking effective information sharing and collaboration mechanisms, which easily leads to resource conflicts, path blockages, and overall inefficiency.

[0004] Therefore, there is an urgent need for an intelligent management system that can proactively sense and dynamically adapt to environmental changes, learn and utilize external knowledge in real time, and efficiently coordinate heterogeneous resources from multiple parties, so as to effectively solve the technical problems of static passivity, lagging knowledge updates, and weak collaboration capabilities in existing systems. Summary of the Invention

[0005] This invention aims to address the problems of static passivity, lagging knowledge updates, and weak collaborative capabilities in existing technologies. Embodiments of this invention provide an intelligent management and scheduling system based on multi-agent and retrieval-enhanced generation, comprising:

[0006] The task agent layer includes multiple task agents, which are used to receive, analyze and initially process tasks, and generate an initial scheduling scheme based on a reinforcement learning model.

[0007] The coordinating agent layer includes at least one coordinating agent, which monitors the task execution status of the task agent layer, models the dependencies and conflicts between tasks through graph neural networks, and initiates optimization requests.

[0008] The knowledge agent layer includes at least one knowledge agent, which integrates a retrieval enhancement and generation module to respond to the query requests of the coordinating agent, retrieve external knowledge bases, and generate optimized scheduling strategies.

[0009] The communication module enables asynchronous communication between agents based on message queues.

[0010] The retrieval enhancement generation module includes:

[0011] The retrieval unit is used to convert query requests into semantic vectors and perform similarity searches in the knowledge base;

[0012] The generation unit is used to concatenate the search results with real-time status data into prompt words, which are then input into a large language model to generate a structured scheduling scheme.

[0013] In some embodiments, the knowledge base is a dynamically updated vector database that stores at least one of the following: historical scheduling cases, industry standards, and equipment manuals.

[0014] In some embodiments, the coordinating agent uses a graph neural network (GNN) to model the task agent and its task relationships to detect at least one of path conflicts and resource preemption.

[0015] In some embodiments, the task agent, the coordinating agent, and the knowledge agent communicate with each other via JSON-formatted messages, which contain at least the task ID, status code, and query instruction.

[0016] In some embodiments, the above system also supports deployment in at least one scenario among smart campus, smart warehousing, and smart retail, and achieves adaptation by defining scenario-specific intelligent agents and knowledge bases.

[0017] On the other hand, this application also provides an intelligent management and scheduling method based on multi-agent and retrieval enhancement generation, comprising the following steps:

[0018] Step S1: The task agent receives the task request, breaks down the task, and generates an initial scheduling scheme based on the reinforcement learning model.

[0019] Step S2: The coordinating agent monitors the task execution status, detects conflicts through a graph neural network, and initiates an optimization query to the knowledge agent;

[0020] Step S3: The knowledge agent calls the retrieval enhancement generation module to retrieve the knowledge base and generate optimization strategies;

[0021] Step S4: Coordinate the intelligent agents to integrate and optimize strategies, and issue update instructions to the task intelligent agents;

[0022] Step S5: The task agent executes the updated plan and feeds back the execution results to the system, forming a closed-loop control.

[0023] In some embodiments, the retrieval process of the retrieval enhancement generation module in step S3 uses Sentence-BERT for semantic encoding, and the generation process uses the LLaMA-2 large language model.

[0024] In some embodiments, the optimization strategy includes at least one of device scheduling instructions, path adjustment instructions, and resource allocation instructions.

[0025] On the other hand, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, to implement any of the intelligent management and scheduling methods based on multi-agent and retrieval enhancement generation as described above.

[0026] On another aspect, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements any of the intelligent management and scheduling methods based on multi-agent and retrieval-enhanced generation as described above.

[0027] The above technical solution has the following technical effects:

[0028] This application presents an intelligent management and scheduling system and method based on multi-agent and retrieval-enhanced generation (RAG). By constructing a hierarchical architecture of task agents, coordination agents, and knowledge agents, and introducing the core mechanism of "conflict-driven knowledge enhancement," it deeply integrates the autonomous decision-making of reinforcement learning, the forward-looking conflict prediction of graph neural networks, and the dynamic knowledge reasoning capabilities of retrieval-enhanced generation (RAG). This systematically solves the core problems of static passivity, lagging knowledge updates, and weak collaborative capabilities in existing technologies. Specifically, the system achieves a paradigm shift from passive response to proactive prediction and optimization, proactively avoiding potential conflicts and significantly improving the quality and efficiency of scheduling decisions. By drawing experience and norms from dynamically updated external knowledge bases on demand and in real-time, the system possesses strong environmental adaptability and continuous evolution capabilities, enabling it to make highly compliant and optimal decisions even in the face of unknown or unexpected situations. Simultaneously, the asynchronous communication mechanism based on message queues ensures efficient and reliable collaboration among multiple agents, thereby improving overall operational efficiency, resource utilization, and system robustness in complex and dynamic environments. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the structure of an intelligent management and scheduling system based on multi-agent and retrieval enhancement according to an embodiment of this application;

[0030] Figure 2 This is a flowchart of another embodiment of the intelligent management and scheduling method based on multi-agent and retrieval enhancement generation in this application;

[0031] Figure 3 This is a schematic diagram of another embodiment of this application. Detailed Implementation

[0032] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0033] Currently, in complex and dynamic scenarios such as smart restaurants, modern industrial parks, and intelligent warehouses, although automated equipment has been introduced, existing scheduling systems are generally limited by static rules, isolated decision-making, and passive response modes. They are unable to cope with real-time changes in task flow, environment, and resources, resulting in low efficiency, frequent conflicts, and a lack of adaptability.

[0034] In view of this, this application proposes an intelligent management and scheduling system and method based on multi-agent and retrieval-augmented generation, the core of which lies in constructing a "conflict-driven knowledge enhancement" closed loop. Logically, the system is divided into a task agent layer, a coordination agent layer, and a knowledge agent layer. The task agents, as distributed execution units, generate preliminary scheduling schemes for received tasks based on reinforcement learning models. The coordination agent acts as the system's "nerve center," continuously modeling the states and interactions of all task agents through graph neural networks, constructing a global dynamic relationship graph, and thereby proactively predicting potential risks such as resource contention and path conflicts. This predicted "conflict signal" serves as a precise trigger condition, driving the knowledge agent to activate its integrated Retrieval-Augmented Generation (RAG) module. This module then retrieves relevant cases and specifications from a dynamically updated domain knowledge base, integrates real-time contextual information, and uses a large-scale language model to generate structured, executable optimization strategies. The coordination agent translates this strategy into specific instructions and issues them to the relevant task agents. The results of the entire process will be fed back to the system to continuously optimize the decision-making model of the task agent and the prediction model of the coordinating agent, thereby forming a complete intelligent closed loop of "perception-prediction-knowledge-decision-execution-feedback-learning", ultimately enabling the system to have the ability to self-optimize and continuously evolve in complex dynamic environments.

[0035] like Figure 1As shown, one embodiment of this application provides an intelligent management and scheduling system based on multi-agent and retrieval-enhanced generation, comprising:

[0036] The task agent layer includes multiple task agents, which are used to receive, analyze and initially process tasks, and generate an initial scheduling scheme based on a reinforcement learning model.

[0037] The coordinating agent layer includes at least one coordinating agent, which monitors the task execution status of the task agent layer, models the dependencies and conflicts between tasks through graph neural networks, and initiates optimization requests.

[0038] The knowledge agent layer includes at least one knowledge agent, which integrates a retrieval enhancement and generation module to respond to the query requests of the coordinating agent, retrieve external knowledge bases, and generate optimized scheduling strategies.

[0039] The communication module enables asynchronous communication between agents based on message queues.

[0040] The retrieval enhancement generation module includes:

[0041] The retrieval unit is used to convert query requests into semantic vectors and perform similarity searches in the knowledge base;

[0042] The generation unit is used to concatenate the search results with real-time status data into prompt words, which are then input into a large language model to generate a structured scheduling scheme.

[0043] The above embodiments achieve intelligent scheduling through the deep integration of a hierarchical multi-agent system and retrieval-enhanced generation technology, constructing a "conflict-driven knowledge enhancement" closed-loop architecture. Specifically, the task agent layer acts as the execution end, making autonomous decisions and preliminary plans based on a reinforcement learning model; the coordination agent layer acts as the central hub, using graph neural networks to dynamically model global task relationships, proactively predicting rather than passively responding to potential conflicts; once a conflict is predicted, the retrieval-enhanced generation module of the knowledge agent layer is immediately triggered. This module obtains relevant experience and norms from a dynamic knowledge base through semantic retrieval and uses the reasoning capabilities of a large language model to generate an optimization strategy that integrates real-time context; finally, the strategy is fed back to the execution layer through a message queue-based communication module, forming a complete closed loop of "perception-prediction-knowledge-decision-execution-feedback".

[0044] The system described in the above embodiment realizes a paradigm shift from passive response to proactive prediction and optimization. Through the forward-looking conflict prediction of graph neural networks and the dynamic knowledge injection of RAG, it can avoid resource conflicts and path blockages in advance, thereby significantly improving the quality of scheduling decisions and overall operational efficiency. Due to the system's strong environmental adaptability and continuous evolution capabilities, its dynamic knowledge base and closed-loop learning mechanism based on execution feedback enable the system to absorb external knowledge in real time and optimize its internal models. Even when facing unknown scenarios or unexpected situations, it can still make highly compliant and optimal decisions. This solves the "information silo" problem in complex dynamic environments, effectively improving resource utilization, collaborative efficiency, and robustness.

[0045] In some embodiments, the knowledge base is a dynamically updated vector database that stores at least one of the following: historical scheduling cases, industry standards, and equipment manuals.

[0046] The above embodiments utilize semantic coding models (such as Sentence-BERT) to convert unstructured or semi-structured documents such as historical scheduling cases, industry standards, and equipment manuals into semantic vectors in a high-dimensional space for storage and indexing. This upgrades the traditional static, structured knowledge base into a dynamic database based on vector representation. When a knowledge agent performs a retrieval, the query request is also converted into a vector, and the most relevant knowledge fragments are found in the vector space through efficient similarity calculation (such as cosine similarity), thereby achieving accurate knowledge retrieval based on semantic similarity rather than keyword matching. This embodiment aims to solve the problems of outdated knowledge and inaccurate retrieval caused by the lagging updates and mechanical retrieval methods of traditional knowledge bases. It provides a continuously evolving knowledge foundation that can be "deeply understood" by the agent, ensuring that system decisions are always based on the latest and most relevant domain knowledge.

[0047] The system described in the above embodiments possesses powerful knowledge preservation and continuous evolution capabilities. The dynamic update mechanism allows newly generated scheduling cases and revised industry standards to be incorporated into the knowledge system in real time, fundamentally overcoming the drawback of traditional systems where decision-making quality declines due to knowledge stagnation. Furthermore, vector-based semantic retrieval significantly improves the accuracy and efficiency of knowledge retrieval. The system can intelligently and quickly locate the most relevant solutions and historical experiences for the current conflict scenario from massive amounts of knowledge, avoiding the limitations and information overload of keyword retrieval. This provides a reliable information foundation for generating high-quality, highly compliant optimization strategies from large-scale language models. Thus, the entire intelligent management and scheduling system becomes a learning system that continuously accumulates experience and becomes increasingly intelligent over time.

[0048] In some embodiments, the coordinating agent uses a graph neural network (GNN) to model the task agent and its task relationships to detect at least one of path conflicts and resource preemption.

[0049] In this embodiment, the coordinating agent abstracts the entire multi-agent system into a dynamic graph-structured data model. Each task agent acts as a node in the graph, and its state information (such as position, speed, and resource load) constitutes node features. The interactions or potential relationships between agents (such as spatial proximity and resource competition) constitute edges in the graph. Edge features can represent physical distances, communication delays, or competition for the same scarce resource (such as shared charging stations or narrow passages), while edge weights can dynamically represent the probability or cost of conflict. This graph is a "snapshot" of the entire system at a certain moment. The coordinating agent uses a graph neural network (GNN) to model this dynamic graph. Through its core message passing mechanism, each node can aggregate information from its neighboring nodes, thereby learning the global, structured dependencies of the system. The purpose of this modeling approach is to transcend the limitations of traditional methods that can only analyze independent events, and instead proactively and forward-lookingly detect potential conflicts hidden in complex interactions, such as path intersections and resource contention, from the perspective of the overall system's interconnectedness, thereby achieving a fundamental shift from passive response to proactive early warning.

[0050] The above embodiments endow the system with powerful global situational awareness and forward-looking risk prediction capabilities, enabling it to identify risks several decision cycles before a conflict occurs. This provides a valuable time window for subsequent knowledge enhancement and optimization decisions, significantly reducing system efficiency losses and resource waste caused by conflict. Furthermore, the GNN-based modeling method exhibits excellent scalability to system scale. Regardless of the increase or decrease in the number of agents in the system, their interaction relationships can be effectively analyzed within a unified graph structure, allowing the system to adapt to increasingly complex and large-scale application scenarios. This embodiment constitutes the core perception link of the "conflict-driven" mechanism, providing indispensable and accurate predictive signals for the closed-loop intelligent optimization of the entire system.

[0051] In some embodiments, the task agent, the coordinating agent, and the knowledge agent communicate with each other via JSON-formatted messages, which contain at least the task ID, status code, and query instruction.

[0052] This embodiment designs a standardized, structured, and lightweight communication protocol for heterogeneous multi-agent systems. The protocol uses JSON (JavaScript Object Notation) as the message encapsulation format, stipulating that each message must include core fields such as task ID, status code, and query command. The "task ID" is used to uniquely identify and track the task chain; the "status code" accurately conveys the execution stage or abnormal situation of the agent (such as "in execution" or "conflict warning"); and the "query command" carries a structured request initiated by the coordinating agent to the knowledge agent. This embodiment aims to solve the common problems of inconsistent communication protocols, semantic ambiguity, and excessive interface coupling in distributed agent systems. It establishes an efficient, unambiguous, and loosely coupled information interaction mechanism between the three heterogeneous modules—task agent, coordinating agent, and knowledge agent—laying the foundation for reliable collaboration and efficient decoupling of the entire system.

[0053] This embodiment significantly improves the ease of system integration and the reliability of communication. JSON, as a universal format easily readable by both humans and machines, reduces the debugging and integration costs between different modules, while its clearly defined structured fields completely eliminate the risk of instruction misunderstandings that may arise from natural language or custom protocols. Furthermore, this standardized communication method provides a solid foundation for the system's scalability and maintainability. New agents can easily access the system by understanding and generating standard JSON messages without modifying the existing communication framework, enabling the system to flexibly adapt to business growth and technological iteration. It also helps ensure the efficient and accurate transmission of information flow in the "conflict-driven" closed loop, ensuring that the intelligent management and scheduling system based on multi-agent and retrieval enhancement can achieve precise perception, collaborative decision-making, and closed-loop optimization.

[0054] In some embodiments, the above system also supports deployment in at least one scenario among smart campus, smart warehousing, and smart retail, and achieves adaptation by defining scenario-specific intelligent agents and knowledge bases.

[0055] This embodiment of the system employs a three-layer intelligent agent architecture of "task-coordination-knowledge" and a "conflict-driven" closed-loop mechanism, which effectively improves the system's modularity and adaptability, making it a general-purpose technical solution independent of specific scenarios. When adapting to specific scenarios, functional customization is achieved by defining scenario-specific intelligent agent roles (e.g., defining a "classroom resource intelligent agent" in a smart campus and an "AGV scheduling intelligent agent" in a smart warehouse) and constructing corresponding domain knowledge bases (e.g., campus timetables, warehouse layout diagrams), without changing the core workflow and communication mechanism of the system. Therefore, by defining scenario-specific intelligent agents and knowledge bases for adaptation, the system effectively solves the technical problems of long development cycles, poor versatility, and difficulty in reusing traditional customized systems. This provides a powerful and rapidly portable intelligent foundation, enabling the same technical solution to be efficiently applied to complex and dynamic scenarios with varying requirements.

[0056] The above embodiments, by "injecting" corresponding intelligent agent definitions and knowledge bases into different scenarios, allow a scheduling system successfully validated in smart retail to be quickly reused in smart warehousing or smart campuses, significantly reducing the cost and time of repetitive development. Furthermore, this embodiment achieves a balance between versatility and domain specificity, maintaining the stability and technical advantages of the core architecture while deeply integrating with the business logic and knowledge system of specific industries through scenario-based configuration, thereby outputting high-quality scheduling decisions in various scenarios. This transforms the invention from a single system solution into a highly adaptable intelligent platform that can empower multiple industries, greatly expanding the application scope and commercial value of the technology.

[0057] On the other hand, such as Figure 2 This application also provides an intelligent management and scheduling method based on multi-agent and retrieval enhancement generation, comprising the following steps:

[0058] Step S1: The task agent receives the task request, breaks down the task, and generates an initial scheduling scheme based on the reinforcement learning model.

[0059] Step S2: The coordinating agent monitors the task execution status, detects conflicts through a graph neural network, and initiates an optimization query to the knowledge agent;

[0060] Step S3: The knowledge agent calls the retrieval enhancement generation module to retrieve the knowledge base and generate optimization strategies;

[0061] Step S4: Coordinate the intelligent agents to integrate and optimize strategies, and issue update instructions to the task intelligent agents;

[0062] Step S5: The task agent executes the updated plan and feeds back the execution results to the system, forming a closed-loop control.

[0063] The technical principle of the method in this embodiment lies in constructing an intelligent closed-loop control process of "perception-prediction-knowledge-decision-execution-feedback". This control process begins with the task agent's distributed autonomous decision-making based on reinforcement learning. Then, the coordinating agent uses a graph neural network to model the global state of the system and predict conflicts, using this as a precise trigger signal to drive the knowledge agent to acquire and generate optimization strategies that integrate historical experience and real-time context from a dynamic knowledge base through retrieval-enhanced generation technology. Finally, decision iteration and model optimization are completed through instruction issuance and result feedback. By deeply connecting and integrating the autonomy of reinforcement learning, the predictive power of graph neural networks, and the knowledge-based nature of retrieval-enhanced generation, this method fundamentally solves the technical defects of traditional scheduling methods, such as static and isolated decision-making, lagging knowledge updates, and difficulties in multi-party collaboration. This enables the system to transform from passive response to proactive optimization, and from experience solidification to continuous evolution.

[0064] Through the aforementioned closed-loop process, this method achieves significant technical effects: First, by proactively detecting conflicts and injecting knowledge on demand, it greatly improves the quality and foresight of scheduling decisions, effectively avoiding potential conflicts and reducing system oscillations and resource waste. Second, this method endows the system with strong environmental adaptability and continuous learning capabilities. Its closed-loop feedback mechanism allows the decision-making model of the task agent and the prediction model of the coordinating agent to continuously optimize based on execution results, thus becoming increasingly intelligent in complex dynamic environments. Finally, this method tightly coordinates dispersed agents into an organic whole, significantly improving the overall collaborative efficiency, resource utilization, and operational robustness of the system when dealing with high-concurrency, multi-task scenarios.

[0065] In some embodiments, the retrieval process of the retrieval enhancement generation module in step S3 uses Sentence-BERT for semantic encoding, and the generation process uses the LLaMA-2 large language model.

[0066] This embodiment precisely selects technologies and assigns functional roles to the two core stages of the retrieval enhancement generation module. In the retrieval stage, the Sentence-BERT semantic encoding model is used to convert query requests and knowledge base documents into high-dimensional vectors. Its purpose is to leverage its superior semantic understanding capabilities to achieve accurate and efficient knowledge retrieval based on semantic similarity rather than simple keyword matching, ensuring that the recalled knowledge fragments are semantically highly relevant to the current conflict scenario. In the generation stage, a large-scale language model like LLaMA-2 is employed. Its purpose is to utilize its powerful natural language understanding, logical reasoning, and text generation capabilities to deeply integrate the retrieved relevant knowledge with real-time state data, reasoning and constructing a complete, reasonable, and executable optimized scheduling scheme. This combination aims to collaboratively solve the two key problems of "accurate retrieval" and "high-quality generation," jointly improving the overall performance of the RAG module.

[0067] Through the aforementioned principles and division of labor, this embodiment ensures the accuracy of knowledge augmentation and the high quality of decision support from the outset. Sentence-BERT ensures a high semantic relevance between the retrieved experience cases and the current conflict scenario, effectively filtering out irrelevant information. LLaMA-2, on the other hand, generates logically coherent, executable, and insightful strategies, significantly surpassing rigid solutions generated by simple templates or rules. Furthermore, this embodiment's method forms an efficient "specialist and generalist collaboration" working mode. Sentence-BERT, as a specialist, ensures retrieval efficiency and accuracy, while LLaMA-2, as a generalist, ensures the generalization and innovation capabilities of the strategies. The combination of the two enhances the overall intelligence level while optimizing the utilization of computing resources. Thus, the RAG module is no longer a simple "document search and stitching" module, but becomes an "intelligent brain" capable of deep semantic understanding and creative problem-solving, thereby improving the intelligence level and reliability of the system's decision-making to a certain extent.

[0068] In some embodiments, the optimization strategy includes at least one of device scheduling instructions, path adjustment instructions, and resource allocation instructions.

[0069] This embodiment concretizes the abstract "optimization strategy" output by the retrieval enhancement generation module into a series of structured control instructions that can be directly recognized and executed by the system. These instructions are designed according to the core dimensions of the scheduling scenario, mainly covering equipment scheduling instructions (such as assigning / switching specific equipment), path adjustment instructions (such as replanning the travel route), and resource allocation instructions (such as reallocating computing resources, materials, or human resources). This embodiment helps to solve the problem of "disconnect between decision-making and execution" that is common in the output of large language models, thereby seamlessly and unambiguously transforming the optimization scheme generated by the knowledge agent based on semantic understanding and reasoning into atomic commands that can be accurately operated at the execution layer, thus building a "last mile" bridge from intelligent decision-making to physical execution.

[0070] The above embodiments enhance the executability and real-time response of optimization strategies, enabling intelligent decisions to be unambiguously parsed by the task agent and immediately drive physical devices or software resources. This achieves instantaneous and precise transformation from "thinking" to "action," significantly improving scheduling efficiency. Furthermore, this structured instruction system enhances the determinism and controllability of intelligent management and scheduling methods based on multi-agent and retrieval-enhanced generation, avoiding the ambiguity and vagueness that may arise from natural language descriptions. This allows the entire closed-loop method of "conflict prediction - knowledge enhancement - decision optimization" to operate efficiently, significantly strengthening the method's real-time intervention capability and overall operational efficiency in complex dynamic environments.

[0071] like Figure 3This application also provides an electronic device, including a processor 301, a memory 302, a bus 303, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-mentioned intelligent management and scheduling method based on multi-agent and retrieval enhancement generation.

[0072] Furthermore, as an executable solution, the aforementioned heterogeneous embodied intelligent device collaborative management device based on digital twins can be a computer unit, which can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer unit may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described computer unit structure is merely an example and does not constitute a limitation on the computer unit. It may include more or fewer components, or combine certain components, or use different components. For example, the computer unit may also include input / output devices, network access devices, buses, etc., which are not limited in this embodiment of the invention.

[0073] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the computer unit, connecting various parts of the entire computer unit via various interfaces and lines.

[0074] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer unit by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0075] In some embodiments, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the methods described above in the embodiments of the present invention.

[0076] If the modules / units integrated in the computer unit are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0077] In some embodiments, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0078] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. An intelligent management and scheduling system based on multi-agent and retrieval-enhanced generation, characterized in that, include: The task agent layer includes multiple task agents, which are used to receive, analyze and initially process tasks, and generate an initial scheduling scheme based on a reinforcement learning model. The coordinating agent layer includes at least one coordinating agent, which is used to monitor the task execution status of the task agent layer, model the dependency and conflict relationships between tasks through graph neural networks, predict potential task conflicts, and generate and initiate optimization query requests for the conflict based on the predicted potential conflicts. The knowledge agent layer includes at least one knowledge agent, which integrates a retrieval enhancement generation module. This module is activated only when it receives an optimization query request initiated by the coordinating agent for a specific potential conflict, in order to retrieve an external knowledge base and generate an optimization scheduling strategy for the specific potential conflict. The communication module enables asynchronous communication between agents based on message queues. The retrieval enhancement generation module includes: The retrieval unit is used to convert query requests into semantic vectors and perform similarity searches in the knowledge base; The generation unit is used to concatenate the search results with real-time status data into prompt words, which are then input into a large language model to generate a structured scheduling scheme.

2. The system according to claim 1, characterized in that, The knowledge base is a dynamically updated vector database that stores at least one of the following: historical scheduling cases, industry standards, and equipment manuals.

3. The system according to claim 1, characterized in that, The coordinating agent uses a graph neural network (GNN) to model the task agent and its task relationships to detect at least one of path conflicts and resource preemption.

4. The system according to claim 1, characterized in that, The task agent, coordinating agent, and knowledge agent communicate with each other via JSON format messages, which contain at least the task ID, status code, and query command.

5. The system according to claim 1, characterized in that, The system also supports deployment in at least one of the following scenarios: smart campus, smart warehousing, and smart retail, and achieves adaptation by defining scenario-specific intelligent agents and knowledge bases.

6. An intelligent management and scheduling method based on multi-agent and retrieval-enhanced generation, characterized in that, Includes the following steps: Step S1: The task agent receives the task request, breaks down the task, and generates an initial scheduling scheme based on the reinforcement learning model. Step S2: The coordinating agent monitors the task execution status, detects conflicts through a graph neural network, predicts potential task conflicts, and generates and sends an optimization query request to the knowledge agent based on the predicted potential conflicts. Step S3: In response to the optimization query request, the knowledge agent calls the retrieval enhancement generation module to retrieve the knowledge base and generate an optimization strategy; Step S4: Coordinate the intelligent agents to integrate and optimize strategies, and issue update instructions to the task intelligent agents; Step S5: The task agent executes the updated plan and feeds back the execution results to the system, forming a closed-loop control.

7. The method according to claim 6, characterized in that, The retrieval process of the retrieval enhancement generation module described in step S3 uses Sentence-BERT for semantic encoding, and the generation process uses the LLaMA-2 large language model.

8. The method according to claim 6, characterized in that, The optimization strategy includes at least one of equipment scheduling instructions, path adjustment instructions, and resource allocation instructions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 6 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 6 to 8.

Images