Multi-robot chemical synthesis real-time scheduling control method based on MAS

Abstract

The invention relates to a multi-robot chemical synthesis real-time scheduling control method based on MAS, and the method enables chemical synthesis resources to be mapped into intelligent agents with different functional attributes, and carries out the deep improvement through employing an MAS optimization scheduling mechanism. By carefully constructing an MAS real-time scheduling architecture model, optimizing a CNP negotiation mechanism and designing an efficient real-time scheduling method, efficient collaborative scheduling of diversified small-batch reaction tasks, multi-type robot workstations and multiple mobile robots is realized. Therefore, the overall efficiency and quality of chemical synthesis are remarkably improved, and the cost and the safety risk are reduced.

Description

Technical Field

[0001] The present application relates to the field of process industry control technology, and in particular to a multi-robot chemical synthesis real-time scheduling control method based on MAS. Background Art

[0002] With the widespread application of automation and machine learning algorithms in the field of chemistry, laboratory synthetic chemistry has begun to adopt intelligent unmanned systems from material discovery to synthesis planning. The system consists of virtual screening, automatic experiments and data feedback, and is completely executed by algorithms and robots without human intervention. This change not only reduces costs and risks, but also frees researchers from repetitive work and allows them to focus on more innovative research. Although some AI-based chemical skill systems or robots have been launched, it is still challenging to achieve comprehensive scientific research automation.

[0003] In 2022, the University of Science and Technology of China successfully developed a full-process machine chemist, initially realizing a new paradigm of intelligent chemistry. This full-process machine chemist can automatically read literature in the cloud database, propose experimental plans, and complete the entire experimental process on 14 workstations. Its computational brain uses machine learning and Bayesian optimization to analyze data and provide new hypotheses for subsequent research; although intelligent chemical synthesis systems have made significant progress, they still face the challenge of designing efficient, stable and scalable multi-agent systems in multi-robot collaborative scenarios to achieve real-time scheduling and control of chemical synthesis. MAS needs to consider the communication, collaboration and decision-making mechanisms between agents to ensure close cooperation to complete complex experimental steps, and process data in real time and dynamically adjust task allocation. Due to the existence of parallel tasks and dynamic environmental factors in multi-robot systems, traditional scheduling methods are difficult to adapt, and there are problems such as simple mechanisms, insufficient efficiency, high costs and high safety risks. Summary of the invention

[0004] Based on this, it is necessary to provide a MAS-based multi-robot chemical synthesis real-time scheduling control method to address the above technical problems.

[0005] A multi-robot chemical synthesis real-time scheduling control method based on MAS, the method comprising:

[0006] Build the JADE middle layer; the JADE middle layer is used to provide a JADE operating environment to realize the registration, management and interaction of intelligent agents. Each running instance in the JADE operating environment is a container, and each container contains multiple intelligent agents. All active container sets constitute a platform. There is a main container in the platform, which includes the intelligent agent management system AMS, the directory service provider DF and several intelligent agents. All other containers in the platform are registered in the main container immediately after startup.

[0007] The chemical synthesis resources of the multi-robot autonomous chemical synthesis process are mapped into several agents with different functional attributes, and all agents form a multi-agent layer;

[0008] Based on the real-time data of the MAS platform, the PCA real-time scheduling method based on the relay collaboration model of the improved process relationship matrix and the reaction transfer collaborative agent real-time scheduling method based on bargaining game are adopted to realize the scheduling and control of autonomous chemical synthesis, so that the system can respond quickly and adapt to different chemical reaction requirements.

[0009] The above-mentioned MAS-based multi-robot chemical synthesis real-time scheduling control method maps chemical synthesis resources into intelligent entities with different functional attributes, and uses the MAS optimization scheduling mechanism for in-depth improvement. By carefully constructing the MAS real-time scheduling architecture model, optimizing the CNP negotiation mechanism, and designing an efficient real-time scheduling method, efficient collaborative scheduling of diversified small-batch reaction tasks, multi-type robot workstations, and multiple mobile robots is achieved. In this way, the overall efficiency and quality of chemical synthesis are significantly improved, and costs and safety risks are reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of a flow chart of a real-time scheduling control method for multi-robot chemical synthesis based on MAS in one embodiment;

[0011] Figure 2 A schematic diagram of the core architecture of JADE in an embodiment;

[0012] Figure 3 A schematic diagram of the overall architecture of a MAS based on JADE in an embodiment;

[0013] Figure 4 A schematic diagram of a physical agent model in another embodiment;

[0014] Figure 5 A behavior diagram of a reaction processing agent in another embodiment;

[0015] Figure 6 A behavior diagram of a reaction transfer agent in another embodiment;

[0016] Figure 7 A behavior diagram of a reaction baton agent in another embodiment;

[0017] Figure 8 A behavior diagram of a collaborative agent for reaction processing in another embodiment;

[0018] Fig. 9 A behavior diagram of a reaction-transfer collaborative agent in another embodiment;

[0019] Fig.10A behavior diagram of a reaction monitoring agent in another embodiment;

[0020] Fig.11 is a diagram showing the operation process of reaction R in another embodiment;

[0021] Fig.12 is a flow chart of F-CNP in another embodiment;

[0022] Fig.13 Improved CNP flow chart in another embodiment;

[0023] Fig.14 A game tree representation of a three-player game in another embodiment;

[0024] Fig.15 Another embodiment is a three-person bargaining game problem;

[0025] Fig.16 The figure is a flowchart of a bargaining game real-time scheduling algorithm in another embodiment. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0027] In one embodiment, Figure 1 As shown, a multi-robot chemical synthesis real-time scheduling control method based on MAS is provided, and the method comprises the following steps:

[0028] Step 100: Build the JADE middle layer; the JADE middle layer is used to provide a JADE operating environment to realize the registration, management and interaction of intelligent agents. Each running instance in the JADE operating environment is a container, and each container contains multiple intelligent agents; all active container sets constitute a platform; there is a main container in the platform, which includes the intelligent agent management system AMS, the directory service provider DF and several intelligent agents; all other containers in the platform are registered in the main container immediately after startup.

[0029] Specifically, the implementation technology of MAS involves selecting programming languages ​​and frameworks, and establishing efficient communication protocols to ensure smooth collaboration between agents. The design considers collaborative algorithms, negotiation mechanisms, and task allocation to enhance the collaborative efficiency between agents. The integration of intelligent decision-making and learning capabilities enables agents to better adapt to changing environments. While ensuring scalability and flexibility, the system architecture optimizes the dynamic management, resource allocation, and conflict resolution of new agents. These comprehensive technologies together create the excellent adaptability and collaboration capabilities of MAS.

[0030] JADE is a pure Java-implemented intelligent agent development framework that runs on a Java virtual machine, taking advantage of Java's cross-platform nature. The framework follows the FIPA specification, aims to simplify the construction of multi-agent systems, and provides debugging and deployment tool support. FIPA is an international organization that promotes interoperability of intelligent agents. JADE supports the distribution of the agent platform on multiple machines and provides a remote graphical user interface for configuration control. Its communication architecture is efficient and flexible, supporting multiple agent queue access modes. JADE has been widely used in multiple companies and academic fields, demonstrating excellent scalability and flexibility.

[0031] The core architecture of JADE is as follows Figure 2 As shown in the figure, the basic component unit is an independent entity called an agent, which performs tasks through message interaction. These agents run on a platform that provides basic messaging services. The platform consists of containers that can be distributed on different hosts to achieve distributed deployment. Each container can carry multiple agents, among which the main container has special responsibilities: it not only serves as the initial container of the platform and is responsible for the registration of other containers, but also has two core agents built in, AMS (Agent Management System) and DF (Directory Facilitator). AMS is the management center of the platform, which is responsible for the startup, termination and shutdown of the agent as a whole. DF is similar to the yellow pages service, helping agents publish and find the required services.

[0032] Starting a new main container is equivalent to building a new platform. The communication between agents is transparent, and communication is achieved through asynchronous message passing that follows the Agent Interaction Language (ACL) specification established by FIFA. JADE supports distributed platforms, containers can run across hosts, dynamically manage agent registration and deregistration, and ensure cross-container and cross-platform communication. It is highly flexible and scalable. Therefore, JADE has been widely used in many fields such as industrial manufacturing, communications, and medical care, with strong practicality and broad prospects.

[0033] Whether the agents are located on the same or different platforms or containers, they can communicate seamlessly. Communication uses asynchronous message passing that follows the ACL specification, including sender, receiver, communication behavior and content. For example, INFORM messages are used to notify facts, and REQUEST messages are used to request actions. The 22 communication behaviors defined by FIPA enrich the interaction between agents and improve accuracy and flexibility.

[0034] JADE provides a series of convenient features for agent developers, which helps to create and manage multi-agent systems more easily. It provides a complete agent development kit, allowing developers to build systems according to actual needs at a higher level of abstraction. This not only reduces the difficulty of implementation and improves development efficiency, but also ensures the high quality of multi-agent systems.

[0035] The core features of JADE include the following: (1) Distributed Agent Platform: It supports the distribution of Agent platform on multiple hosts. Each host runs an independent Java virtual machine, so that Agent runs in the form of Java thread. This decentralization ensures the independent operation of Agent on different hosts, bringing diversity to the execution environment. (2) Efficient message processing: In JADE, messages exist in the form of Java objects, which simplifies the interaction between Agents. Its intelligent message conversion mechanism can automatically convert messages into ACL format as needed, providing flexibility and diversity for Agent communication. (3) Rich FIPA interaction protocol library: JADE provides FIPA interaction protocol library, allowing developers to flexibly select appropriate interaction protocols according to specific needs, enhancing the customizability and adaptability of the system. (4) Comprehensive Agent management platform: The platform is equipped with management services such as AMS, DF, ACL, etc., supporting Agent lifecycle management, directory services and communication channels. This integrated management platform ensures the smooth operation and efficient collaboration of multi-agent systems. (5) Convenient monitoring and debugging tools: Through the intuitive graphical interface, developers can view and edit Agent message content in real time, providing convenience for the development and debugging process. (6) User-friendly platform management GUI: Developers can easily manage Agents and containers through the graphical user interface of the remote host, perform operations such as creation, destruction, suspension and restart, and realize intuitive monitoring and adjustment of the operation status of the entire Agent system. (7) Support for multi-behavior and multi-threaded execution: JADE allows each Agent to execute multiple behaviors at the same time and run in parallel in a multi-threaded environment. This design enhances the diversity and flexibility of Agent behaviors and the concurrency of the system when handling multiple tasks.

[0036] These features together form a powerful and flexible development framework, simplifying the creation and management process of multi-agent systems and improving development efficiency and system quality.

[0037] Step 102: The chemical synthesis resources of the multi-robot autonomous chemical synthesis process are mapped into a number of intelligent agents with different functional attributes, and all the intelligent agents form a multi-agent layer.

[0038] Specifically, the MARS strategy uses multi-agent technology to enable machines to feedback status information to the system and request tasks in real time. Through continuous task-machine interaction, it ensures that tasks can be efficiently assigned to the best-performing machine. Since the MARS strategy is completely driven by real-time information, it can significantly reduce the planning and execution deviations in traditional strategies and improve overall scheduling efficiency. The common practice for interconnecting manufacturing equipment with other units in complex manufacturing control systems is to build software-defined models of physical equipment. These models can interact with the corresponding manufacturing equipment in real time to ensure smooth connections between devices.

[0039] Based on the MARS strategy, the overall architecture of real-time scheduling control of the JADE-based multi-robot chemical synthesis experimental platform MAS was designed. Figure 3 As shown. It consists of two layers: the JADE middle layer and the multi-agent layer. The JADE middle layer provides a JADE runtime environment that enables agent registration, management, and interaction. Each running instance in the JADE runtime environment is called a container, which can contain multiple agents. The set of active containers is called a platform. There is a special container in the platform that is always active, called the main container. All other containers are registered to the main container immediately after startup. Once the platform is activated, the JADE default agents, including the agent management system AMS and the directory service provider DF, will be instantiated. The AMS agent acts as a supervisor to control the use of the platform by other agents; while the DF agent provides the default yellow pages service on the platform. In this study, there is only one main container on the platform, which acts as the execution environment of the agent.

[0040] In the MAS system of multi-robot autonomous chemical synthesis, chemical synthesis resources are mapped to agents with different functional attributes. These agents constitute the multi-agent layer, which realizes real-time scheduling and control based on the real-time data of the laboratory platform. Specifically, the agent layer covers reaction processing agent (RPA), reaction transfer agent (RTA), reaction baton agent (RBA), reaction processing collaboration agent (PCA), reaction transfer collaboration agent (TCA) and reaction monitoring agent (RMA).

[0041] Different agents play unique roles in MAS. Physical agents are responsible for the execution phase of chemical synthesis, such as the chemical robot workstation RPA, which is responsible for processing reaction operations, and the mobile robot RTA, which is responsible for transferring between reaction operations. Service agents provide global optimization support for physical agents, such as RBA, which replaces traditional IoT tracking methods by generating reaction batons. PCA and TCA coordinate the scheduling and processing of reaction processes, while RMA is responsible for monitoring the reaction process and providing real-time information support. The entire multi-agent layer is designed to achieve efficient scheduling and control of autonomous chemical synthesis through collaborative work. This layer is based on real-time data from the MAS platform, enabling the system to respond quickly and adapt to different chemical reaction requirements. Each type of agent can be implemented in the JADE platform. The main functions of these agents are described below.

[0042] (1)RPA and RTA are responsible for capturing real-time data from automatic identification devices such as RFID and processing complex real-time data so that it can be understood as meaningful synthetic information. Through this process, the relevant intelligent agents can obtain the status and available capacity of synthetic resources in real time. This real-time data processing method helps ensure that the intelligent agent system has instant understanding of the synthesis process, thereby improving the system's management efficiency of resources and tasks.

[0043] (2) RBA is used to capture the information of all tasks and generate corresponding batons. The baton is an innovative process relationship matrix that is specifically used to record entity labels and related process information. Through the generation of the baton, the effective capture and transmission of task information is achieved, providing a more efficient way for task coordination and collaboration in multi-agent systems. The introduction of this new process relationship matrix aims to improve the efficiency of information management and task allocation in the system.

[0044] (3) PCA and TCA are responsible for providing negotiation scheduling algorithms for RPA and RTA respectively. By perceiving real-time reaction information, they optimally schedule the start and completion time of each operation of each task. The two collaborative agents work together to help achieve efficient scheduling of reaction tasks and ensure the smooth progress of the entire synthesis process. This collaboration-based scheduling method plays an important role in improving synthesis efficiency and resource utilization.

[0045] (4) RMA is responsible for real-time capture and processing of reaction execution information, and responding to related synthesis information queries. During the synthesis process, RMA timely tracks and traces disturbances and changes in the reaction process. By monitoring the reaction process, it ensures the smooth progress of the reaction and can respond to any potential problems and changes in a timely manner, thereby improving the reliability and efficiency of the synthesis.

[0046] Step 104: Based on the real-time data of the MAS platform, the PCA real-time scheduling method based on the relay collaboration model of the improved process relationship matrix and the TCA real-time scheduling method based on the bargaining game are adopted to realize the scheduling and control of autonomous chemical synthesis, so that the system can quickly respond to and adapt to different chemical reaction requirements.

[0047] Specifically, PCA focuses on the real-time scheduling of RPA robot workstations, continuously receives the real-time capability data of RPA, the available operations of the baton, and the real-time synthesis information of RMA, and then generates the task processing queue. The PCA real-time scheduling method based on the relay cooperation model of the improved process relationship matrix uses the relay cooperation model of the process relationship matrix and the improved CNP to optimize the collaboration and information transmission between components, thereby achieving an efficient negotiation solution.

[0048] Aiming at the optimization problems of RTA in terms of task completion time, key machine load and energy consumption, a TCA real-time scheduling method based on bargaining game is proposed by integrating the concepts of dynamic game and bargaining game. Firstly, a bargaining game model of MOP is constructed, and the RTA path planning and movement sequence of each game stage are optimized by using CA* algorithm, a dynamic game tree is generated, and SPNE is solved by reverse induction. Subsequently, the bargaining mechanism for maximizing joint benefits on the game tree is further refined, thereby improving the real-time scheduling strategy and equilibrium solution solution method.

[0049] In the above-mentioned MAS-based multi-robot chemical synthesis real-time scheduling control method, the method maps chemical synthesis resources into intelligent entities with different functional attributes, and uses the MAS optimization scheduling mechanism for in-depth improvement. By carefully constructing the MAS real-time scheduling architecture model, optimizing the CNP negotiation mechanism, and designing an efficient real-time scheduling method, efficient collaborative scheduling of diversified small-batch reaction tasks, multi-type robot workstations, and multiple mobile robots is achieved. In this way, the overall efficiency and quality of chemical synthesis are significantly improved, and costs and safety risks are reduced.

[0050] In one embodiment, the intelligent agents in step 100 include: entity agents and service agents; entity agents include: reaction processing agents and reaction processing collaborative agents; service agents include: reaction baton agents, reaction processing collaborative agents, reaction transfer collaborative agents and reaction monitoring agents; reaction processing agents and reaction transfer agents are used to capture real-time data from automatic identification devices and process complex real-time data so that it can be understood as meaningful synthetic information; reaction baton agents are used to capture information of all tasks and generate corresponding batons; reaction processing agents and reaction transfer agents are used to provide negotiation scheduling algorithms for reaction processing agents and reaction transfer agents, and by sensing real-time reaction information, optimally schedule the start time and completion time of each operation of each character; reaction monitoring agents are used to capture and process reaction execution information in real time, and respond to related synthetic information queries.

[0051] Specifically, (1) Response Processing Agent RPA

[0052] In traditional manufacturing workshops, due to the existence of heterogeneous hierarchical control systems, processing entities cannot be effectively interconnected. With the continuous development of technology, the combination of IoT manufacturing technology and MAS provides an effective solution to overcome this problem, enabling processing entities to work together and share information more flexibly. As mentioned above, the physical agent RPA and RTA are software-defined manufacturing devices with basic functions such as perception, interaction, analysis and execution. It can sense its status, interact with other agents in MAS, and process tasks in a timely manner according to the real-time status. The physical agent hardware is an embedded computer equipped with an operating system and has various physical interfaces. At the software level, it includes an adaptation layer, an intelligent analysis layer, and an information development layer based on JADE.

[0053] Entity-agent models such as Figure 4 As shown in the figure, the model integrates physical entities such as chemical workstations and mobile robots, and realizes seamless connection and real-time information interaction between entities through a unified software package, ensuring efficient scheduling and exception handling of the RMA platform, while optimizing communication between units. In order to cope with complex functional requirements, the intelligent agent adopts a multi-threaded architecture to collaboratively execute core tasks such as control, monitoring, interaction and decision-making.

[0054] The core functions of the adaptation layer are to realize device interconnection, control underlying devices, action-driven operation and information acquisition, solve the incompatibility and communication format problems of heterogeneous systems, promote the collaborative operation of the main control system and manufacturing equipment modules, improve production efficiency, and promote the process of digitalization and intelligence. The intelligent decision-making layer is the core of the intelligent individual, using edge computing to analyze equipment and environmental information and generate extensible markup language reports. Its functions include real-time status analysis, parameter adjustment, task selection, and monitoring of intelligent agents and workshop information to decide whether to accept the task. After reaching an agreement, the adaptation layer is driven to operate, and different devices have different mechanisms. The information development layer based on JADE is the hub of the self-organizing Internet of Things multi-robot chemical synthesis, which ensures smooth interaction between intelligent agents. By formulating message exchange guidelines, intelligent agents can accurately transmit information. This layer is stably encapsulated in each intelligent agent to establish connections between intelligent agents.

[0055] RPA is an intelligent agent designed for chemical robot workstations. It is responsible for performing chemical reaction operations, real-time data capture and complex processing, converting reaction data into valuable information, and transmitting it to the information management platform RMA in real time. With the message transmission mechanism provided by JADE, RPA can autonomously publish its functions and status in the multi-agent system and achieve efficient communication and collaboration with other intelligent agents.

[0056] RPA behavior design Figure 5 As shown in the figure. After startup, RPA first initializes its behavior. When idle, RPA triggers the Yellow Pages registration service behavior; when the state changes, it triggers the real-time state behavior, including handling machine failures and responding to operation completion. There are three states in each loop step: idle state, waiting state, and working state.

[0057] 1) In idle state, RPA automatically registers the corresponding yellow page services, including the yellow page services corresponding to the executable response operation types and the yellow page services of the machine in normal state

[0058] 2) In the available state, RPA will respond to the negotiation behavior of PCA until it successfully matches the reaction operation task. Once the match is successful, RPA waits to receive the baton behavior and switches to the waiting state; if the negotiation fails, it enters the next round of loop steps. In the waiting state, RPA continues to wait until it receives the baton passed by RTA, and then enters the working state.

[0059] 3) In the working state, RPA will adjust its own state according to the completion of the task. If the task is successfully completed, RPA will pass the baton to start the next operation, and then turn to the idle state, waiting to participate in the next round of response negotiation. If the task is not completed, RPA will continue to work and enter the next cycle step. When the chemical workstation machine fails, RPA will immediately cancel the corresponding yellow page service, and after receiving the baton, send the abnormal baton to RMA for timely processing.

[0060] This state flow design enables RPA to dynamically adjust its state according to real-time needs and negotiate with other agents, so as to adapt to complex working environments more flexibly. The entire process enables RPA to make intelligent and coordinated responses in different situations, improving the system's adaptability and overall efficiency.

[0061] (2) Response Transfer Agent RTA

[0062] RTA is the intelligent agent representative of the composite mobile robot. Its main task is to efficiently transfer the reagent tray to the chemical workstation to ensure a smooth transition of the reaction operation. Similar to RPA, RTA, as another physical agent, has the ability to capture real-time data and perform complex processing, converting this data into meaningful reaction processing information. Through the Yellow Pages service, RTA can autonomously publish its functions and status, allowing other agents to easily find, understand and communicate with it effectively. This agent design enables RTA to play a key role in collaborative work and automation of laboratory operations. Together with RPA, it builds an efficient physical multi-agent system, which promotes the intelligent and collaborative development of chemical laboratories. RTA's behavioral design is as follows: Figure 6 As shown. The differences from RPA are:

[0063] 1) Based on the RPA behavior design, RTA omits the waiting state. This is because when RTA successfully receives the baton behavior, it will immediately switch to the working state according to the baton information and quickly transfer the reagent tray to the corresponding chemical workstation.

[0064] 2) In addition to the task negotiation behavior, the negotiation behavior of the cyclic walking behavior also includes obstacle avoidance behavior.

[0065] It is used to prevent the idle composite mobile robot from blocking the RTA that is performing the mobile task. At the same time, the RTA work behavior also includes searching for idle RTA yellow pages services and requesting obstacle avoidance behavior to quickly transport the reactants to the designated location.

[0066] These differences further emphasize the specific role of RTA in mobile robotic systems and its unique contribution to chemical laboratory automation.

[0067] (3) Reaction Baton Agent (RBA)

[0068] In the reaction process, the baton serves as a key medium to efficiently carry the process relationship matrix within the MAS and update the process information in real time. At the end of the process, the RTA sends the reagent tray to the recycling area, and the baton is handed over to the RBA for management. This fully records and effectively controls the entire processing process, ensuring the efficient operation of the system and the accuracy of information.

[0069] RBA Figure 7 Its behavioral design and two core behaviors are fully demonstrated in the . After completing the initialization, RBA switches to the mode of receiving and processing reaction tasks in real time, and performs the behaviors of receiving and sending batons at the same time. Once a new task arrives, RBA can quickly and accurately receive it, and generate a precisely matched reaction baton based on the task details, and send it to PCA for further processing. To ensure the stability of the system and its excellent execution and adaptability, a number of tests were carefully planned and implemented in the cyclic walking behavior stage. These tests specifically include the introduction of single-order reactions at critical moments and the rapid generation of corresponding batons to fully verify the performance and reliability of the system.

[0070] (4) Processing collaborative agent PCA

[0071] When PCA successfully receives the reaction baton from other agents in MAS, it takes on the important task of determining all operations in the next stage. By carefully scanning the reaction baton, PCA can fully understand the upcoming tasks and use the yellow page service to extensively search for RPAs that are capable of these operations. Based on the improved contract network protocol, PCA constructs a relay cooperation model, which is based on the positive process relationship matrix table and integrates two modules: problem formulation and solution. In order to make up for the limitations of RPA greedy bidding and short-sighted decision-making, an initial solution generation strategy that considers load balancing is adopted in the solution module. This model can perceive the information of MAS in real time, calculate and screen the best bidding set, and achieve the optimal scheduling of tasks under the global load balancing of RPA.

[0072] To improve computing and communication efficiency, PCA is designed as a service agent independent of RPA. It can perform distributed scheduling optimization based on the real-time status of the entity agent, which is more flexible and efficient than the traditional dynamic scheduling system. PCA works closely with RPA to efficiently arrange reaction tasks and ensure the smoothness of the chemical reaction synthesis process.

[0073] Figure 8The behavioral design of PCA is depicted, including its core responsibilities and behavioral logic throughout the scheduling process. After startup, PCA is first initialized, and then receives the response baton in real time to comprehensively collect the required operation information. In each cycle, PCA uses the information in the baton to quickly find the corresponding RPA in the Yellow Pages service, request its status information from RMA and initiate a tender. After negotiation and scheduling, when the task is successfully assigned, the baton carries the complete scheduling information and passes it to TCA for subsequent processing. At the same time, PCA returns to standby status. This design not only ensures the efficient collaboration between PCA and RPA, but also ensures the smooth operation of the entire system through real-time interaction between PCA and RMA.

[0074] (5) Transferring Collaborative Agents (TCA)

[0075] After receiving the transfer request of the reaction baton, TCA will immediately interpret the information and use the yellow page service to search for currently idle RTAs based on the pickup and delivery locations required for all current baton operations. This step ensures that TCA has a comprehensive understanding of the reaction tasks that need to be transferred and can quickly find a mobile robot that can perform the transfer task.

[0076] Under the framework of MAS's contract network negotiation agreement, TCA will initiate a tender to the RTA that can perform the transfer task and collect relevant tender information in detail. To ensure optimal scheduling, TCA uses an embedded bargaining game algorithm to handle multi-objective optimization problems. Based on key factors such as task completion time, RTA load and energy consumption, TCA formulates the best task scheduling and path planning for RTA to achieve efficient and balanced operation of the system.

[0077] Fig. 9 The behavior design process of TCA is shown. Once TCA is generated and initialized, it triggers the behavior of receiving the reaction baton in real time to receive the task to be transferred. In each loop step, TCA determines the RTA that can perform the transfer task by searching the Yellow Pages service and sends a tender request to it to obtain real-time information. Unlike the negotiation scheduling of PCA, TCA directly applies the operation information in the real-time received baton behavior to its negotiation scheduling process. During this process, TCA matches the task with the RTA, plans the path in sequence, and prioritizes the baton information with interference. After the negotiation scheduling is completed, TCA triggers the behavior of sending the baton in real time to transfer the baton to the corresponding RTA. The RTA that receives the baton will perform the corresponding transfer operation according to the information in it. Subsequently, TCA waits for the next instruction to coordinate and promote the system to achieve efficient operation in the complex reaction transfer process.

[0078] TCA and PCA together constitute the key components that work together in the system, promoting the coordination and collaborative development of the entire MAS.

[0079] (6) Response Monitoring Agent RMA

[0080] In the MAS-based multi-robot chemical synthesis system, a single agent is limited by local information and limited knowledge and is unable to cope with complex tasks alone. Therefore, the collaboration between agents becomes crucial and is the core of completing reaction tasks and dealing with abnormal situations. However, when dealing with interference, distributed agents often fall into local optimality due to insufficient utilization of global information. To compensate for this defect, the system introduces the RMA mechanism to enhance the utilization of global information. Fig.10 Design details for specific actions regarding RMA.

[0081] RMA is generated immediately after MAS is started. Its core function is to receive and process the status information of the agent and the reaction baton process information in real time, and provide instant response queries. In addition, RMA is also equipped with an abnormal event handling mechanism to specifically deal with emergencies in the agent status and baton process, ensuring that the system can be restored in time and maintain stable operation.

[0082] When a chemical robot workstation fails, RMA intervenes quickly, regardless of whether the RPA has received the task baton. It first deregisters the RPA's online status and widely disseminates this update through the Yellow Pages service. If the failure occurs during the execution of a task by the RPA, RMA will reset the task flow, regenerate the baton, and reallocate the task within the MAS. If the RPA fails while on standby, RMA will immediately take over the failed baton information to ensure that the current task can be quickly and seamlessly transferred to other available RPAs for continued execution. In addition, to meet the needs of emergency insertion tasks, RMA will immediately issue a broadcast notification, and PCA will give priority to matching and processing such urgent matters for RPA.

[0083] In one of the embodiments, the reaction baton carries an optimized reaction process relationship matrix to record the core information of the chemical reaction synthesis in real time; the core information includes at least: raw material consumption, processing details, and process information; the reaction process relationship matrix is ​​constructed according to the process characteristics in order to express the dependency relationship between the steps in the chemical reaction process; the element value of the reaction process relationship matrix is ​​Oij, and when Oij is a positive integer, it means that operation i is the predecessor operation of operation j, and the smaller the value, the higher the priority; if Oij is 0, there is no precedence relationship between the two.

[0084] The reaction baton-based transfer collaboration process specifically includes: after MAS receives the order, the reaction baton agent RBA generates a new reaction baton and passes it to the reaction processing collaborative agent PCA, thereby starting the entire task process; the reaction processing collaborative agent PCA negotiates with the reaction processing agent RPA to select the best processing workstation based on the requirements and process standards of the chemical reaction; after selecting the workstation, the reaction processing collaborative agent PCA hands the reaction baton to the reaction transfer collaborative agent TCA, and the reaction transfer collaborative agent TCA is responsible for coordinating the reaction transfer agent RTA to ensure that the reagent tray can be accurately and efficiently transferred to the selected reaction processing agent RPA; the reaction processing agent RPA starts to execute the chemical reaction operation after receiving the reaction baton, and returns the reaction baton to the reaction processing collaborative agent PCA after completion for subsequent operation or monitoring; through continuous negotiation, transfer and execution of operations, until all chemical reaction operations are completed, it finally marks the end of the entire task.

[0085] Specifically, with the assistance of service agents, MAS has achieved comprehensive optimization of functions. RBA uses a reaction baton to replace traditional tracking, greatly improving monitoring and control efficiency; PCA accurately dispatches RPA to ensure efficient execution and resource optimization; TCA focuses on coordinating RTA to ensure smooth operation and system coordination; RMA monitors the reaction process in real time and provides comprehensive real-time information support. The close coordination of these service agents significantly improves the overall performance of the system, allowing physical agents to work together more intelligently and efficiently in complex reactions, thereby improving synthesis efficiency and controllability.

[0086] In the research field of manufacturing MAS, previous studies are usually based on a premise: there are intelligent terminals in the manufacturing environment, such as machine agents or task agents. However, the rationality of these assumptions has not been fully verified. In fact, in order to ensure that each manufacturing resource can effectively participate in the negotiation process of task allocation, each resource needs to be encapsulated as an independent agent. With the continuous expansion of manufacturing entities and task scales, existing task agents face severe challenges in terms of computational complexity and communication load. The core of the Internet of Things lies in interconnection, while the tracking and display of manufacturing process information mainly relies on RFID. This patent innovatively proposes the concept of "baton", which carries an optimized reaction process relationship matrix and records entity tags and process information in real time. Through the "baton" transmission between multiple agents, this patent abandons the traditional RFID information reading and writing method and constructs a relay cooperation model based on positive process relationship improvement.

[0087] In order to clearly express the dependencies between the steps in the chemical reaction process, it is necessary to construct a positive process relationship matrix based on the process characteristics. When processing tasks, a positive operation relationship matrix is ​​constructed, and its element value is defined as ij . When Oij When it is a positive integer, it means that operation i is the previous operation of operation j, and the smaller the value, the higher the priority; if O ij If it is 0, there is no precedence relationship between the two.

[0088] In order to intuitively display the operational relationship of a certain reaction R, the corresponding directed graph and matrix table (such as Fig.11 and Table 1). After analysis, the process sequence is: O1, O5, O2, O6, O3, O7, O8, O9, O10, O4. This sequence provides a strong basis for optimizing the reaction process and improving the synthesis efficiency.

[0089] Although the task agent may have a certain impact on the system performance, the lack of an independent management unit is not a good solution. Therefore, the system adopts an improved relay cooperative control method to manage tasks and chemical reactions more effectively. The reaction baton, which transmits state information, records the core information of chemical reaction synthesis, such as raw material consumption, processing details, process information, etc. This information is organized into an n-dimensional matrix and uniformly managed and generated by RBA.

[0090] Table 1 Reaction R operation process relationship matrix

[0091] Reaction Operation O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 O1 0 1 0 0 0 0 0 0 0 0 O2 0 0 2 0 0 0 0 0 0 0 O3 0 0 0 3 0 0 0 0 0 0 O4 0 0 0 0 0 0 0 0 0 0 O5 0 5 0 0 0 0 0 0 0 0 O6 0 0 6 0 0 0 0 0 0 0 O7 0 0 0 7 0 0 0 0 0 0 O8 0 0 0 0 0 0 0 0 8 0 O9 0 0 0 0 0 0 0 0 0 9 O10 0 0 0 10 0 0 0 0 0 0

[0092] By introducing the reaction baton as a relay to unify the key information of task and reaction, the overall coordination and efficiency of the system are improved. Although this method may have a slight impact on performance, reasonable baton management enables the system to execute tasks and reaction processes more orderly and achieve more accurate control and recording. This method not only ensures the rationality of task management, but also simplifies the complexity of the system to a certain extent.

[0093] In the reaction process, the baton serves as a key medium to efficiently carry the process relationship matrix within the MAS and update the process information in real time. At the end of the process, the RTA sends the reagent tray to the recycling area, and the baton is handed over to the RBA for management. This fully records and effectively controls the entire processing process, ensuring the efficient operation of the system and the accuracy of information.

[0094] In a specific embodiment, the normal reaction process can be achieved through interactive methods, such as process information sharing and collaboration between agents participating in task allocation negotiation. However, in a narrow knowledge domain, the decision-making of the agent is not accurate enough, so it is difficult to easily achieve the optimal scheduling result. The functional and service agents in the MAS work together to achieve real-time scheduling under limited synthetic resource conditions. In addition, when an abnormal event occurs, the MAS can accurately identify the type of disturbance and handle it through real-time collaboration. The MAS uses a distributed control strategy to verify the negotiation results, thereby ensuring that the self-organizing scheduling results of the MAS develop in a positive direction during the task allocation reaction process. At each time t, the problem-solving process is described as follows.

[0095] Step A1: When a new task arrives, RBA generates a corresponding reaction relay baton, records the task number and process relationship information matrix, and realizes synchronous information tracking within MAS.

[0096] Step A2: PCA aims to achieve real-time scheduling of RPAs. At each time step t, its input includes real-time capability information from the current RPA, available operations from the baton received, and real-time reaction synthesis information from the RMA. The task of PCA is to generate a task queue for the robot workstation based on this information, and to reach a contract with the corresponding RPA to write the scheduling result information into the baton. Once PCA completes the scheduling, it hands over the scheduled reaction baton to TCA for subsequent processing.

[0097] Step A3: TCA aims to achieve real-time scheduling of RTA. At each time step t, the task of TCA is to generate the task queue of the mobile robot based on the current information and the baton information received from TCA, and to reach a contract with the corresponding RTA to write the scheduling result information into the baton. Once the PCA scheduling is completed, it will hand over the corresponding reaction baton to the RTA for transfer.

[0098] Step A4: RPA receives the baton transferred from RTA and verifies it with its own contract. After RPA completes the corresponding reaction operation, it hands the reaction baton that has completed the current operation back to PCA for processing. This cycle continues until all the reaction process operations recorded by the reaction baton are completed.

[0099] Step A5: By querying the global information of RMA, PCA and TCA negotiate and dispatch appropriate RPA and RTA to execute the current process according to optimization goals, such as completion time and load balancing. When handling abnormal events, RMA needs to correct and broadcast the internal information of the reaction baton, and then incorporate it into the corresponding process to achieve rapid processing.

[0100] In the real-time scheduling model of self-organizing MAS, both normal processing and exception processing adopt event-driven rescheduling. The implementation of real-time rescheduling can significantly reduce the impact on the unexecuted part of the initial scheduling result.

[0101] In one of the embodiments, the process of collaboration between the functions and service agents in the MAS specifically includes: when a new task arrives, the RBA generates a corresponding reaction relay baton, records the task number and process relationship information matrix, and realizes synchronous information tracking within the MAS; PCA is used to realize real-time scheduling of the RPA. At each time step t, its input includes real-time capability information from the current RPA, available operations from the received baton, and real-time reaction synthesis information from the RMA. The task of the PCA is to generate a task queue for the robot workstation based on this information, and to reach a contract with the corresponding RPA to write the scheduling result information into the baton. Once PCA has completed scheduling, it will hand over the scheduled reaction baton to TCA for subsequent processing; TCA is used to implement real-time scheduling of RTA; RPA receives the reaction baton transferred from RTA and verifies it with its own contract; after RPA performs the corresponding reaction operation, it hands over the reaction baton that has completed the current operation back to PCA for processing; this cycle continues until all reaction process operations recorded by the reaction baton are completed; by querying the global information of RMA, PCA and TCA negotiate and schedule appropriate RPA and RTA to execute the current process based on the optimization goals; when handling abnormal events, RMA corrects and broadcasts the internal information of the reaction baton, and then incorporates the corrected internal information of the reaction baton into the corresponding process to achieve rapid processing.

[0102] In one embodiment, the steps of the contract network protocol improved based on MAS include: the initiator issues a request for proposals, specifies the task content and execution conditions in detail, and waits for responses from other agents; each agent submits its own proposal based on the CFP, and the initiator then collects and evaluates these proposals; the initiator selects a suitable agent to perform the task based on the evaluation results, and sends an acceptance notification to the selected agent, and notifies the unselected agents at the same time; the selected agent performs the task according to its own proposal; after the task is completed, the agent reports the result to the initiator; if it fails to be completed, the initiator is notified in time for follow-up processing.

[0103] In one embodiment, the steps of the PCA real-time scheduling method of the relay cooperation model based on the improved process relationship matrix include: constructing a relay cooperation model based on the positive process relationship matrix table; the relay cooperation model includes a main objective function, a secondary objective function and constraints; the main objective function is:

[0104]

[0105] The secondary objective function is:

[0106]

[0107] C i =S i +t ik 1≤i≤n,1≤k≤p (3)

[0108] The constraints are:

[0109] min{S i}&min{F k} (4)

[0110] S i ≥S j +t jk , Y ik =1,Y jk =1 (5)

[0111] S i ≥max(C l ), X li =1 (6)

[0112] Among them, f and T are two minimization functions. Function f is obtained by finding C i (Operation i Completion time) and D i (Operation i The maximum value of the difference between the time span) is defined by minimizing the f function and D i The sum of the positive differences of , so that each operation is completed as close as possible to its corresponding time span; function T is the sum of the positive differences of operation O 1。。。 O n The maximum value of the completion time of T is to minimize the operation O 1。。。 O n completion time, so that each operation is completed as soon as possible; C i For operation O i The completion time, F k P k The idle time period, p is the total number of available RPAs, P k is the kth RPA among all available RPAs, 1≤k≤p, S i For operation O i The start processing time, Y jk For operation O i Processing on Pk, t ik Operation O for Pk i processing time, n is the total number of reaction operations, Di is the time span of operation Oi, X li For operation O l It is operation Oi The previous process.

[0113] According to the relay cooperation model based on the positive process relationship matrix, a CNP solution method considering load balancing is adopted to optimize the cooperation and information transmission between components, thereby achieving an efficient negotiation solution.

[0114] Specifically, formula (1) reflects the minimum completion time or completion time of a task. The goal of optimization is to shorten the completion time of the product as much as possible, which is the goal of comprehensive scheduling optimization. Specifically, formula (1) aims to minimize the completion time of any operation exceeding the sum of its Makespan, thereby promoting a schedule in which operations are completed as close to their corresponding Makespan as possible. Formula (2) shows that, under the constraints of formula (3) and formula (4), each operation can start processing as early as possible to shorten the static idle time on the equipment. This helps to improve resource utilization and overall efficiency. Formula (5) states that operations of the same entity must be processed in sequence, while formula (6) shows that each process can only start after the previous process is completed. These constraints help ensure that tasks are executed in the correct order and schedule, ensuring the orderliness and controllability of the production process.

[0115] The JADE platform supports basic information exchange and further integrates the enhanced FIPA Contract Network Interaction Protocol (F-CNP), which greatly facilitates developers to build customized information interaction protocols. Compared with the original contract network IP model, F-CNP innovatively adds rejection and confirmation communication links to enrich the interaction process. In this model, the initiator assumes management responsibilities, entrusts other intelligent agents to perform specific tasks, and is committed to optimizing key performance indicators such as price, completion time, and fairness of task allocation. Each task attracts multiple participants to actively submit suggestions, and those who do not submit must be explicitly rejected. Subsequently, the manager conducts detailed negotiations with the participants who have submitted suggestions to ensure efficient execution of the task. The F-CNP flow chart is shown below. Fig.12 shown.

[0116] Fig.12 The whole picture of F-CNP is clearly presented. Its core steps can be summarized as: initiation and preparation, submission and evaluation, selection and notification, task execution, completion and feedback.

[0117] (1) Initiation and preparation: The initiator publishes a request for proposals (CFP), specifies the task content and execution conditions, and waits for responses from other agents.

[0118] (2) Submission and evaluation: Agents submit their proposals based on the CFP, and the initiator then collects and evaluates these proposals.

[0119] (3) Selection and notification: The initiator selects appropriate agents to perform the task based on the evaluation results and sends them acceptance notifications, while also notifying the unselected agents.

[0120] (4) Task execution: The selected agents perform tasks according to their proposals and try their best to complete them according to the promised conditions.

[0121] (5) Completion and feedback: After the task is completed, the agent reports the result to the initiator; if it fails to complete the task, the agent promptly notifies the initiator for follow-up processing.

[0122] In order to overcome the defects of classic CNP in extensive bidding, heavy communication load between tasks and limited local optimization capabilities, this patent introduces the yellow page service, sets a bidding time limit, and adopts a concurrent mechanism for multi-task bidding. These measures effectively avoid blind bidding, strengthen the linkage between tasks, and improve the scheduling efficiency of CNP across multiple steps of optimization. After implementing these improvements on the basis of the F-CNP framework, taking PCA negotiation scheduling as an example, the improved CNP negotiation process is as follows Fig.13 As displayed.

[0123] By incorporating yellow page services, bidding time limits, and multi-task concurrent bidding strategies, and supplemented by an event-triggered task queue mechanism, the improved CNP interaction protocol demonstrates a higher level of intelligence, operational efficiency, and flexibility in the field of real-time scheduling, providing solid support for the collaborative operation of multi-agent systems. The details of each improvement measure are described below.

[0124] (1) Yellow Pages Service helps physical agents demonstrate their capabilities. Through the Yellow Pages service, all agents can enjoy the search and registration Yellow Pages service. Physical agents RPA and RTA can clearly demonstrate their own capabilities and services, which facilitates the search and targeted negotiation of other service agents, greatly improves the transparency and accessibility of information, effectively avoids blind bidding, and significantly improves the intelligence and efficiency of the negotiation system.

[0125] (2) Bidding time setting to optimize the negotiation process To achieve load balancing, the collaborative agents use the Yellow Pages service to bid accurately and set deadlines to control the negotiation progress and reduce waiting time, thereby improving the efficiency and real-time nature of the negotiation. This reduces the system communication load, avoids traffic redundancy, and effectively prevents unnecessary communication overhead in large multi-agent systems.

[0126] (3) Multi-task bidding concurrent mechanism improves bidding efficiency. In each round of bidding, the collaborative agent selects the entity agent information of the agent with relevant processing capabilities from the yellow page service based on the current operation status and clarifies the bidding object. Each entity agent formulates a detailed planning plan based on the bidding information and decides whether to participate in the bidding after strict screening. After the bidding is completed, the plan is submitted to the collaborative agent for bidding screening. This concurrent mechanism ensures the efficient response of the collaborative agent to multi-task bidding, further improving the efficiency and accuracy of the bidding process.

[0127] (4) Event-based task queue triggering mechanism enhances response flexibility. Changes in the task queue, including the insertion of urgent orders, priority adjustments, or task cancellations, are considered abnormal events and trigger corresponding strategies to ensure that the production process is not interrupted. The insertion or cancellation of new tasks may cause the state of the response in the logistics queue or buffer to be uncertain. To deal with this situation, the response mechanism will reschedule the task queue, taking into account the processing information of the work-in-progress and the real-time status of the task queue. After the abnormal event is published to the MAS, the execution unit can quickly adjust the task queue to effectively respond to changes in the response process.

[0128] In one of the embodiments, a new initial solution generation strategy is adopted in the CNP solution method considering load balancing to make up for the shortcomings of greedy bidding and short-sighted strategies, and ensure that PCA can fully consider global information and load balancing between RPAs when making decisions; the new initial solution generation strategy is: when an emergency order task occurs, the emergency order task has the highest priority and is processed by the device with the smallest load first, and the generated initial solution satisfies the following conditions:

[0129] |L a -L b |<max{t ik} (7)

[0130] Among them, t ik is the processing time of operation Oi on the device Pk with the smallest current load, P k =f min , the load value during the scheduling period is Lmin, L min =L min +t ik , Lmin is the minimum processing load, L a , L b are the load values ​​of entity a and entity b respectively.

[0131] Specifically, when multiple agents need to establish a bidding mechanism to screen suitable processing entities due to their greedy nature in task selection, this mechanism often derives many low-quality solutions based on the current state, thereby slowing down the overall processing progress. To make up for the shortcomings caused by greedy bidding and short-sighted decision-making, this embodiment proposes a novel initial solution generation strategy. This strategy emphasizes that when processing collaborative agent PCA for decision-making, it is necessary to comprehensively consider global information and the load balance of each entity. By implementing this strategy, we can more accurately select the appropriate scheduling solution, thereby optimizing the overall performance. It can be achieved through P k =f min and L min =L min +t ik Select an appropriate scheduling effect; where Pk represents the device with the smallest current load, and its load value during scheduling is Lmin. When an emergency task occurs, this task has the highest priority and is processed by the device with the smallest load. At the same time, assuming that operation O i On device P k The processing time on ik If the generated initial solution satisfies the condition formula (7), load balancing can be achieved.

[0132] After adding the i-th task, the manufacturing system consisting of entities a and b is load balanced. For any manufacturing entity b except fmin, LL≤0 is satisfied. After the k-th operation, the difference between the load added by task j in fmin and the load of any other entity b is:

[0133] |L min +t ik -L b |≤t ik ≤max{t ik} (8)

[0134] It can be seen from formula (2.10) that the load difference between entity fmin and other entities satisfies load balance after adding new tasks, and the load difference between other entities is not affected by the newly added tasks.

[0135]

[0136] S i = {U i (S 1 ), U i (S 2 ), ..., U i (S k )} (10)

[0137]

[0138] In formula (9), T(Oij, Pk) represents the processing time required to perform operation Oij on entity Pk. If Oij is processed on device Pk, the corresponding Vijk value is 1, otherwise it is 0. In formula (10), Si represents the set of alternative bids considering global device load balancing. In this application, U * (s k ) represents the maximum bidding value of the contract network protocol.

[0139] With the active process relationship matrix and relay cooperation mechanism, the negotiation solution aimed at achieving RPA load balancing and minimizing the reaction task completion time can be accurately calculated. * (s k ). The global load balancing index parameter GLB, as shown in formula (11), measures the global load balancing effect by evaluating the maximum workload difference among all work types, where m represents the number of chemical workstation types. This mechanism comprehensively considers load balancing and task completion time, providing solid support for production process optimization.

[0140] In one embodiment, the real-time scheduling method based on bargaining game includes the following specific steps:

[0141] Step 1: Construct a bargaining game model of MOP, use the CA* algorithm to optimize the RTA path planning and movement sequence of each game stage, generate a dynamic game tree, and solve the SPNE solution through reverse induction.

[0142] Specifically, when seeking the equilibrium solution of the bargaining game, solving SPNE is crucial. In non-cooperative static games, Nash equilibrium, as a recognized effective strategy, indicates that players are unwilling to change their actions under the established strategy. However, in order to further explore the strategic choices of players in non-cooperative dynamic games with complete information, it is particularly important to introduce the concept of extended solution of SPNE. This concept covers dimensions such as reaction task completion time, RTA maximum load and total energy consumption, and clearly displays the decision-making order and player information mastery in a tree structure. Fig.14 As shown in the figure, the tree diagram of the three-player game clearly shows the game process. The game tree contains n stages in which the three players take turns to make decisions.

[0143] In this game, the three objective functions f1(x), f2(x) and f3(x) correspond to the three players and are called the players' strategy processes.

[0144] Each player in each stage is associated with a specific RTA, and the path information for executing this transfer operation needs to be considered when selecting a reaction action. The problem of strategy selection and game tree construction also involves an extension of the multi-agent path planning (MAPF) problem, namely the improved multi-agent pickup and delivery (MAPD) ​​problem. In this negotiation game, players act in a certain order, and the choice of each player will affect the strategy of subsequent players. The core goal of each player is to find an action plan that best suits their own interests and maximizes returns through negotiation. In each negotiation stage, the RTA associated with the player needs to select one from multiple possible paths while considering the shortest transportation distance to ensure the realization of the overall negotiation goal. To meet this challenge, the collaborative A* algorithm is introduced, which plays a key role in constructing the strategy choices of all players in each stage.

[0145] The cooperative A* algorithm (CA*) performs single-machine path planning for each AGV in order of priority, and incorporates the planned high-priority path information as a constraint into subsequent planning. The algorithm decomposes the multi-agent path planning problem into a series of interrelated single-machine path planning problems. At each stage, the robot performs single-machine path planning according to the priority order, and the previous robot's path planning information is added as a constraint to the subsequent planning. Specifically, based on the single-machine A* search, CA* adds a constraint table to store and integrate path planning information. Whenever a robot's path planning is completed, its path information is added to the constraint table, and together with the existing constraints, it constitutes the time-limited obstacle constraint for the next robot's path planning. This process is repeated until the planning schemes of all robots are completed. In addition, the CA* algorithm uses Manhattan distance as a heuristic function in the single-machine A* search. Although it may not work well in complex environments, it has a faster execution speed.

[0146] In each player's game phase, the CA* algorithm is used to select task strategies and plan corresponding paths according to their goals and assigned RTA. The planned path information is then included in the constraint table and used together with the previous constraints to construct the time-limited obstacle constraints in the subsequent robot path planning until the player completes all stages of the game. The pseudo code of the dynamic game tree construction algorithm based on CA* is shown below.

[0147] Algorithm 1 Pseudo code of dynamic game tree construction algorithm based on CA*:

[0148] Input: rootNode: the initial node of the decision tree; stage: the current decision stage; players: the list of players participating in the decision; tasks: the list of optional tasks; constraints: the set of constraints for task selection;

[0149] Output: dynamic game tree T; process: build Game Tree (rootNode, stage, players, tasks, constraints);

[0150] 1: If stage is equal to the length of players or tasks is empty;

[0151] 2: Return;

[0152] 3: End if;

[0153] 4:currentPlayer players[stage] / / Get the players in the current stage / / Let the current player choose the task according to the constraints;

[0154] 5:selectedTasks selects tasks (currentPlayer,tasks,constraints)

[0155] 6: Execute for each task in selectedTasks;

[0156] 7:path CA*path planning (task, constraints) / / Plan a path for the current task / / Update constraints based on the planned path

[0157] 8:updatedConstraints update constraints (constraints, path) / / Create a new child node

[0158] 9:childNode creates a node (node.info, path, updated Constraints);

[0159] 10: Add childNode to the child node list of node / / Add the child node to the current node;

[0160] 11:remainingTasks removes the selected tasks (tasks, task) / / Update the remaining task list;

[0161] 12: Recursively call buildGameTree(childNode,stage+1,players,remainingTasks,updatedConstraints) / / Recursively build a subtree for the next player;

[0162] 13: End the loop.

[0163] In this process, Algorithm 1 significantly reduces the complexity of search and improves the computational efficiency by focusing on the optimal finite strategy of the objective function. This not only simulates the limited rationality of players, but also avoids redundant over-computation, making the algorithm more efficient and practical when dealing with large-scale game scenarios. In this process, Algorithm 1 significantly reduces the complexity of search and improves the computational efficiency by focusing on the optimal finite strategy of the objective function. This not only simulates the limited rationality of players, but also avoids redundant over-computation, making the algorithm more efficient and practical when dealing with large-scale game scenarios.

[0164] Using the backward induction method, we explore the SPNE solution based on the successful construction of the dynamic game tree T. After the strategy allocation in each stage, we calculate the returns of each party and perform the backward induction based on the player's utility to determine the SPNE. The pseudo code for solving the SPNE by the backward induction method is as follows.

[0165] Algorithm 2: Reverse induction method to solve SPNE

[0166] Input: Dynamic game tree T

[0167] Output: spne_path as a subgame perfect Nash equilibrium (SPNE) process: SPNE_Searching(T);

[0168] / / Initialize variables;

[0169] 1: Let k be the number of stages of the game tree T;

[0170] 2: Let current_stage = k / / Traverse each stage of the game tree from the last stage forward;

[0171] 3: Execute the loop when current_stage is greater than or equal to 1;

[0172] / / Find the player fi in the current stage current_stage;

[0173] 4: Let fi = get the current stage player (current_stage, T) / / Calculate the number of strategy combinations of player fi in stage current_stage;

[0174] 5: Let num_profiles = calculate the number of strategy combinations (fi, current_stage, T) / / traverse each strategy combination to find the combination with the maximum benefit;

[0175] 6: Set max_payoff_profile = null 7: Set max_payoff = negative infinity;

[0176] 8: For h, execute a loop from 1 to num_profiles; / / Calculate the return of the current strategy combination;

[0177] 9: Let current_payoff = evaluation_payoff(fi,current_stage,h,T) / / compare and update the maximum payoff and the corresponding strategy combination;

[0178] 10: If current_payoff is greater than max_payoff, execute

[0179] ax_payoff = current_payoff max_payoff_profile = h; end if

[0180] 11: End loop / / End the strategy combination traversal of the current stage / / Record the path of the maximum profit strategy combination and delete other paths;

[0181] 12: Record the maximum profit path (max_payoff_profile, current_stage, T);

[0182] 13: Delete non-maximum benefit strategy combination (current_stage, T) / / Move forward to the previous stage;

[0183] 14:current_stage=current_stage-1;

[0184] 15: End loop / / End traversal of all stages

[0185] 16: Let spne_path = extract_full_path (T) / / Extract the complete SPNE path from the record.

[0186] The main process of solving SPNE in Algorithm 2 is described as follows.

[0187] Initialization phase: Determine the total number of stages k of the game tree T and set the current stage to the last stage, i.e., the kth stage.

[0188] Stage-by-stage analysis: For each current stage, determine the player fi who needs to make a decision. Calculate the number of strategy combinations that player fi can choose in this stage.

[0189] Find the optimal strategy: Traverse all strategy combinations of player fi and evaluate the potential benefits of each combination. Find the strategy combination with the maximum benefit and record its information and the corresponding maximum benefit.

[0190] Record and prune: Record the optimal strategy combination path of the current stage. Then, prune the game tree to remove all path branches of non-optimal strategy combinations in the current stage.

[0191] Move the stage forward: Move the current stage forward one stage and repeat steps 2 to 4 until all stages are traversed and the starting point of the game tree is reached.

[0192] Extract results: Extract the complete path from the starting point to the end of the game tree from the record. This path is the desired SPNE.

[0193] Output results: The final output is the extracted SPNE path, which represents the optimal decision sequence at each stage of the entire game process.

[0194] Step 2: Determine the agreement value that maximizes the sum of the benefits of multiple players and calculate the current maximum joint benefit; if the current maximum joint benefit is the default benefit, the SPNE solution is the equilibrium solution of the bargaining game and the agreement is reached; otherwise, proceed to step 3.

[0195] Step 3: Calculate the profit difference Qi and corresponding tolerance value e* of each player under the current agreement. When Qi is a negative value, if the player's tolerance value e* exceeds his maximum tolerance eimax, abandon the current agreement and return to step 2 for renegotiation; otherwise, continue to step 4 to screen feasible strategies for maximizing joint profits.

[0196] Step 4: Calculate the final profit of each player under the current protocol strategy based on the protocol surplus Q, default profit d and the player’s bargaining power P.

[0197] Step 5: Calculate the current contract currency transfer T, and determine the transfer amount to achieve the required remaining allocation based on the current contract currency transfer T.

[0198] Step 6: When the players negotiate and obtain the final benefits, calculate the tolerance value of the player whose contract currency transfer is a negative value; if the tolerance value does not exceed its maximum tolerance eimax, the player agrees to the contract; if all players agree, the agreement is reached and the contract ends; otherwise, if the player's tolerance value exceeds eimax, delete the maximum joint benefit and strategy under the current agreement, and return to step 2 for renegotiation.

[0199] Specifically, at each real-time scheduling time t, a real-time scheduling method is activated in the TCA. This method is based on a bargaining game mechanism and aims to accurately allocate task operations to the most appropriate RTA based on the real-time availability information of the RTA. The following are the core steps of the problem solution.

[0200] (a) According to the number of targets, RTAs are assigned to each target player in turn, such as RTA1 to player f1, RTA2 to f2, and so on. If the number of RTAs exceeds the number of targets, they are assigned in a loop, such as RTA4 to f1, until all RTAs have targets.

[0201] (b) The three goals correspond to three players, and each player chooses the action that best benefits its goal and maximizes its reward through negotiation. During the negotiation process, one player or the corresponding RTA makes a decision at each stage and selects the available actions from the TCA.

[0202] (c) Based on the feasible strategy combinations, use Formula (29) to Formula (31) to calculate the utility functions u1(s), u2(s), and u3(s) of each player.

[0203] (d) Find the equilibrium solution of the bargaining game and allocate the available actions in the TCA to the most suitable RTA in an optimal way.

[0204] (e) Repeat the above steps as time goes by until all tasks are assigned. If an abnormal event (such as machine failure, order change, etc.) occurs, the players or strategies of the negotiation game can be adjusted to respond in time and reduce the impact.

[0205] SPNE has become a universal solution for solving N-player non-cooperative dynamic games. It is represented by a combination of strategies, each player holds one of them, and any player cannot improve the expected return by unilaterally changing the strategy. Unlike dynamic games, bargaining games focus more on the creation and redistribution of value, and the agreement reached through negotiation usually allows players to obtain higher returns than before the negotiation. Although Watson provides a standard solution to the negotiation problem, it is limited to the two-player scenario. In view of this, an algorithm based on Watson's solution is proposed to find the equilibrium solution of multi-player negotiation games.

[0206] Let V be the set of alternative plan benefit vectors of the players in a given bargaining problem. This set expresses in detail the possible benefits that the players may obtain under different strategy choices. The specific mathematical expression is shown in formula (13).

[0207]

[0208] V={u i (s 1 ),u i (s 2 ), ..., u i (s k ), ..., u i (s n )} i=1,2,3 (13)

[0209] Let d be the payment vector corresponding to the default outcome, which belongs to the payoff vector set V and can be accurately expressed by equation (14). The d vector essentially reveals the possible outcomes faced by the players when they fail to reach an effective agreement. In the specific problem discussed in this paper, the determination of d depends on two key factors, ui(s*) and s*. Among them, s* represents the SPNE solution of the bargaining game, while ui(s*) specifically represents the payoff of the players under this equilibrium solution.

[0210] d = {d1, d2, d3}, d i =u i (s * ) (14)

[0211]

[0212] Let u * (s k ) is the joint benefit of the maximum agreement in the bargaining game, and its expression is shown in equation (16). In some cases, the default benefit is the maximum benefit, that is, u * (s k ) is the default benefit. Q represents the agreement surplus, which is the difference between the joint benefit and the benefit when no agreement is reached. The specific calculation is shown in formula (17). Qi refers to the difference between the benefits of each player under the maximum joint benefit and the benefit when no agreement is reached. The specific calculation is shown in formula (17).

[0213]

[0214] Q i =u i * (s k )-d i i=1,2,3 (17)

[0215] It is meaningful to use the bargaining weight P to summarize the bargaining power of players, where the weight Pi of player fi is constrained by formula (19). This weight can be regarded as the proportion of the surplus value Q obtained by player fi in the negotiation. In this patent, each player Pi is set to 1 / 3.

[0216] P={P i} i=1,2,3 (18)

[0217]

[0218] The standard negotiation solution is a mathematical representation of efficiency and proportional distribution, ensuring that each participant can obtain a corresponding share of the residual value on the basis of the default return. Once the agreement contract is reached, each player will obtain the final benefit according to formula (20)

[0219]

[0220] Fig.15 A bargaining game scenario involving three players is outlined, where players make strategic trade-offs between maximizing joint payoffs and maintaining default payoffs, aiming to reach consensus and determine the final payoff.

[0221] Consider the set T of player contract currency transfer situations, and its calculation method is detailed in formula (22). This set is mainly used to evaluate the k After reaching the agreement, the absolute difference between the actual payment of each player fi and the income under SPNE. With this metric, we can deeply analyze the actual performance of the protocol, the strategic choices of the players and the stability of the protocol.

[0222] T={T i} i=1,2,3 (21)

[0223]

[0224] In the bargaining game, we pay special attention to the negative value of Ti, because it directly reveals the potential adverse impact of the protocol on the player fi and becomes an important indicator for evaluating the player's tolerance. This application introduces three key concepts: the tolerance value ei of the player fi under the condition of maximizing the joint benefit * , the tolerance value ei when the agreement is actually reached, and its maximum tolerance value x. According to the importance of each player's objective function, the corresponding maximum tolerance value x = [0.1, 0.2, 0.2] is set for each player. The initial screening is carried out through formula (13), and the final bargaining agreement is confirmed through formula (24). The validity of the agreement depends on whether the contract money transfer falls within the tolerance range of player fi, as shown in formula (25).

[0225]

[0226] According to the pseudo code of the three-player bargaining game solution algorithm, we can find the negotiation solution and the equilibrium state of the game, where the final payoff of player fi is The contract currency transfer between players is T. The overall process of the real-time scheduling algorithm for bargaining games is as follows: Fig.16 shown.

[0227] In one embodiment, the bargaining game model for constructing MOP is:

[0228] G={F i ; S i ; U i} i=1,2,...,n (26)

[0229] Among them, Fi represents the set of players participating in the game, S i represents the strategy or action taken by player i, U i is the reward function of player i.

[0230] Specifically, the essence of the real-time scheduling problem of TCA based on MAS is to solve a multi-objective optimization problem. Such problems usually follow a general pattern, as shown in equation (27), which requires finding the best balance between multiple conflicting or mutually constrained objectives. The optimization process requires weighing the priority and importance of different objectives to reach the best solution that meets all objective requirements.

[0231]

[0232] The decision variable is represented by x, and its variable space is X. The cost function is represented by (x), where i ranges from 1 to k. At the same time, (x) and h(x) together constitute constraints. The bargaining game theory model contains three elements: players, strategies, and benefits. Players compete and cooperate with each other. In order to apply this theory to multi-objective optimization problems (MOPs), a corresponding MOP model is constructed.

[0233] The bargaining game theory model for MOP adopts the dynamic game theory method to find the optimal solution. The model describes the MOP involving k players as equation (27), where the strategy space S is equivalent to the variable space X, with the accompanying payoff function (x). The problem is regarded as a negotiation game, the player interaction is the game process, and the equilibrium solution of the game is the optimal solution of the MOP. Therefore, this module elaborates on the negotiation game, the various elements of the solution, including players, strategies, payoff design, and equilibrium solution, aiming to provide a comprehensive framework to consider the strategic interaction between players and find the optimal decision-making plan.

[0234] Both bargaining games and dynamic games involve the core element of subgame perfect Nash equilibrium (SPNE), which is the cornerstone of non-cooperative dynamic game theory. Nash equilibrium means that each player chooses the best strategy combination given the strategies of others. In a dynamic environment with perfect information, although there may be more than one Nash equilibrium, not all equilibria are trustworthy. Subgame perfect Nash equilibrium is an equilibrium state that has been screened to exclude untrustworthy threats or promises, thereby ensuring the stability and credibility of the strategy. In short, SPNE represents an equilibrium state in dynamic games that is both trustworthy and stable. It is defined as: in each subgame of the game, the strategies of the players constitute a Nash equilibrium.

[0235]

[0236] The strategy set S*, as the game defined in equation (28), is a set of strategies. If it not only forms a Nash equilibrium in the original game, but also maintains a Nash equilibrium state in all subgames, then S* can be considered as a SPNE solution to the game. Therefore, for the MOP problem, it can be regarded as a bargaining game involving k players, in which the decision strategy space S of each player is consistent with the variable space X of the entire problem. The utility function of each player is represented by (). Therefore, the SPNE solution can be regarded as a valid solution to the MOP.

[0237] This application regards the real-time scheduling problem of TCA based on MAS as an N-person game problem with complete information, in which game theory aims to reasonably distribute profits through negotiation. To build a game theory model, three core elements must be clarified: players, strategies, and rewards. Specifically, the model can be expressed as formula (26).

[0238] The present application relates to a negotiation game, in which Fi represents the set of players participating in the game. Three objective functions are set, one for each player, for a total of three. This setting can be easily extended to an N-person game model. During the game, the players take actions in turn, and their choices influence each other, forming a strategic interaction. Among them, Si represents the strategy or action taken by player i, specifically the first reaction transfer operation to be processed at each time point t. And Ui, as the reward function of player i, closely corresponds to the three objective functions set in this patent. In each stage of the game, an executable RTA will be assigned to the player, and the players will act in turn and select strategies from their respective feasible strategy sets. Finally, by solving the game equilibrium solution, the negotiation scheduling of multi-objective optimization problems is achieved.

[0239] In one embodiment, the TCA real-time scheduling problem is specifically described as follows: Faced with a task set containing n reaction transfer tasks, these tasks need to be performed by m mobile robots RTA to transfer reagent trays. Reaction task i contains a total of ni current operation sets. For the current operation Oij in task i, it is necessary to select an RTA with corresponding capabilities to move and transfer it to the corresponding robot workstation RPA to complete the operation.

[0240] The core goal of the TCA real-time scheduling problem is to optimally allocate mobile operations at the current time t, while reasonably arranging the execution order and path planning of RTA. Based on the symbols listed in Table 4, the MOP mathematical model is constructed for this problem, where the mathematical model of MOP is:

[0241]

[0242] The constraints are:

[0243] C ij =TCij +D ij (32)

[0244] C i，j -C i，j-1 ≥t i，j，k x i，j，k +D ij j=2,...,n j (33)

[0245] C ij >0 i=1,...,n (34)

[0246]

[0247] Where f1 is the maximum completion time of the reaction task, f2 is the workload of the RTA key machine, f3 is the total energy consumption of the RTA task execution, TCij is the completion time of Oij's transfer from Mk, Dij is the RPA processing time, Oij is the jth operation of reaction task i, Dij is the span of Oij in RPA, and x ijk is the decision variable for the kth RTA to be selected to perform the jth operation of reaction task i, which is used to determine whether the kth RTA is selected to perform the jth operation of reaction task i. i,j is the completion time of Oij, P lk is the Mk idle power, t lk is the total idle time of Mk, n j is the total number of operations of reaction task j, E is the total energy consumption of Mk, m is the total number of RTAs, n is the total number of current reactions, ni is the total number of operations of reaction task i, and C M is the longest completion time of RPA, W M is the RTA with the heaviest workload, W k M k The workload, x ij If Mk is selected to execute Oij, it is 1, if not selected, it is 0, t ijk is the processing time of Oij on Mk.

[0248] In formulas (29) to (31), three key objective functions are accurately characterized: f1 reflects the maximum completion time of the reaction task, f2 measures the workload of the RTA key machine, and f3 focuses on the total energy consumption of the RTA execution task. These functions are flexibly adjusted as the scheduling results change. By minimizing these objectives, the optimal scheduling effect is ensured at each time point t.

[0249] Furthermore, formula (32) provides another way to express the operation Oij with TCij (RTA transfer completion time) and Dij (RPA processing time) as parameters, which helps to gain a deeper understanding of the execution status of tasks on the machine and plays a key role in optimizing scheduling.

[0250] In order to ensure the rationality of scheduling, inequalities (33) and (34) are introduced. As prerequisites between operations, they are indispensable constraints in the scheduling process and help maintain the order and logic between tasks. At the same time, constraint formula (35) clearly stipulates that each operation can only be assigned to one machine, which not only ensures the effectiveness of scheduling, but also ensures the rational use of resources.

[0251] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0252] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0253] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A multi-robot chemical synthesis real-time scheduling control method based on MAS, characterized in that: The method comprises: Construct a JADE middle layer; the JADE middle layer is used to provide a JADE operating environment to realize the registration, management and interaction of intelligent agents. Each running instance in the JADE operating environment is a container, and each container contains multiple intelligent agents. All active container sets constitute a platform. There is a main container in the platform, which includes the intelligent agent management system AMS, the directory service provider DF and several intelligent agents. All other containers in the platform are registered in the main container immediately after startup. The chemical synthesis resources of the multi-robot autonomous chemical synthesis process are mapped into several agents with different functional attributes, and all agents form a multi-agent layer; Based on the real-time data of the MAS platform, the PCA real-time scheduling method based on the relay collaboration model of the improved process relationship matrix and the TCA real-time scheduling method based on the bargaining game are adopted to realize the scheduling and control of autonomous chemical synthesis, so that the system can respond quickly and adapt to different chemical reaction requirements.

2. The method according to claim 1, characterized in that The intelligent agents include: entity intelligent agents and service intelligent agents; The entity agent includes: a reaction processing agent and a reaction transfer agent; The service agent includes: a reaction baton agent, a reaction processing cooperation agent, a reaction transfer cooperation agent and a reaction monitoring agent; The reaction processing agent and the reaction transfer agent are used to capture real-time data from automatic identification equipment and process complex real-time data so that the data can be understood as meaningful synthetic information; The reaction baton agent is used to capture the information of all tasks and generate corresponding batons; The reaction processing agent and the reaction transfer agent are used to provide a negotiation scheduling algorithm for the reaction processing agent and the reaction transfer agent, and optimally schedule the start time and completion time of each operation of each character by sensing real-time reaction information; The reaction monitoring agent is used to capture and process reaction execution information in real time, and respond to related synthetic information queries.

3. The method according to claim 2, characterized in that The baton carries an optimized reaction process relationship matrix and records the core information of chemical reaction synthesis in real time; the core information includes at least: raw material consumption, processing details, and process information; The reaction process relationship matrix is ​​constructed according to the process characteristics in order to express the dependency relationship between the steps in the chemical reaction process; the element value of the reaction process relationship matrix is ​​0 ij , when O ij When it is a positive integer, it means that operation i is the previous operation of operation j, and the smaller the value, the higher the priority; if O ij If it is 0, there is no precedence relationship between the two; The reaction baton-based transfer collaboration process specifically includes: After the MAS receives the order, the reaction baton agent generates a new reaction baton and passes it to the reaction processing collaborative agent, thus starting the entire task process; The reaction processing collaborative agent negotiates with the reaction processing agent to select the best processing workstation according to the requirements and process standards of the chemical reaction; After selecting the workstation, the reaction processing collaborative agent hands over the reaction baton to the reaction transfer collaborative agent, which is responsible for coordinating the reaction transfer agents to ensure that the reagent tray can be accurately and efficiently transferred to the selected reaction processing agent; After receiving the reaction baton, the reaction processing agent starts to perform the chemical reaction operation, and after completion, returns the reaction baton to the reaction processing collaborative agent for subsequent operation or monitoring; Through continuous negotiation, transfer and execution of operations, all chemical reaction operations are processed, which finally marks the end of the entire task.

4. The method according to claim 1, characterized in that: The functions and service agents in MAS work together in the following ways: When a new task arrives, the reaction baton agent generates a corresponding reaction relay baton, records the task number and process relationship information matrix, and realizes synchronous information tracking within the MAS; The reaction processing collaborative agent is used to realize the real-time scheduling of the reaction processing agent. At each time step t, its input includes the real-time capability information from the current reaction baton agent, the available operations from the baton received, and the real-time reaction synthesis information from the reaction monitoring agent. The task of the reaction processing collaborative agent is to generate the task queue of the robot workstation based on this information, and to reach a contract with the corresponding reaction processing agent to write the scheduling result information into the baton. When the reaction processing collaborative agent completes the scheduling, it will hand over the completed reaction baton to the reaction transfer collaborative agent for subsequent processing. Adopting reaction transfer cooperative agent to realize real-time scheduling of reaction transfer agent; The reaction processing agent receives the reaction baton transferred from the reaction transfer agent and verifies it with its own contract; after the reaction processing agent completes the corresponding reaction operation, it hands the reaction baton that has completed the current operation back to the reaction processing collaborative agent for processing; this cycle continues until all the reaction process operations recorded by the reaction baton are completed; By querying the global information of the reaction monitoring agent, the reaction processing collaborative agent and the reaction transfer collaborative agent negotiate and schedule appropriate reaction processing agents and reaction transfer agents to execute the current process according to the optimization goal; when handling abnormal events, the reaction monitoring agent corrects and broadcasts the internal information of the reaction baton, and then incorporates the corrected internal information of the reaction baton into the corresponding process to achieve rapid processing.

5. The method according to claim 4, characterized in that The steps of the MAS-based improved contract network protocol include: The initiator publishes a request for proposals, specifies the task content and execution conditions, and waits for responses from other agents; Each agent submits its own proposal according to the CFP, and the initiator then collects and evaluates these proposals; The initiator selects a suitable agent to perform the task based on the evaluation results, sends an acceptance notification to the selected agent, and notifies the unselected agents at the same time; The selected agent performs the task according to its own proposal; After the task is completed, the agent reports the result to the initiator; if it cannot be completed, the agent notifies the initiator in time for follow-up processing.

6. The method according to claim 1, characterized in that The steps of the PCA real-time scheduling method of the relay cooperation model based on the improved process relationship matrix include: A relay cooperation model based on a positive process relationship matrix table is constructed, wherein the relay cooperation model includes a main objective function, a secondary objective function and constraints; the main objective function is: The secondary objective function is: C i =S i +t ik 1≤i≤n,1≤k≤p The constraints are: my{S i }&min{F k } S i ≥S j +t jk ,AND ik =1,Y jk =1 S i ≥max(C l ),Xl i =1 Among them, f and T are two minimization functions, C i For operation O i The completion time, F k P k The idle time period, p is the total number of available reaction processing agents, P k is the kth reaction processing agent among all available reaction processing agents, 1≤k≤p, S i For operation O i The start processing time, Y jk For operation O i Processing on Pk, t ik Operation O for Pk i processing time, n is the total number of reaction operations, Di is the time span of operation Oi, X li For operation O l It is operation O i The previous process; According to the relay cooperation model based on the positive process relationship matrix table, a CNP solution method considering load balancing is adopted to optimize the cooperation and information transmission between components, thereby realizing an efficient negotiation solution.

7. The method according to claim 6, characterized in that A new initial solution generation strategy is adopted in the CNP solution method considering load balancing to make up for the shortcomings of greedy bidding and short-sighted strategies, and ensure that the reaction processing collaborative agent can fully consider the global information and the load balancing between reaction processing agents when making decisions; the new initial solution generation strategy is: when an emergency insertion task occurs, the emergency insertion task has the highest priority and is processed by the device with the smallest load first, and the generated initial solution satisfies the following conditions: |L a -L b |<max{t ik } Among them, t ik is the processing time of operation Oi on the device Pk with the smallest current load, P k =fmin, the load value during the scheduling period is Lmin, L min =L min +t ik , Lmin is the minimum processing load, L a , L b are the load values ​​of entity a and entity b respectively.

8. The method according to claim 1, characterized in that The specific steps of the real-time scheduling method based on bargaining game include: Step 1: Construct a bargaining game model of MOP, use the CA* algorithm to optimize the path planning and movement sequence of the reaction transfer agent in each game stage, generate a dynamic game tree, and solve the SPNE solution through reverse induction; Step 2: Determine the agreement value that maximizes the sum of the benefits of multiple players and calculate the current maximum joint benefit; if the current maximum joint benefit is the default benefit, the SPNE solution is the equilibrium solution of the bargaining game and the agreement is reached; otherwise, proceed to step 3; Step 3: Calculate the profit difference and corresponding tolerance value of each player under the current agreement. If the profit difference is negative, and the player's tolerance value exceeds his maximum tolerance, abandon the current agreement and return to step 2 for renegotiation; otherwise, continue to step 4 to screen feasible strategies to maximize joint profits; Step 4: Calculate the final profit of each player under the current protocol strategy based on the protocol surplus Q, default profit and the player's bargaining power; Step 5: Calculate the current contract money transfer T, and determine the transfer amount to achieve the required remaining allocation based on the current contract money transfer T; Step 6: When the players negotiate and obtain the final benefits, calculate the tolerance value of the player whose contract currency transfer is a negative value; if the tolerance value does not exceed its maximum tolerance, the player agrees to the contract; if all players agree, the agreement is reached and ended; otherwise, if the player's tolerance value exceeds, delete the maximum joint benefit and strategy under the current agreement, and return to step 2 for renegotiation.

9. The method according to claim 8, characterized in that The bargaining game model for constructing MOP is: G={F i ;S i ;U i } i=1,2,...,n Among them, Fi represents the set of players participating in the game, S i represents the strategy or action taken by player i, U i is the reward function of player i.

10. The method according to claim 9, characterized in that The mathematical model of MOP is: The constraints are: C ij =TC ij +D ij C i，j -C i，j-1 ≥t i，j,k x i，j，k +D ij j=2,…,n j C ij >0 i=1,…,n Where f1 is the maximum completion time of the reaction task, f2 is the workload of the key machine of the reaction transfer agent, f3 is the total energy consumption of the reaction transfer agent to perform the task, TCij is the completion time of Oij from Mk, Dij is the processing time of the reaction processing agent, Oij is the jth operation of reaction task i, Dij is the span of Oij in the reaction processing agent, and x ijk is the decision variable for the kth RTA to be selected to perform the jth operation of reaction task i, C i,j is the completion time of Oij, P lk is the Mk idle power, t lk is the total idle time of Mk, n j is the total number of operations of reaction task j, E is the total energy consumption of Mk, m is the total number of reaction transfer agents, n is the total number of current reactions, ni is the total number of operations of reaction task i, C M is the maximum completion time of the reaction processing agent, W M is the reactive transfer agent with the heaviest workload, W k M k The workload, xi j If Mk is selected to execute Oij, it is 1, if not selected, it is 0, t ijk is the processing time of Oij on Mk.

Images