Multi-agent collaborative task planning method, related device, equipment and storage medium

Abstract

The invention discloses a multi-agent collaborative task planning method, and is applied to the technical field of artificial intelligence. The method comprises the steps of decomposing a task into a plurality of sub-tasks through semantic recognition and generating corresponding semantic coding vectors; meanwhile, a preset agent resource library is called, and quantitative evaluation capability vectors of all agents in multiple skill dimensions are obtained; dynamically allocating the most adaptive target agent to execute the corresponding subtask based on matching calculation of the subtask coding vector and the agent capability vector; then parallelly driving the target agent to execute the subtasks, fusing processing results output by the target agent, and integrating to generate a task response text; and finally returning the response text to the user. According to the method, the task is split into the coding vectors corresponding to the sub-tasks through semantic recognition, and dynamic matching is performed in combination with the multi-dimensional capability vector of each agent, so that adaptation of task requirements and agent resources is realized, and the resource scheduling efficiency and execution reliability of a multi-agent system in a complex task scene are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a multi-agent collaborative task planning method, related devices, equipment, and storage medium. Background Art

[0002] Mission planning technology, dedicated to automatically decomposing complex objectives and coordinating resources to generate feasible execution sequences, is a key enabler for the implementation of AI applications. As application scenarios expand into complex and dynamic fields such as intelligent office work, data annotation assistance, and content review, higher requirements are placed on the processing efficiency, environmental adaptability, and coordination capabilities of planning systems. An ideal mission planning system must be able to efficiently handle high-dimensional information, dynamically changing environmental constraints, and effectively manage collaboration among multiple execution units.

[0003] Current mainstream solutions mainly include traditional optimization algorithms and multi-agent systems (MAS). Traditional algorithms perform well in static environments with clear structures, but when dealing with large-scale, high-dimensional tasks or tasks with dynamic interference, they often face the dilemma of excessive computational burden and difficulty in responding to real-time changes. Multi-agent systems share task complexity through a distributed architecture, theoretically improving the ability to handle complex tasks. However, when coordinating the work of multiple agents, existing MAS solutions often lack efficient resource scheduling mechanisms, which can easily lead to problems such as insufficient resource utilization or local overload. At the same time, such systems generally lack effective ways to respond to and adjust to environmental changes or new constraints.

[0004] Overall, existing technologies have significant shortcomings: First, when faced with complex tasks that require the collaboration of multiple execution units, there is a lack of efficient resource coordination mechanisms, which can easily lead to low overall operating efficiency or improper resource allocation; second, there is a general lack of adaptability to dynamic environments. The system finds it difficult to effectively perceive changes during task execution and correct the planned path in real time, resulting in poor robustness in the face of environmental disturbances or new constraints, and the execution results are likely to deviate from the expected goals. These shortcomings limit the reliable application of task planning technology in complex real-world scenarios. Therefore, a new method is urgently needed to solve such problems. Summary of the Invention

[0005] The embodiments of the present application provide a multi-agent collaborative task planning method, related devices, equipment and storage medium, which solve the problems of inefficient resource scheduling and weak environmental adaptability in multi-agent systems.

[0006] On one hand, the present application provides a multi-agent collaborative task planning method, comprising:

[0007] Receive task request text;

[0008] Based on semantic recognition, the task in the task request text is split into P subtasks, where P subtasks correspond to P subtask encoding vectors, P ≥ 1;

[0009] Obtain M agents, where M agents correspond to M capability vectors. The capability vectors are used to represent the quantitative performance of the agents in K skill dimensions, where M ≥ P ≥ 1 and K ≥ 1.

[0010] Determine P target agents for processing the P subtasks based on the P subtask encoding vectors and M capability vectors;

[0011] Call P target agents to execute P subtasks, integrate the P subtask processing results obtained by the P target agents executing the P subtasks, and generate a task response text;

[0012] Sends a response message containing the task response text.

[0013] Another aspect of the present application provides a multi-agent collaborative task planning device, comprising:

[0014] Receiving module, used for receiving task request text;

[0015] A semantic recognition module is used to split the task in the task request text based on semantic recognition to obtain P subtasks, where P subtasks correspond to P subtask encoding vectors, P ≥ 1;

[0016] An acquisition module is used to acquire M agents, where M agents correspond to M capability vectors. The capability vectors are used to represent the quantitative performance of the agents in K skill dimensions, where M ≥ P ≥ 1 and K ≥ 1.

[0017] A determination module, configured to determine P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors;

[0018] A generation module is used to call P target agents to execute P subtasks, integrate the P subtask processing results obtained by the P target agents executing the P subtasks, and generate a task response text;

[0019] The sending module is used to send response information including task response text.

[0020] On the other hand, the present application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-mentioned methods when executing the computer program.

[0021] Another aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned methods when the computer program is executed by a processor.

[0022] Another aspect of the present application provides a computer program product, including a computer program, which implements the above-mentioned methods when executed by a processor.

[0023] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0024] In an embodiment of the present application, a task planning method for multi-agent collaboration is provided. The method receives a task request text and splits it into coding vectors corresponding to multiple sub-tasks based on semantic recognition. At the same time, dynamic matching is performed in combination with the multi-dimensional capability vectors of each agent to achieve precise adaptation of task requirements and agent resources. High-quality task response text is generated through collaborative execution and result integration, which ultimately improves the resource scheduling efficiency and execution reliability of the multi-agent system in complex task scenarios, and effectively solves the problems of insufficient coordination ability and poor robustness of traditional planning methods in dynamic collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is an interface diagram of the smart assistant provided in the embodiment of the present application;

[0026] Figure 2 This is an interface diagram of the multi-agent collaboration platform provided in an embodiment of the present application;

[0027] Figure 3 This is an interface diagram of the customer service robot provided in an embodiment of the present application;

[0028] Figure 4 This is an interface diagram of the image content review product provided in an embodiment of the present application;

[0029] Figure 5 An architectural diagram of the multi-agent collaborative task planning method provided in an embodiment of the present application applied to an intelligent assistant scenario;

[0030] Figure 6 A flowchart of a multi-agent collaborative task planning method provided in an embodiment of the present application;

[0031] Figure 7 A flowchart of an image recognition content review scenario of a manual annotation assistance system provided in an embodiment of the present application;

[0032] Figure 8 A schematic diagram of an agent task matching algorithm provided in an embodiment of the present application;

[0033] Figure 9 A schematic diagram of a similarity value calculation process provided in an embodiment of the present application;

[0034] Figure 10A schematic diagram of the calculation process of the skill dimension value provided in an embodiment of the present application;

[0035] Figure 11 A schematic diagram of the process of dynamic adjustment and resource scheduling provided in an embodiment of the present application;

[0036] Figure 12 A schematic diagram of the process of splitting subtasks provided in an embodiment of the present application;

[0037] Figure 13 A task diagram corresponding to the travel arrangements provided in the embodiment of this application;

[0038] Figure 14 A flowchart of the task graph generation process provided in an embodiment of the present application;

[0039] Figure 15 A schematic diagram of the embedded coding process provided in an embodiment of the present application;

[0040] Figure 16 A schematic diagram of closed-loop iterative optimization provided in an embodiment of the present application;

[0041] Figure 17 A schematic diagram of user feedback provided in an embodiment of the present application;

[0042] Figure 18 A schematic diagram of a flow chart for generating a feedback signal according to an embodiment of the present application;

[0043] Figure 19 A flowchart of a multi-agent collaborative task planning method provided in an embodiment of the present application;

[0044] Figure 20 A structural diagram of a multi-agent collaborative task planning device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] The embodiment of the present application provides a task planning method for multi-agent collaboration, which significantly improves the dynamic coordination ability and execution reliability of the multi-agent system for complex tasks through task semantic decomposition, agent capability vector matching and collaborative execution integration mechanism.

[0046] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0047] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0048] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0049] Before introducing the specific method of this application, the application scenario of the multi-agent collaborative task planning method in this application is first exemplified. Of course, it should be noted that the following application scenario is only an example and is not limited to this.

[0050] In a possible implementation, the method provided in the embodiment of the present application can be applied in an intelligent assistant scenario. Figure 1 As shown, Figure 1 The figure shows the interface diagram of the intelligent assistant. Figure 1 In (A), the user enters the task request "Please help me arrange a business trip from location A to location B, including air ticket reservation, hotel recommendation, and meeting reminder setting" in the intelligent assistant platform. 111.

[0051] like Figure 1 In (C), we first receive the task request text "Please help me arrange a business trip from location A to location B, including air ticket booking, hotel recommendations, and meeting reminder settings." Based on semantic recognition, we split the task in the task request text into three subtasks, as shown in Table 1. Each subtask corresponds to a subtask encoding vector:

[0052] Table 1

[0053] Subtask number Subtask content Subtask encoding vector Subtask 1 Book a flight [0.83,0.92,0.12,0.95] Subtask 2 Book a hotel [0.78,0.85,0.63,0.71] Subtask 2 Set a schedule reminder [0.35,0.97,0.88,0.59]

[0054] Get M predefined agents with specific capabilities. Each agent corresponds to a functional module, and each functional module corresponds to a capability vector, as shown in Table 2:

[0055] Table 2

[0056] Agent Role Functional modules Capability Vector Flight Agent Flight query and booking interface call [0.94,0.75,0.08,0.98] Hotel Agent Hotel database search, sorting and recommendation [0.87,0.91,0.79,0.82] Calendar Agent Calendar API operations, reminder settings [0.29,0.95,0.93,0.81] GeneralAgent General dialogue understanding, task coordination [0.62,0.68,0.77,0.54] ··· ··· ···

[0057] Based on the three subtask encoding vectors and M capability vectors, three target agents are determined to handle the three subtasks: the flight agent is used to handle subtask 1, the hotel agent is used to handle subtask 2, and the calendar agent is used to handle subtask 3. These three target agents are called to perform the three subtasks respectively, and the three subtask processing results are obtained:

[0058] Subtask 1 result: {Departure: Airport T3, Destination: Airport B, Flight number: xxxx, Time: xxxx(year)-xx(month)-xx(day)xx:xx(take-off time)-xx:xx(estimated landing time), Price: ￥xxx}

[0059] Subtask 2 result: {Hotel: B xxxx 9 o'clock, Room type: xxxx, Price: ￥xxx / night, Location: xxxxxxxxxxx}

[0060] Subtask 3 result: {Item: Business trip meeting, Time: xx:xx-xx:xx, Reminder: 30 minutes in advance, Location: xxxx}

[0061] Integrate the processing results of the three subtasks to generate the task response text:

[0062]

[0063] Send a response message including the task response text. Figure 1 (B) in the example shows the task response text 121.

[0064] In the intelligent assistant scenario, the method provided in the embodiment of the present application accurately decomposes the user's business trip tasks into three sub-tasks: flight booking, hotel recommendation, and schedule setting through semantic recognition, and uses vectorized coding technology to quantify the core requirement characteristics of each sub-task. Combined with the quantitative ability vectors of predefined intelligent agents in multi-dimensional skill fields, dynamic matching of tasks and resources is achieved, effectively ensuring that the professional capabilities of flight agents in real-time ticketing processing, hotel agents in matching geographic locations with user preferences, and calendar agents in cross-platform schedule synchronization are accurately adapted to the requirements of their respective sub-tasks; during the task execution process, structured intermediate result collection and natural language generation technology are used to achieve seamless integration of multi-agent output results, and finally generate a user-friendly response text containing a complete itinerary, price details and location correlation information, significantly improving the processing efficiency and execution reliability of complex multi-module collaborative tasks.

[0065] In a possible implementation, the method provided in the embodiment of the present application can be applied in document generation on a multi-agent collaboration platform. Figure 2 As shown, Figure 2 The interface diagram of the multi-agent collaboration platform is shown in FIG. Figure 2 In (A), the user enters the task request text: "Please generate the company's annual summary document, covering performance analysis, team achievement presentation and three-year business plan." 211.

[0066] like Figure 2 In (C), we first receive the task request text: "Please generate a company annual summary document covering performance analysis, team achievements, and three-year business plan." The task is broken down based on the semantic recognition module. As shown in Table 3, the decomposition yields three subtasks and generates corresponding encoding vectors:

[0067] Table 3

[0068] Subtask number Subtask content Subtask encoding vector Subtask 1 Performance data analysis [0.83,0.92,0.12,0.95] Subtask 2 Structured presentation of team achievements [0.78,0.85,0.63,0.71] Subtask 2 Business planning strategy generation [0.35,0.97,0.88,0.59]

[0069] At the same time, we obtain M predefined document processing agents (M≥3), whose capability vectors quantify the performance of each agent in the skill dimension, as shown in Table 4:

[0070] Table 4

[0071] Agent Role Capability Vector Data Analysis Agent [0.94,0.75,0.08,0.98] Results Summary Agent [0.87,0.91,0.79,0.82] Planning Strategy Agent [0.29,0.95,0.93,0.81] General Document Coordination Agent [0.62,0.68,0.77,0.54]

[0072] According to the matching relationship between the subtask encoding vector and the capability vector, the target agent allocation scheme is determined:

[0073] The data analysis agent handles subtask 1 (performance data analysis);

[0074] Results summary agent processing subtask 2 (structured presentation of team results);

[0075] Planning strategy agent processing subtask 3 (business planning strategy generation);

[0076] Call the target agent to perform the subtask and obtain the intermediate result:

[0077] Subtask 1 output: standardized performance indicator set and visual chart data;

[0078] Output of Subtask 2: List of cross-departmental project deliverables and key milestones;

[0079] Output of Subtask 3: A three-year development path framework based on market forecasts;

[0080] The integration engine converts the intermediate results into structured document elements and generates task response text:

[0081]

[0082] Send a response message including the task response text. Figure 2 (B) in the figure displays the task response text 221.

[0083] In the document generation scenario of a multi-agent collaborative platform, the method provided in the embodiment of the present application decomposes complex document creation tasks into sub-task modules that can be described by quantifiable vectors through semantic analysis, and utilizes the specialized division of labor mechanism of multiple agents (data analysis agents focus on numerical processing, results summarization agents are good at information structuring, and planning strategy agents specialize in logical framework construction) to achieve precise resource scheduling in the document generation process. Each agent ensures the adaptability of task requirements and execution capabilities based on the dynamic matching of encoding vectors and capability vectors; through the standardized output of intermediate results and a unified integration engine, logically coherent and formatted professional documents are generated, effectively solving the pain points of content fragmentation and low efficiency of traditional document tools when processing multi-dimensional and long-term reports, and significantly improving the accuracy and automation level of enterprise-level complex document generation.

[0084] In one possible implementation, the method provided in the embodiment of the present application can be applied in a customer service robot. Figure 3 As shown, Figure 3 The interface diagram of the customer service robot is shown in FIG. Figure 3 In (A), the user enters the task request text: "My newly purchased smart device cannot connect to WiFi. How can I quickly repair it? Also, how can I apply for the three-year warranty?" 311.

[0085] like Figure 3 In (C), we first receive the task request text "My newly purchased smart device cannot connect to WiFi. How can I quickly repair it? Also, how can I apply for a three-year warranty?" The task is decomposed based on the semantic recognition module. As shown in Table 5, the decomposition yields two subtasks and generates corresponding encoding vectors, as shown in Table 5:

[0086] Table 5

[0087] Subtask number Subtask content Subtask encoding vector Subtask 1 Equipment fault diagnosis [0.83,0.92,0.12,0.95] Subtask 2 Warranty Policy Analysis [0.78,0.85,0.63,0.71]

[0088] Obtain the preset M customer service agents and their capability vectors (see Table 6). The vector dimensions are mapped to skills such as knowledge retrieval, process guidance, and communication expression:

[0089] Table 6

[0090] Agent Role Capability Vector Troubleshooting Agent [0.94,0.75,0.08,0.98] Policy Service Agent [0.87,0.91,0.79,0.82] Emergency Response Agent [0.29,0.95,0.93,0.81] General Coordination Agent [0.62,0.68,0.77,0.54]

[0091] According to the matching mechanism between subtask vector and capability vector, the target agent allocation is determined:

[0092] Fault diagnosis agent processing subtask 1, execution result: output network configuration self-check steps and driver update plan;

[0093] Policy service agent processes subtask 2. Execution result: Output warranty scope description and online application link.

[0094] Convert subtask results into natural language response text:

[0095]

[0096] Send a response message including the task response text. Figure 3 (B) in the example shows a task response text 321 .

[0097] In the customer service robot scenario, the method provided in the embodiment of this application breaks down complex customer service requests into independently executable subtasks (such as troubleshooting and policy consultation) through semantic parsing, and uses a vectorized matching mechanism to enable professional intelligent agents to accurately respond to segmented needs: the fault diagnosis agent generates targeted repair solutions based on device protocol knowledge, and the policy service agent accurately extracts warranty terms and generates process guidance. Through the standardized integration of multi-agent output results, a clearly structured and highly operational solution is formed, which solves the defects of traditional customer service systems in responding disjointedly and fragmenting information when facing multi-level problems, and significantly improves the processing efficiency and user experience of complex consultations.

[0098] In a possible implementation, the method provided in the embodiment of the present application can be applied in image content review. Figure 4 As shown, Figure 4 The following figure shows the interface diagram of the image content review product. Figure 4 In (A), the user enters the task request text: "Please help me review whether these images meet the publishing specifications." 411.

[0099] like Figure 4In (C), we first receive the task request text "Please help me review whether these images meet the release specifications." Based on the semantic recognition module, we decompose the task into three subtasks and generate corresponding encoding vectors, as shown in Table 7:

[0100] Table 7

[0101] Subtask number Subtask content Subtask encoding vector Subtask 1 Sensitive content identification [0.83,0.92,0.12,0.95] Subtask 2 Copyright compliance verification [0.78,0.85,0.63,0.71] Subtask 3 Image quality standard assessment [0.35,0.97,0.88,0.59]

[0102] Get the preset M audit agents, whose capability vectors are mapped to skill dimensions such as image analysis and rule base matching (as shown in Table 8):

[0103] Table 8

[0104] Agent Role Capability Vector Features Sensitive content detection agent [0.94,0.75,0.08,0.98] Copyright tracing agent [0.87,0.91,0.79,0.82] Quality Assessment Agent [0.29,0.95,0.93,0.81] Comprehensive judgment agent [0.62,0.68,0.77,0.54]

[0105] According to the dynamic matching of subtask vectors and capability vectors, the target agent allocation is determined:

[0106] Sensitive content detection agent processes subtask 1 → outputs violation marks (such as the coordinates of the location of the violation element);

[0107] Copyright tracing agent processing subtask 2 → output copyright matching score and suspected source link;

[0108] The quality assessment agent processes subtask 3 → outputs a resolution compliance report and blur area annotations;

[0109] Transform multi-source results into structured audit reports:

[0110]

[0111] Send a response message including the task response text. Figure 4 (B) in the example shows a task response text 421.

[0112] In image content review scenarios, semantic parsing breaks down the image review task into three specialized subtasks: violation identification, copyright verification, and quality assessment. A vector matching mechanism enables specialized intelligent agents to precisely respond to specific needs: a sensitive content detection agent rapidly locates offending elements, a copyright traceability agent accurately compares image feature libraries, and a quality assessment agent objectively quantifies technical parameters. By standardizing and integrating intermediate results, a comprehensive report containing risk identification and compliance recommendations is generated, addressing the shortcomings of traditional manual review in complex image scenarios, such as low efficiency, strong subjectivity, and high rates of missed detections.

[0113] See also Figure 5 , Figure 5The following diagram shows an architecture diagram of a multi-agent collaborative task planning method provided in an embodiment of the present application applied to an intelligent assistant scenario. The scenario includes a user terminal 510, a network 520, and a server 530.

[0114] Terminal 510 includes a human-computer interaction screen, a processor, and memory. The human-computer interaction screen is used to display the intelligent assistant platform's interface image 5101, task request text 5102, and task response text 5103. It also provides a human-computer interaction interface to receive user-entered human-computer interaction operations such as task request text. The processor is used to generate interaction instructions in response to these human-computer interaction operations and send these interaction instructions to the server.

[0115] The terminal 510 involved in this application includes but is not limited to mobile phones, tablet computers, laptop computers, desktop computers, intelligent voice interaction devices, virtual reality devices, smart home appliances, vehicle-mounted terminals, aircraft, etc.

[0116] A client 511 is running in the terminal 510. The client 511 takes an intelligent assistant client as an example. The client 511 is deployed on the terminal 510. The client 511 can be run on the terminal 510 in the form of a browser, or in the form of an independent application (APP) or a small program.

[0117] Network 520 utilizes standard communication technologies and / or protocols, typically the Internet, but may be any network, including but not limited to Bluetooth, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a mobile network, a private network, or any combination of virtual private networks. In some embodiments, customized or proprietary data communication technologies may be used in place of or in addition to the aforementioned data communication technologies.

[0118] The server 530 includes a processor. The server 530 involved in this application can be an independent physical server, or a server cluster or distributed system composed of at least one physical server, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence (AI) platforms.

[0119] A user enters a task request via a terminal: "Please arrange a business trip from location A to location B, including air ticket booking, hotel recommendations, and meeting reminder settings." 5102. In step S501, terminal 510 sends the task request text to server 530. In step S502, server 530 receives the task request text: "Please arrange a business trip from location A to location B, including air ticket booking, hotel recommendations, and meeting reminder settings." In step S503, server 530, based on semantic recognition, splits the task in the task request text into three subtasks, each corresponding to a subtask encoding vector. In step S504, server 530 obtains M predefined agents with specific capabilities. Each agent corresponds to a functional module, and each functional module corresponds to a capability vector. In step S505, server 530 determines three target agents for handling the three subtasks based on the three subtask encoding vectors and the M capability vectors. In step S506, server 530 invokes the three target agents to execute the three subtasks, respectively, obtaining the three subtask processing results. In step S507, server 530 integrates the three subtask processing results to generate a task response text. In step S508, server 530 sends a response message including the task response text to terminal 510. Terminal 510 displays task response text 5103.

[0120] The method provided in the embodiment of the present application accurately decomposes the user's business trip task into three sub-tasks: flight booking, hotel recommendation, and schedule setting through semantic recognition, and uses vectorized coding technology to quantify the core requirement characteristics of each sub-task. Combined with the quantitative ability vectors of predefined intelligent agents in multi-dimensional skill fields, dynamic matching of tasks and resources is achieved, effectively ensuring that the professional capabilities of flight agents in real-time ticketing processing, hotel agents in matching geographic locations with user preferences, and calendar agents in cross-platform schedule synchronization are accurately adapted to the requirements of their respective sub-tasks; during the task execution process, structured intermediate result collection and natural language generation technology are used to achieve seamless integration of multi-agent output results, and ultimately generate a user-friendly response text containing a complete itinerary, price details, and location-related information, significantly improving the processing efficiency and execution reliability of complex multi-module collaborative tasks.

[0121] See also Figure 6 , Figure 6 A flowchart of a multi-agent collaborative task planning method is shown. It should be noted that the multi-agent collaborative task planning method provided in the embodiment of the present application can be applied to a server, and the embodiment of the present application is not limited thereto. The method includes:

[0122] S610: Receive a task request text.

[0123] It can be understood that a task request text refers to a description of task requirements entered by the user through the interface, such as the user's needs expressed in natural language. Initial task input is received, enabling the system to begin processing user intent. For example, in an intelligent assistant scenario, a task request text might be, "Please arrange a business trip from location A to location B, including airfare reservations, hotel recommendations, and meeting reminders." In this case, the task in the task request text is travel arrangement. In a multi-agent collaboration platform scenario, a task request text might be, "Please generate a company annual summary document covering performance analysis, team achievement presentation, and three-year business plan." In this case, the task in the task request text is document generation. In a customer service robot scenario, a task request text might be, "My newly purchased smart device can't connect to WiFi. How can I quickly fix it? Also, how can I apply for a three-year warranty?" In this case, the task in the task request text is after-sales service. In an image content review scenario, a task request text might be, "Please review these images to see if they meet publishing specifications." In this case, the task in the task request text is image recognition content review.

[0124] S620: Based on semantic recognition, split the task in the task request text to obtain P subtasks.

[0125] Among them, P subtasks correspond to P subtask encoding vectors, P ≥ 1.

[0126] It is understandable that semantic recognition refers to the deep understanding of the underlying intent and logical structure of task request text through natural language processing technology. Its core lies in converting human language into structured semantic representations that can be operated by machines. Specifically, it includes three layers of meaning:

[0127] 1. Intent parsing and logical deconstruction: Pre-trained language models (such as BERT and GPT) are used to capture key entities, actions, and relationships in the task. For example, for the text "Analyze the air quality in Location A over the past five years and predict trends," the system needs to identify: the main task objective (air quality analysis + trend prediction), the constraints (region: Shanghai; time: the past five years), and the implicit requirements (requires associated statistical analysis and time series prediction capabilities). Ultimately, it decouples the complex task into independently executable atomic operation units (such as "Data collection: Air quality indicators in Location A from 2019 to 2023," "Statistical modeling: Indicator trend fitting," and "Forecast output: Trend chart for the next three years").

[0128] 2. Semantic vector generation mechanism:

[0129] The subtask encoding vector is a mathematical representation of semantics, which is usually generated through the following steps:

[0130] 1) Use sentence embedding models (such as Sentence-BERT) to map subtask description text into high-dimensional vectors;

[0131] 2) Preserving semantic similarity in vector space (e.g., the vectors for “detecting illegal content in images” and “identifying sensitive visual elements” are close);

[0132] 3) Injecting domain knowledge (e.g., the “image quality standard assessment” vector in image content review will be associated with the image resolution skill dimension).

[0133] 3. Dynamic adaptability:

[0134] The system can adaptively handle two types of key scenarios: 1) fuzzy statement analysis: for example, "handle complaints as quickly as possible" automatically breaks down into the chain subtasks of "complaint classification → priority sorting → customer service assignment"; 2) cross-domain connection: for example, "preparing a smart home market competitor report" simultaneously triggers heterogeneous subtasks such as "patent analysis", "product parameter collection", and "user review sentiment analysis".

[0135] Semantic recognition enables machines to go beyond keyword matching and understand task context like humans.

[0136] Subtasks are decomposed, independent operational units. Subtask encoding vectors are digital representations generated by a semantic processing model that capture the semantic characteristics of the task. For example, in an intelligent assistant scenario, semantic recognition can be used to decompose the task request text "Please help me arrange a business trip from location A to location B, including air ticket reservations, hotel recommendations, and meeting reminder settings" into three subtasks: booking a flight, booking a hotel, and setting a calendar reminder. The encoding vector for booking a flight is [0.83, 0.92, 0.12, 0.95], the encoding vector for booking a hotel is [0.78, 0.85, 0.63, 0.71], and the encoding vector for setting a calendar reminder is [0.35, 0.97, 0.88, 0.59].

[0137] By using semantic recognition technology to analyze task structures, complex tasks are broken down into logically related but independently executable modules. This improves task processing efficiency. By breaking down complex problems into units that can be executed in parallel, overall processing time and resource usage are reduced, while also improving task adaptability because each subtask can be optimized independently.

[0138] S630. Obtain M intelligent agents.

[0139] Among them, M agents correspond to M capability vectors, and the capability vectors are used to represent the quantitative performance of the agent in K skill dimensions, M≥P≥1, K≥1.

[0140] It can be understood that the agent represents an independent processing unit, and the capability vector is a numerical sequence that maps the skill level of the agent in multiple dimensions.

[0141] Several types of agent roles (Role-based Agents) with specific capabilities are pre-defined, and each role has different functional modules and skill levels. Each agent role is also equipped with a capability vector, which represents its skill performance in different dimensions (such as information acquisition, logical reasoning, interface calls, etc.). In a multi-agent system (MAS), each agent is assigned a role based on its functions and capabilities. Role definition is key to the successful operation of the system. Each role represents the specific function or task processing capability of the agent. For example, in the intelligent assistant scenario, the pre-defined agent roles include:

[0142] Flight Agent: Responsible for flight booking.

[0143] Hotel Agent: Responsible for hotel recommendations and reservations.

[0144] Calendar Agent: responsible for arranging schedule reminders.

[0145] The capabilities of each agent role are represented by its capability vector, which includes the agent's capabilities in multiple dimensions, such as information acquisition, logical reasoning, and interface calls. The capability vector Ci can be expressed as:

[0146] Ci=[IRi,LRi,IIi];

[0147] in:

[0148] IRi: represents the agent's ability in information retrieval.

[0149] LRi: represents the agent's ability in logical reasoning.

[0150] IIi: Indicates the agent's ability in interface interaction.

[0151] By loading the agent library with predefined capability dimensions, resource management is optimized, ensuring a dynamic match between agent capabilities and task requirements and avoiding redundant deployment. This significantly improves system scalability and flexibility, especially in multi-skill scenarios, and adapts to changing load demands.

[0152] S640: Determine P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors.

[0153] It is understood that the target agent refers to the agent selected to perform a specific subtask. A vector matching algorithm, such as calculating similarity, is used to select the most suitable agent. Specifically, the matching is performed by calculating the similarity between the encoding vector of the subtask and the ability vector of the agent.

[0154] Preferably, a matching algorithm that combines semantic matching and capability scoring can be used when performing matching. Given a subtask tj and an agent ai, each subtask tj has an encoding vector Tj, and each agent ai has a capability vector Ci. The matching algorithm can be expressed as:

[0155] MatchScore(tj,ai)=α·cos(Tj,Ci)+(1-α)·Scapability(ai);

[0156] in:

[0157] cos(Tj,Ci) is the cosine similarity between the subtask encoding vector Tj and the agent capability vector Ci.

[0158] Scapability(ai) is the comprehensive ability score of the intelligent agent ai, which indicates its ability to complete the task.

[0159] α∈[0,1] is a weight coefficient used to balance the influence of semantic matching and ability score.

[0160] Automatic matching reduces manual intervention and error rates, thereby enhancing the accuracy and robustness of task processing.

[0161] S650: Call P target agents to execute P subtasks, integrate the P subtask processing results obtained by the P target agents executing the P subtasks, and generate a task response text.

[0162] As you can understand, invocation initiates the execution process, subtask processing results are output by each agent, and integration combines these results into a coherent text response. P target agents are invoked to execute P subtasks in parallel, obtaining P subtask processing results. These P subtask processing results are then integrated to generate a natural language task response text. This integration process avoids information silos and ensures comprehensive responses. Concurrent execution also accelerates overall task completion and improves system throughput.

[0163] S660: Send response information including the task response text.

[0164] As you can understand, the response information is the output containing the final conclusion, and sending refers to transmitting it back to the user through the interface. This achieves a closed-loop interaction with the user, analyzing their role as the coordinator at the end of the process. This concise output simplifies understanding, especially in high-frequency tasks, effectively reducing wait times and enhancing user experience and system credibility.

[0165] For example, see Figure 7 , Figure 7 The flowchart shows the image recognition content review scenario of the manual annotation assistance system. The user enters the task request text "Please help me review whether these images meet the publishing specifications." After receiving the request, the task is split into three subtasks based on semantic recognition, including sensitive element detection, copyright verification, and quality assessment. Each subtask generates a corresponding semantic vector. The preset agent library loads multiple capability vectors covering image analysis skill dimensions, such as object detection and similarity calculation capabilities. Specific agents are assigned to perform each subtask through vector matching. For example, the sensitive content detection agent handles subtask 1 (sensitive element detection) and outputs violation marks (such as the location coordinates of the violating element); the copyright tracing agent handles subtask 2 (copyright verification) and outputs a copyright matching score and suspected source links; and the quality assessment agent handles subtask 3 (quality assessment) and outputs a resolution compliance report and fuzzy area annotations. These outputs are integrated to generate a text report containing violation warnings and improvement suggestions. Finally, the report is sent back to the user interface as a response message, providing an intuitive review conclusion.

[0166] The method provided in the embodiment of the present application achieves efficient and accurate multi-agent resource allocation through intelligent parsing splitting and dynamic matching mechanisms, reducing human errors and response delays; the system's natural language interaction simplifies operational complexity, and the integration mechanism ensures output consistency, especially improving reliability and scalability in complex scenarios, and ultimately enhancing user experience with closed-loop feedback to support the large-scale deployment of high-throughput applications.

[0167] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, determining P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors includes:

[0168] Based on the P subtask encoding vectors and M capability vectors, the semantic matching values ​​between the P subtasks and the M agents are calculated to obtain P×M semantic similarity values.

[0169] Based on the scores of each agent in the K skill dimensions, the ability values ​​of the M agents to perform P subtasks are calculated to obtain P×M skill dimension values;

[0170] Based on the P×M semantic similarity values ​​and P×M skill dimension values, calculate the M matching values ​​corresponding to each of the P subtasks;

[0171] According to the M matching values ​​corresponding to each of the P subtasks, P target intelligent agents for processing the P subtasks are determined, wherein the matching value corresponding to the target intelligent agent is the maximum value among the M matching values.

[0172] It's understandable that semantic similarity represents the strength of the association between subtask requirements and agent capability descriptions in semantic space. Essentially, it measures geometric relationships, such as cosine similarity, between semantic encoding vectors and capability vectors using a vector space model. For example, a subtask vector representing "visual anomaly detection" might match an agent with a capability vector labeled "image defect recognition." Despite differing textual descriptions, the semantics closely match. This improves the generalization of matching, making it particularly robust to the diversity of user task descriptions (such as synonyms and term variants).

[0173] The skill dimension value reflects the quantitative ability level of the agent in the professional field required for a specific subtask. By traversing the K skill dimensions of each agent (such as natural language processing, image analysis, data modeling, etc.), the comprehensive ability score is calculated weighted by the subtask requirements.

[0174] The matching value serves as the final decision-making basis, generating a composite indicator by integrating semantic similarity values ​​with skill dimension values. A normalization algorithm balances the influencing factors of the two heterogeneous data types (for example, assigning higher weight to skill dimension values ​​to ensure professionalism). By constructing a multi-objective optimization mechanism, we retain the advantages of semantic intent understanding while maintaining the rigid capacity constraints of the skill dimension, addressing the bias that can arise from a single indicator.

[0175] Finally, agents are selected based on the matching scores of the agents corresponding to each subtask. For each subtask, only the single agent with the highest matching score is selected to execute. This avoids the increased coordination costs caused by multiple agents working on the same subtask in parallel, optimizing resource utilization while ensuring matching quality.

[0176] For example, in the scenario of intelligent assistant travel arrangement, for the subtask "book a flight", its semantic vector Tj is compared with the capability vectors of each agent, and the scores are as shown in Table 9:

[0177] Table 9

[0178] Agent Role Semantic similarity value Skill dimension value Matching Value Flight Agent 0.92 0.95 0.93 Hotel Agent 0.35 0.88 0.56 Calendar Agent 0.28 0.92 0.54

[0179] Finally, the flight agent with the highest score is selected to perform the subtask.

[0180] For example, a matching algorithm that combines semantic matching and capability scoring can be used when performing matching. Given a subtask tj and an agent ai, each subtask tj has an encoding vector Tj, and each agent ai has a capability vector Ci. The matching algorithm can be expressed as:

[0181] MatchScore(tj,ai)=α·cos(Tj,Ci)+(1-α)·Scapability(ai);

[0182] in:

[0183] cos(Tj,Ci) is the cosine similarity between the subtask encoding vector Tj and the agent capability vector Ci.

[0184] Scapability(ai) is the comprehensive ability score of the intelligent agent ai, which indicates its ability to complete the task.

[0185] α∈[0,1] is a weight coefficient used to balance the influence of semantic matching and ability score.

[0186] In the scenario of intelligent assistant travel arrangement, for the subtask "book a flight", the subtask encoding vector is TFlightBooking = [0.85, 0.75, 0.6], and the flight agent's capability vector is represented as CFlightAgent = [0.8, 0.9, 0.7]. Calculate the cosine similarity:

[0187]

[0188] The agent's capability score Scapability(ai) is calculated as follows:

[0189]

[0190] Where: C i，k is the score of the agent ai on the kth skill dimension. k is the weight of each skill dimension, reflecting the importance of the skill in performing the subtask.

[0191] Agent role definition and capability modeling clarify each agent's capabilities and task requirements, enabling more precise task allocation. Using a weighted matching algorithm based on semantic similarity and capability scores, the system selects the most appropriate agent for each subtask, improving system efficiency and accuracy. The capability scoring mechanism helps continuously evaluate and optimize agent performance, enabling the system to flexibly respond to dynamically changing environments.

[0192] The combination of semantic similarity and capability-scoring weighted matching embodies several innovations. First, a dynamic weight coefficient α is introduced to automatically adjust the importance of semantics and capabilities in matching based on task requirements and environmental changes, enhancing the system's flexibility. Second, capability scoring employs a multidimensional comprehensive model that considers the agent's performance across multiple skill dimensions, improving the accuracy of task assignment. Furthermore, by enhancing the interpretability of capability-task matching, the system provides a transparent basis for matching, making it easier for users and developers to understand the decision-making process. Furthermore, the capability transfer mechanism allows agents to switch roles across tasks, enhancing adaptability. Finally, combined with a dynamic feedback mechanism, the system can continuously optimize task assignment strategies based on user feedback, achieving self-adjustment and long-term optimization. These innovations effectively improve the system's efficiency, accuracy, and robustness in complex and dynamic task scenarios.

[0193] See also Figure 8 , Figure 8 A schematic diagram of the agent-task matching algorithm is shown. The initial stage receives four input parameters: P subtask encoding vectors and P skill dimension weight sets on the task side, and M capability vectors and M skill dimension scores on the agent side. Subsequently, through two parallel processing channels, the semantic matching calculation channel generates a P×M semantic similarity value matrix based on the cosine similarity between the subtask vectors and the capability vector. The capability assessment channel weightedly aggregates the agent skill scores according to the task weights to generate a P×M skill dimension value matrix. These two matrices are then normalized and fused, and the final P×M matching value matrix is ​​synthesized through linear weighting. In the final stage, the maximum matching value index of each subtask row vector is taken, and the P optimal target agents are determined and output, forming a dynamic allocation scheme that balances the depth of semantic understanding and the accuracy of professional capabilities.

[0194] The method provided in the embodiment of the present application effectively eliminates the challenges of ambiguity and diversity in natural language descriptions through semantic matching, while capability scoring builds a solid defense mechanism for professional and technical thresholds; the weighted fusion of the two avoids the risk of insufficient capabilities that may be caused by pure semantic matching, and prevents the deviation in intention understanding caused by mechanical reliance on skill scoring; ultimately, the intelligent agent decision-making based on optimized sorting forms a flexible resource scheduling capability, which can significantly improve the task success rate and reduce the error tolerance rate.

[0195] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, the semantic matching values ​​between the P subtasks and the M agents are calculated based on the P subtask encoding vectors and the M capability vectors to obtain P×M semantic similarity values, including:

[0196] Calculate the cosine similarity between the P subtask encoding vectors and the M capability vectors to obtain P×M semantic similarity values.

[0197] It can be understood that the semantic similarity value refers to the quantitative correlation between the subtask encoding vector and the agent capability vector obtained by the cosine similarity algorithm. The essence of this calculation is to measure the directional consistency of two vectors in a high-dimensional semantic space. When the angle between the two vectors approaches zero, the semantic similarity value approaches the theoretical maximum value. By converting human language descriptions into computable geometric relationships through mathematical modeling, the expression difference problem caused by synonyms, near-synonyms and term variants is effectively overcome. It reduces the dependence on manual annotation and enables the system to automatically identify semantically equivalent non-complete matching expressions, thereby improving the generalization and adaptability of agent matching, especially when dealing with complex industry terminology or multi-language mixed scenarios, significantly reducing communication costs.

[0198] For example, in the scenario of intelligent assistant travel arrangement, for the subtask "book a flight", the subtask encoding vector is TFlightBooking = [0.85, 0.75, 0.6], and the flight agent's capability vector is represented as CFlightAgent = [0.8, 0.9, 0.7]. The cosine similarity is calculated as follows:

[0199]

[0200] For example, see Figure 9 , Figure 9 The diagram shows the similarity calculation process. Inputs are a matrix consisting of P subtask encoding vectors and a matrix consisting of M agent capability vectors. The core calculation module then performs matrix-based batch calculations on the two vector sets using the cosine similarity function, outputting a semantic similarity matrix with P rows and M columns. The final output port presents a structured semantic similarity matrix, whose row vectors correspond to subtask indices and column vectors to agent identifiers. Each element in the matrix, ranging from [-1, 1], represents the strength of the semantic association between a specific subtask and the agent, completing the computational chain from vector input to quantitative matching.

[0201] The method provided in the embodiment of the present application converts fuzzy semantic associations into continuous numerical values ​​that can be quantified and compared, avoiding the waste of resources caused by binary matching; at the same time, it ensures the accuracy of intention understanding in multilingual and cross-industry scenarios, enables the intelligent agent to respond to needs at the semantic essence level, greatly enhances the robustness of the system in processing non-standardized task requests, and provides basic computational support for efficient decision-making in dynamic environments.

[0202] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, the ability values ​​of the M agents to perform P subtasks are calculated based on the scores of each agent in K skill dimensions, and P×M skill dimension values ​​are obtained, including:

[0203] Obtain M skill dimension subsets corresponding to the M agents, where the skill dimension subset includes K skill scores, which are used to represent the quantitative performance of the agents in the skill dimension;

[0204] Obtain a set of P skill dimension weights corresponding to the P subtasks, where the set of P skill dimension weights includes K skill dimension weights, and the skill dimension weights are used to represent the importance of the skill dimension when the agent performs the subtask;

[0205] Based on the M skill dimension subsets and the P skill dimension weight sets, P×M skill dimension values ​​are calculated.

[0206] It can be understood that each agent includes K skill dimensions, each skill dimension corresponds to a skill dimension score, and each agent corresponds to a skill dimension subset set, which includes K skill dimension scores corresponding to K skill dimensions. The skill dimension subset set is essentially a quantitative representation of the agent's ability dimension, and each skill score represents a quantitative assessment result in a specific professional field. These scores are typically derived from historical task performance records or professional ability certifications, forming a multi-dimensional ability matrix. Structured data storage enables the computational management of agent capabilities. For example, a data analysis agent has scores of 8.7, 6.3, and 9.2 in the three dimensions of "statistical modeling," "real-time computing," and "visualization," respectively. This converts abstract capabilities into machine-processable discrete variables, providing a data foundation for precise decision-making.

[0207] The skill dimension weights define the contribution of each skill to completing each subtask. Weight assignments must be consistent with the essential characteristics of the task. For example, in the medical image analysis subtask, "pathology feature identification" has a weight of 0.85, while "report formatting" has a weight of only 0.15. This ensures that demand-side and supply-side capabilities are aligned, avoiding resource mismatches caused by high-capability agents being assigned to less relevant tasks and ensuring that key capability elements are prioritized.

[0208] Using a weighted fusion algorithm, centered around a dot product operation, the agent's skill scores are multiplied by the set of task weights, then added up. For example, when a subtask requires high timeliness, even if an agent's professional score is excellent but its response speed weight falls short, its overall dimensional score will be significantly reduced. This has the beneficial effect of establishing the ability to search for optimal solutions under multiple constraints. Through mathematical operations, it automatically eliminates agents that don't meet core requirements, creating a hard filter for capabilities.

[0209] For example, the agent's capability score Scapability(ai) is calculated as follows:

[0210]

[0211] Where: Ci，k is the score of the agent ai on the kth skill dimension. k is the weight of each skill dimension, reflecting the importance of the skill in performing the subtask.

[0212] See also Figure 10 , Figure 10 The figure shows a schematic diagram of the skill dimension value calculation process. The input channel receives M sets of skill dimension subsets (each set contains K skill dimension scores) for each agent, which are mapped to form an M-row, K-column agent skill matrix. Simultaneously, the input channel also receives P sets of skill dimension weights for each subtask (each set contains K weight coefficients), which are mapped to form a P-row, K-column task weight matrix. These two matrices enter the multiplication module (labeled with the weighted summation formula), where element-by-element multiplication and dimensional summation operations generate the final P-row, M-column skill dimension value matrix. Each element value in the output matrix represents a quantitative score of the comprehensive ability of a specific agent to perform a specific subtask, completing the conversion process from raw ability data to task-oriented ability assessment.

[0213] The method provided in this application embodiment uses skill sets to establish objective capability benchmarks on the supply side, while weight sets define capability priority rules on the demand side. These two are mathematically integrated to generate a scalable decision-making basis. This design eliminates the subjective bias of manual evaluation while breaking through the rigid limitations of fixed matching rules. It can automatically identify the agent resources that best meet the essential requirements of the task, significantly improving the success rate and timeliness of key tasks.

[0214] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, after calling P target agents to perform P subtasks, the method further includes:

[0215] Perform load detection on P target agents to obtain P load information and P response times corresponding to the P target agents executing P subtasks;

[0216] When the load information corresponding to the first target intelligent agent among the P target intelligent agents is greater than the load threshold, or the response duration corresponding to the first target intelligent agent is greater than the duration threshold, the first backup intelligent agent is determined from the MP intelligent agents, wherein the matching value corresponding to the first backup intelligent agent is the second largest value among the M matching values.

[0217] The first backup agent is called to execute the subtask executed by the first target agent.

[0218] It's understood that load information refers to real-time metrics such as the number of concurrently processed tasks and resource utilization by the agent, while response time refers to the time between receiving a task and returning preliminary results. By continuously collecting runtime data, such as monitoring agent memory usage and CPU peaks every 300 milliseconds, and establishing a preventative monitoring mechanism, an alert is automatically triggered when an agent simultaneously processes tasks exceeding a threshold, preventing system collapse caused by overload and ensuring critical business continuity.

[0219] The first backup agent represents the second-best candidate for capability matching. The system dynamically extracts suboptimal technical paths from the original matching value matrix. For example, if the primary agent times out due to hardware failure, the system immediately calls a backup agent whose semantic similarity and capability dimension scores reach 92% of the original matching value. This creates seamless failover capabilities and reduces decision latency by pre-calculating alternative solutions. In real-time systems like autonomous driving, switching times can be compressed to hundreds of milliseconds.

[0220] Agent switching involves the dynamic replacement and state synchronization of running instances. For example, in a video content review scenario, if the primary agent times out after completing 80% of sensitive frame detection, the backup agent can continue analyzing the remaining frames based on the shared detection history. This ensures the atomicity and integrity of task execution, eliminates data consistency issues caused by node failures, and maintains the credibility of the results.

[0221] For example, during operation, if an agent is overloaded or responds late, the system will trigger a dynamic reallocation mechanism to reschedule tasks in the following ways:

[0222] 1) Load balancing detection: monitor the number of tasks currently processed by each agent and the response time.

[0223] 2) Backup role switching: If the primary role is unavailable, the suboptimal matching role will be enabled to take over the task.

[0224] 3) Priority queue management: Dynamically adjust the execution order according to task urgency and user feedback.

[0225] See also Figure 11 , Figure 11 The diagram shows the dynamic adjustment and resource scheduling process. Multiple target agents execute subtasks in parallel, continuously collecting load metrics (CPU utilization, memory usage, and number of concurrent tasks) and response time data from each node. When any agent's load metric exceeds a preset threshold or its response time times out, the system automatically marks the node as faulty and activates the backup resource pool. The initial matching value calculation results are then retrieved to locate and activate the backup agent with the second-best matching value (i.e., the suboptimal matching role). Finally, state snapshot synchronization enables seamless takeover, ensuring continuous task execution.

[0226] The method provided in this application embodiment achieves proactive awareness of resource bottlenecks through real-time load monitoring; ensures manageable service degradation through rapid switching based on pre-ordered alternatives; and maintains continuous execution through the integration of task state synchronization technology. These elements work together to overcome the limitations of traditional failover strategies, avoiding the response lag of polling detection while addressing the computational overhead of full rematching, significantly improving task success rates.

[0227] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, based on semantic recognition, the task in the task request text is split into P subtasks, including:

[0228] Perform semantic recognition on the task request text to obtain task request semantic information;

[0229] Perform entity analysis on the semantic information of the task request to obtain P subtasks.

[0230] It's understandable that task request semantic information refers to the structured intent expression parsed by a deep language model. Essentially, it's an abstract reconstruction of the user's original description. By identifying the core action objects and their modifiers in the text, for example, "Please arrange a business trip from location A to location B, including airfare reservations, hotel recommendations, and meeting reminders," it can be parsed into a structured semantic framework encompassing temporal and spatial scope, comparison subjects, and data visualization requirements. This eliminates ambiguity and redundant information in natural language, laying the cognitive foundation for precise task decomposition. This ensures consistent understanding of intent, especially when dealing with industry abbreviations or polysemous terms.

[0231] The entity analysis phase focuses on extracting executable atomic units, or subtasks, from the semantic framework. This analysis involves traversing the syntactic dependency tree of semantic information and decoupling the action chain. For example, within a travel arrangement task, three independent subtasks—"book a flight," "book a hotel," and "set a schedule reminder"—can be separated. This algorithmic decomposition transforms complex requirements into modular units that can be operated in parallel, ensuring the integrity of business logic while improving system resource utilization efficiency.

[0232] See also Figure 12 , Figure 12 The following diagram shows the process flow for subtask splitting. The task request text is received from the user, and natural language processing techniques are used to parse the task text, extracting the core intent and contextual information, generating structured task request semantic information. This semantic information is then deeply processed to identify and decompose action entities, object entities, and dependency relationships. Finally, the task is decoupled and output into P atomic subtasks, completing the modular splitting of the task.

[0233] The method provided in the embodiment of the present application uses semantic recognition to construct a machine-understandable intent expression space, overcoming the discreteness and ambiguity of human language expression; entity analysis implements domain-driven structured cutting in this space to generate atomic operation units that conform to the intelligent agent capability model. The two-stage collaboration not only retains the complete business context of the original requirements, but also achieves precise decoupling and allocation of execution resources. When faced with complex scenarios such as cross-time zone collaboration and multilingual conversion, it can avoid systemic coupling risks through modular decomposition, while improving task parallelism while ensuring the independence of each subtask execution context, providing scalable underlying logical support for large-scale intelligent agent collaboration.

[0234] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, the method further includes:

[0235] According to the semantic information of the task request, the relationship between the P subtasks is analyzed to obtain the execution order relationship of the P subtasks;

[0236] Generate a task graph based on the relationship between the P subtasks and the execution order of the P subtasks, wherein the task graph includes P nodes and edge relationships between the P nodes, the P nodes correspond to the P subtasks, and the edge relationship between the P nodes corresponds to the relationship between the execution order of the P subtasks;

[0237] Call P target agents to perform P subtasks, including:

[0238] According to the task graph, P target agents are called to perform P subtasks.

[0239] It's understood that the execution order relationship refers to the temporal logic and data dependency rules that must be followed between subtasks. Semantic role labeling and contextual reasoning techniques are used to analyze the predecessor and successor constraints of the action chain. For example, when analyzing the task "recommend a hotel and set a calendar reminder after booking a flight," the system identifies that the date and location information generated by the flight booking constitutes an input dependency for the hotel recommendation, and both results serve as prerequisites for generating the reminder. This ensures strict compliance of business processes, avoids data gaps or logical errors caused by out-of-order execution, and eliminates business risks caused by out-of-order execution in critical scenarios such as financial transactions.

[0240] During the task graph construction phase, abstract sequential rules are converted into a graph data structure. Nodes are mapped to specific subtask entities, and directed edges represent execution order constraints. This step requires differentiating between strong and weak dependencies. For example, a strong dependency (forbidding hotel recommendations when flights are undecided) is represented by a solid edge, while a weak dependency (allowing concurrent attraction searches when recommending hotels) is marked with a dashed edge.

[0241] The graph-driven scheduling process dynamically triggers agent calls based on the topological sorting of the task graph. Figure 13 , Figure 13 The task diagram for travel arrangements is shown. Figure 13 In (A), an event-driven mechanism is used to monitor node state changes. For example, only when the flight agent completes the flight booking subtask will the hotel agent be activated to start the hotel booking subtask. When the hotel booking subtask is completed, the calendar agent will be activated to start the schedule reminder subtask. Figure 13 In step (B), the flight agent and hotel agent execute their respective subtasks simultaneously. Only after both agents complete their respective subtasks (the flight agent executes the flight reservation subtask, and the hotel agent executes the hotel reservation subtask) does the calendar agent activate to begin the schedule reminder subtask. This achieves an optimal balance between resource utilization and execution efficiency, shortening task completion time in manufacturing process control while ensuring data integrity.

[0242] See also Figure 14 , Figure 14 The flowchart of the task graph generation process is shown. It begins by parsing the semantic information of the task request, extracting P atomic operation units through subtask relationship analysis, and then determining the execution order dependencies between them. Next, a task graph structure is constructed, consisting of P nodes (representing subtasks) and directed edges (representing the execution order). During the drive execution phase, the system traverses the task graph nodes in topological order. When it detects that a node's pre-dependency meets the ready condition, it immediately calls the matching target agent to execute the subtask. After execution is completed, the task graph status is automatically updated, and the next ready node is cyclically detected, forming a dynamic closed loop of "status detection-agent call-graph update", thereby ensuring that complex task chains can achieve efficient collaborative operation under the premise of strictly adhering to the execution order constraints.

[0243] The method provided in the embodiment of the present application uses sequential relationship analysis to accurately capture real business constraints and ensure the basic principle of "triggering actions only after data is ready"; task graph modeling generates an execution framework that is both visual and verifiable, and ensures logical completeness through graph traversal algorithms; graph-driven scheduling achieves event-triggered resource allocation with millisecond-level response.

[0244] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, after obtaining P subtasks, the method further includes:

[0245] Encode the P subtasks in the task request text to obtain P subtask encoding vectors.

[0246] It can be understood that the subtask encoding vector refers to the conversion of the subtask text in the user's original task description into a numerical representation in a high-dimensional semantic space through pre-trained language models (such as BERT and ChatGPT). By capturing the domain knowledge, operation objects, and contextual associations implicit in the subtask description, for example, encoding "book a flight" as [0.83, 0.92, 0.12, 0.95], this breaks through the limitations of traditional keyword matching and brings tasks with similar semantics but different expressions (such as "book a flight" and "flight query and ticket booking interface call") close in distance in the vector space, establishing a computable semantic benchmark for subsequent precise agent matching.

[0247] See also Figure 15 , Figure 15 The figure shows a schematic diagram of the embedding encoding process. The encoding module vectorizes each subtask description after task decomposition, ultimately outputting P subtask encoding vectors. The encoding process uses a pre-trained language model for high-dimensional semantic mapping. Each output vector corresponds one-to-one with an atomic task, forming a machine-processable vectorized representation that provides structured input for subsequent agent matching.

[0248] The method provided in the embodiment of the present application, on the one hand, retains the complete semantic context of the user's original request, and on the other hand, realizes cross-task capability correlation calculation through vector operations, providing a scalable quantitative basis for subsequent intelligent agent matching and improving the accuracy of task allocation.

[0249] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, after sending the response information including the task response text, the method further includes:

[0250] Generate feedback request text based on the task response text;

[0251] Send feedback request text;

[0252] Receive feedback response text corresponding to the feedback request text;

[0253] Generate target reward information based on the feedback response text;

[0254] A training sample pair is constructed according to the target reward information and the task response text, and at least one target agent among the P target agents is optimized according to the training sample pair.

[0255] Understandably, feedback request text refers to interactive content proactively generated by the system to solicit user feedback. Its design requires dynamic customization based on the key characteristics of task response text. Targeted questions are constructed using a template engine and context-aware algorithms. For example, when the system returns a response containing flight and hotel information, a follow-up question like "Please rate the suitability of the flight time and hotel location." This improves the accuracy of feedback collection and user engagement, and by focusing on core business nodes, guides users to provide high-value information, anchoring key directions for subsequent optimization.

[0256] The process of receiving feedback response text requires real-time parsing of heterogeneous input formats. Different forms of user evaluation, such as text comments and multiple-choice responses, are integrated into the semantic analysis pipeline. When users mix subjective complaints with industry jargon, the system can identify core entities such as "flight too early" and "hotel location." This ensures the effective integration of diverse feedback and maintains the integrity of evaluation information in complex scenarios such as cross-border transactions.

[0257] During the target reward information generation phase, semantic parsing results are converted into reward and penalty signals based on preconfigured quantization rules. For example, referring to a reward value table, "the hotel is too far away" is mapped to a local penalty of -0.8. A weighted strategy is employed to balance multi-dimensional feedback. For example, when a user simultaneously provides a moderate rating and clear improvement suggestions, a composite reward value is synthesized according to predefined rules. A machine-executable feedback incentive mechanism is established to accurately convey quality assessment results through quantized signals, significantly reducing the cost of manual intervention.

[0258] Training samples represent the entire lifecycle of tasks associated with the construction phase. Each sample includes the original task request, the agent's decision path, the response output, and the corresponding reward value tag. This creates a high-quality supervisory signal source, providing clear-cut training material for strategy optimization and significantly improving optimization efficiency.

[0259] Agent optimization is implemented through online iterations within a reinforcement learning framework, using a proximal policy optimization algorithm to adjust policy network parameters based on reward signals. For example, if the hotel recommendation agent receives a -0.8 penalty, it immediately triggers a fine-tuning of the network weights for that node.

[0260] See also Figure 16 , Figure 16The schematic diagram of the closed-loop iterative optimization is shown. First, a customized feedback request is automatically generated based on the task response text and sent to the user. Upon receiving the user feedback response text, the semantic parsing engine is activated to map it to a preset feedback type and determine a quantified target reward value. The associated specific subtask is then locked through a responsibility traceability mechanism. The reward value and subtask identifier are then combined to generate structured target reward information. This information is then combined with the original execution records to construct training sample pairs, driving the target agent to optimize parameters. Finally, the updated capability characteristics are fed back to the system core, forming an autonomous evolutionary closed loop of generating requests → collecting feedback → quantifying incentives → assigning responsibilities → targeted optimization → capability updating. This allows the agent system to dynamically improve task processing accuracy and adaptability through continuous interaction.

[0261] For example, see Figure 17 , Figure 17 A schematic diagram of user feedback is shown. In the scenario of intelligent assistant travel arrangement, Figure 17 In (A), the feedback can be in the form of text comments, such as Figure 17 In (B), the feedback form can be selective feedback. In addition, the user's feedback behavior can be obtained through the tracking design, such as the user's click behavior, stay time, modification operation, etc. The embodiment of the present application does not limit the feedback form.

[0262] After the task response text is displayed on the terminal, a concise feedback interface is provided to the user to guide the user to evaluate the task response text. The feedback can include:

[0263] 1. Text comments (e.g., "the arrangement is unreasonable," "the hotel is too far away");

[0264] 2. Selective feedback (e.g., “Are you satisfied?”, “What needs improvement?”);

[0265] 3. Behavioral feedback (such as click behavior, dwell time, modification operations, etc.).

[0266] Evaluation can be based on a rating system. Specifically, based on semantic recognition, the user's feedback response text is processed to extract key information and classify it. The feedback types are shown in Table 10:

[0267] Table 10

[0268] Feedback Type Example Description Overall satisfaction "Overall satisfied" or "Not very satisfied" Subtask quality assessment "The flight arrangements were poor" and "the hotel recommendations were inappropriate" Improvement Suggestions "Choose a later flight next time" or "Change hotel" Emotional expression "Very disappointing", "Excellent"

[0269] At the same time, the feedback response text is converted into structured data and the specific subtasks or execution paths are marked.

[0270] The complete execution path of each subtask (i.e., which agent specifically executes which subtasks) is recorded and associated with the user's feedback response text. For example, the task execution path in the intelligent assistant travel arrangement scenario is: flight path → hotel agent → calendar agent. Combined with the user's feedback response text, it is recognized that the user's feedback is mainly about the first subtask (flight reservation). Based on the parsed feedback response text, a corresponding local reward signal (Local Reward Signal) is generated to guide the specific agent's subsequent behavioral adjustments.

[0271] The reward signal is configurable, as shown in Table 11:

[0272] Table 11

[0273] Feedback Type Reward Value illustrate High satisfaction +10 Indicates that the current decision is reasonable Moderate satisfaction +0.5 Indicates that basic needs are met Low satisfaction -0.5 Indicates an obvious problem Clear improvement suggestions -0.8 Indicates decision adjustment Multiple negative feedback accumulation -1.0 Indicates a serious error

[0274] The Proximal Policy Optimization (PPO) algorithm is used to update the agent's decision-making strategy using task execution trajectories and feedback rewards.

[0275] The specific training process is as follows:

[0276] 1. Construct training samples: Each sample contains task input, execution path, node actions, and corresponding rewards.

[0277] 2. Policy network forward propagation: Input the task description and output the probability distribution of agent selection for each subtask.

[0278] 3. Loss calculation and backpropagation: Optimize the strategy based on feedback rewards to increase the probability of making better decisions in similar scenarios in the future.

[0279] 4. Model version management: Save the model version after each update and record the training log for auditing.

[0280] The updated policy model is deployed to the online environment and used for new task allocation decisions. At the same time, the system continues to collect feedback on new tasks, forming a closed-loop optimization process.

[0281] The method provided in the embodiments of this application uses dynamically generated feedback requests to capture users' core concerns, heterogeneous feedback parsing technology to express differences, structured reward mapping to establish a computable improvement scale, full-link sample construction to ensure optimization accuracy, and online reinforcement learning to complete minute-by-minute policy iteration. This entire process breaks through the resource bottleneck of traditional A / B testing, converting user satisfaction into algorithm optimization power while preserving business logic. It can reduce the recurrence rate of critical errors, while enabling the system to accumulate domain knowledge during continuous operation, gradually building an intelligent decision-making system with human-level expert judgment.

[0282] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, generating target reward information based on the feedback response text includes:

[0283] Based on semantic recognition, mapping the feedback response text to at least one preset feedback type to obtain a target feedback type corresponding to the feedback response text, wherein semantic information of the feedback response text is the same as semantic information of the preset feedback type;

[0284] Based on the mapping information between feedback type and reward information, determine the target reward value according to the target feedback type;

[0285] Based on semantic recognition, the feedback response text is matched with P subtasks to determine L target subtasks. The feedback response text is used to provide feedback on the processing results of the L subtasks corresponding to the L target subtasks, where P ≥ L ≥ 1.

[0286] Generate target reward information based on the target reward value and L target subtasks.

[0287] It's understood that the preset feedback type refers to a predefined standard feedback classification system, while the target feedback type is the category to which the user feedback belongs, determined by the system through semantic matching. Natural language understanding models are used to semantically align user feedback with the preset types, for example, mapping "inaccurate chart data" to the "error result" category. This establishes a standardized feedback processing benchmark, transcending the ambiguity of natural language expression and providing a unified framework for subsequent quantification, particularly ensuring consistent classification when processing multilingual feedback.

[0288] The reward value determination mechanism converts feedback based on pre-set feedback types and numerical mapping rules. Essentially, it establishes a configurable conversion channel between "feedback semantics" and "reward and punishment intensity." A hierarchical matching strategy is employed. For example, when user feedback includes both "fast response speed" (mapped to +0.5) and "missing key data" (mapped to -0.8), the system automatically synthesizes a composite reward value. This transforms subjective evaluations into machine-processable reinforcement learning signals, precisely driving system optimization through a quantitative incentive mechanism.

[0289] The target subtask location phase enables feedback problem tracing, linking feedback text to atomic tasks through cross-modal matching technology. Through entity co-occurrence analysis and intent association mapping, for example, when a user points out "the hotel location is incorrectly labeled," the system automatically targets the "geographic information annotation" subtask. This builds the ability to locate the root cause of faults, accurately breaking down global feedback into specific responsible modules and avoiding the ineffective spread of optimization resources.

[0290] The reward information synthesis phase encapsulates discrete data into structured training instructions. This requires integrating metadata such as subtask identifiers, reward and penalty intensities, and timestamps to generate an input format suitable for the reinforcement learning framework. This forms a closed-loop optimization pathway, enabling single user feedback to accurately trigger parameter updates for specific agents, improving optimization efficiency in high-frequency iteration scenarios.

[0291] See also Figure 18 , Figure 18 The schematic diagram shows the feedback signal generation process. When a user submits a feedback response, semantic parsing is first performed to extract the key intent. The parsed results are semantically matched against a pre-set feedback type library. After determining the target feedback type, the reward mapping rules are queried to output a quantitative reward value. Furthermore, subtask association analysis is performed to locate the specific target subtasks involved in the feedback. Finally, the reward value and subtask identifiers are integrated into structured target reward information, forming a closed-loop mapping from subjective evaluation to precise optimization drive.

[0292] The method provided in the embodiments of this application uses type mapping to overcome the bottleneck of natural language understanding, reward conversion to establish an objective quantitative scale, task positioning to achieve accurate attribution of problems, and information synthesis to build plug-and-play reinforcement signals. It converts user feedback information into time-consuming parameter adjustments, while using modular positioning to avoid systemic fluctuations during the optimization process. This allows for the continuous evolution of intelligent agent capabilities while ensuring the stability of core functions, ultimately achieving a spiral of user experience and system performance.

[0293] In an optional embodiment of the multi-agent collaborative task planning method provided in the above embodiment of the present application, a training sample pair is constructed based on the target reward information and the task response text, and at least one target agent among the P target agents is optimized based on the training sample pair, including:

[0294] Create training sample pairs based on L target subtasks, L subtask processing results corresponding to the L target subtasks, and target reward information;

[0295] Based on the training sample pairs, L target agents are optimized to perform L target subtasks.

[0296] In the method provided in the embodiment of the present application, the target subtask refers to an atomic-level operation unit that is decomposed and executed independently. The subtask processing result is the data or decision conclusion produced by the intelligent agent after execution, and the target reward information is the reward and punishment signal value generated quantitatively based on user feedback. By establishing a triple mapping relationship to form a structured training sample (such as <scheduled flight task, output scheduled flight number and time, reward value -0.5>), the abstract user experience is converted into a machine-processable supervision signal. By breaking the lag of traditional offline training, real-time linkage between business scenarios and algorithm optimization is achieved. For example, when a user complains that "the hotel is too far from the airport", the system can immediately locate the corresponding subtask and generate a training sample with a geographical constraint identifier.

[0297] The agent optimization process uses training samples to adjust policy network parameters. "At least one target agent" indicates support for a modular update strategy. This step uses a gradient backpropagation mechanism combined with reward signals to modify behavioral policies (for example, the flight agent reduces the weight of "lowest price first" recommendations after receiving consecutive negative rewards), establishing a backtracking channel for reward value influencing factors.

[0298] See also Figure 19 , Figure 19 The flowchart of the multi-agent collaborative task planning method is shown. For user-entered task request text, natural language parsing technology is used to identify intent and perform semantic modeling to generate a structured semantic vector. The task is then structurally decomposed into subtasks, and a task dependency graph is constructed to clarify the execution order. Agent role models are simultaneously defined and multidimensional capability vectors are quantified and generated. Based on the collaborative calculation of semantic and capability vectors, a matching algorithm is used to dynamically select the optimal agent combination. Finally, task execution is driven within a load balancing framework, and resources are dynamically scheduled and adaptively adjusted based on real-time monitoring data. This forms a closed-loop multi-agent collaborative task planning method, from semantic understanding and intelligent matching to flexible scheduling.

[0299] The method provided in the embodiment of the present application converts fragmented feedback into causally clear supervision data in the training sample construction phase to form a traceable task optimization path; and in the intelligent agent tuning phase, achieves targeted capability enhancement through local strategy updates.

[0300] On the other hand, this application provides a multi-agent collaborative task planning device. Figure 20 In (A), the multi-agent collaborative task planning device 2000 includes:

[0301] Receiving module 2001, used for receiving task request text;

[0302] Semantic recognition module 2002, used for splitting the task in the task request text based on semantic recognition to obtain P subtasks, wherein the P subtasks correspond to P subtask encoding vectors, P ≥ 1;

[0303] An acquisition module 2003 is used to acquire M agents, where the M agents correspond to M capability vectors, and the capability vectors are used to represent the quantitative performance of the agents in K skill dimensions, where M ≥ P ≥ 1 and K ≥ 1.

[0304] A determination module 2004 is configured to determine P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors;

[0305] The generation module 2005 is used to call P target agents to execute P subtasks, integrate the P subtask processing results obtained by the P target agents executing the P subtasks, and generate a task response text;

[0306] The sending module 2006 is configured to send a response message including the task response text.

[0307] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application, please refer to Figure 20 In (B), the multi-agent collaborative task planning device 2000 further includes:

[0308] The calculation module 2007 is used to calculate the semantic matching values ​​between the P subtasks and the M agents based on the P subtask encoding vectors and the M capability vectors, and obtain P×M semantic similarity values;

[0309] The calculation module 2007 is further configured to calculate the ability values ​​of the M agents to perform P subtasks based on the scores of each agent in the K skill dimensions, thereby obtaining P×M skill dimension values;

[0310] The calculation module 2007 is further configured to calculate M matching values ​​corresponding to each of the P subtasks based on the P×M semantic similarity values ​​and the P×M skill dimension values;

[0311] The determination module 2004 is further used to determine P target agents for processing the P subtasks based on the M matching values ​​corresponding to each of the P subtasks, wherein the matching value corresponding to the target agent is the maximum value among the M matching values.

[0312] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0313] The determination module 2004 is further configured to calculate the cosine similarities between the P subtask encoding vectors and the M capability vectors to obtain P×M semantic similarity values.

[0314] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0315] The acquisition module 2003 is further configured to acquire M skill dimension subsets corresponding to the M agents, wherein the skill dimension subsets include K skill scores, and the skill scores are used to represent the quantitative performance of the agents in the skill dimensions;

[0316] The acquisition module 2003 is further configured to acquire P skill dimension weight sets corresponding to the P subtasks, wherein the P skill dimension weight sets include K skill dimension weights, and the skill dimension weights are used to represent the importance of the skill dimension when the agent performs the subtask;

[0317] The calculation module 2007 is further configured to calculate P×M skill dimension values ​​based on the M skill dimension subsets and the P skill dimension weight sets.

[0318] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application, please refer to Figure 20 In (C), the multi-agent collaborative task planning device 2000 further includes:

[0319] Processing module 2008 is used to detect the load conditions of P target agents and obtain P load information and P response times corresponding to the P target agents executing P subtasks;

[0320] Determination module 2004 is also used to determine the first backup intelligent agent from MP intelligent agents when the load information corresponding to the first target intelligent agent among P target intelligent agents is greater than the load threshold, or the response duration corresponding to the first target intelligent agent is greater than the duration threshold, wherein the matching value corresponding to the first backup intelligent agent is the second largest value among the M matching values.

[0321] The processing module 2008 is further configured to call the first standby agent to execute the subtask that the first target agent is to execute.

[0322] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0323] The processing module 2008 is further used to perform semantic recognition on the task request text to obtain task request semantic information;

[0324] The processing module 2008 is further configured to perform entity analysis on the semantic information of the task request to obtain P subtasks.

[0325] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0326] The processing module 2008 is further configured to perform a relationship analysis on the P subtasks according to the task request semantic information to obtain the execution order relationship of the P subtasks;

[0327] The generating module 2005 is further configured to generate a task graph based on the relationship between the P subtasks and the execution order of the P subtasks, wherein the task graph includes P nodes and edge relationships between the P nodes, the P nodes correspond to the P subtasks, and the edge relationship between the P nodes corresponds to the relationship between the execution order of the P subtasks;

[0328] The processing module 2008 is further used to call P target agents to execute P subtasks according to the task graph.

[0329] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0330] The processing module 2008 is further configured to encode the P subtasks in the task request text to obtain P subtask encoding vectors.

[0331] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application, please refer to Figure 20 In (D), the multi-agent collaborative task planning device 2000 further includes:

[0332] The generating module 2005 is further used to generate a feedback request text according to the task response text;

[0333] The sending module 2006 is further used to send the feedback request text;

[0334] The receiving module 2001 is further configured to receive a feedback response text corresponding to the feedback request text;

[0335] The generating module 2005 is further used to generate target reward information according to the feedback response text;

[0336] The training module 2009 is used to construct a training sample pair according to the target reward information and the task response text, and optimize at least one target agent among the P target agents according to the training sample pair.

[0337] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0338] The processing module 2008 is further configured to map the feedback response text to at least one preset feedback type based on semantic recognition to obtain a target feedback type corresponding to the feedback response text, wherein the semantic information of the feedback response text is the same as the semantic information of the preset feedback type;

[0339] The determination module 2004 is further configured to determine a target reward value according to a target feedback type based on mapping information between feedback type and reward information;

[0340] The determination module 2004 is further configured to match the feedback response text with the P subtasks based on semantic recognition to determine L target subtasks, wherein the feedback response text is used to provide feedback on the processing results of the L subtasks corresponding to the L target subtasks, where P≥L≥1;

[0341] The generation module 2005 is further configured to generate target reward information according to the target reward value and the L target subtasks.

[0342] In an optional embodiment of the multi-agent collaborative task planning device provided in the above embodiment of the present application,

[0343] The training module 2009 is further configured to create training sample pairs based on the L target subtasks, the L subtask processing results corresponding to the L target subtasks, and the target reward information;

[0344] The training module 2009 is also used to optimize the L target intelligent agents used to perform the L target subtasks based on the training sample pairs.

[0345] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0346] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0347] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0348] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0349] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0350] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a server or terminal device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes various media that can store computer programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0351] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-agent collaborative task planning method, characterized in that: include: Receive task request text; Based on semantic recognition, split the task in the task request text to obtain P subtasks, wherein the P subtasks correspond to P subtask encoding vectors, P ≥ 1; Obtain M agents, wherein the M agents correspond to M capability vectors, and the capability vectors are used to represent the quantitative performance of the agents in K skill dimensions, where M ≥ P ≥ 1 and K ≥ 1; Determining P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors; Calling the P target agents to execute the P subtasks, integrating the P subtask processing results obtained by the P target agents executing the P subtasks, and generating a task response text; Sending a response message including the task response text.

2. The task planning method according to claim 1, wherein: The step of determining P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors includes: Calculate semantic matching values ​​between the P subtasks and the M agents based on the P subtask encoding vectors and the M capability vectors to obtain P×M semantic similarity values; Calculate the ability of the M agents to perform the P subtasks based on the scores of each of the M agents in the K skill dimensions, and obtain P×M skill dimension values; Calculate M matching values ​​corresponding to each of the P subtasks according to the P×M semantic similarity values ​​and the P×M skill dimension values; According to the M matching values ​​corresponding to each of the P subtasks, P target agents for processing the P subtasks are determined, wherein the matching value corresponding to the target agent is the maximum value among the M matching values.

3. The task planning method according to claim 2, wherein: The step of calculating semantic matching values ​​between the P subtasks and the M agents based on the P subtask encoding vectors and the M capability vectors to obtain P×M semantic similarity values ​​includes: The cosine similarities between the P subtask encoding vectors and the M capability vectors are calculated to obtain P×M semantic similarity values.

4. The task planning method according to claim 2, wherein: The step of calculating the ability of the M agents to perform the P subtasks based on the scores of each of the M agents in the K skill dimensions to obtain P×M skill dimension values ​​includes: Obtaining M skill dimension subsets corresponding to the M agents, wherein the skill dimension subsets include K skill scores, and the skill scores are used to represent the quantitative performance of the agents in the skill dimensions; Obtaining P skill dimension weight sets corresponding to the P subtasks, wherein the P skill dimension weight sets include K skill dimension weights, and the skill dimension weights are used to represent the importance of the skill dimension when the agent performs the subtasks; According to the M skill dimension subsets and the P skill dimension weight sets, P×M skill dimension values ​​are calculated.

5. The task planning method according to claim 2, wherein: After calling the P target agents to perform the P subtasks, the method further includes: Performing load detection on the P target agents to obtain P load information and P response times corresponding to the P target agents executing the P subtasks; If the load information corresponding to the first target agent among the P target agents is greater than the load threshold, or the response time corresponding to the first target agent is greater than the time threshold, a first backup agent is determined from the MP agents, wherein the matching value corresponding to the first backup agent is the second largest value among the M matching values; The first backup agent is called to execute the subtask executed by the first target agent.

6. The task planning method according to claim 1, wherein: Based on semantic recognition, the task in the task request text is split into P subtasks, including: Performing semantic recognition on the task request text to obtain task request semantic information; Perform entity analysis on the semantic information of the task request to obtain P subtasks.

7. The task planning method according to claim 6, wherein: The method further comprises: Performing a relationship analysis on the P subtasks according to the task request semantic information to obtain a relationship of execution orders corresponding to the P subtasks; Generate a task graph based on the relationship between the P subtasks and the execution order of the P subtasks, wherein the task graph includes P nodes and edge relationships between the P nodes, the P nodes correspond to the P subtasks, and the edge relationships between the P nodes correspond to the relationship between the execution order of the P subtasks; The calling of the P target agents to perform the P subtasks includes: According to the task graph, the P target agents are called to perform the P subtasks.

8. The task planning method according to claim 1, wherein: After obtaining P subtasks, the following steps are also included: The P subtasks in the task request text are encoded to obtain P subtask encoding vectors.

9. The task planning method according to claim 1, wherein: After sending the response information including the task response text, the method further includes: generating a feedback request text according to the task response text; Sending the feedback request text; Receive a feedback response text corresponding to the feedback request text; generating target reward information according to the feedback response text; A training sample pair is constructed according to the target reward information and the task response text, and at least one target agent among the P target agents is optimized according to the training sample pair.

10. The task planning method according to claim 9, wherein: Generating target reward information according to the feedback response text includes: Based on semantic recognition, mapping the feedback response text to at least one preset feedback type to obtain a target feedback type corresponding to the feedback response text, wherein semantic information of the feedback response text is the same as semantic information of the preset feedback type; Determining a target reward value according to the target feedback type based on mapping information between feedback type and reward information; Based on semantic recognition, the feedback response text is matched with the P subtasks to determine L target subtasks, wherein the feedback response text is used to provide feedback on the processing results of the L subtasks corresponding to the L target subtasks, and P≥L≥1; Generate target reward information based on the target reward value and the L target subtasks.

11. The task planning method according to claim 10, wherein: The step of constructing a training sample pair according to the target reward information and the task response text, and optimizing at least one of the P target agents according to the training sample pair, includes: Creating training sample pairs according to the L target subtasks, the L subtask processing results corresponding to the L target subtasks, and the target reward information; Based on the training sample pairs, L target agents for performing the L target subtasks are optimized.

12. A multi-agent collaborative task planning device, characterized in that: include: Receiving module, used for receiving task request text; A semantic recognition module, configured to split the task in the task request text based on semantic recognition to obtain P subtasks, wherein the P subtasks correspond to P subtask encoding vectors, and P ≥ 1; An acquisition module is configured to acquire M agents, wherein the M agents correspond to M capability vectors, and the capability vectors are used to represent the quantitative performance of the agents in K skill dimensions, where M ≥ P ≥ 1 and K ≥ 1; a determination module, configured to determine P target agents for processing the P subtasks based on the P subtask encoding vectors and the M capability vectors; a generation module, configured to call the P target agents to execute the P subtasks, integrate the P subtask processing results obtained by the P target agents executing the P subtasks, and generate a task response text; The sending module is used to send the response information including the task response text.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the multi-agent collaborative task planning method described in any one of claims 1 to 11 are implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-agent collaborative task planning method described in any one of claims 1 to 11 are implemented.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the multi-agent collaborative task planning method described in any one of claims 1 to 11 are implemented.

Images