Task planning method and device, task processing method and device and computer equipment

Abstract

The invention relates to a task planning method and device, a task processing method and device, computer equipment, a computer readable storage medium and a computer program product. The task planning method comprises the steps that when a target task split from a user task is executed, feasible follow-up paths of the target task are planned through multiple professional agents, and candidate solutions generated by the multiple professional agents are obtained; performing quantitative evaluation on the plurality of candidate solutions to obtain respective scheme evaluation results of the plurality of candidate solutions; determining a first solution of the target task based on the respective scheme evaluation results of the plurality of candidate solutions; and generating a task solution of the user task according to the first solution of the target task. By adopting the method, the task planning decision accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a task planning method, apparatus, computer equipment, computer-readable storage medium, and computer program product, as well as a task processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of computer technology, large language models have emerged. Large language models refer to a new generation of artificial intelligence models with large-scale natural language understanding and generation capabilities, capable of handling tasks using planning abilities. The planning ability of large language models refers to their ability to automatically decompose executable steps and design solutions for complex problems.

[0003] In related technologies, the planning mode of large language models is that the large language model is responsible for the "think-action" cycle of the entire task. When encountering a planning fork point, the large language model generates a one-time solution based on its internal knowledge.

[0004] However, the quality of task planning and decision-making in related technologies depends entirely on the current state and preferences of the large language model, which is highly uncertain and random, resulting in low accuracy in task planning and decision-making. Summary of the Invention

[0005] Therefore, it is necessary to provide a task planning method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of task planning decisions, as well as a task processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product, in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a task planning method, including:

[0007] When the target task extracted from the user task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task, and obtain the candidate solutions generated by each of the multiple professional intelligent agents.

[0008] Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions;

[0009] Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined.

[0010] Based on the first solution of the target task, a task solution for the user task is generated.

[0011] Secondly, this application also provides a task planning apparatus, comprising:

[0012] The first task execution module is used to, when executing a target task extracted from a user task, plan feasible subsequent paths for the target task through multiple professional intelligent agents, and obtain candidate solutions generated by each of the multiple professional intelligent agents.

[0013] The first quantitative evaluation module is used to perform quantitative evaluation on multiple candidate solutions respectively, and obtain the evaluation results of each of the multiple candidate solutions.

[0014] The first solution selection module is used to determine the first solution for the target task based on the evaluation results of the multiple candidate solutions.

[0015] The first solution generation module is used to generate a task solution for the user task based on the first solution for the target task.

[0016] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0017] When the target task extracted from the user task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task, and obtain the candidate solutions generated by each of the multiple professional intelligent agents.

[0018] Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions;

[0019] Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined.

[0020] Based on the first solution of the target task, a task solution for the user task is generated.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0022] When the target task extracted from the user task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task, and obtain the candidate solutions generated by each of the multiple professional intelligent agents.

[0023] Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions;

[0024] Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined.

[0025] Based on the first solution of the target task, a task solution for the user task is generated.

[0026] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0027] When the target task extracted from the user task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task, and obtain the candidate solutions generated by each of the multiple professional intelligent agents.

[0028] Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions;

[0029] Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined.

[0030] Based on the first solution of the target task, a task solution for the user task is generated.

[0031] The aforementioned task planning method, apparatus, computer equipment, computer-readable storage medium, and computer program product, when executing a target task decomposed from a user task, utilizes multiple specialized intelligent agents to plan feasible subsequent paths for the target task, obtaining candidate solutions generated by each agent. At planning bifurcation points, multiple agents are introduced to perform parallel reasoning on the target task, leveraging collective intelligence to reduce the uncertainty and randomness of task planning decisions. By quantitatively evaluating each candidate solution, evaluation results are obtained, enabling objective assessment of multiple candidate solutions. Based on these evaluation results, a first solution for the target task is determined, and a task solution for the user task is generated. Throughout this process, at planning bifurcation points, the parallel reasoning of multiple specialized intelligent agents and the quantitative evaluation of their respective candidate solutions reduce uncertainty and randomness in task planning decisions, improving the quality and accuracy of these decisions.

[0032] Sixthly, this application provides a task processing method, including:

[0033] When a user task is received, the user task is broken down into multiple tasks to be executed, and the multiple tasks to be executed are executed sequentially.

[0034] During the sequential execution of the plurality of tasks to be executed, if the current task to be executed is a multi-path planning type task, the current task to be executed is determined as the target task;

[0035] When the target task is executed, multiple specialized intelligent agents plan feasible subsequent paths for the target task, thereby obtaining candidate solutions generated by each of the multiple specialized intelligent agents.

[0036] Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions;

[0037] Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined.

[0038] Based on the first solution of the target task, a task solution for the user task is generated, and the task solution is fed back to the sender of the user task.

[0039] Seventhly, this application also provides a task processing apparatus, comprising:

[0040] The second task execution module is used to, upon receiving a user task, decompose the user task into multiple tasks to be executed, and execute the multiple tasks to be executed sequentially. During the sequential execution of the multiple tasks to be executed, if the current task to be executed is a multi-path planning type task, the current task to be executed is determined as the target task. When the target task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task respectively, and obtain the candidate solutions generated by the multiple professional intelligent agents.

[0041] The second quantitative evaluation module is used to perform quantitative evaluation on multiple candidate solutions respectively, and obtain the evaluation results of each of the multiple candidate solutions.

[0042] The second solution selection module is used to determine the first solution for the target task based on the evaluation results of the multiple candidate solutions.

[0043] The second solution generation module is used to generate a task solution for the user task based on the first solution for the target task, and to feed back the task solution to the sender of the user task.

[0044] Eighthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described task processing method.

[0045] Ninthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described task processing method.

[0046] In a tenth aspect, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described task processing method.

[0047] The aforementioned task processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product, upon receiving a user task, decompose the user task into multiple tasks to be executed, and execute these tasks sequentially. This enables the planning and decomposition of the user task. During the sequential execution of these tasks, if the current task is a multi-path planning type task, it is identified as the target task. It can also identify target tasks where planning forks occur during task execution. When the target task is reached, multiple specialized intelligent agents plan feasible subsequent paths for the target task, obtaining candidate solutions generated by each agent. When encountering planning forks, multiple specialized intelligent agents can be introduced to perform parallel reasoning on the target task, utilizing collective intelligence to reduce the uncertainty and randomness of task planning decisions. By quantitatively evaluating multiple candidate solutions, evaluation results are obtained for each solution. This quantitative evaluation method allows for objective evaluation of multiple candidate solutions. Based on the evaluation results of each candidate solution, a first solution for the target task can be determined, and a task solution for the user task can be generated based on this first solution. Throughout the process, when encountering planning bifurcation points, the uncertainty and randomness of task planning decisions can be reduced by using parallel reasoning by multiple specialized intelligent agents and quantitative evaluation of the candidate solutions generated by each of the specialized intelligent agents. This improves the quality of task planning decisions and increases their accuracy. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a diagram illustrating the application environment of the task planning method in one embodiment;

[0050] Figure 2 This is a flowchart illustrating a task planning method in one embodiment;

[0051] Figure 3 This is a schematic diagram of the process for obtaining model evaluation scores in one embodiment;

[0052] Figure 4 This is a schematic diagram of the process for obtaining rule evaluation scores in one embodiment;

[0053] Figure 5 This is a flowchart illustrating the calculation of confidence level in one embodiment;

[0054] Figure 6 This is a schematic diagram of a user decision-making interface in one embodiment;

[0055] Figure 7 This is a schematic diagram comparing the evaluation data of the schemes in one embodiment;

[0056] Figure 8 This is an application environment diagram of a task processing method in one embodiment;

[0057] Figure 9 This is a flowchart illustrating a task processing method in one embodiment;

[0058] Figure 10 This is a flowchart illustrating the task planning method in another embodiment;

[0059] Figure 11 This is a schematic diagram of the user decision-making interface in another embodiment;

[0060] Figure 12 This is a core architecture diagram of a task planning method in one embodiment;

[0061] Figure 13 This is a flowchart illustrating the workflow of a specialized intelligent agent pool in one embodiment.

[0062] Figure 14 Here is a flowchart of the voting consensus process in one embodiment;

[0063] Figure 15 Here is a flowchart of the feedback optimization process in one embodiment;

[0064] Figure 16 Here is a flowchart of the task planning system in one embodiment;

[0065] Figure 17 This is a structural block diagram of a task planning device in one embodiment;

[0066] Figure 18 This is a structural block diagram of a task processing device in one embodiment;

[0067] Figure 19 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0069] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0070] In order to clearly describe the technical solution of this application and facilitate understanding of the technical solution of this application, the key concepts involved in this application will be explained below.

[0071] 1. Objectives and tasks.

[0072] A target task refers to an execution task with multiple feasible follow-up paths, meaning a task with multiple strategy options. A feasible follow-up path refers to a logically sound and executable solution direction for the target task, based on its objectives and context; in other words, a solution for the target task. It is understandable that different feasible follow-up paths differ in their methods, perspectives, or specific content, thus constituting different strategy choices.

[0073] 2. Professional intelligent agent.

[0074] A specialized intelligent agent refers to an intelligent agent focused on a specific domain or perspective, capable of generating independent solutions or plans for a specific problem. In this embodiment, the specialized intelligent agent can generate independent solutions, i.e., candidate solutions, for the target task. It is understood that the specialized intelligent agent is built upon a large language model.

[0075] 3. Evaluation results of the plan.

[0076] The evaluation result refers to the objective reflection of the merits of candidate solutions after a quantitative evaluation. For example, the evaluation result could be a solution evaluation score, or a solution evaluation priority.

[0077] 4. Confidence level.

[0078] Confidence level refers to a quantitative or qualitative measure of the degree of certainty regarding the reliability, rationality, and effectiveness of a primary solution. In other words, a higher confidence level indicates stronger reliability, rationality, and effectiveness, meaning the solution can be trusted. Conversely, a lower confidence level indicates weaker reliability, rationality, and effectiveness, meaning the solution cannot be trusted.

[0079] The task planning method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on other network servers. When a user submits a user task through terminal 102, when the execution reaches the target task derived from the user task, server 104 will use multiple specialized intelligent agents to plan feasible subsequent paths for the target task, obtaining candidate solutions generated by each agent. These candidate solutions will be quantitatively evaluated to obtain evaluation results. Based on these evaluation results, a first solution for the target task will be determined, and a task solution for the user task will be generated based on this first solution.

[0080] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0081] In one exemplary embodiment, such as Figure 2 As shown, a task planning method is provided. This embodiment illustrates the method applied to a server, but it is understood that the method can also be applied to a terminal, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes steps 202 to 208. Wherein:

[0082] Step 202: When the target task extracted from the user task is executed, multiple professional intelligent agents plan feasible subsequent paths for the target task, and obtain candidate solutions generated by each of the multiple professional intelligent agents.

[0083] In this context, a user task refers to a query request issued by a user to obtain specific information, typically accompanied by a user question. A user question is a question entered by the user to query or retrieve specific content. For example, a user question can be multiple keywords, a question, or a description in natural language.

[0084] The target task refers to an execution task with multiple feasible follow-up paths, meaning there are multiple strategy options available. A feasible follow-up path refers to a logically sound and executable solution to the target task, based on its objectives and context; in other words, a solution for the target task. It is understandable that different feasible follow-up paths differ in their methods, perspectives, or specific content, thus constituting different strategy choices.

[0085] Understandably, the emergence of a target task signifies a planning fork, indicating multiple feasible subsequent paths. Based on planning these feasible paths and obtaining multiple candidate solutions, a solution is selected from these candidate solutions. For example, taking the target task as "which scenic spots to visit," different feasible subsequent paths can be planned from the perspectives of opening hours, geographical location, and user interest matching, resulting in multiple candidate solutions. The final solution is then selected from these candidate solutions. Planning forks refer to a scenario where, during the execution of a user task, at a specific node (i.e., a specific task to be executed), multiple feasible subsequent paths or strategies exist, and a single model cannot directly determine the optimal solution.

[0086] Among them, the task to be executed refers to the standardized, executable task unit extracted from the user task through the reasoning and planning capabilities of the large language model. It is represented in a structured form and usually includes key information such as task title, task description information, and dependencies (execution order). Since it is a step in the overall planning process, it can also be called a "planning step".

[0087] In this context, a specialized intelligent agent refers to an agent focused on a specific domain or perspective, capable of generating independent solutions or plans for specific problems. In this embodiment, the specialized intelligent agent can generate independent solutions, or candidate solutions, for the target task. It is understood that the specialized intelligent agent is built upon a large language model, which refers to a new generation of artificial intelligence models possessing large-scale natural language understanding and generation capabilities.

[0088] For example, the specific domains or perspectives mentioned here could be search, analysis, programming, design, creativity, etc., and the specialized intelligent agents could be strategic intelligent agents, creative intelligent agents, domain expert intelligent agents, etc. Strategic intelligent agents are designed with task processing prompts prioritizing the path to achieving the goal, resource allocation, and risk avoidance, favoring robust and executable solutions. Creative intelligent agents are designed with task processing prompts designed to encourage divergent thinking, break conventional constraints, and strive to discover novel and non-obvious solutions. Domain expert intelligent agents are configured with detailed knowledge bases and thinking patterns for specific domains (such as shopping and programming), primarily used to provide professional and authoritative solutions.

[0089] For example, when a user task is received, the server will break down the user task into multiple tasks to be executed and execute the multiple tasks to be executed in sequence. When the target task that was broken down from the user task is executed, the server will call multiple professional intelligent agents. Through the multiple professional intelligent agents, the feasible subsequent paths of the target task are planned respectively, and the candidate solutions generated by the multiple professional intelligent agents are obtained.

[0090] In practical applications, during the sequential execution of multiple tasks, the server monitors the execution process and determines whether a planning fork is encountered. If a planning fork is encountered, it indicates that the task to be executed is a multi-path planning type task with multiple possible subsequent paths. The server will then identify the task to be executed as the target task and trigger a multi-agent decision-making process, calling multiple specialized agents. Through these specialized agents, feasible subsequent paths for the target task are planned respectively, resulting in candidate solutions generated by each of the specialized agents.

[0091] In practical applications, when executing a task, the server implicitly evaluates for the existence of a fork in the road. If a fork exists, it is considered to have encountered a strategic fork. This is mainly based on "common sense" and "reasoning patterns" learned from massive amounts of data. For example, if the problem to be executed is not linear and there are multiple reasonable choices, a fork can be considered to exist. If the task to be executed reaches a certain step where a strategic choice needs to be made, a fork can also be considered to exist.

[0092] In practical applications, when multiple specialized intelligent agents are invoked, the server will send a unified context and instructions to each specialized intelligent agent, so that each specialized intelligent agent can plan a feasible subsequent path for the target task based on the unified context and instructions, and obtain the candidate solutions generated by each specialized intelligent agent.

[0093] Step 204: Quantitatively evaluate the multiple candidate solutions to obtain the evaluation results for each candidate solution.

[0094] Quantitative evaluation refers to the objective analysis and comparison of candidate solutions through clearly defined indicators and quantitative methods. The evaluation result refers to the objective reflection of the merits of the candidate solutions after quantitative evaluation. For example, the evaluation result could be a solution evaluation score, or a solution evaluation priority.

[0095] For example, when multiple professional intelligent agents generate their own candidate solutions, the server will perform quantitative evaluation on each candidate solution to achieve an objective evaluation of the multiple candidate solutions and obtain the evaluation results of each candidate solution.

[0096] In practical applications, the server can perform quantitative evaluation on multiple candidate solutions based on predefined quantitative evaluation rules, obtaining rule evaluation scores for each candidate solution, and using these scores as the solution evaluation results. Alternatively, the server can use a large language model, based on predefined quantitative evaluation dimensions, to perform quantitative evaluation on multiple candidate solutions, obtaining model evaluation scores for each candidate solution, and using these scores as the solution evaluation results.

[0097] In specific applications, the server can also perform quantitative evaluation on multiple candidate solutions based on predefined quantitative evaluation rules to obtain the rule evaluation scores of each candidate solution. Then, using a large language model, based on predefined quantitative evaluation dimensions, it can perform quantitative evaluation on multiple candidate solutions to obtain the model evaluation scores of each candidate solution. Finally, based on the rule evaluation scores and model evaluation scores of each candidate solution, the solution evaluation results of each candidate solution can be obtained.

[0098] Step 206: Based on the evaluation results of multiple candidate solutions, determine the first solution for the target task.

[0099] The first solution refers to the solution to the target task selected from multiple candidate solutions, that is, the optimal candidate solution determined after quantitative evaluation of multiple candidate solutions.

[0100] For example, the server will determine the first solution for the target task from among the multiple candidate solutions based on the solution evaluation results of each candidate solution.

[0101] In practical applications, the solution evaluation results include solution evaluation scores. The server can then determine the candidate solution with the highest solution evaluation score from among multiple candidate solutions based on their respective solution evaluation scores, and use it as the first solution for the target task.

[0102] In practical applications, the solution evaluation results include solution evaluation priorities. The server can then determine the candidate solution with the highest evaluation priority from among multiple candidate solutions based on their respective evaluation priorities, and use it as the first solution for the target task.

[0103] Step 208: Generate a task solution for the user task based on the first solution for the target task.

[0104] Among them, the task solution refers to the problem answer generated through systematic planning and reasoning in response to the user task, which can achieve the user task's goal. It is the final answer with a complete structure and operability.

[0105] For example, given a first solution to the target task, the server can generate a task solution for the user task based on the first solution to the target task.

[0106] In practical applications, the server first evaluates the confidence level of the first solution to the target task. If the confidence level of the first solution indicates that the first solution can be determined as the target solution of the target task, the server will directly use the first solution as the target solution of the target task, integrate the target solution of the target task into the task solution of the user task, and continue to execute the subsequent tasks of the target task among multiple tasks to be executed, so as to further generate the task solution of the user task.

[0107] In practical applications, if the confidence level of the first solution indicates that the first solution cannot be determined as the target solution for the target task, the server needs to give the user the right to choose the solution, that is, request the user to provide feedback and let the user choose whether to determine the first solution as the target solution for the target task.

[0108] In a specific application, when requesting user feedback, the server generates a user decision request, sends the user decision request back to the user, obtains the user's decision result based on the user's decision request, determines whether to use the first solution of the target task as the target solution of the target task based on the user's decision result, and further generates the task solution of the user task.

[0109] In a specific application, when determining the confidence level of a first solution to a target task, the server compares the confidence level of the first solution with a predefined confidence threshold. If the confidence level of the first solution is greater than the predefined confidence threshold, it can be considered that the first solution is the target solution for the target task. If the confidence level of the first solution is less than or equal to the predefined confidence threshold, it can be considered that the first solution is not the target solution for the target task. The predefined confidence threshold can be configured according to the actual application scenario. For example, the predefined confidence threshold could be 0.75.

[0110] The aforementioned task planning method, when encountering the target task decomposed from the user task, utilizes multiple specialized intelligent agents to plan feasible subsequent paths for the target task, resulting in candidate solutions generated by each agent. At planning fork points, multiple agents are introduced to perform parallel reasoning on the target task, leveraging collective intelligence to reduce the uncertainty and randomness of task planning decisions. By quantitatively evaluating each candidate solution, evaluation results are obtained, enabling objective assessment of multiple candidate solutions. Based on these evaluation results, a first solution for the target task is determined, and a task solution for the user task is generated from this first solution. Throughout this process, at planning fork points, the parallel reasoning of multiple specialized intelligent agents and the quantitative evaluation of their respective candidate solutions reduce uncertainty and randomness in task planning decisions, improving the quality and accuracy of these decisions.

[0111] In an exemplary embodiment, multiple candidate solutions are quantitatively evaluated to obtain evaluation results for each candidate solution, including:

[0112] Based on predefined quantitative evaluation rules, multiple candidate solutions are quantitatively evaluated separately to obtain the rule evaluation scores of each candidate solution.

[0113] Using a large language model, based on predefined quantitative evaluation dimensions, multiple candidate solutions are quantitatively evaluated separately to obtain the model evaluation scores for each candidate solution.

[0114] For each candidate solution, the solution evaluation result is obtained based on the rule evaluation score and model evaluation score of the candidate solution.

[0115] The quantitative evaluation rules refer to the evaluation rules that must be followed when quantitatively evaluating candidate solutions. These rules can be configured according to the actual application scenario. For example, specific quantitative evaluation rules can be predefined solution scoring rules, which predefine multiple quantitative evaluation indicators, the scoring scale for each indicator, and the weight allocation for multiple indicators. The rule-based evaluation score refers to the evaluation score determined based on the quantitative evaluation rules for the candidate solution.

[0116] The quantitative evaluation dimensions refer to the evaluation dimensions set when quantitatively evaluating candidate solutions, and can be configured according to the actual application scenario. For example, specific quantitative evaluation dimensions may include solution feasibility, solution execution efficiency, consistency between the solution and the overall goal, solution innovation, and solution risk assessment. The model evaluation score refers to the evaluation score output by the large language model based on the quantitative evaluation dimensions when quantitatively evaluating candidate solutions.

[0117] For example, during the quantitative evaluation, the server performs quantitative evaluation on multiple candidate solutions based on predefined quantitative evaluation rules, obtaining rule evaluation scores for each candidate solution. Then, using a large language model, the server performs quantitative evaluation on multiple candidate solutions based on predefined quantitative evaluation dimensions, obtaining model evaluation scores for each candidate solution. Finally, for each candidate solution, the solution evaluation result can be obtained based on the rule evaluation score and model evaluation score of the candidate solution.

[0118] In practical applications, for each candidate solution, when using predefined quantitative evaluation rules to quantitatively evaluate the candidate solution, the server can use multiple quantitative evaluation indicators in the quantitative evaluation rules, the scoring scale of each quantitative evaluation indicator, and the weight allocation of multiple quantitative evaluation indicators to score the candidate solution and obtain the rule evaluation score of the candidate solution.

[0119] In practical applications, for each candidate solution, when the candidate solution is quantitatively evaluated by the large language model, the prompt word template of the large language model will require it to summarize and score the candidate solution based on the predefined quantitative evaluation dimensions. The score generated is the model evaluation score of the candidate solution.

[0120] When requiring a large language model to provide scores, a scoring range must be specified, which can be configured according to the actual application scenario. For example, the scoring range can be from 1 to 10 points. It should be noted that this scoring range is the same as the scoring range used when using predefined quantitative evaluation rules to quantitatively evaluate candidate solutions; that is, the scoring range used for quantitative evaluation using quantitative evaluation rules and for quantitative evaluation using a large language model should be consistent.

[0121] In a specific application, such as Figure 3 As shown, for each candidate solution, when performing quantitative evaluation using a large language model, if there is only one predefined quantitative evaluation dimension, the large language model will score the candidate solution based on that single dimension, obtaining a model evaluation score for the candidate solution. If there are multiple predefined quantitative evaluation dimensions, the large language model can score the candidate solution based on each of the multiple dimensions, obtaining a quantitative evaluation dimension score for each dimension, and then using these scores to obtain the final model evaluation score for the candidate solution.

[0122] In a specific application, such as Figure 3 As shown, after obtaining the scores for each of the multiple quantitative evaluation dimensions, the server can acquire the dimensional weights for each of these dimensions. Using these dimensional weights, the server performs a weighted calculation on the scores for each of the multiple quantitative evaluation dimensions to obtain the model evaluation score for the candidate solution. The dimensional weights for each of the multiple quantitative evaluation dimensions can be configured according to the actual application scenario; the dimensional weights for different quantitative evaluation dimensions can be the same or different.

[0123] In practical applications, for each candidate solution, the server can calculate a solution evaluation score based on the candidate solution's rule evaluation score and model evaluation score, and then use the solution evaluation score as the solution evaluation result. The server can also determine the solution evaluation priority of the candidate solution based on its rule evaluation score and model evaluation score, and then use this priority as the solution evaluation score.

[0124] In a specific application, when determining the evaluation priority of candidate solutions, the server can combine the rule evaluation scores and model evaluation scores of the candidate solutions with predefined priority ranking rules to determine the evaluation priority. The predefined priority ranking rules can be configured according to the actual application scenario. For example, a predefined priority ranking rule could be to first sort by rule evaluation scores, and then, for at least two candidate solutions with the same rule evaluation score, sort by model evaluation score. Another example is that a predefined priority ranking rule could be to first sort by model evaluation scores, and then, for at least two candidate solutions with the same model evaluation score, sort by rule evaluation scores.

[0125] In this embodiment, by combining quantitative evaluation of quantitative evaluation rules and quantitative evaluation of large language models, an objective evaluation of candidate solutions can be achieved, resulting in accurate evaluation results of candidate solutions.

[0126] In an exemplary embodiment, based on predefined quantitative evaluation rules, multiple candidate solutions are quantitatively evaluated separately to obtain rule evaluation scores for each candidate solution, including:

[0127] For each candidate solution, a quantitative evaluation is performed on the candidate solution based on multiple quantitative evaluation indicators in the predefined quantitative evaluation rules, and the quantitative evaluation scores corresponding to each of the multiple quantitative evaluation indicators are obtained.

[0128] Based on the evaluation scores of the various quantitative evaluation indicators, the rule evaluation scores of the candidate solutions are obtained.

[0129] Among them, quantitative evaluation indicators refer to predefined evaluation metrics when quantitatively evaluating candidate solutions based on quantitative evaluation rules. For example, quantitative evaluation indicators may specifically include the number of steps in the solution, the estimated time taken, and the cost of the solution. The quantitative indicator evaluation score refers to the evaluation score determined based on the quantitative evaluation indicators when quantitatively evaluating candidate solutions.

[0130] For example, for each candidate solution, the server will perform quantitative evaluation on the candidate solution based on multiple quantitative evaluation indicators in the predefined quantitative evaluation rules, obtain the quantitative indicator evaluation scores corresponding to each of the multiple quantitative evaluation indicators, and then obtain the rule evaluation score of the candidate solution based on the quantitative indicator evaluation scores corresponding to each of the multiple quantitative evaluation indicators.

[0131] In practical applications, for each quantitative evaluation metric, when quantitatively evaluating candidate solutions based on the metric, the server obtains the scoring scale of the quantitative evaluation metric, and then uses the scoring scale to quantitatively evaluate the candidate solutions, obtaining the corresponding quantitative evaluation score. The scoring scale defines a specific scoring method, which can be configured according to the actual application scenario. For example, if the scoring method defined in the scoring scale can be a deduction item, then the quantitative evaluation method can be to start from the maximum score of the quantitative evaluation metric, deduct points from the candidate solutions based on the deduction items, and obtain the corresponding quantitative evaluation score.

[0132] In a specific application, such as Figure 4 As shown, the quantitative evaluation indicators can specifically include the number of steps in the solution. The server can determine the quantitative indicator evaluation score corresponding to the number of steps in the solution by comparing the number of steps in the solution with the average number of steps in similar solutions. The average number of steps in similar solutions can be obtained statistically. Further, as... Figure 4 As shown, the quantitative evaluation indicators can specifically include the estimated time consumption of the solution. The server can then determine the quantitative indicator evaluation score corresponding to the estimated time consumption by comparing it with the average time consumption of similar solutions. The average time consumption of similar solutions can be obtained through statistical analysis. Furthermore, as... Figure 4 As shown, specific quantitative evaluation metrics can include solution cost. The server can determine the solution cost by calculating the ratio of search time to execution time in candidate solutions, and then determine the corresponding quantitative evaluation score for that solution cost. Finally, as... Figure 4 As shown, after obtaining the quantitative evaluation scores corresponding to the number of steps in the solution, the estimated time consumption of the solution, and the cost of the solution, the server will obtain the rule evaluation score of the candidate solution based on these three quantitative evaluation scores.

[0133] Understandably, a higher time consumption ratio indicates a higher solution cost, and a higher solution cost corresponds to a lower evaluation score. For example, the server can determine the evaluation score corresponding to the solution cost based on a predefined time consumption ratio and score correspondence table. In this table, different time consumption ratios correspond to different evaluation scores, and the correlation is negative. This time consumption ratio and score correspondence table can be configured according to the actual application scenario.

[0134] In practical applications, the full score of each quantitative evaluation indicator can be different. The sum of the full scores of multiple quantitative evaluation indicators is the full score of the quantitative evaluation rule. Based on the quantitative evaluation scores of each of the multiple quantitative evaluation indicators, the rule evaluation score of the candidate solution can be obtained by directly accumulating the quantitative evaluation scores of each of the multiple quantitative evaluation indicators.

[0135] In practical applications, the full score of each quantitative evaluation indicator can be the same. Thus, the full score of each of the multiple quantitative evaluation indicators can be the same as the full score of the quantitative evaluation rule. Based on the quantitative evaluation scores corresponding to each of the multiple quantitative evaluation indicators, the server will obtain the weight allocation of the multiple quantitative evaluation indicators. Based on the weight allocation of the multiple quantitative evaluation indicators, the quantitative evaluation scores corresponding to each of the multiple quantitative evaluation indicators will be weighted and calculated to obtain the rule evaluation score of the candidate solution.

[0136] In this embodiment, multiple quantitative evaluation indicators can be used to quantitatively evaluate candidate solutions from multiple perspectives, thereby achieving an objective evaluation of candidate solutions based on quantitative evaluation rules and obtaining accurate rule evaluation scores for candidate solutions.

[0137] In an exemplary embodiment, for each candidate solution, a solution evaluation result is obtained based on the candidate solution's rule evaluation score and model evaluation score, including:

[0138] For each candidate solution, the rule evaluation score and model evaluation score of the candidate solution are weighted and calculated based on predefined rule evaluation weights and model evaluation weights to obtain the solution evaluation result of the candidate solution;

[0139] Task planning methods also include:

[0140] Once the user's decision results are obtained, the rule evaluation weights and model evaluation weights are adjusted accordingly.

[0141] The rule evaluation weight refers to the weight set for quantitative evaluation based on quantitative evaluation rules, and can be configured according to the actual application scenario. The model evaluation weight refers to the weight set for quantitative evaluation based on a large language model, and can also be configured according to the actual application scenario. It should be noted that when configuring, the sum of the rule evaluation weight and the model evaluation weight can be set to 1. For example, if the rule evaluation weight is 0.3, then the model evaluation weight is 0.7, meaning that the result of quantitative evaluation based on the large language model is the primary reference.

[0142] For example, for each candidate solution, the server will perform a weighted calculation on the rule evaluation score and model evaluation score of the candidate solution based on predefined rule evaluation weights and model evaluation weights to obtain the solution evaluation score of the candidate solution, and then use the solution evaluation score of the candidate solution as the solution evaluation result of the candidate solution.

[0143] For example, by utilizing the evaluation results of multiple candidate solutions, a first solution for the target task can be determined from among them. Based on this, the server determines the confidence level of the first solution. If the confidence level indicates that the first solution cannot be determined as the target solution for the target task, the server needs to delegate the choice of solution to the user, i.e., request user feedback, allowing the user to choose whether to determine the first solution as the target solution for the target task. When requesting user feedback, the server generates a user decision request, sends it back to the user, and obtains the user's decision result based on the user's decision request. Based on the user's decision result, the server determines whether to adopt the first solution for the target task as the target solution for the target task, and further generates the task solution for the user's task. Upon obtaining the user's decision result, the server adjusts the rule evaluation weights and model evaluation weights accordingly to optimize them.

[0144] In specific applications, when the user's decision result is represented by the first solution to the target task as the target solution, it indicates that the current rule evaluation weight and model evaluation weight are relatively reasonable. The server can appropriately increase the larger weight of the rule evaluation weight and model evaluation weight, and decrease the smaller weight of the rule evaluation weight and model evaluation weight.

[0145] In specific applications, when the user's decision result is represented by the first solution to the target task as the basic solution to the target task, it indicates that the current rule evaluation weights and model evaluation weights are somewhat unreasonable. The server can increase the smaller weight of the rule evaluation weights and model evaluation weights and decrease the larger weight of the rule evaluation weights and model evaluation weights.

[0146] In practical applications, if the user's decision result representation does not take the first solution of the target task as the target solution, it indicates that the current rule evaluation weights and model evaluation weights are relatively unreasonable, and the server needs to significantly adjust the rule evaluation weights and model evaluation weights.

[0147] In this embodiment, by using rule evaluation weights and model evaluation weights to perform weighted calculations on the rule evaluation scores and model evaluation scores, the evaluation results of candidate solutions can be accurately determined. When the user's decision results are obtained, the rule evaluation weights and model evaluation weights can be adjusted according to the user's decision results, thereby optimizing the quantitative evaluation of candidate solutions and enhancing the objectivity of the quantitative evaluation.

[0148] In an exemplary embodiment, a task solution for the user task is generated based on a first solution for the target task, including:

[0149] The confidence level of the first solution is determined based on the evaluation results of the first solution for the target task and the evaluation results of the remaining solutions among multiple candidate solutions.

[0150] If the confidence level of the first solution is greater than a predefined confidence level threshold, the first solution will be used as the target solution for the target task.

[0151] Based on the target solution for the target task, generate a task solution for the user task.

[0152] Confidence level refers to a quantitative or qualitative measure of the degree of certainty regarding the reliability, rationality, and effectiveness of the first solution. Understandably, a higher confidence level indicates stronger reliability, rationality, and effectiveness of the first solution, meaning it can be trusted. Conversely, a lower confidence level indicates weaker reliability, rationality, and effectiveness, meaning it cannot be trusted.

[0153] For example, after obtaining the first solution to the target task, the server needs to evaluate the reliability of the first solution. At this point, the server determines the confidence level of the first solution based on the evaluation results of the first solution and the evaluation results of the remaining solutions among multiple candidate solutions. Then, it compares the confidence level of the first solution with a predefined confidence threshold. If the confidence level of the first solution is greater than the predefined confidence threshold, the first solution is considered reliable. The server then uses the first solution as the target solution for the target task and generates a task solution for the user task based on the target solution. The predefined confidence threshold can be configured according to the actual application scenario. For example, the predefined confidence threshold could be 0.75.

[0154] In practical applications, when the first solution is used as the target solution for the target task, the server will integrate the target solution of the target task into the task solution of the user task, and continue to execute the subsequent tasks of the target task among multiple tasks to be executed, so as to further generate the task solution of the user task.

[0155] In this embodiment, the confidence level of the first solution can be calculated by using the evaluation results of the first solution and the evaluation results of the remaining solutions among multiple candidate solutions. Then, the confidence level of the first solution can be used to accurately determine whether the first solution is credible. If the confidence level of the first solution is greater than a predefined confidence level threshold, the first solution can be directly used as the target solution of the target task, thereby determining the target solution of the target task. Based on the target solution of the target task, the task solution of the user task can be generated.

[0156] In an exemplary embodiment, the solution evaluation result includes a solution evaluation score; based on the solution evaluation result of the first solution for the target task and the solution evaluation results of the remaining solutions among multiple candidate solutions, the confidence level of the first solution is determined, including:

[0157] Based on the evaluation scores of the remaining solutions among multiple candidate solutions, the evaluation scores of the solutions to be compared are determined.

[0158] Calculate the scheme evaluation score of the first solution to the target task, and the score difference between the evaluation scores of the schemes to be compared;

[0159] Calculate the confidence level of the first solution based on the score difference.

[0160] For example, the evaluation results include an evaluation score for the chosen option, such as... Figure 5 As shown, when calculating the confidence score of the first solution, the server first determines the evaluation score of the solution to be compared based on the evaluation scores of the remaining solutions among multiple candidate solutions. Then, it calculates the evaluation score of the first solution for the target task, as well as the score difference between the evaluation scores of the solutions to be compared. Using the score difference and a predefined confidence calculation formula, the confidence score of the first solution is calculated. This predefined confidence calculation formula can be configured according to the actual application scenario. It can be understood that the confidence score of the first solution can be a value between 0 and 1, with a confidence score closer to 0 indicating lower confidence and a confidence score closer to 1 indicating higher confidence.

[0161] In specific applications, such as Figure 5As shown, when determining the evaluation score of the solution to be compared, the server can sort the evaluation scores of the remaining solutions among multiple candidate solutions, and select the highest evaluation score as the evaluation score of the solution to be compared. Alternatively, the server can calculate the average evaluation score of the remaining solutions among multiple candidate solutions, and use the resulting average evaluation score as the evaluation score of the solution to be compared.

[0162] In practical applications, a predefined formula for calculating confidence level is used: Confidence Level = (Highest Score / 10) * (Δscore / Highest Score). Here, the highest score refers to the evaluation score of the first solution, and Δscore refers to the evaluation score of the solution being compared. In essence, this method determines the confidence level of the first solution by comparing its evaluation score with the evaluation scores of the remaining solutions among the candidate solutions. A larger difference indicates a higher confidence level and greater credibility for the first solution, while a smaller difference indicates a lower confidence level and less credibility for the first solution.

[0163] In this embodiment, when the solution evaluation result includes a solution evaluation score, the solution evaluation score to be compared is determined based on the solution evaluation scores of the remaining solutions among multiple candidate solutions. This allows for the accurate calculation of the confidence level of the first solution by utilizing the solution evaluation score of the first solution for the target task and the score difference between the solution evaluation scores to be compared.

[0164] In one exemplary embodiment, the task planning method further includes:

[0165] If the confidence level of the first solution is less than or equal to a predefined confidence threshold, the first solution will be used as the recommended solution, and a user decision request will be generated based on the recommended solution.

[0166] Obtain the user decision results based on the user's decision request;

[0167] Based on the user's decision, generate a task solution for the user's task.

[0168] For example, if the confidence level of the first solution is less than or equal to a predefined confidence threshold, it indicates that the first solution is unreliable, meaning it cannot be directly identified as the target solution for the task. The server needs to grant the user the option to choose whether to accept the first solution as the target solution. In this case, the server will use the first solution as a recommended solution and generate a user decision request based on it to request user feedback, i.e., allow the user to select a solution. Upon receiving the user decision request, the user's terminal will display a user decision interface based on the request. The user can provide feedback on their decision through interactive operations on the user decision interface. After obtaining the user decision feedback based on the user decision request, the server can generate a task solution for the user's task based on the user's decision.

[0169] In practical applications, when generating user decision requests based on recommended solutions, the server organizes multiple candidate solutions into structured, user-friendly options and generates a concise "execution summary" for each recommended solution, highlighting its core ideas, main advantages, and potential risks to assist users in making decisions.

[0170] In a specific application, the user decision interface can be specifically as follows: Figure 6 As shown, this includes details of the recommended solution (i.e., the core idea), its main advantages and potential risks, a comparison of evaluation data from multiple candidate solutions, multiple selection options, and a feedback area. Among these, for example... Figure 6 As shown, the various options include "Approve and continue with this plan," "Modify this plan," and "Reject and replan." "Approve and continue with this plan" indicates the user accepts the recommended plan; "Modify this plan" indicates the user modifies the recommended plan; and "Reject and replan" indicates the user rejects the recommended plan. Furthermore, if the user chooses to modify the recommended plan, they can enter their suggestions in the feedback area; if the user chooses to reject the recommended plan, they can enter user instructions for secondary planning in the feedback area.

[0171] Furthermore, considering multiple candidate solutions as three candidate solutions, namely Solution A, Solution B, and Solution C, with Solution C being an example of the first solution, such as... Figure 7 As shown, the evaluation data comparison can be presented in tabular form, including the scores of multiple candidate solutions on various evaluation metrics. For example, Figure 7 As shown, the various evaluation indicators may include feasibility score, innovation score, user experience score, cost estimate (X, which can be a cost unit), and time efficiency (hours), etc.

[0172] Understandably, by viewing the details (i.e., the core idea), main advantages, and potential risks of the recommended solutions displayed in the user decision-making interface, users can make accurate decisions about whether the recommended solutions are suitable. By comparing the solution evaluation data displayed in the user decision-making interface, users can intuitively compare multiple candidate solutions and make accurate decisions.

[0173] In practical applications, based on the user's decision, the server can determine whether the user chooses to accept the recommended solution, modify the recommended solution, or reject the recommended solution, and then generate a task solution for the user's task based on the user's choice.

[0174] In this embodiment, when the confidence level of the first solution is low, the first solution can be used as a recommended solution for the user to decide whether to accept the recommended solution. Based on the obtained user decision results, a task solution for the user task is generated. By letting the user make the final decision, human-machine collaboration is achieved, ensuring the correctness of the target task decision.

[0175] In one exemplary embodiment, a task solution for the user task is generated based on the user's decision, including:

[0176] If the user's decision result indicates acceptance of the recommended solution, the first solution will be taken as the target solution for the target task.

[0177] Based on the target solution for the target task, generate a task solution for the user task.

[0178] For example, based on the user's decision, the server can determine whether the user chooses to accept the recommended solution, modify the recommended solution, or reject the recommended solution. If the user's decision indicates acceptance of the recommended solution, the server will take the first solution as the target solution of the target task, integrate the target solution of the target task into the task solution of the user task, and continue to execute the subsequent tasks of the target task among multiple tasks to be executed, so as to further generate the task solution of the user task.

[0179] In practical applications, when multiple options are provided to the user, the user's decision result can include the selected option. Based on the user's selection, the server can determine whether the user chooses to accept the recommended solution, modify the recommended solution, or reject the recommended solution. In a specific application, the multiple options may include a first option indicating that the user chooses to accept the recommended solution, a second option indicating that the user chooses to modify the recommended solution, and a third option indicating that the user chooses to reject the recommended solution and instructs for replanning. If the user selects the first option, the server can determine that the user's decision result indicates acceptance of the recommended solution.

[0180] In this embodiment, when the user's decision result indicates acceptance of the recommended solution, the first solution is taken as the target solution for the target task. This allows the determination of the target solution for the target task to be achieved using user feedback, ensuring the correctness of the target task decision. Furthermore, based on the target solution for the target task, a task solution for the user task can be generated.

[0181] In one exemplary embodiment, a task solution for the user task is generated based on the user's decision, including:

[0182] When modifying the recommended solution based on the user's decision result representation, the target solution for the target task is generated based on the modified recommended solution based on the user's decision result.

[0183] Based on the target solution for the target task, generate a task solution for the user task.

[0184] For example, when the user's decision result indicates a modification to the recommended solution, the server extracts the user's feedback from the decision result, modifies the recommended solution based on the feedback, generates a target solution for the target task, integrates this target solution into the user's task solution, and continues to execute the subsequent tasks of the target task among multiple pending tasks to further generate the user's task solution. Specifically, the user's feedback can be the modification suggestions entered by the user through the user decision interface.

[0185] In practical applications, when modifying the recommended solution based on the modification suggestions, the server can input the modification suggestions and the recommended solution into the large language model, allowing the large language model to modify the recommended solution based on the modification suggestions and output the target solution for the target task.

[0186] In this embodiment, when the recommended solution is modified based on the user's decision result, the recommended solution can be modified based on the user's decision result. This allows the determination of the target solution for the target task using user feedback, ensuring the correctness of the target task decision. Furthermore, a task solution for the user task can be generated based on the target solution for the target task.

[0187] In one exemplary embodiment, a task solution for the user task is generated based on the user's decision, including:

[0188] If the user's decision result indicates rejection of the recommended solution, obtain the user's instructions from the user's decision result;

[0189] Based on user instructions, perform secondary task planning to generate task solutions for the user's tasks.

[0190] For example, if the user's decision result indicates rejection of the recommended solution, the server will extract the user's feedback instructions from the user's decision result, and perform secondary task planning based on the user's instructions to generate a task solution for the user's task.

[0191] In practical applications, secondary task planning is performed based on user instructions, i.e., task replanning. At this point, the server can break down the user task based on the user instructions, resulting in multiple tasks to be executed sequentially, thus generating a task solution for the user's instructions. It's understandable that if a user chooses to reject the recommended solution, it means the user is not satisfied with it. In this case, the user can provide feedback containing their decision, allowing the server to understand their needs. Upon receiving the user instructions, the server can analyze them to understand the user's requirements and use them to accurately break down the user task.

[0192] In a specific application, while executing multiple tasks sequentially, the server still monitors the execution process and determines whether a planning fork is encountered. If a planning fork is encountered, it indicates that there are multiple possible subsequent paths for the task to be executed. The server then identifies the task to be executed as the target task and triggers a multi-agent decision-making process. This process involves calling multiple specialized agents to execute the target task, obtaining candidate solutions generated by each agent, and then quantitatively evaluating each candidate solution to obtain its own evaluation result. Based on the evaluation results of each candidate solution, the first solution for the target task is determined, and a task solution for the user task is generated based on the first solution for the target task.

[0193] In this embodiment, when the user's decision result indicates rejection of the recommended solution, the user's instructions are obtained from the user's decision result, and secondary task planning is performed using the user's instructions to generate a task solution for the user's task. This can utilize user feedback to achieve accurate generation of task solutions and ensure the correctness of the target task decision.

[0194] In one exemplary embodiment, a task solution for the user task is generated based on the target solution for the target task, including:

[0195] When a user task is broken down into multiple tasks to be executed, the subsequent tasks of the target task in the multiple tasks to be executed will continue to be executed.

[0196] Until multiple pending tasks are completed, a task solution for the user task is generated based on the target solution of the target task and the second solution of the remaining pending tasks among the multiple pending tasks.

[0197] For example, when the server executes the target task extracted from the user task, it first determines the target solution of the target task. If the user task is divided into multiple tasks to be executed, the server continues to execute the subsequent tasks of the target task in the multiple paths of tasks to be executed until the multiple tasks to be executed are completed. Based on the target solution of the target task and the second solution of the remaining tasks to be executed in the multiple tasks to be executed, the server generates the task solution of the user task.

[0198] In practical applications, during the execution of subsequent tasks, if there are no multi-path subsequent tasks, the server can directly generate a second solution for the subsequent tasks and integrate the target solution for the objective task with the second solutions for the remaining tasks to generate the user's task solution. If there are multi-path subsequent tasks, multiple specialized agents need to be introduced to perform parallel reasoning on the multi-path subsequent tasks, determine the second solutions for the multi-path subsequent tasks, and then integrate the target solution for the objective task with the second solutions for the remaining tasks to generate the user's task solution.

[0199] In a specific application, when multiple specialized intelligent agents are introduced to perform parallel reasoning on subsequent tasks with multiple paths, if the user needs to choose a solution and the user rejects the recommended solution, then a secondary task planning needs to be performed based on the user's instructions to generate a task solution for the user's task.

[0200] In this embodiment, user tasks can be processed by decomposition. Based on the target solution of the target task, the subsequent tasks to be executed can be processed by continuing to execute the subsequent tasks to be executed among multiple tasks to be executed. When multiple tasks to be executed are completed, the task solution of the user task is generated according to the target solution of the target task and the second solution of the remaining tasks to be executed among multiple tasks to be executed. This enables the generation of task solutions for user tasks.

[0201] In one exemplary embodiment, the task planning method further includes:

[0202] Determine the task domain to which the user's task belongs;

[0203] Based on the task domain, the metadata tags of multiple candidate agents, and predefined scheduling strategies, agent matching is performed to select multiple specialized agents from the multiple candidate agents.

[0204] Once the user's decision is obtained, the scheduling strategy is adjusted accordingly.

[0205] The task domain refers to the specific knowledge category, professional discipline, or activity type to which the user task belongs. It defines the scope of core knowledge, methodologies, and tools required to solve the user task, and can specifically be one of the predefined domains. Predefined domains can be configured according to actual application scenarios. For example, predefined domains may specifically include search, analysis, programming, and design.

[0206] Metadata tags are labels used to describe the capabilities and characteristics of candidate agents. For example, metadata tags can specifically describe the role definition of a candidate agent. For instance, metadata tags could be labels such as "strategist," "creative," or "domain expert."

[0207] The scheduling strategy refers to the policy used to describe how to schedule agents, which can be the matching relationship between the agent's metadata tags and the task domain. For example, the scheduling strategy can take the form of: metadata tag 1 - task domain 1, task domain 2; metadata tag 2 - task domain 2, task domain 3; metadata tag 3 - task domain 3, task domain 4.

[0208] For example, when multiple specialized agents need to be introduced to perform parallel reasoning on a target task, the server first determines the task domain to which the user task belongs, and then performs agent matching based on the task domain, the metadata tags of each of the multiple candidate agents, and the predefined scheduling strategy, so as to select multiple specialized agents from the multiple candidate agents to perform parallel reasoning on the target task, obtain multiple candidate solutions, and then use the solution evaluation results of each of the multiple candidate solutions to determine the first solution for the target task from the multiple candidate solutions.

[0209] For example, based on this, the server determines the confidence level of the first solution. If the confidence level of the first solution indicates that it cannot be determined as the target solution for the target task, the server needs to give the choice of solution to the user, that is, request user feedback and let the user choose whether to determine the first solution as the target solution for the target task. In the case of requesting user feedback, the server generates a user decision request, sends the user decision request back to the user, and obtains the user's decision result based on the user's decision request. Based on the user's decision result, the server determines whether to adopt the first solution for the target task as the target solution for the target task, and further generates the task solution for the user's task. Upon obtaining the user's decision result, the server adjusts the scheduling strategy according to the user's decision result to optimize the scheduling strategy.

[0210] In practical applications, the task domain of a user task can be obtained through semantic analysis using a large language model. Based on the determined task domain, the server determines the matching metadata tags from a predefined scheduling strategy, and then matches the matching metadata tags with the metadata tags of multiple candidate agents to select multiple specialized agents from the candidate agents to perform parallel reasoning on the target task.

[0211] In practical applications, if the user's decision represents the first solution to the target task as the target solution itself, the current scheduling strategy is relatively reasonable, and the server does not need to adjust it. If the user's decision represents the first solution to the target task as the basic solution to the target task, the current scheduling strategy is somewhat unreasonable, and the server can appropriately adjust the matching relationship between metadata tags and task domains in the scheduling strategy, such as adding or removing task domains matched by each metadata tag. If the user's decision does not represent the first solution to the target task as the target solution, the current scheduling strategy is relatively unreasonable, and the server needs to significantly adjust the matching relationship between metadata tags and task domains in the scheduling strategy, such as reducing the number of task domains matched by each metadata tag.

[0212] In this embodiment, by determining the task domain to which the user task belongs, agent matching can be performed using the task domain, the metadata tags of multiple candidate agents, and predefined scheduling strategies. This enables accurate selection of multiple specialized agents, allowing multiple specialized agents to perform parallel reasoning on the target task. Collective intelligence is then used to reduce the uncertainty and randomness of task planning decisions. After obtaining the user's decision results, the scheduling strategy can be adjusted to optimize it, thereby achieving further accurate selection of multiple specialized agents.

[0213] In an exemplary embodiment, multiple specialized intelligent agents plan feasible subsequent paths for the target task, resulting in candidate solutions generated by each agent, including:

[0214] By using multiple specialized intelligent agents, and based on the task processing prompts of each agent, feasible subsequent paths for the target task are planned, resulting in candidate solutions generated by each agent.

[0215] Task planning methods also include:

[0216] Based on the user's decision, the task processing prompts for multiple specialized intelligent agents are adjusted accordingly.

[0217] Task processing prompts refer to the prompts set for specialized intelligent agents to handle tasks, primarily guiding them in their execution. For example, task processing prompts may include processing steps and agent role definitions. Specifically, processing steps might instruct the agent where to obtain data and how to process it. Agent role definitions allow for the assignment of roles to the specialized intelligent agents. It's understandable that the role definitions for different agents will be completely different, and consequently, the task processing prompts for different specialized intelligent agents will also be entirely different.

[0218] For example, based on the identification of multiple specialized intelligent agents, the server will, through these agents and based on their respective task processing prompts, plan feasible subsequent paths for the target task, resulting in candidate solutions generated by each agent. Then, using the evaluation results of these candidate solutions, the server will determine the first solution for the target task from among them. It is understandable that, since the task processing prompts can be different, the candidate solutions generated by the multiple specialized intelligent agents can be completely different.

[0219] For example, based on this, the server determines the confidence level of the first solution. If the confidence level of the first solution indicates that it cannot be determined as the target solution for the target task, the server needs to give the choice of solution to the user, that is, request user feedback and let the user choose whether to determine the first solution as the target solution for the target task. When requesting user feedback, the server generates a user decision request, sends the user decision request back to the user, and obtains the user's decision result based on the user's decision request. Based on the user's decision result, the server determines whether to adopt the first solution for the target task as the target solution for the target task, and further generates the task solution for the user's task. Upon obtaining the user's decision result, the server adjusts the task processing prompts for each of the multiple specialized intelligent agents according to the user's decision result to optimize the task processing prompts.

[0220] In practical applications, if the user's decision representation uses the first solution to the target task as the target solution, the current scheduling strategy is reasonable, and the server does not need to adjust the task processing prompts. If the user's decision representation uses the first solution to the target task as the basic solution, the current task processing prompts are somewhat unreasonable, and the server can adjust them appropriately, such as adding or removing processing steps. If the user's decision representation does not use the first solution to the target task as the target solution, the current task processing prompts are unreasonable, and the server needs to significantly adjust them, such as redesigning the processing steps or defining the agent roles.

[0221] In this embodiment, for each specialized intelligent agent, task processing prompts can be used to guide the agent in accurately planning the target task. By obtaining the user's decision results and adjusting the task processing prompts for each of the multiple specialized intelligent agents accordingly, the prompts can be optimized, enabling the agents to further process the target task more accurately.

[0222] In one exemplary embodiment, the task planning method further includes:

[0223] When a user task is received, it is broken down into multiple tasks to be executed, and these tasks are executed sequentially.

[0224] When executing multiple tasks in sequence, if the current task is a multi-path planning type task, the current task will be identified as the target task.

[0225] For example, when a user task is received, the server breaks it down into multiple tasks to be executed and executes them sequentially. During the sequential execution of these tasks, the server monitors the execution process and determines whether a planning fork is encountered. If a planning fork is encountered, it indicates that the current task to be executed is a multi-path planning task with multiple possible subsequent paths. The server then identifies the current task to be executed as the target task. If no planning fork is encountered, it indicates that the current task to be executed is a single-path planning task with only a single subsequent path. The server then generates a solution in one go based on its internal knowledge.

[0226] In this embodiment, by breaking down the user task into multiple tasks to be executed, and in the process of executing multiple tasks in sequence, if the current task to be executed is a multi-path planning type task, the current task to be executed can be identified as the target task, and the feasible subsequent paths of the target task can be further planned, so as to realize simultaneous execution and planning, which can improve the efficiency of task planning and decision-making.

[0227] The task processing method provided in this application embodiment can be applied to, for example, Figure 8 In the application environment shown, terminal 802 communicates with server 804 via a network. A data storage system can store the data that server 804 needs to process. The data storage system can be set up independently, integrated into server 804, or placed in the cloud or on other network servers. When a user task is received from terminal 802, server 804 breaks it down into multiple tasks to be executed and executes them sequentially. During the sequential execution of these tasks, if the current task is a multi-path planning type task, it is identified as the target task. When the target task is reached, multiple specialized intelligent agents plan feasible subsequent paths for the target task, obtaining candidate solutions generated by each agent. These candidate solutions are then quantitatively evaluated to obtain evaluation results. Based on these evaluation results, a first solution for the target task is determined. A task solution for the user task is generated based on the first solution and fed back to terminal 802.

[0228] Among them, terminal 802 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 804 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0229] In one exemplary embodiment, such as Figure 9As shown, a task processing method is provided. This embodiment illustrates the method applied to a server, but it is understood that the method can also be applied to a terminal, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes steps 902 to 912. Wherein:

[0230] Step 902: When a user task is received, the user task is broken down into multiple tasks to be executed, and the multiple tasks to be executed are executed in sequence.

[0231] For example, when a user task is received, the server will break the user task down into multiple tasks to be executed, and execute the multiple tasks to be executed in sequence.

[0232] Step 904: In the process of executing multiple tasks in sequence, if the current task to be executed is a multi-path planning type task, the current task to be executed is determined as the target task.

[0233] For example, during the sequential execution of multiple tasks to be executed, the server will monitor the execution process and determine whether a planning fork is encountered. If a planning fork is encountered, it means that the current task to be executed is a multi-path planning type task with multiple optional subsequent paths. The server will then determine the current task to be executed as the target task.

[0234] Step 906: When the target task is executed, multiple professional intelligent agents plan feasible subsequent paths for the target task, and obtain candidate solutions generated by each of the multiple professional intelligent agents.

[0235] For example, when the target task is executed, the server will call multiple specialized intelligent agents. Through these agents, feasible subsequent paths for the target task will be planned, resulting in candidate solutions generated by each agent.

[0236] In practical applications, when multiple specialized intelligent agents are invoked, the server will send a unified context and instructions to each specialized intelligent agent, so that each specialized intelligent agent can plan a feasible subsequent path for the target task based on the unified context and instructions, and obtain the candidate solutions generated by each specialized intelligent agent.

[0237] Step 908: Quantitatively evaluate the multiple candidate solutions to obtain the evaluation results for each candidate solution.

[0238] For example, when multiple professional intelligent agents generate their own candidate solutions, the server will perform quantitative evaluation on each candidate solution to achieve an objective evaluation of the multiple candidate solutions and obtain the evaluation results of each candidate solution.

[0239] Step 910: Based on the evaluation results of multiple candidate solutions, determine the first solution for the target task.

[0240] For example, the server will determine the first solution for the target task from among the multiple candidate solutions based on the solution evaluation results of each candidate solution.

[0241] In practical applications, the solution evaluation results include solution evaluation scores. The server can then determine the candidate solution with the highest solution evaluation score from among multiple candidate solutions based on their respective solution evaluation scores, and use it as the first solution for the target task.

[0242] In practical applications, the solution evaluation results include solution evaluation priorities. The server can then determine the candidate solution with the highest evaluation priority from among multiple candidate solutions based on their respective evaluation priorities, and use it as the first solution for the target task.

[0243] Step 912: Based on the first solution of the target task, generate a task solution for the user task and send the task solution back to the sender of the user task.

[0244] For example, if a first solution to the target task is obtained, the server can generate a task solution for the user task based on the first solution to the target task, and feed the task solution back to the sender of the user task.

[0245] In practical applications, taking a user task as a query task as an example, the task solution for a query task is actually the query result, and the server will return the query result to the sender of the query task.

[0246] The above task processing method, upon receiving a user task, decomposes the user task into multiple tasks to be executed, and executes these tasks sequentially. This enables the planning and decomposition of the user task. During the sequential execution of these tasks, if the current task is a multi-path planning type task, it is identified as the target task. The method also identifies target tasks where planning branches occur during task execution. When the target task is reached, multiple specialized agents plan feasible subsequent paths for the target task, resulting in candidate solutions generated by each agent. At planning branch points, multiple specialized agents are introduced to perform parallel reasoning on the target task, leveraging collective intelligence to reduce the uncertainty and randomness of task planning decisions. Quantitative evaluation of each candidate solution yields its own evaluation result, allowing for objective evaluation. Based on these evaluation results, a first solution for the target task is determined, and a task solution for the user task is generated. Throughout the process, when encountering planning bifurcation points, the uncertainty and randomness of task planning decisions can be reduced by using parallel reasoning by multiple specialized intelligent agents and quantitative evaluation of the candidate solutions generated by each of the specialized intelligent agents. This improves the quality of task planning decisions and increases their accuracy.

[0247] In an exemplary embodiment, taking the task planning method of this application applied to a task planning system based on a large language model as an example, the task planning scheme of this application will be described. Figure 10 As shown, the user inputs a user task: "Give me a week-long travel plan on the theme of XX, including visits to well-known companies, exchanges with industry professionals, and unique cultural activities." The master intelligence agent in the task planning system will break down the user task into multiple tasks to be executed, such as "transportation," "accommodation," "visits," "exchanges," and "culture," and execute these tasks sequentially. When executing the task of "visits," the planning branch will occur regarding "which companies to visit" due to the large number of companies available and the need to consider factors such as opening hours, geographical location, and user preference matching. At this point, "visits" is actually the target task. The master intelligence agent will initiate a multi-professional intelligence agent vote based on a voting consensus mechanism. Through multiple professional intelligence agents, feasible subsequent paths for the target task of "visits" will be planned, resulting in candidate solutions generated by each professional intelligence agent. Figure 10As shown, there can be three professional agents. Agent A's candidate solution A is to suggest the most geographically compact and time-efficient tour route; Agent B's candidate solution B is to suggest focusing on visiting companies where you have connections to arrange internal exchanges; and Agent C's candidate solution C is to suggest adding more innovative cultural landmarks and activities. Based on these three candidate solutions, the task planning system initiates a voting and decision-making process, performing weighted voting (i.e., quantitative evaluation) on each of the three candidate solutions to obtain their respective evaluation results. Based on these evaluation results, the first solution for the target task is determined as candidate solution C. Since the confidence level of candidate solution C is lower than a predefined confidence threshold, a human feedback mode is activated, allowing the user to make further choices. At this point, the task planning system displays candidate solution C as the recommended solution to the user and provides a selection prompt. For example, the user prompt could be: "A recommended travel plan has been generated for you. The current plan includes more innovative cultural landmarks and activities. Do you want to modify it?" Users can choose to accept the recommended plan, modify and supplement the recommended plan, or reject the recommended plan. If the user modifies and supplements the plan by "wanting to add a visit to XXX", the task planning system will modify the recommended plan, integrate the activity "wanting to add a visit to XXX" into the recommended plan, and continue to execute the subsequent pending tasks according to the user's choice.

[0248] The master agent is a centralized agent responsible for task decomposition, process coordination, aggregating opinions from multiple specialized agents, and making the final decision. Specialized agents are sub-agents focused on specific domains or perspectives (such as search, analysis, programming, and creativity), scheduled by the master agent, and generating independent solutions or plans for specific problems. The voting consensus mechanism is a decision-making strategy that gathers opinions from multiple independent agents and selects the best solution based on preset rules (such as majority voting, weighted scoring, and consistency judgment) to reduce the uncertainty and randomness of a single source. The human feedback mode refers to an interactive mode where, when an automated system encounters decision points with high uncertainty, high risk, or where consensus cannot be reached, the process is proactively interrupted, and the choice is handed over to the user for judgment.

[0249] In practical applications, when displaying recommended solutions to users, the user decision-making interface can specifically be as follows: Figure 11As shown, the system includes details of candidate solution C (i.e., its core idea), its main advantages and potential risks, a comparison of evaluation data for the three candidate solutions, multiple selection options, and a feedback area. These options include "Approve and continue this solution," "Modify this solution," and "Reject and replan." "Approve and continue this solution" indicates the user accepts the recommended solution, "Modify this solution" indicates the user modifies the recommended solution, and "Reject and replan" indicates the user rejects the recommended solution. Furthermore, if the user chooses to modify the recommended solution, they can enter modification suggestions in the feedback area (e.g., "We hope to add visits to XXX"). If the user chooses to reject the recommended solution, they can enter user instructions for secondary planning in the feedback area. The comparison of solution evaluation data can be presented in tabular form, including the scores of each of the three candidate solutions on multiple evaluation indicators. For example, as shown... Figure 11 As shown, the various evaluation indicators may include feasibility score, innovation score, user experience score, cost estimate (X, which can be a cost unit), and time efficiency (hours), etc.

[0250] Understandably, by viewing the details (i.e., the core idea), main advantages, and potential risks of the recommended solutions displayed in the user decision-making interface, users can make accurate decisions about the suitability of the recommended solutions. Through the comparison of solution evaluation data displayed in the user decision-making interface, users can intuitively compare multiple candidate solutions and make accurate decisions. In other words, users have the right to know and the right to choose at key decision points, and the final plan is more in line with their personalized needs, while experiencing the collaboration rather than control of artificial intelligence.

[0251] In related technologies, task planning systems are mostly based on single-model decision-making or simple linear processes. Typical examples include: First, the single-agent planning mode, where a large language model is responsible for the entire task's "think-act" cycle. When encountering a fork in the road, this large language model generates a solution based on its internal knowledge, and the decision quality depends entirely on the current state and preferences of this single model, lacking comparison and verification mechanisms. Second, the ensemble prompt mode, which requires the large language model to generate multiple possible solutions at once through a single prompt word and performs a simple comparison. This method increases diversity to some extent, but it is essentially still a "self-game" of a single model, lacking a structured evaluation and selection mechanism, and the final decision may still have randomness.

[0252] In summary, the relevant technologies suffer from the following drawbacks: First, the decision-making source is singular: relying on a single model or a one-time generation, the model-generated plan is heavily influenced by its initial state and built-in preferences, resulting in unstable output, susceptibility to bias, and a tendency to get trapped in local optima rather than the globally optimal path. Second, there is a lack of quantitative evaluation: the selection of solutions lacks transparent and quantifiable evaluation standards, making it more like a "black box" selection process. Third, it cannot effectively utilize human wisdom: at complex or critical decision points, the system cannot proactively seek human assistance. Fourth, the process is rigid: the decision-making process is fixed and cannot be dynamically adjusted according to the difficulty of the decision and the context. Fifth, the failure handling strategy is simplistic: when encountering execution obstacles, the system can usually only attempt a limited number of retries or simple replanning, lacking the ability to utilize collective wisdom to find breakthrough solutions from multiple perspectives.

[0253] Based on this, this application proposes a task planning method applicable to task planning systems. When task planning encounters forks or obstacles, it improves decision-making accuracy through multi-agent voting and human interaction. This method can be integrated as a decision enhancement module into various AI agent products. Users perceive a smarter, more reliable AI assistant that fully respects user opinions in key decisions.

[0254] Specifically, the task planning scheme of this application mainly includes the following key technical points: First, multi-agent parallel reasoning: For planning forks or execution obstacles, multiple professional agents with different expertise and thinking patterns are simultaneously invoked to generate diverse subsequent plan schemes, using collective wisdom to reduce uncertainty and randomness. Second, rule-based voting consensus mechanism: A set of configurable voting and scoring rules is designed to quantitatively evaluate and rank the schemes generated by multiple professional agents in multiple dimensions, select the best one, and automatically filter out the best scheme with the highest confidence. Third, seamless integration of human feedback: When the confidence of the voting result is lower than the threshold or multiple schemes have their own advantages and disadvantages, the system automatically pauses and generates clear options and their advantages and disadvantages analysis, which are then handed over to the user for final decision-making, realizing human-machine collaboration. That is, using human-machine interaction as a backup, when the system cannot reach a high-confidence consensus, human feedback is actively integrated to ensure the correctness of key decisions. Fourth, dynamic workflow engine: The system can adaptively switch according to the decision confidence, intelligently selecting fully automatic, semi-automatic (requiring user confirmation), or fully manual (user decision) execution modes, balancing efficiency and accuracy.

[0255] In practical applications, the task planning method of this application is applied to a task planning system, which is equivalent to extending the original single-agent architecture with a parallel reasoning and decision-making layer. Its core architecture is as follows: Figure 12 As shown, it includes a main control agent, a pool of specialized agents, a voting consensus module, and a human-computer interaction module. The following section will combine... Figure 12 Each module will be introduced separately.

[0256] In a specific application, such as Figure 12 As shown, the master control agent mainly includes a task coordinator and an expert scheduler. The task coordinator is primarily used to receive user tasks, perform initial decomposition, and monitor the entire execution process. When a planning fork (such as multiple optional branches) is identified, a multi-agent decision-making process is triggered. The expert scheduler is mainly used to select and invoke a group (usually 3-5) of the most relevant professional agents from a predefined pool of professional agents based on the domain of the current user task (such as search, analysis, programming, design). The scheduling strategy can be based on metadata tag matching.

[0257] In a specific application, such as Figure 12 As shown, the pool of specialized intelligent agents includes multiple specialized intelligent agents that are invoked. Figure 12 The example uses n instances, namely professional agent 1, professional agent 2, ..., professional agent n. Its implementation mechanism mainly consists of a set of pre-configured instances of a large language model with different system prompts (i.e., task processing prompts) and areas of expertise (specifically, different instances of the same model or different models can be called), and each professional agent has a unique role definition. For example... Figure 12 As shown, the workflow of each specialized intelligent agent is as follows: receiving the unified context and scheduling instructions sent by the master intelligent agent, independently and in parallel generating a complete solution or next step plan, and returning it to the master intelligent agent, that is, returning multiple candidate solutions to the master intelligent agent.

[0258] For example, specialized intelligent agents can be categorized into strategic intelligent agents, creative intelligent agents, and domain expert intelligent agents. Strategic intelligent agents prioritize goal-achieving paths, resource allocation, and risk avoidance in their task prompts, favoring robust and executable solutions. Creative intelligent agents encourage divergent thinking and breaking conventional constraints, aiming to discover novel and non-obvious solutions. Domain expert intelligent agents are configured with detailed knowledge bases and thinking patterns for specific domains (such as shopping or programming), primarily providing professional and authoritative solutions.

[0259] In a specific application, taking three specialized intelligent agents as an example—a strategy agent, a creative agent, and a domain expert agent—the workflow of the specialized agent pool can be as follows: Figure 13As shown, the master control agent can issue strategic tasks, creative tasks, and domain-specific tasks to the strategic agent, creative agent, and domain expert agent respectively through task parsing and scheduling. After executing the strategic task, the strategic agent will generate solution A; after executing the creative task, the creative agent will generate solution B; and after executing the domain-specific task, the domain expert agent will generate solution C. The candidate solutions generated by the three professional agents will be aggregated into the voting consensus module for processing.

[0260] In a specific application, such as Figure 12 As shown, the voting consensus module includes a scheme evaluator and a confidence calculator. The voting consensus process of this module can be specifically described as follows: Figure 14 As shown, the solution evaluator is mainly used to receive candidate solutions generated by all specialized agents (such as...). Figure 14 The example shows N candidate solutions, which are evaluated according to preset, configurable rules. The core evaluation mechanism can be as follows: First, a rule engine is used for scoring, setting rules to score quantifiable metrics (i.e., quantitative evaluation metrics, such as number of steps, estimated time, and cost) (the scoring range can be 1-10). Second, a large language model is used as a judge. A high-performance, neutral large language model (such as GPT-4 (Generative Pre-trained Transformer 4)) is invoked as the "judge," and its prompt word template requires it to summarize and score each solution based on the following quantitative evaluation dimensions (the scoring range can be 1-10). These quantitative evaluation dimensions can include feasibility, efficiency, consistency with the overall goal, innovativeness, and risk assessment. The final score is a weighted average of the rule engine score (i.e., rule evaluation score) and the large language model's evaluation score (i.e., model evaluation score). Figure 14 As shown, the rule evaluation weight can be 0.3, and the model evaluation weight can be 0.7, to obtain the total score for each solution. Based on the total score of each solution, the best solution, i.e., the highest-scoring solution, can be selected by sorting the scores. The confidence calculator mainly calculates the confidence score of the highest-scoring solution. If the confidence score is higher than a preset threshold (i.e., a predefined confidence threshold, such as...), then... Figure 14 If the confidence level is 0.75, the best solution will be automatically adopted. Otherwise, the system will enter the human-computer interaction module to request human feedback; that is, if the confidence level is insufficient, intervention will be requested.

[0261] The confidence level can be calculated as follows: first, calculate the score difference Δscore between the highest-scoring scheme and the second-highest-scoring scheme, and then calculate the confidence level using the confidence level calculation formula (e.g., confidence = (highest score / 10) * (Δscore / highest score)).

[0262] In a specific application, such as Figure 12 As shown, the human-computer interaction module mainly includes a decision point generator, a user interface, and a result parser. The decision point generator is primarily used to organize multiple solutions into structured, user-friendly options when the confidence level is below a preset threshold, and to generate a concise "execution summary" for the highest-scoring option, highlighting its core idea, main advantages, and potential risks. The user interface is mainly used to present the best solution recommended by the task planning system to the user through a graphical interface. The result parser is mainly used to feed back the user's selection or free input (user selection result, i.e., user decision result) to the system, transforming it into instructions for the next step. Simultaneously, this user selection result (i.e., user decision result) can be used as feedback data to optimize the weights or prompts of the corresponding specialized intelligent agent.

[0263] In a specific application, the feedback optimization flowchart can be specifically as follows: Figure 15 As shown, based on the user's decision, the task planning system records feedback data (i.e., the user's decision result, specifically approval / modification / rejection). The system analyzes the feedback data and optimizes the task processing prompts for the professional intelligent agent, adjusts voting weights (i.e., rule evaluation weights and model evaluation weights), and improves the scheduling strategy, thereby obtaining an optimized task planning system that can improve the quality of decision-making when executing new tasks.

[0264] In one exemplary embodiment, the complete workflow of the task planning system from task input to final completion can be specifically as follows: Figure 16 As shown, it includes the following steps:

[0265] 1. Task Input: The user submits a user task to the task planning system, which is then received by the main control agent.

[0266] 2. Planning and Execution: The master control agent formulates an initial plan P0, breaks down the user task into multiple steps (i.e., tasks to be executed), and begins to execute the multiple steps (i.e., tasks to be executed) in the initial plan P0 in sequence.

[0267] 3. Decision point detection: Detect the execution process of multiple steps to determine whether a planning fork has been encountered.

[0268] 4. Multi-agent decision-making: If a planning fork is encountered, it means that there are multiple feasible subsequent paths for the current step, i.e., the current step is the target task. Then, the multi-agent decision-making process is triggered, and multiple professional agents are scheduled to plan the feasible subsequent paths for the current step and generate multiple candidate solutions in parallel.

[0269] 5. Voting and Consensus: The voting consensus module quantitatively evaluates all candidate solutions and calculates the confidence level C of the best solution.

[0270] 6. Dynamic routing:

[0271] a. High-confidence automatic execution: If the confidence level C is higher than the threshold, the task planning system automatically adopts the best solution.

[0272] b. Low-confidence manual intervention: If the confidence level C is lower than the threshold, the human-computer interaction module is invoked to present the best solution recommended by the task planning system to the user (i.e., push the current recommended solution to the user): "Current execution solution X of the task planning system...", in order to request the user to make a decision.

[0273] c. Users have the following options:

[0274] - Approval / Default: The task planning system will adopt this recommended approach.

[0275] - Modification: Users can make adjustments based on the recommended scheme, and the task planning system will adopt the modified scheme.

[0276] - Rejection: The user completely rejects the plan and requests new instructions. The user can enter new instructions or directions, and the task planning system will re-plan based on the new instructions.

[0277] 7. Integration and Cycle: Integrate the adopted new solutions into the original plan and continue execution until all steps are completed, outputting a task solution.

[0278] It should be noted that the task planning method in this application, by introducing a decision-making mechanism of multi-agent voting consensus and structured human feedback, produces the following significant beneficial effects compared to existing single-agent planning systems:

[0279] 1. Significantly improve decision quality and reliability: Through a hybrid evaluation mechanism of "rules + big language model evaluation", multiple solutions are quantitatively scored in multiple dimensions, which effectively reduces the bias and randomness of a single model, selects more reliable solutions from the root, and greatly improves the success rate of tasks.

[0280] 2. Significantly optimizes human-computer interaction efficiency and experience:

[0281] Clear roles: The collaborative boundaries of "system suggestions and human decision-making" are clearly defined. Users feel like they are approving an intelligent draft rather than being presented with a difficult problem. The experience is upgraded from "passive decision-making" to "proactive management", which enhances their trust in the system.

[0282] Simplifying complexity: By presenting only a single recommended solution and its rationale, instead of forcing users to make difficult choices among multiple complex options, the cognitive load and decision fatigue of users are greatly reduced.

[0283] 3. Enhance system robustness and interpretability: When faced with ambiguous requirements or high-risk decisions, the system can provide an optimal draft that has been demonstrated from multiple perspectives and its reasoning basis. The decision-making process changes from a "black box" to an "interpretable white box", becoming more transparent and traceable.

[0284] 4. Continuous optimization capability: The system can record the user's final approval decision (approval, modification of content, reason for rejection). This high-quality human feedback data can be used to continuously optimize the prompts, voting weights, or scheduling strategies of the professional intelligent agent, making the system smarter with use.

[0285] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0286] Based on the same inventive concept, this application also provides a task planning apparatus for implementing the task planning method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more task planning apparatus embodiments provided below can be found in the limitations of the task planning method described above, and will not be repeated here.

[0287] In one exemplary embodiment, such as Figure 17 As shown, a task planning device is provided, including: a first task execution module 1702, a first quantitative evaluation module 1704, a first solution selection module 1706, and a first solution generation module 1708, wherein:

[0288] The first task execution module 1702 is used to, when executing the target task decomposed from the user task, plan the feasible subsequent path of the target task through multiple professional intelligent agents, and obtain the candidate solutions generated by each of the multiple professional intelligent agents.

[0289] The first quantitative evaluation module 1704 is used to perform quantitative evaluation on multiple candidate solutions respectively, and obtain the evaluation results of each candidate solution.

[0290] The first solution selection module 1706 is used to determine the first solution for the target task based on the evaluation results of multiple candidate solutions.

[0291] The first solution generation module 1708 is used to generate a task solution for the user task based on the first solution for the target task.

[0292] The aforementioned task planning device, when executing the target task extracted from the user task, utilizes multiple specialized intelligent agents to plan feasible subsequent paths for the target task, obtaining candidate solutions generated by each agent. At planning fork points, multiple agents are introduced to perform parallel reasoning on the target task, leveraging collective intelligence to reduce the uncertainty and randomness of task planning decisions. By quantitatively evaluating each candidate solution, evaluation results are obtained, enabling objective assessment of multiple candidate solutions. Based on these evaluation results, a first solution for the target task is determined, and a task solution for the user task is generated. Throughout this process, at planning fork points, the parallel reasoning of multiple specialized intelligent agents and the quantitative evaluation of their respective candidate solutions reduce uncertainty and randomness in task planning decisions, improving the quality and accuracy of these decisions.

[0293] In an exemplary embodiment, the first quantitative evaluation module is further configured to perform quantitative evaluation on multiple candidate solutions based on predefined quantitative evaluation rules to obtain rule evaluation scores for each candidate solution, and to perform quantitative evaluation on multiple candidate solutions based on predefined quantitative evaluation dimensions using a large language model to obtain model evaluation scores for each candidate solution, and to obtain a solution evaluation result for each candidate solution based on the rule evaluation score and model evaluation score of the candidate solution for each candidate solution.

[0294] In an exemplary embodiment, the first quantitative evaluation module is further configured to perform quantitative evaluation on each candidate solution based on multiple quantitative evaluation indicators in the predefined quantitative evaluation rules, obtain the quantitative indicator evaluation scores corresponding to each of the multiple quantitative evaluation indicators, and obtain the rule evaluation score of the candidate solution based on the quantitative indicator evaluation scores corresponding to each of the multiple quantitative evaluation indicators.

[0295] In an exemplary embodiment, the first quantitative evaluation module is further configured to perform a weighted calculation on the rule evaluation score and model evaluation score of each candidate solution based on predefined rule evaluation weights and model evaluation weights to obtain the solution evaluation result of the candidate solution, and adjust the rule evaluation weights and model evaluation weights according to the user decision result if the user decision result is obtained.

[0296] In an exemplary embodiment, the first solution generation module is further configured to determine the confidence level of the first solution based on the scheme evaluation results of the first solution for the target task and the scheme evaluation results of the remaining solutions among multiple candidate solutions. If the confidence level of the first solution is greater than a predefined confidence level threshold, the first solution is used as the target solution for the target task, and a task solution for the user task is generated based on the target solution for the target task.

[0297] In an exemplary embodiment, the solution evaluation result includes a solution evaluation score. The first solution generation module is further configured to determine the solution evaluation score to be compared based on the solution evaluation scores of the remaining solutions among multiple candidate solutions, calculate the solution evaluation score of the first solution for the target task, and the score difference between the solution evaluation scores to be compared, and calculate the confidence level of the first solution based on the score difference.

[0298] In an exemplary embodiment, the first solution generation module is further configured to, when the confidence level of the first solution is less than or equal to a predefined confidence level threshold, use the first solution as a recommended solution, generate a user decision request based on the recommended solution, obtain the user decision result fed back according to the user decision request, and generate a task solution for the user task based on the user decision result.

[0299] In an exemplary embodiment, the first solution generation module is further configured to, when the user's decision result indicates acceptance of the recommended solution, use the first solution as the target solution for the target task, and generate a task solution for the user task based on the target solution for the target task.

[0300] In an exemplary embodiment, the first solution generation module is further configured to modify the recommended solution based on the user decision result, generate a target solution for the target task, and generate a task solution for the user task based on the target solution for the target task.

[0301] In an exemplary embodiment, the first solution generation module is further configured to obtain user instructions from the user decision results when the user decision results indicate rejection of the recommended solution, perform secondary task planning based on the user instructions, and generate a task solution for the user task.

[0302] In an exemplary embodiment, the first solution generation module is further configured to, when the user task is decomposed into multiple tasks to be executed, continue to execute the subsequent tasks to be executed of the target task among the multiple tasks to be executed until the multiple tasks to be executed are completed, and generate a task solution for the user task based on the target solution of the target task and the second solutions of the remaining tasks to be executed among the multiple tasks to be executed.

[0303] In an exemplary embodiment, the first task execution module is further configured to determine the task domain to which the user task belongs, perform agent matching based on the task domain, the metadata tags of multiple candidate agents and a predefined scheduling strategy, select multiple professional agents from the multiple candidate agents, and adjust the scheduling strategy based on the user decision results when the user decision results are obtained.

[0304] In an exemplary embodiment, the first task execution module is further configured to plan feasible subsequent paths for the target task based on the task processing prompts of each of the multiple professional intelligent agents, thereby obtaining candidate solutions generated by each of the multiple professional intelligent agents, and adjusting the task processing prompts of each of the multiple professional intelligent agents according to the user's decision results.

[0305] In an exemplary embodiment, the first task execution module is further configured to, upon receiving a user task, decompose the user task into multiple tasks to be executed, and execute the multiple tasks to be executed sequentially, and, in the process of executing the multiple tasks to be executed sequentially, determine the current task to be executed as the target task if the current task to be executed is a multi-path planning type task.

[0306] In one exemplary embodiment, such as Figure 18 As shown, a task processing device is provided, including: a second task execution module 1802, a second quantitative evaluation module 1804, a second solution selection module 1806, and a second solution generation module 1808, wherein:

[0307] The second task execution module 1802 is used to, when receiving a user task, decompose the user task into multiple tasks to be executed, and execute the multiple tasks to be executed in sequence. During the sequential execution of the multiple tasks to be executed, if the current task to be executed is a multi-path planning type task, the current task to be executed is determined as the target task. When the target task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task respectively, and obtain the candidate solutions generated by the multiple professional intelligent agents.

[0308] The second quantitative evaluation module 1804 is used to perform quantitative evaluation on multiple candidate solutions respectively, and obtain the evaluation results of each of the multiple candidate solutions.

[0309] The second solution selection module 1806 is used to determine the first solution for the target task based on the evaluation results of the multiple candidate solutions.

[0310] The second solution generation module 1808 is used to generate a task solution for the user task based on the first solution for the target task, and to feed back the task solution to the sender of the user task.

[0311] The aforementioned task processing device, upon receiving a user task, breaks it down into multiple tasks to be executed and executes them sequentially. This enables the planning and decomposition of the user task. During the sequential execution of these tasks, if the current task is a multi-path planning type task, it identifies that task as the target task. It can also identify target tasks where planning branches occur during task execution. When the target task is reached, multiple specialized agents plan feasible subsequent paths for it, generating candidate solutions for each agent. At planning branch points, multiple agents can be introduced to perform parallel reasoning on the target task, leveraging collective intelligence to reduce the uncertainty and randomness of task planning decisions. By quantitatively evaluating each candidate solution, evaluation results are obtained, allowing for objective evaluation of multiple candidate solutions. Based on these evaluation results, a first solution for the target task is determined, and a task solution for the user task is generated. Throughout the process, when encountering planning bifurcation points, the uncertainty and randomness of task planning decisions can be reduced by using parallel reasoning by multiple specialized intelligent agents and quantitative evaluation of the candidate solutions generated by each of the specialized intelligent agents. This improves the quality of task planning decisions and increases their accuracy.

[0312] The modules in the aforementioned task planning and task processing devices can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0313] In one exemplary embodiment, a computer device is provided, which can be a server or a terminal. Taking the computer device as a server as an example, its internal structure diagram can be as follows: Figure 19 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores metadata tags and other data for multiple candidate intelligent agents. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a task planning method.

[0314] Those skilled in the art will understand that Figure 19 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0315] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0316] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0317] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0318] It should be noted that the data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are all data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0319] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0320] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0321] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A task planning method, characterized in that, The method includes: When the target task extracted from the user task is executed, multiple professional intelligent agents plan the feasible subsequent paths of the target task, and obtain the candidate solutions generated by each of the multiple professional intelligent agents. Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions; Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined. Based on the first solution of the target task, a task solution for the user task is generated.

2. The method according to claim 1, characterized in that, The quantitative evaluation of multiple candidate solutions to obtain the evaluation results of each candidate solution includes: Based on predefined quantitative evaluation rules, multiple candidate solutions are quantitatively evaluated separately to obtain the rule evaluation scores of each candidate solution. Using a large language model, based on predefined quantitative evaluation dimensions, the multiple candidate solutions are quantitatively evaluated to obtain their respective model evaluation scores. For each candidate solution, the solution evaluation result is obtained based on the rule evaluation score and model evaluation score of the candidate solution.

3. The method according to claim 2, characterized in that, The predefined quantitative evaluation rules are used to quantitatively evaluate multiple candidate solutions respectively, obtaining the rule evaluation scores for each of the multiple candidate solutions, including: For each candidate solution, a quantitative evaluation is performed on the candidate solution based on multiple quantitative evaluation indicators in the predefined quantitative evaluation rules, and the quantitative evaluation scores corresponding to each of the multiple quantitative evaluation indicators are obtained. The rule evaluation score of the candidate solution is obtained based on the evaluation scores of the quantitative indicators corresponding to each of the multiple quantitative evaluation indicators.

4. The method according to claim 2, characterized in that, For each candidate solution, the solution evaluation result is obtained based on the rule evaluation score and model evaluation score of the candidate solution, including: For each candidate solution, the rule evaluation score and model evaluation score of the candidate solution are weighted and calculated based on predefined rule evaluation weights and model evaluation weights to obtain the solution evaluation result of the candidate solution; The method further includes: Upon obtaining the user's decision result, the rule evaluation weight and the model evaluation weight are adjusted based on the user's decision result.

5. The method according to claim 1, characterized in that, The step of generating a task solution for the user task based on the first solution for the target task includes: Based on the evaluation results of the first solution to the target task and the evaluation results of the remaining solutions among the multiple candidate solutions, the confidence level of the first solution is determined. If the confidence level of the first solution is greater than a predefined confidence level threshold, the first solution shall be used as the target solution for the target task. Based on the target solution of the target task, a task solution for the user task is generated.

6. The method according to claim 5, characterized in that, The solution evaluation result includes a solution evaluation score; determining the confidence level of the first solution based on the solution evaluation result of the first solution to the target task and the solution evaluation results of the remaining solutions among the multiple candidate solutions includes: Based on the evaluation scores of the remaining solutions among the multiple candidate solutions, the evaluation score of the solution to be compared is determined; Calculate the scheme evaluation score of the first solution to the target task, and the score difference between the evaluation scores of the schemes to be compared; Based on the score difference, the confidence level of the first solution is calculated.

7. The method according to claim 5, characterized in that, The method further includes: If the confidence level of the first solution is less than or equal to the predefined confidence level threshold, the first solution is used as the recommended solution, and a user decision request is generated based on the recommended solution. Obtain the user decision results based on the user decision request; Based on the user's decision, a task solution for the user's task is generated.

8. The method according to claim 7, characterized in that, The step of generating a task solution for the user task based on the user's decision results includes: If the user's decision result indicates acceptance of the recommended solution, the first solution will be used as the target solution for the target task. Based on the target solution of the target task, a task solution for the user task is generated.

9. The method according to claim 7, characterized in that, The step of generating a task solution for the user task based on the user's decision results includes: If the user's decision result characterizes the modification of the recommendation scheme, the recommendation scheme is modified based on the user's decision result to generate the target solution for the target task; Based on the target solution of the target task, a task solution for the user task is generated.

10. The method according to claim 7, characterized in that, The step of generating a task solution for the user task based on the user's decision results includes: If the user's decision result indicates rejection of the recommended solution, a user instruction is obtained from the user's decision result; Based on the user instructions, secondary task planning is performed to generate a task solution for the user task.

11. The method according to claim 5, characterized in that, The process of generating a task solution for the user task based on the target solution of the target task includes: When the user task is broken down into multiple tasks to be executed, the subsequent tasks to be executed for the target task among the multiple tasks to be executed continue to be executed; Until the multiple tasks to be executed are completed, a task solution for the user task is generated based on the target solution of the target task and the second solutions of the remaining tasks to be executed among the multiple tasks to be executed.

12. The method according to any one of claims 1 to 11, characterized in that, The method further includes: Determine the task domain to which the user task belongs; Based on the task domain, the metadata tags of each of the multiple candidate agents, and the predefined scheduling strategy, agent matching is performed, and multiple specialized agents are selected from the multiple candidate agents. If the user's decision result is obtained, the scheduling strategy is adjusted according to the user's decision result.

13. The method according to any one of claims 1 to 11, characterized in that, The step of planning feasible subsequent paths for the target task through the multiple specialized intelligent agents to obtain candidate solutions generated by each of the multiple specialized intelligent agents includes: By using the multiple specialized intelligent agents, and based on the task processing prompts of each of the multiple specialized intelligent agents, feasible subsequent paths for the target task are planned respectively, and candidate solutions generated by each of the multiple specialized intelligent agents are obtained. The method further includes: Upon obtaining the user's decision result, the task processing prompts of the multiple professional intelligent agents are adjusted according to the user's decision result.

14. The method according to any one of claims 1 to 11, characterized in that, The method further includes: When a user task is received, the user task is broken down into multiple tasks to be executed, and the multiple tasks to be executed are executed sequentially. During the sequential execution of the multiple tasks to be executed, if the current task to be executed is a multi-path planning type task, the current task to be executed is determined as the target task.

15. A task processing method, characterized in that, The method includes: When a user task is received, the user task is broken down into multiple tasks to be executed, and the multiple tasks to be executed are executed sequentially. During the sequential execution of the plurality of tasks to be executed, if the current task to be executed is a multi-path planning type task, the current task to be executed is determined as the target task; When the target task is executed, multiple specialized intelligent agents plan feasible subsequent paths for the target task, thereby obtaining candidate solutions generated by each of the multiple specialized intelligent agents. Quantitative evaluations are performed on multiple candidate solutions to obtain the evaluation results for each of the multiple candidate solutions; Based on the evaluation results of the multiple candidate solutions, a first solution for the target task is determined. Based on the first solution of the target task, a task solution for the user task is generated, and the task solution is fed back to the sender of the user task.

16. A task planning device, characterized in that, The device includes: The first task execution module is used to, when executing a target task extracted from a user task, plan feasible subsequent paths for the target task through multiple professional intelligent agents, and obtain candidate solutions generated by each of the multiple professional intelligent agents. The first quantitative evaluation module is used to perform quantitative evaluation on multiple candidate solutions respectively, and obtain the evaluation results of each of the multiple candidate solutions. The first solution selection module is used to determine the first solution for the target task based on the evaluation results of the multiple candidate solutions. The first solution generation module is used to generate a task solution for the user task based on the first solution for the target task.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 15.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 15.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 15.

Images