Task planning method and device based on large language model, equipment and storage medium

Abstract

This invention provides a task planning method, apparatus, device, and storage medium based on a large language model. It receives and parses natural language task requirements through a large language model, generating structured scene feature data. The scene feature data is then input into a self-learning dedicated optimization model to obtain an initial solution for task planning. Starting from the initial solution, a heuristic algorithm is invoked for iterative optimization to generate an executable task planning scheme. The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme. This invention achieves natural language interaction and intelligent scheduling through a large language model, solving the problems of unintelligent human-computer interaction and reliance on manual requirement parsing in traditional systems, thus improving the system's usability and automation level. By constructing a collaborative optimization mechanism of "dedicated optimization model - heuristic algorithm," it effectively overcomes the technical bottlenecks of low iteration efficiency and long solution time caused by random initialization in traditional heuristic algorithms.

Description

Technical Field

[0001] This invention relates to the field of task planning technology, and in particular to a task planning method, apparatus, device, and storage medium based on a large language model. Background Technology

[0002] With the rapid development of industry and information technology, the demand for complex task planning in fields such as engineering scheduling, resource allocation, and intelligent decision-making is increasing. Traditional task planning methods mainly rely on heuristic algorithms, iteratively searching through preset rules. While capable of handling certain optimization problems, they generally suffer from bottlenecks such as high computational complexity, slow convergence speed, and poor adaptability to dynamic environments. Especially in real-world scenarios with multiple tasks, high dimensionality, and strong constraints, the solution efficiency and solution quality of traditional methods are insufficient to meet real-time requirements. Related technologies introduce neural network models to learn optimization rules to accelerate the solution process, but the need for natural language processing and system interaction is difficult to handle efficiently, and the level of human-machine collaboration is insufficient. Dedicated models rely on large amounts of labeled data for training, resulting in limited generalization ability in practical applications and difficulty in covering complex and ever-changing planning scenarios. Furthermore, a systematic solution is still lacking for how to organically coordinate cutting-edge large language models, traditional optimization algorithms, and dedicated models to construct an efficient and adaptive solution chain. Summary of the Invention

[0003] This invention provides a task planning method, apparatus, device, and storage medium based on a large language model, which addresses the shortcomings of existing technologies in planning complex tasks, such as unintelligent interaction, low solution efficiency, weak model generalization, and poor system coordination.

[0004] This invention provides a task planning method based on a large language model, comprising: The system receives and parses natural language task requirements through a large language model, generating structured scene feature data. The scene feature data is input into a self-learning optimization model to obtain an initial solution for task planning; Starting with the initial solution, a heuristic algorithm is invoked for iterative optimization to generate an executable task planning scheme; The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

[0005] According to the task planning method based on a large language model provided by the present invention, the self-learning dedicated optimization model adopts a lightweight neural network architecture, which sequentially includes an input layer, a feature extraction layer and an output layer. Low-rank matrices A and B are connected in parallel on at least one fully connected layer of the feature extraction layer and / or the output layer; The weights of the self-learning dedicated optimization model are: ,in For the pre-trained weight matrix, To update the weight matrix online.

[0006] According to the task planning method based on a large language model provided by the present invention, the training method of the self-learning dedicated optimization model includes: Collect scenario environment data, historical optimization data, and performance feedback data for task planning scenarios; The scene environment data, historical optimization data, and performance feedback data are preprocessed to obtain preprocessed scene environment data. Using the preprocessed scene environment data as input and the high-quality solutions obtained after optimization by the heuristic algorithm as labels, training sample pairs are constructed. Based on the training sample pairs, the parameters of the initial self-learning dedicated optimization model are optimized using gradient descent until the first loss function converges, thus obtaining the pre-trained weight matrix. .

[0007] According to the task planning method based on a large language model provided by the present invention, the preprocessing of the scene environment data, historical optimization data, and performance feedback data to obtain preprocessed scene environment data includes: The scene environment data, historical optimization data, and performance feedback data are subjected to missing value correction, outlier removal, and standardization processing to obtain preprocessed scene environment data.

[0008] According to the task planning method based on a large language model provided by the present invention, the self-learning dedicated optimization model performs online parameter updates based on feedback from the executable task planning scheme, including: Set up a data collection window, collect executable task planning schemes from the converged output of the heuristic algorithm within the data collection window, and generate self-learning data samples. When the number of self-learning data samples reaches a preset threshold, the fine-tuning operation of the self-learning dedicated optimization model is triggered.

[0009] According to the task planning method based on a large language model provided by the present invention, the fine-tuning operation includes: The executable task planning scheme, obtained from the convergence of the collected heuristic algorithms, is input into the pre-trained self-learning optimization model. The mean squared error is used as the second loss function, and the parallel low-rank matrices A and B are updated via backpropagation. , ; Where η is the fine-tuning learning rate and L is the mean squared error; The updated weight matrix is ​​obtained based on the updated parallel low-rank matrices A and B. .

[0010] The task planning method based on a large language model provided by the present invention further includes: Real-time monitoring of the solution time, resource utilization, and convergence accuracy of the generated executable task planning scheme; Based on the monitoring results, the iteration threshold of the heuristic algorithm or the learning rate of the online parameter update is dynamically adjusted.

[0011] The present invention also provides a task planning device based on a large language model, comprising: The first generation module is used to receive and parse the natural language task requirements through a large language model and generate structured scene feature data. The input module is used to input the scene feature data into the self-learning dedicated optimization model to obtain the initial solution for task planning; The second generation module is used to take the initial solution as the starting point, call the heuristic algorithm to perform iterative optimization, and generate an executable task planning scheme. The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the task planning method based on a large language model as described in any of the preceding claims.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the task planning method based on a large language model as described above.

[0014] This invention provides a task planning method, apparatus, device, and storage medium based on a large language model. It receives and parses natural language task requirements using a large language model to generate structured scene feature data. This scene feature data is then input into a self-learning dedicated optimization model to obtain an initial solution for task planning. Starting from this initial solution, a heuristic algorithm is invoked for iterative optimization to generate an executable task planning scheme. The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme. This invention achieves natural language interaction and intelligent scheduling through a large language model, solving the problems of unintelligent human-computer interaction and reliance on manual requirement analysis in traditional systems, thus improving the system's usability and automation level. By constructing a collaborative optimization mechanism of "dedicated optimization model - heuristic algorithm," the high-quality initial solution output by the model guides the algorithm to converge quickly, effectively overcoming the technical bottlenecks of low iteration efficiency and long solution time caused by random initialization in traditional heuristic algorithms. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is one of the flowcharts of the task planning method based on a large language model provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the low-rank adaptive fine-tuning process provided in an embodiment of the present invention; Figure 3 This is the second flowchart of the task planning method based on a large language model provided in the embodiments of the present invention; Figure 4 This is a functional structure diagram of the task planning device based on a large language model provided in an embodiment of the present invention; Figure 5 This is a functional structure diagram of the electronic device provided in the embodiments of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Figure 1 A flowchart of the task planning method based on a large language model provided in the embodiments of the present invention is shown below. Figure 1 As shown, the task planning method based on a large language model provided in this embodiment of the invention includes: Step 101: Receive and parse the natural language task requirements through a large language model to generate structured scene feature data; This invention utilizes a large language model to perform task parsing, user question answering, and end-to-end scheduling in natural language form, autonomously organizing input data and driving subsequent planning processes.

[0019] Step 102: Input the scene feature data into the self-learning dedicated optimization model to obtain the initial solution for task planning; This invention constructs a self-learning lightweight model, which takes task planning scenario environment data (such as task parameters, constraints, resource configuration information, etc.) as input and high-quality planning schemes obtained by heuristic algorithm iteration as labels to complete pre-training, so that it learns the task planning logic and solution space characteristics; Step 103: Starting from the initial solution, call the heuristic algorithm to perform iterative optimization and generate an executable task planning scheme; The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

[0020] Traditional task planning often relies on heuristic algorithms for iterative search. While it can solve complex optimization problems such as task prioritization, resource matching, and process integration to some extent, it suffers from high time complexity and slow iterative convergence, making it difficult to meet the real-time and efficiency requirements of task planning in real-world scenarios. Especially in complex scenarios with multi-task parallelism, high-dimensional constraints, and dynamic changes in requirements, the solution efficiency and the quality of the planning scheme further decline, failing to adapt to dynamically changing task planning needs. Meanwhile, although existing intelligent task planning methods incorporate machine learning techniques, they suffer from inefficient natural language task interaction processing, insufficient generalization ability of specialized models, and poor synergy between models and traditional algorithms. Furthermore, model updates consume significant computational resources, making it difficult to achieve efficient self-learning iteration.

[0021] The task planning method based on a large language model provided in this invention receives and parses natural language task requirements through a large language model, generating structured scene feature data. This scene feature data is then input into a self-learning dedicated optimization model to obtain an initial solution for task planning. Starting from this initial solution, a heuristic algorithm is invoked for iterative optimization to generate an executable task planning scheme. The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme. This invention achieves natural language interaction and intelligent scheduling through a large language model, solving the problems of unintelligent human-computer interaction and reliance on manual requirement parsing in traditional systems, thus improving the system's usability and automation level. By constructing a collaborative optimization mechanism of "dedicated optimization model - heuristic algorithm," the high-quality initial solution output by the model guides the algorithm to converge quickly, effectively overcoming the technical bottlenecks of low iteration efficiency and long solution time caused by random initialization in traditional heuristic algorithms. This invention effectively solves the problems of low efficiency, insufficient model generalization ability, and low level of system interaction and scheduling intelligence of traditional heuristic algorithms by organically combining the intelligent scheduling capability of large language models, the collaborative planning of dedicated models and heuristic algorithms, and the efficient self-learning mechanism of low-rank adaptive algorithms. It significantly improves the solution efficiency and solution quality of complex task planning problems, and provides efficient, flexible and intelligent optimization solutions for various general-purpose task planning scenarios.

[0022] Based on any of the above embodiments, this invention selects a basic model with strong natural language understanding and task scheduling capabilities (such as a large language model based on the Transformer architecture) to construct a scheduling and interaction hub. This model supports receiving task planning requirements and scenario questions in natural language form, extracting core task parameters (such as resource type, constraints, and optimization objectives) through semantic parsing, and generating standardized data input formats. At the same time, it presets scheduling rules and, based on task complexity and real-time requirements, autonomously triggers subsequent processes such as dedicated optimization model inference, heuristic algorithm iteration, and model self-learning updates to achieve fully automated scheduling across the entire chain.

[0023] Based on any of the above embodiments, the self-learning dedicated optimization model adopts a lightweight neural network architecture, which sequentially includes an input layer, a feature extraction layer, and an output layer. The feature extraction layer employs a fully connected network or a lightweight neural network structure to meet the demands of rapid inference. The model input is the preprocessed feature vector of the scene environment data. ,in, For a single sample feature vector, The input sample size is denoted as ; the output is the initial solution vector for the task planning optimization problem, corresponding to the initial search starting point of the heuristic algorithm.

[0024] Low-rank matrices A and B are connected in parallel on at least one fully connected layer of the feature extraction layer and / or the output layer; The weights of the self-learning dedicated optimization model are: ,in For the pre-trained weight matrix, To update the weight matrix online.

[0025] As shown in Figure 2, in at least one fully connected layer of the linear layer of the self-learning dedicated optimization model, such as the feature extraction layer and / or the output layer, low-rank matrices A and B are connected in parallel. The model pre-training weight matrix is... Weight update amount during fine-tuning Final model weights During fine-tuning, only the low-rank matrices A and B are trained, while the remaining weight parameters are frozen. This significantly reduces computational resource consumption while enabling efficient iteration of model parameters, ensuring that the model can quickly absorb optimization experience from real-world scenarios.

[0026] This invention uses the inference results of a dedicated model as the initial planning scheme for a heuristic algorithm, enabling efficient algorithm startup and significantly reducing iteration convergence time. By continuously collecting high-quality planning schemes output by the heuristic algorithm, a low-rank adaptive fine-tuning technique is adopted. A low-rank matrix is ​​connected in parallel in the linear layer of the dedicated model and the remaining weights are frozen. The low-rank matrix is ​​trained only, enabling the model to self-learn and update with low computational resource consumption, and continuously improving generalization ability.

[0027] Based on any of the above embodiments, the training method of the self-learning dedicated optimization model includes: Step 201: Collect scene environment data, historical optimization data, and performance feedback data under the task planning scenario; Step 202: Preprocess the scene environment data, historical optimization data, and performance feedback data to obtain preprocessed scene environment data; Step 203: Using the preprocessed scene environment data as input and the high-quality solution obtained after optimization by the heuristic algorithm as the label, construct training sample pairs; Step 204: Based on the training sample pairs, optimize the parameters of the initial self-learning dedicated optimization model using gradient descent until the first loss function converges, thus obtaining the pre-trained weight matrix. .

[0028] This invention embodiment performs model pre-training based on a preprocessed dataset: Scene environment data is used as input, and the corresponding high-quality solutions obtained from the iterative convergence of the heuristic algorithm are used as labels to construct training sample pairs (X, Y), where Y is the optimal solution label vector. The training loss function is defined as mean squared error (MSE), with the following formula: in, The true label (a high-quality solution from a heuristic algorithm). For the model's predicted output, This represents the number of training samples.

[0029] The gradient descent method is used to optimize the model parameters. Hyperparameters such as learning rate and number of iterations are set. Through multiple rounds of training, the model fully learns the optimization logic and solution space characteristics of the heuristic algorithm until the loss function converges, thus obtaining a pre-trained self-learning optimization model.

[0030] Based on any of the above embodiments, the scene environment data, historical optimization data, and performance feedback data are preprocessed to obtain preprocessed scene environment data, including: The scene environment data, historical optimization data, and performance feedback data are subjected to missing value correction, outlier removal, and standardization processing to obtain preprocessed scene environment data.

[0031] In this embodiment of the invention, multiple types of data are collected in the task planning scenario, including scenario environment data (such as total resources, constraints, task requirement parameters, etc.), historical optimization data (past resource allocation schemes, heuristic algorithm iteration process data, etc.), and performance feedback data (evaluation indicators such as the execution efficiency and resource utilization rate of historical optimization schemes).

[0032] Data cleaning includes removing or correcting missing and outlier values. For missing values ​​in the data... Mean imputation can be used: Where N is the number of non-missing values ​​in the dataset. For other non-missing data points. If outliers exist, they can be detected and removed using the standard deviation method, with the following formula: in, The standardized value. The mean of the data. represents the standard deviation of the data. When a certain threshold is exceeded, it is considered This is an outlier.

[0033] Transform data of different dimensions to the same scale. The method is min-max normalization, and its formula is: The preprocessed dataset can be represented as: .

[0034] Based on any of the above embodiments, the self-learning dedicated optimization model updates parameters online based on feedback from the executable task planning scheme, including: Set up a data collection window, collect executable task planning schemes from the converged output of the heuristic algorithm within the data collection window, and generate self-learning data samples. When the number of self-learning data samples reaches a preset threshold, the fine-tuning operation of the self-learning dedicated optimization model is triggered.

[0035] In this embodiment of the invention, the fine-tuning operation includes: The executable task planning scheme, obtained from the convergence of the collected heuristic algorithms, is input into the pre-trained self-learning optimization model. The mean squared error is used as the second loss function, and the parallel low-rank matrices A and B are updated via backpropagation. , ; Where η is the fine-tuning learning rate and L is the mean squared error; The updated weight matrix is ​​obtained based on the updated parallel low-rank matrices A and B. .

[0036] In this embodiment of the invention, a heuristic algorithm suitable for non-convex optimization problems of task planning, such as genetic algorithm, particle swarm optimization algorithm, simulated annealing algorithm, etc., is selected. The initial solution output by the dedicated optimization model is received, and iterative search is carried out based on preset iteration rules (such as iteration number threshold, convergence accuracy threshold). The quality of the solution is continuously optimized until the convergence condition is met, and a high-quality task planning optimization solution is output.

[0037] The low-rank adaptive fine-tuning technique enables model self-learning and updating. By paralleling a low-rank matrix in the linear layer of the self-learning dedicated model, model iteration can be completed by training only a small number of low-rank parameters, which greatly reduces the computational resource consumption of fine-tuning. At the same time, it continuously absorbs the experience of high-quality solutions from heuristic algorithms, constantly improving the model's generalization ability, enabling the system to quickly adapt to complex and ever-changing real-world resource management scenarios.

[0038] The task planning method based on a large language model provided by the present invention further includes: Real-time monitoring of the solution time, resource utilization, and convergence accuracy of the generated executable task planning scheme; Based on the monitoring results, the iteration threshold of the heuristic algorithm or the learning rate of the online parameter update is dynamically adjusted.

[0039] In this embodiment of the invention, the system's operating status and output performance are monitored in real time: key evaluation indicators (such as solution time, resource utilization, and solution convergence accuracy) are set, and the entire process operation data is recorded through a data acquisition module; the indicator data is analyzed periodically to evaluate the system's solution efficiency and solution quality. If the indicators decline (such as increased solution time or decreased resource utilization), the system hyperparameters are adjusted (such as heuristic algorithm iteration threshold, model fine-tuning learning rate, and low-rank matrix dimension); when significant changes in the scenario lead to insufficient model generalization ability, the training dataset is reconstructed, and the dedicated optimization model is pre-trained and fine-tuned a second time to ensure that the system continuously adapts to dynamically changing task planning requirements.

[0040] like Figure 3 As shown in the embodiment of the present invention, the task planning method based on a large language model specifically includes: the user submits task requirements in natural language, the LLM in the agent acts as the core of the interaction, and the agent conducts multiple rounds of dialogic clarification and parsing with the user through preset system instructions, and finally transforms the fuzzy requirements into structured task parameters and constraints, completing the key conversion from natural language to machine-understandable instructions, and obtaining structured instructions.

[0041] Subsequently, the structured instructions are sent into the task planning environment, triggering an efficient two-stage solution mechanism: First, a pre-trained lightweight dedicated optimization model is invoked to quickly infer a high-quality initial planning scheme based on historical experience. This provides a high-quality search starting point for subsequent optimization and overcomes the efficiency bottleneck of random initialization in traditional heuristic algorithms. Then, using this initial solution as a "warm start" signal, a heuristic algorithm (such as a genetic algorithm) is invoked to perform a fine iterative search, seeking the optimal or near-optimal final solution in the solution space composed of the specific scenario environment, resource pool, and task parameters.

[0042] After a solution is generated, its key performance indicators, such as solution time, resource utilization, and solution quality, are automatically calculated, forming performance feedback data. This feedback, together with the corresponding scenario data, constitutes new experience samples, which flow into the system's data buffer pool. This closed-loop design endows the system with self-evolution capabilities: on the one hand, new samples can be used for online low-rank fine-tuning of the lightweight model, achieving rapid adaptation; on the other hand, continuously accumulated data can fuel the model's periodic strong-oriented training, thereby continuously improving the system's overall generalization performance and solution efficiency in complex, dynamic task planning scenarios.

[0043] The overall architecture of this invention has strong versatility and flexibility. It is not limited to specific types of task planning scenarios or single heuristic algorithms. It can replace the dedicated model structure, heuristic algorithm type and fine-tune parameters according to actual needs. It can adapt to various high-dimensional, multi-constraint non-convex optimization problems of task planning and has a wide range of applications.

[0044] The task planning method based on a large language model provided in this invention uses the large language model as the core scheduling and interaction hub. It can receive task requests in natural language form and answer scenario-related questions. Simultaneously, it autonomously organizes input data and schedules subsequent multi-stage optimization links to perform task planning optimization. Addressing the non-convex optimization problem commonly found in task planning, a dedicated lightweight model is constructed and pre-trained. The training process uses the scenario environment data of the optimization problem as input and the iterative optimization results of traditional heuristic algorithms as labels, fully learning the optimization logic and rules of heuristic algorithms. In the actual inference stage, the system uses the inference results of the lightweight model as the initial solution of the heuristic algorithm, achieving efficient startup of the heuristic algorithm, significantly accelerating its iterative convergence speed, and quickly obtaining high-quality task planning optimization solutions. To further improve the system's adaptability to complex real-world scenarios, high-quality optimization solutions after heuristic algorithm convergence are continuously collected. A low-rank adaptive fine-tuning technique is used to self-learn and update the lightweight model. A low-rank matrix is ​​connected in parallel in the linear layer of the model, and the remaining weight parameters are frozen. Only the low-rank matrix is ​​trained specifically, effectively improving the model's generalization performance while ensuring fine-tuning efficiency. The large language model can autonomously complete the scheduling and execution of the entire process according to the actual task requirements. Through the organic combination of multi-stage optimization and self-learning mechanism, it effectively solves the problem of high time complexity of traditional heuristic algorithms, significantly improves the efficiency and quality of solving task planning and optimization problems, and provides efficient and universal intelligent optimization solutions for various task planning scenarios.

[0045] The task planning device based on a large language model provided by the present invention is described below. The task planning device based on a large language model described below and the task planning method based on a large language model described above can be referred to in correspondence.

[0046] Figure 4 A functional structure diagram of the task planning device based on a large language model provided in an embodiment of the present invention is shown below. Figure 4 As shown, the task planning device based on a large language model provided in this embodiment of the invention includes: The first generation module 401 is used to receive and parse the natural language task requirements through a large language model and generate structured scene feature data. The input module 402 is used to input the scene feature data into the self-learning dedicated optimization model to obtain the initial solution for task planning; The second generation module 403 is used to take the initial solution as the starting point, call a heuristic algorithm to perform iterative optimization, and generate an executable task planning scheme. The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

[0047] The task planning device based on a large language model provided in this invention receives and parses natural language task requirements through a large language model, generating structured scene feature data; inputs the scene feature data into a self-learning dedicated optimization model to obtain an initial solution for task planning; using the initial solution as a starting point, iterative optimization is performed by calling a heuristic algorithm to generate an executable task planning scheme; wherein, the self-learning dedicated optimization model updates its parameters online based on the feedback from the executable task planning scheme. This invention achieves natural language interaction and intelligent scheduling through a large language model, solving the problems of unintelligent human-computer interaction and reliance on manual requirement parsing in traditional systems, and improving the usability and automation level of the system; by constructing a collaborative optimization mechanism of "dedicated optimization model - heuristic algorithm", the high-quality initial solution output by the model guides the algorithm to converge quickly, effectively overcoming the technical bottlenecks of low iteration efficiency and long solution time caused by random initialization in traditional heuristic algorithms.

[0048] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540. The processor 510, communication interface 520, and memory 530 communicate with each other via the communication bus 540. The memory 530 includes computer programs, an operating system, and acquired data. The processor 510 can call logical instructions in the memory 530 to execute a task planning method based on a large language model. This method includes: receiving and parsing natural language task requirements through a large language model to generate structured scene feature data; inputting the scene feature data into a self-learning dedicated optimization model to obtain an initial solution for task planning; using the initial solution as a starting point, iteratively optimizing using a heuristic algorithm to generate an executable task planning scheme; wherein the self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

[0049] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the task planning method based on a large language model provided by the above methods. The method includes: receiving and parsing natural language task requirements through a large language model to generate structured scene feature data; inputting the scene feature data into a self-learning dedicated optimization model to obtain an initial solution for task planning; using the initial solution as a starting point, calling a heuristic algorithm for iterative optimization to generate an executable task planning scheme; wherein the self-learning dedicated optimization model updates its parameters online based on the feedback from the executable task planning scheme.

[0051] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0052] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A task planning method based on a large language model, characterized in that, include: The system receives and parses natural language task requirements through a large language model, generating structured scene feature data. The scene feature data is input into a self-learning optimization model to obtain an initial solution for task planning; Starting with the initial solution, a heuristic algorithm is invoked for iterative optimization to generate an executable task planning scheme; The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

2. The task planning method based on a large language model according to claim 1, characterized in that, The self-learning dedicated optimization model adopts a lightweight neural network architecture, which includes an input layer, a feature extraction layer, and an output layer in sequence. Low-rank matrices A and B are connected in parallel on at least one fully connected layer of the feature extraction layer and / or the output layer; The weights of the self-learning dedicated optimization model are: ,in For the pre-trained weight matrix, To update the weight matrix online.

3. The task planning method based on a large language model according to claim 2, characterized in that, The training method for the self-learning dedicated optimization model includes: Collect scenario environment data, historical optimization data, and performance feedback data for task planning scenarios; The scene environment data, historical optimization data, and performance feedback data are preprocessed to obtain preprocessed scene environment data. Using the preprocessed scene environment data as input and the high-quality solutions obtained after optimization by the heuristic algorithm as labels, training sample pairs are constructed. Based on the training sample pairs, the parameters of the initial self-learning dedicated optimization model are optimized using gradient descent until the first loss function converges, thus obtaining the pre-trained weight matrix. .

4. The task planning method based on a large language model according to claim 3, characterized in that, The process of preprocessing the scene environment data, historical optimization data, and performance feedback data to obtain preprocessed scene environment data includes: The scene environment data, historical optimization data, and performance feedback data are subjected to missing value correction, outlier removal, and standardization processing to obtain preprocessed scene environment data.

5. The task planning method based on a large language model according to claim 3, characterized in that, The self-learning dedicated optimization model updates parameters online based on feedback from the executable task planning scheme, including: Set up a data collection window, collect executable task planning schemes from the converged output of the heuristic algorithm within the data collection window, and generate self-learning data samples. When the number of self-learning data samples reaches a preset threshold, the fine-tuning operation of the self-learning dedicated optimization model is triggered.

6. The task planning method based on a large language model according to claim 5, characterized in that, The fine-tuning operation includes: The executable task planning scheme, obtained from the convergence of the collected heuristic algorithms, is input into the pre-trained self-learning optimization model. The mean squared error is used as the second loss function, and the parallel low-rank matrices A and B are updated via backpropagation. ， ； Where η is the fine-tuning learning rate and L is the mean squared error; The updated weight matrix is ​​obtained based on the updated parallel low-rank matrices A and B. .

7. The task planning method based on a large language model according to claim 6, characterized in that, Also includes: Real-time monitoring of the solution time, resource utilization, and convergence accuracy of the generated executable task planning scheme; Based on the monitoring results, the iteration threshold of the heuristic algorithm or the learning rate of the online parameter update is dynamically adjusted.

8. A task planning device based on a large language model, characterized in that, include: The first generation module is used to receive and parse the natural language task requirements through a large language model and generate structured scene feature data. The input module is used to input the scene feature data into the self-learning dedicated optimization model to obtain the initial solution for task planning; The second generation module is used to take the initial solution as the starting point, call the heuristic algorithm to perform iterative optimization, and generate an executable task planning scheme. The self-learning dedicated optimization model updates its parameters online based on feedback from the executable task planning scheme.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the task planning method based on a large language model as described in any one of claims 1 to 7.

10. A non-transitory readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the task planning method based on a large language model as described in any one of claims 1 to 7.

Images