Double-layer double-target optimization scheduling method based on large language model

Abstract

The invention relates to a double-layer double-objective optimization scheduling method based on a large language model, and the method comprises the steps: building a double-layer double-objective optimization model in a server, and building an upper-layer model and a lower-layer model for an AISTS problem based on a plurality of different variables, different constraint conditions and different objective functions; and prompting a dual-objective optimization model of adjustment and algorithm evolution to form a two-stage method based on LLM, the method can overcome the defects that a traditional algorithm needs to design the algorithm manually and adjust parameters manually, the calculation cost is high, the efficiency is low, complex problems are difficult to effectively decompose and coordinate double-layer optimization, a dynamic collaborative optimization framework is lacked, and the flexibility of algorithm combination is low.

Description

Technical Field

[0001] The present invention relates to the field of task planning and scheduling, and particularly to a two-layer and two-objective optimization scheduling method based on a large language model. Background Art

[0002] With the rapid development of satellite technology, Agile Imaging Satellites (AIS) play an increasingly important role in fields such as earth observation, resource exploration, and disaster management. Compared with traditional satellites, agile imaging satellites have stronger attitude control capabilities and can be adjusted in all three degrees of freedom (roll, pitch, yaw) directions, thus achieving a wider observation range and higher imaging accuracy. However, this high flexibility brings complexity to task scheduling. Especially in the case of multi-satellite and large-scale task scheduling, how to efficiently allocate tasks and optimize task scheduling has become a technical problem to be solved urgently.

[0003] The Agile Imaging Satellite Task Scheduling (AISTS) problem belongs to the NP-hard optimization problem. In the AISTS problem, satellites need to reasonably arrange the order and time of task execution within a given time window to maximize the task completion rate and meet multiple constraints, such as satellite energy, storage limitations, task priorities, and the available time window of the satellite. Task scheduling not only involves the resource allocation problem of multiple satellites but also needs to be optimized among multiple tasks and multiple time windows.

[0004] Currently, the research on solving the AISTS problem is mainly divided into two stages: task allocation and single-satellite scheduling. In the task allocation stage, it is necessary to allocate suitable satellites to each task. In the single-satellite scheduling stage, it is necessary to determine the task execution order of each satellite in different time windows. Although some studies have proposed phased scheduling methods, how to efficiently solve these problems, especially in the case of large-scale, multi-satellite, and complex task scheduling, remains a challenging topic.

[0005] At present, the research on the AISTS problem faces the following issues: 1) The AISTS problem is a large-scale, multi-task, and complex-constraint problem, and the efficiency of traditional algorithms needs to be improved. Most existing solutions rely on traditional optimization algorithms such as genetic algorithms, simulated annealing, and local search algorithms. Although these methods have achieved good results in dealing with small-scale problems, in the face of large-scale, multi-task, and complex constraints, traditional algorithms often face problems such as excessive computing time, getting trapped in local optimal solutions, and being unable to effectively adapt to the rapid changes in task requirements. In addition, with the dynamic changes in task scheduling requirements, traditional optimization algorithms are difficult to quickly adjust and adapt, resulting in unsatisfactory effects in practical applications. 2) There is a lack of effective usage methods to directly apply large language models (LLMs) to the AISTS problem. With the rapid development of artificial intelligence technology, especially the breakthroughs of large language models (LLMs) in natural language processing and algorithm generation, more and more research has started to attempt to use LLMs for algorithm design and optimization. Through its powerful natural language understanding and generation capabilities, an LLM can automatically generate efficient optimization algorithms and improve algorithm performance through self-evolution. However, the complexity and professionalism of the AISTS problem make it difficult for existing LLM models to be directly applied to this field. How to transform the AISTS problem into a format suitable for LLM optimization and how to design appropriate prompts to guide the LLM to generate efficient scheduling algorithms are still the core challenges in current research. Summary of the Invention

[0006] In view of the above problems in the prior art, the present invention provides a two-layer and two-objective optimization scheduling method based on a large language model, which overcomes the deficiencies of traditional algorithms that require manual algorithm design and parameter adjustment, high computational cost and low efficiency, difficulty in effectively decomposing complex problems and coordinating two-layer optimization, lack of a dynamic collaborative optimization framework, and low flexibility in algorithm combination by designing a two-layer and two-objective optimization model and providing a two-stage method based on an LLM.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A two-layer and two-objective optimization scheduling method based on a large language model performs the following steps:

[0009] S1. Establish a two-layer and two-objective optimization model in the server;

[0010] The two-layer and two-objective optimization model is divided into two sub-models: an outer layer and an inner layer;

[0011] The outer layer sub-model is used for prompt optimization and algorithm evolution, and the inner layer sub-model is used for solving specific task scheduling problems;

[0012] S2. Establish the upper - layer model and lower - layer model for the AISTS problem based on multiple different variables, different constraint conditions, and different objective functions respectively;

[0013] S3. A dual - objective optimization model for hint adjustment and algorithm evolution;

[0014] S4. A two - stage method based on LLM;

[0015] S41. Input the scenario sc, the initial hint set p0, and the initial algorithm set AL0;

[0016] S42. Execute Algorithm 2 based on p0, AL0, and sc to obtain the fitness value Fit0 of p0 and the optimal algorithm in ;

[0017] S43. ; P p ←p0; AL p ←AL0; that is, assign to al * , assign p0 to P p , and assign AL0 to AL p ;

[0018] S44. Select the top n hints from P p according to the fitness Fit p , and complete the evolutionary strategy on P p to generate the offspring hint p0;

[0019] S45. If the termination condition is not met, return to step 42; otherwise, output al * , that is, the optimal algorithm on the scenario sc;

[0020] Among them, let P be the hint set; p i represents a single hint; dl pi represents the algorithm generated by p i ; Fit m represents the set of fitness values, where m is the number of iterations; fit i is a single fitness value;

[0021] Hint initialization: Manually design hints or randomly generate hints as the initial population: P0 = {p1, p2,..., p N}, and generate the fitness set:

[0022] Hint selection: Use the roulette wheel selection or tournament selection method to select a certain number of hints p t-1 from the population P r1 , p r2,...,p rk As the parent prompt;

[0023] The algorithm 2 is as follows:

[0024] Step 1: Input the scenario sc, the prompt set p0, and the algorithm set AL0;

[0025] Step 2: Initialize by setting Assigning the empty set to al *

[0026] Step 3: For each prompt p ∈ P, perform the following steps;

[0027] Step 4: Assign the algorithm population AL to the current prompt P, i.e., AL p ← AL.

[0028] Step 5: Perform an evolutionary operation using the prompt P with the LLM to obtain the offspring algorithm set AL0

[0029] Step 6: For each algorithm al in the offspring algorithm set AL0, evaluate its fitness under the scenario sc to obtain the fitness set Fit0;

[0030] Step 7: Select the algorithm with the optimal fitness from the offspring algorithm set AL0 If Then update the optimal algorithm That is Assign to al * ;

[0031] Step 8: If the termination condition is not met, return to Step 5; otherwise, output the optimal algorithm al * , the offspring fitness set Fit0, and the offspring algorithm set Aμ0..

[0032] Furthermore, the outer sub-model generates effective prompt words through the Prompt Optimization Framework (PEF) to guide the evolution of the inner sub-model algorithm.

[0033] Furthermore, the inner sub-model automatically designs specific algorithms to solve the AISTS problem through the Algorithm Evolution Framework (AEF).

[0034] Furthermore, the upper layer model for the AISTS problem is the task allocation stage, which assigns appropriate satellites to each task and reasonably distributes tasks among satellites to ensure that the tasks are completed within the given time window.

[0035] Furthermore, the lower layer model for the AISTS problem is the single satellite scheduling stage, which determines the task execution order for each satellite within the specified time window and optimizes the usage efficiency of satellite resources.

[0036] Furthermore, the Question and Negotiation Algorithm (QNA) is adopted during the process of assigning tasks to different satellites;

[0037] Each task is assigned a priority value; QNA preferentially assigns high-priority tasks; if two tasks have the same priority, the sorting of the tasks with the same priority will be random;

[0038] Ensure that a satellite with greater resource availability is selected; if two satellites have the same number of available resources, the sorting of the two satellites with the same number of available resources will be random;

[0039] Each satellite has tasks already arranged during the negotiation process; QNA preferentially considers assigning tasks to the satellite with the fewest accumulated tasks and the richest remaining resources.

[0040] The present invention has the following beneficial effects:

[0041] 1) Automated algorithm design. The proposed two-layer BOM in the present invention combines prompt optimization and algorithm evolution, enabling the design of the algorithm to be automatically completed by the LLM. Through the prompt optimization and evolution framework, the LLM can adaptively adjust and generate better algorithms, significantly reducing manual intervention and the complexity of algorithm design.

[0042] 2) Improved algorithm efficiency. This method combines the advantages of evolutionary computation, and improves the effectiveness of the algorithm through continuous iterative optimization, reflecting the effectiveness and superiority of the method.

[0043] 3) Two-layer optimization framework. By decomposing the AISTS problem into two sub-problems, namely task assignment and individual AISTS problems, and optimizing them separately, this hierarchical optimization method effectively reduces the complexity of the problem. The outer-layer optimization is responsible for prompt optimization and algorithm evolution, while the inner-layer optimization solves specific scheduling tasks, enabling the algorithm to better adapt to diverse and complex task scheduling requirements.

[0044] 4) High adaptability and scalability. This method can handle satellite task scheduling problems of various scales and different complexities. Through the prompt evolution framework of the LLM, the system can generate highly adaptable solutions in different scenarios, with strong scalability and generality.

[0045] 5) Time and cost savings. The automated algorithm design and evolution process reduce manual participation and improve the modeling efficiency. This means that time and costs can be significantly saved when solving complex problems, especially in large-scale satellite task scheduling, where efficient solutions can be quickly generated.

[0046] 6) Cross - layer co - evolution. The algorithm evolution of the outer layer and the inner layer cooperate with each other, forming a closed - loop of collaborative optimization. The outer layer is responsible for optimizing the prompts and algorithm design, and the inner layer evolves according to these prompts to obtain the actual scheduling algorithm, making the algorithm evolution process more efficient and generating better scheduling schemes. Brief Description of the Drawings

[0047] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0048] Figure 1 It is a framework for evolving the optimal algorithm combination for the AISTS problem in the double - layer BOM framework.

[0049] Figure 2 It is a negotiation Q&A - based allocation algorithm framework.

[0050] Figure 3 It is a genetic algorithm crossover framework.

[0051] Figure 4 It is a genetic algorithm mutation framework.

[0052] Figure 5 It is a reinforcement learning allocation algorithm framework. Detailed Description of the Embodiments

[0053] The following describes the exemplary embodiments of the present application with reference to the accompanying drawings. Various details of the embodiments of the present application are included to assist understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described here without departing from the scope and spirit of the present application. Similarly, for clarity and conciseness, the description of well - known functions and structures is omitted below.

[0054] To enable those skilled in the art of this technology to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall within the scope of protection of the present invention.

[0055] The following describes exemplary embodiments of the present application in conjunction with the accompanying drawings. Various details of the embodiments of the present application are included to assist understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Similarly, descriptions of well-known functions and structures are omitted in the following description for clarity and conciseness.

[0056] The present invention provides an infrared small and weak target detection method based on adaptive spatio-temporal tensor and weighted tensor average rank approximation, including: adaptive spatio-temporal domain infrared tensor block construction, low-rank and sparse decomposition model, image reconstruction component, as Figure 1 shown. The invention can improve the target detection and background suppression capabilities simultaneously to meet the urgent needs of infrared (IR) target detection technology in complex scenarios.

[0057] S1. Establish a two-layer double-objective optimization model in the server.

[0058] The two-layer double-objective optimization model proposed by the present invention is divided into two sub-models: an outer layer and an inner layer. The outer layer is used for prompt optimization and algorithm evolution, while the inner layer focuses on solving specific task scheduling problems. The outer layer generates efficient prompts through a Prompt Optimization Framework (PEF) to guide the evolution of the inner layer algorithm. The inner layer automatically designs specific algorithms to solve the AISTS problem through an Algorithm Evolution Framework (AEF). Figure 1 Shows the overall process of evolving the optimal algorithm combination for the AISTS problem in the proposed two-layer BOM framework.

[0059] Figure 1 The two-layer optimization model (BOM) based on LLM optimizes the task scheduling problem by automatically adjusting prompts and co-evolving algorithms, and finally generates the optimal task assignment and scheduling algorithms.

[0060] S2. Based on multiple different variables, different constraint conditions, and different objective functions, establish an upper-layer model and a lower-layer model for the AISTS problem respectively.

[0061] The upper-layer model and the lower-layer model are aimed at the agile imaging satellite task scheduling problem

[0062] Next, the mathematical expressions of the variables, constraint conditions, and objective functions in the AISTS problem will be listed first to establish the model in a standardized manner.

[0063] First is the upper-layer model of the AISTS problem, which is the task assignment stage, that is, assigning appropriate satellites to each task and reasonably distributing tasks among satellites to ensure that the tasks are completed within the given time window.

[0064] S21. Variables involved in the upper-layer model:

[0065] S: The set of all satellites;

[0066] s i : The i-th satellite, s i ∈S;

[0067] T: The set of all tasks;

[0068] t i : The i-th task, t i ∈T;

[0069] a ij : A 0-1 variable indicating whether task t i is assigned to satellite s j (1 - Yes, 0 - No);

[0070] W ij : The set of Visible Time Windows (VTWs) between each task t i and satellite s i ;

[0071] w ijk : The k-th VTw between task t i and satellite s i , w ijk ∈W ij ;

[0072] e i : The power consumed by each task t i ;

[0073] g i : The storage space consumed by each task t i ;

[0074] E j : The maximum power of each satellite s j ;

[0075] G j : The maximum storage space of each satellite s j ;

[0076] S22. Constraints involved in the upper-layer model

[0077] Equation (1) indicates that each task is assigned at most once;

[0078] Equation (2) indicates that the cumulative power of all assigned tasks cannot exceed the maximum capacity of the satellite.

[0079] Equation (3) indicates that the cumulative storage of all assigned tasks cannot exceed the maximum capacity of the satellite.

[0080] Equation (4) indicates that the decision variable is binary.

[0081] Equation (5) indicates that if there are no VTWs between task t i and satellite s j , then a ij must be zero.

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] S23. The objective function involved in the upper-layer model

[0088] Equation (6) represents the objective of the upper layer, i.e., the cumulative objective of all satellites

[0089]

[0090] The lower-layer model of the AISTS problem is the single-satellite scheduling stage, which determines the task execution order for each satellite within the specified time window and optimizes the utilization efficiency of satellite resources.

[0091] S24. Variables involved in the lower-layer model

[0092] b ijk : A 0-1 variable indicating whether task ti is executed within the specific time window wijkwijk (1 - yes, 0 - no);

[0093] h i : Represents the priority of task ti;

[0094] c ijk : Represents the exact start time of each task;

[0095] u ijk : Represents the start time;

[0096] v ijk : Represents the end time;

[0097] Δt ii′ : Represents the transition time between two consecutive observation tasks t i and t i ';

[0098] y ii′ : A 0-1 variable indicating whether task t iand t i Whether it is adjacent (1 - yes, 0 - no);

[0099] α j bijk : Represents time b ijk and b i′jk′ 's roll angle;

[0100] β: Represents the pitch angle;

[0101] γ: Represents the yaw angle.

[0102] S25. Constraints involved in the lower - layer model

[0103] Equation (7) represents the conversion time δ between two consecutive observation tasks ti and ti′ ii' , where a1 = 1.5, a2 = 2, a3 = 2.5, a4 = 3 represent different conversion speeds;

[0104] Equation (8) represents the attitude angle of the measurement conversion;

[0105] Equation (9) represents that each satellite can select at most one VTW for each task;

[0106] Equation (10) represents that the observation time should be set within the selected VTW;

[0107] Equation (11) represents that when the satellite completes one task and switches to the next task, the transition time should be satisfied;

[0108] Equations (12) and (13) represent that bijk and yii′ are binary variables;

[0109] Equation (14) represents that the conversion time must be greater than zero.

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] S26. Objective function involved in the lower - layer model

[0119] Equation (15) represents the lower-level objective function, i.e., the total priority of the tasks completed in a satellite.

[0120]

[0121] As mentioned above, a two-layer programming model of AISTS has been established, and the mathematical expressions of decision variables, constraint conditions, and objective functions in the model are given in a standardized manner, which helps relevant departments understand the essential connotation of the AISTS problem and thus better guide the establishment of the next model.

[0122] S3. Bi-objective optimization model for hint adjustment and algorithm evolution

[0123] The present invention proposes a BOM based on LLM to complete the automatic algorithm design (AAD) for the AISTS problem. This model is a two-layer optimization framework that combines hint adjustment optimization and algorithm evolution, enabling the co-evolution of hints and algorithms through dynamic interaction between the upper and lower layers. The overall framework includes two core components: the upper-layer hint adjustment framework and the lower-layer AE framework, which are responsible for optimizing hints and designing algorithms respectively. Through iterative optimization of the upper layer, the hints are optimized to drive the evolution of the lower-layer algorithms, and finally high-quality algorithms for the AIST scenario are generated. In both the upper and lower layers of this LLM-based BOM, an evolutionary framework is adopted. The overall goal of this BOM is to design an effective algorithm to solve the AISTS scenario. The upper layer is used to optimize the LLM hints, and the lower layer generates algorithms based on the upper-layer hints.

[0124] In the lower layer, the present invention introduces an algorithm evolution framework for AAD in the AISTS scenario. The AE framework integrates evolutionary computing, and its core logic focuses on iteratively enhancing the algorithm population using a specific evolutionary strategy based on LLM, thereby generating high-quality algorithms. Let SC be the set of all possible scenarios, and AL be the set of all algorithms. Define a scoring function R: SC × AL ← R, which is used to evaluate the performance of applying algorithm al ∈ AL in scenario sc ∈ SC. For a given scenario sc ∈ SC, a target function Fi is defined in the lower layer, which maps from the algorithm space A to the set of real numbers:

[0125] F(al) = R(sc, al) #(16)

[0126] where Fi represents the fitness function for a specific scenario sc i For scenario sc i , the optimal algorithm al * can be defined as the algorithm that maximizes the function Fi:

[0127]

[0128] Finally, the present invention can define a mapping φ: SC → AL, from the scene set SC to the algorithm set AL. This mapping assigns the optimal algorithm al * (sc) to each scene sc ∈ SC, such that:

[0129]

[0130] Its components include population initialization, selection, crossover, mutation, population update, and fitness evaluation. Algorithm initialization: Both manually designing algorithms and using LLM to generate algorithms are feasible methods for population initialization: AL0 = {al1, al2, …, al N}.

[0131] Algorithm selection: Use the roulette wheel method to select individuals, and select n entities as the parent candidates for subsequent evolutionary operations, denoted as AL t = {al1, al2, …, al n}.

[0132] Algorithm evolution: In the current iteration, the input of the crossover operation consists of the parent candidate set AL t , and then m new entities are generated, denoted as O c = {O1, O2, ..., O m}. During this process, the LLM will create new algorithms based on the clues provided by the upper-layer model.

[0133]

[0134] In the mutation operation, the algorithm is adjusted by the large language model (LLM) to generate a new algorithm. The design clues are similar to the crossover operation, but in this case, only one input entity al k is used as the parent.

[0135]

[0136] Algorithm update: For each individual al in the updated population AL i , use the fitness function F i to calculate its fitness. To maintain a consistent population size, individuals with lower fitness should be removed, and individuals with higher fitness should be retained. This step-by-step screening process aims to improve the overall quality of the population.

[0137] At the upper layer, the present invention introduces PEF for prompt adjustment. Specifically, the workflow of DPOF starts from an initial population consisting of multiple prompts. Based on this, the framework iteratively adopts an evolutionary strategy, calls the LLM according to the feedback from the lower layer, and thus generates new optimized prompts. These optimized prompts are then passed to the lower layer to guide the generation of more powerful algorithms. Let P be the prompt set; p iRepresents a single prompt; Represents the algorithm generated by p i ; Fit m Represents a set of fitness values, where m is the number of iterations; fit i Is a single fitness value; N is the population size; Evo(·) is a carefully designed prompt generation evolutionary operator. The main operations of the upper layer are defined as follows:

[0138] Prompt initialization: Manually design prompts or randomly generate prompts as the initial population: P0 = {p1, p2,..., p N}, and generate the fitness set:

[0139] Prompt selection: Use the roulette wheel selection or tournament selection method to select a certain number of prompts p t-1 from the population P r1 , p r2 ,..., p rk as the parent prompts.

[0140] Prompt evolution: Use the LLM to execute the evolutionary strategy on the selected parent prompts, including crossover and mutation, to generate new prompts:

[0141] p′ ← Evo(p,…, p r ) #(21)

[0142] Prompt update: After each new prompt is evolved, it is necessary to calculate the fitness value according to the lower layer model, and update the prompt set and score according to the assigned score.

[0143] P m ← {P m -1, p i} Fit m ← {Fit m -1, fit i} #(22)

[0144] Use the LLM to re-examine the evolution steps and further optimize the prompt template through an iterative loop until the predefined number of iterations is reached. In the current population, select the prompt with the highest fitness as the optimal prompt, which is the goal of the upper layer:

[0145]

[0146] S4. Two-stage method based on LLM

[0147] To solve the BOM, a two-stage method based on LLM is adopted to perform prompt adjustment and AE respectively. The following algorithm 1 provides the algorithm framework of the two-stage method:

[0148] Step 41: Input the scenario sc, the initial prompt set p0, and the initial algorithm set AL0;

[0149] Step 42: Execute Algorithm 2 based on p0, AL0, and sc to obtain the fitness value Fit0 of p0 and the optimal algorithm in ;

[0150] Step 43: ; P p ← p0; AL p ← AL0; about to assign to al * , assign p0 to P p , assign AL0 to AL p ;

[0151] Step 44: Select the top n prompts from P p according to the fitness Fit p , and complete the evolutionary strategy on P p to generate the offspring prompt p0;

[0152] Step 45: If the termination condition is not met, return to Step 42; otherwise output al * , that is, the optimal algorithm on the scenario sc.

[0153] A. Prompt Evolution Framework

[0154] To explain the problem and the specific evolutionary strategy to the LLM, the initial prompt consists of two parts: problem description and evolutionary operation.

[0155] The mutation operation is completed through the following steps:

[0156] The LLM first selects two initial prompts as parental templates, namely "Prompt 1" and "Prompt 2"; at the same time, the LLM establishes the current optimal prompt, called "Prompt 3"; based on these two parental templates, the LLM extracts key concepts and terms;

[0157] Then it compares and identifies the differences between these keywords to understand the unique contributions of each parental template; the LLM modifies the selected keywords by introducing new features or adjusting existing features to enhance the adaptability and innovativeness of the prompt;

[0158] Finally, the LLM combines the mutated keywords with "Prompt 3" to form a preliminary new prompt.

[0159] The crossover operation is completed through the following steps:

[0160] The LLM selects a template prompt that includes the problem description, work content, and requirements;

[0161] The LLM combines this basic prompt with new prompts through a crossover process, involving the exchange of specific parts and the introduction of mutations, ultimately creating offspring prompts.

[0162] Based on the collaboration of crossover and mutation operations, the final prompt generated through the above process, i.e., the offspring prompt, integrates the best features from different sources, aiming to improve the effectiveness of task allocation and the quality of solutions.

[0163] Since the prompts in the algorithm evolution strategy play a crucial role in improving algorithm performance, the present invention designs additional evolutionary components for specific crossover and mutation strategies. The basic algorithm crossover and mutation prompts evolve within the prompt evolution framework, ultimately generating effective algorithm crossover and mutation prompts. The generated prompts are used to generate more effective algorithms in the lower layer.

[0164] The evolution framework basically includes an initial population and an evolution strategy. In the initial algorithm population, the present invention selects the Contract Net Algorithm (CNA) and the Question and Answer Negotiation Algorithm (QNA) as upper-layer algorithms, and the First Insertion Algorithm (FIA), the Random Insertion Algorithm (RIA), and the Adaptive Large Neighborhood Search (ALNS) algorithm for solving the AISTS problem in the lower layer. It should be noted that these algorithms are composed of some heuristic rules, which are operators that need to be modified. Generally, in the overall framework, all operators are combined together, and each operator is applied with a defined probability. The heuristic rules evolve between generations.

[0165] The evolution strategy is guided by the designed prompts in the upper layer. These strategies mainly include crossover and mutation. After modification, the new individual algorithms of the new generation are evaluated in the upper and lower layers of AISTS through Equations 1 and 9. The algorithms with higher objectives are selected to enter the next generation, while the poorer algorithms are eliminated.

[0166] After iterative evolution, the optimal algorithm with the maximum objective is obtained. The same process is also used for task allocation. Two optimal algorithms jointly solve the AISTS problem.

[0167] The following Algorithm 2 shows the lower layer of the LLM-based BOM, i.e., the algorithm evolution framework.

[0168] Step 1: Input scenario sc, prompt set p0, algorithm set AL0;

[0169] Step 2: Initialize and set Assign the empty set to al *

[0170] Step 3: For each prompt p ∈ P, execute the following steps;

[0171] Step 4: Assign the algorithm population AL to the current prompt P, i.e., AL p ← AL.

[0172] Step 5: Use the LLM to perform an evolutionary operation through prompt P to obtain the offspring algorithm set AL0

[0173] Step 6: For each algorithm al in the offspring algorithm set AL0, evaluate its fitness under the scenario sc to obtain the fitness set Fit0

[0174] Step 7: Select the algorithm with the optimal fitness from the offspring algorithm set AL0 If Then update the optimal algorithm That is Assign to al * ;

[0175] Step 8: If the termination condition is not met, return to Step 5; otherwise, output the optimal algorithm al * , the offspring fitness set Fit0, and the offspring algorithm set AL0.

[0176] Generally speaking, the advantages of the method of the present invention can be summarized as follows:

[0177] 1) Automated algorithm design. The proposed two-layer BOM in the present invention combines prompt optimization and algorithm evolution, enabling the design of algorithms to be completed automatically by the LLM. Through prompt optimization and the evolutionary framework, the LLM can adaptively adjust and generate better algorithms, significantly reducing manual intervention and the complexity of algorithm design.

[0178] 2) Improve algorithm efficiency. This method combines the advantages of evolutionary computing and improves the effectiveness of the algorithm by continuously iteratively optimizing the algorithm, reflecting the effectiveness and superiority of the method.

[0179] 3) Two-layer optimization framework. By decomposing the AISTS problem into two sub-problems, namely task allocation and a single AISTS problem, and optimizing them separately, this hierarchical optimization method effectively reduces the complexity of the problem. The outer layer optimization is responsible for prompt optimization and algorithm evolution, while the inner layer optimization solves specific scheduling tasks, enabling the algorithm to better adapt to diverse and complex task scheduling requirements.

[0180] 4) High adaptability and scalability. This method can handle satellite task scheduling problems of various scales and different complexities. Through the prompt evolution framework of the LLM, the system can generate highly adaptable solutions in different scenarios, with strong scalability and generality.

[0181] 5) Save time and cost. The automated algorithm design and evolutionary process reduce manual intervention and improve the modeling efficiency. This means that significant time and cost savings can be achieved when solving complex problems, especially in large-scale satellite mission scheduling, where efficient solutions can be generated quickly.

[0182] 6) Cross-level co-evolution. The algorithm evolution of the outer layer and the inner layer cooperate with each other to form a closed-loop of collaborative optimization. The outer layer is responsible for optimizing the prompts and algorithm design, and the inner layer evolves according to these prompts to obtain the actual scheduling algorithm, making the algorithm evolution process more efficient and generating better scheduling solutions.

[0183] The following is a specific application of the solution of the present invention. Since there is a lack of publicly available benchmark scenario data for the AISTS problem, the present invention uses the following parameters to generate task scenarios to construct a representative experimental environment. The period of each scenario is set to 24 hours, and the target points are distributed around the world to simulate real earth observation requirements. The satellite system is created based on the Walker constellation model, which is a widely used satellite constellation design for global coverage. The parameters of the Walker constellation include the number of satellites, the number of orbital planes, and the distribution of satellites within the orbital plane, which can ensure uniform coverage of the earth.

[0184] In the experiment, the orbital parameters of each satellite are set as follows:

[0185] Semi-major axis (a): 7200 km, which determines the size of the satellite orbit.

[0186] Eccentricity (e): 0.000627, indicating the ellipticity of the orbit, close to a circular orbit.

[0187] Inclination (i): 96.576 degrees, indicating the angle between the orbital plane and the earth's equatorial plane, close to a polar orbit, suitable for global coverage.

[0188] Argument of perigee (ω): 0 degrees, indicating the angle between the perigee of the orbit and the ascending node.

[0189] Right ascension of ascending node (Ω): 175.72 degrees, indicating the angle between the ascending node of the orbit and the vernal equinox.

[0190] Horizontal approach angle (m): 0.075 degrees, used to adjust the distribution of satellites within the orbital plane.

[0191] These parameters ensure that the satellite system can efficiently cover the global scope and provide diverse observation opportunities for mission scheduling.

[0192] The generation of the task scenario is based on the following assumptions:

[0193] Task cycle: The cycle of each scenario is 24 hours, simulating the task scheduling requirements within a day.

[0194] Target distribution: Target points are randomly distributed across the globe, covering different regions such as land, ocean, and polar regions, to reflect diverse requirements in practical applications.

[0195] Task attributes: Each task has specific priority, time window, energy consumption, and storage requirements, which need to be fully considered during task assignment and scheduling processes.

[0196] As is well known, computing time and solution quality are contradictory; less computing time rarely leads to higher solution quality. Moreover, excessive computing time is unacceptable. Therefore, the present invention sets specific time limits (1, 3, 5, 10 seconds) to evaluate the single scheduling algorithm. At the same time, the present invention sets up small-scale, medium-scale, and large-scale multi-satellite systems, which contain 5, 50, and 100 satellites respectively, and also stipulates that the number of tasks is 10, 25, 50, and 100 to evaluate these algorithms.

[0197] To more comprehensively evaluate the performance of the method proposed by the present invention, the present invention selects the currently optimal algorithm combination for the double-layer optimization model as the competitor. These algorithm combinations represent the advanced level in the current research on the AISTS problem.

[0198] For the task assignment phase, there is:

[0199] (1) Negotiation Q&A-based assignment algorithm

[0200] The Q&A negotiation algorithm (QNA) can negotiate with each AEOS before putting forward specific task requirements. This strategy follows some simple but effective heuristic rules that are widely used in multi-system scheduling problems. It can improve objectivity and be compatible with the Q&A model.

[0201] Through the Q&A negotiation process, QNA comprehensively considers various factors regarding tasks and satellites. It realizes the efficient assignment of tasks to satellites through a "negotiation" method. QNA adopts a greedy assignment rule, seriously considering task priority, the availability of satellite resources, and the number of tasks received by each satellite. In addition, there is a certain degree of randomness in the task assignment strategy. The primary goal of this strategy is to assign tasks to satellites with high priority, sufficient available resources, and the least number of tasks received. The specific assignment rules are as follows:

[0202] 1) Each task is assigned a priority value. QNA gives priority to assigning high-priority tasks. If two tasks have the same priority, their sorting will be random.

[0203] 2) Each satellite has its unique set of available resources, including power and storage space. QNA arranges satellites based on the available resources of the satellites to ensure that satellites with greater resource availability are selected. If two satellites have the same amount of available resources, their sorting will be random.

[0204] 3) Each satellite has an assigned task during the negotiation process. QNA gives priority to assigning tasks to the satellite with the fewest accumulated tasks and the richest remaining resources (VTW length). This rule ensures the fair distribution of tasks among satellites, avoids overloading a few satellites with too many tasks, and makes full use of the available resources.

[0205] Figure 2 Shows the process of assigning multiple tasks to three satellites. In each assignment, the leader negotiates with each satellite, evaluates their current capabilities, and selects the most suitable candidate. Since the assignment order affects the entire task assignment, tasks are sorted by higher priority before negotiation. In addition, a single-satellite scheduling is performed after each task assignment, and a specific assignment does not change after negotiation. Therefore, the Q&A negotiation algorithm not only improves the speed and efficiency of the upper layer but also makes the task load more balanced among satellites.

[0206] (2) Allocation strategy based on genetic algorithm

[0207] Genetic algorithm (GA), as a classic evolutionary algorithm, has been widely applied in many fields such as combinatorial optimization, machine learning, signal processing, etc. Many studies have demonstrated the excellent advantages of GA in certain optimization problems.

[0208] Therefore, the present invention applies it to the evaluation of the upper-layer EA effect. The basic framework of the genetic algorithm includes several key components: population initialization, three operators (selection, crossover, mutation), and termination criteria. I adjusted it for task assignment:

[0209] 1) Population initialization: Randomly assign tasks to each satellite with the task acceptance right (VTW).

[0210] 2) Selection: Select individual solutions with high priority and set them as the objects of the next iteration.

[0211] 3) Crossover: Recombine and modify parts of the task assignment from two parent solutions and generate offspring.

[0212] 4) Mutation: Assign some tasks from the original solution to other satellites, change the parent solution, and generate offspring.

[0213] 5) Termination criteria: Reach the defined maximum number of iterations, or the maximum priority cannot be further improved.

[0214] Figure 3 Describes the crossover operation in the genetic algorithm. Two problems demonstrate different task assignments, where the index represents the task and the corresponding value represents the assigned satellite. For example, the first value of 3 means that task 1 is assigned to satellite 3. In addition, the process traverses each task and decides whether to perform gene crossover according to the specified crossover probability. Figure 4 Demonstrates the mutation operation, traversing each task and randomly changing the assigned task according to the mutation probability. It should be emphasized that these changes are made when there is a task acceptance right between the task and the satellite. Otherwise, an incorrect problem will be formed.

[0215] (III) Allocation strategy based on reinforcement learning

[0216] In the classic adaptive large neighborhood search algorithm for imaging satellite intelligent task scheduling, scholars designed several heuristic allocation operators based on information such as time windows and attitude angles. The key lies in selecting the appropriate allocation operator through an adaptive mechanism to improve the task allocation efficiency of the algorithm. Although heuristic rules help in selecting allocation operators, designing efficient operators still requires reasoning and careful selection like humans. To solve this problem, an adaptive allocation method is adopted to dynamically adjust the decision-making process to cope with environmental changes.

[0217] The adaptive allocation method is essentially a decision-making process that involves selecting the most suitable allocation operator to execute tasks. In this context, reinforcement learning (RL) is used to guide the selection of allocation operators. Regarding the task allocation process as a sequential decision-making process, a relevant Markov decision process (MDP) can be established for task allocation. The MDP at the upper allocation level can be modeled as a quadruple (S, A, R, V), where S represents the finite state set, denoted as S = {S i |i = 1, 2,..., n}, where n is the number of states. Each state S i can be defined by a triple (AS i , TS i , TS i ), where the state vector AS i contains N elements, where N represents the number of satellites, and each element is a state representing a sub-vector obtained through a single-satellite scheduling algorithm. is defined as 0 if the algorithm ends at the maximum number of iterations and 1 if it is still in continuous iteration. The scalar TS i represents the completion ratio, defined as the ratio of the number of scheduled tasks to the total number of tasks. TS iIt should be defined as 0, 1, 2, 3, representing a regression rate of 0%-25%, 25%-50%, 50%-75%, and 75%-100% respectively. Similarly, PS i also represents a scalar used to define whether the overall planned profit of the task is improved. If the overall profit is not increased, it is 0; if the overall profit is obtained in the current iteration, it is 1. From the above state definitions, I can find that the size of the state set is 2N * 4 * 2 = 2N+3.

[0218] 1) The action set A = {ai|i = 1, 2, …, 5} includes options selected from the above various allocation operators.

[0219] 2) R represents the short-term reward. In this paper, R refers to the total profit margin of all satellite scheduling plans after selecting an allocation operator.

[0220] 3) The Q-table updated using the Bellman equation is used as the value function defining V.

[0221] Subsequently, the Q-learning algorithm is used to optimize this decision-making problem. The decision-making process involves determining which tasks to assign to which satellite in a specific state to achieve a transition to the next state. By interacting with the environment (which can also be regarded as a single scheduling level), reinforcement learning enables the system to understand the long-term consequences of various decisions in different states and optimize its behavior to maximize the cumulative reward. This results in an optimal task allocation plan.

[0222] Figure 5 Shows the interaction process of the two-layer framework in reinforcement learning, where the problem layer provides the task allocation plan generated by the selected allocation operator to the answer layer. The answer layer provides various information required for reinforcement learning to the problem layer, and finally completes the interaction process. The specific process of the reinforcement learning-based allocation algorithm (RLA) can be described in detail as the following steps:

[0223] 1) Initialization: Define the state space and action space for satellite task allocation and initialize the Q-table.

[0224] 2) Action selection: Based on the Q-table and the current state, select an action. In the earlier stage, the ∈-greedy policy can be used to explore unknown state-action pairs; as learning progresses, this exploration can be gradually reduced to make full use of the learned experience.

[0225] 3) Execute the action: Assign tasks to satellites according to the selected action.

[0226] 4) Reward and state information collection: Execute the single-satellite scheduling algorithm to obtain the reward (i.e., the total profit margin in the current situation) and state information.

[0227] 5) Update the Q-table: Use the Bellman update formula of Q-learning to update the Q-table with the reward obtained in this step and the maximum Q-value of the next step.

[0228] 6) Check the stopping condition: If the predefined number of learning rounds has been reached or other stopping learning conditions are met, stop learning; if not, return to step 2) and repeat the process.

[0229] 7) Policy determination: The final Q-table determines the policy for satellite task allocation.

[0230] For the single-satellite scheduling phase, there is:

[0231] (I) Late Acceptance Algorithm

[0232] The Late Acceptance Algorithm (LAA) is a metaheuristic method. LAA incorporates the concept of "late acceptance", considering solutions from a certain time (step) ago, enabling it to escape local optima. LAA not only retains the asymptotic convergence of the traditional hill-climbing algorithm but also has the ability to escape local optima. Therefore, it performs well in solving many well-established problems in the field of operations research. Before presenting the algorithm framework of LAA, it is necessary to delve into the algorithm operators responsible for generating neighboring solutions. A neighboring solution refers to a solution derived by making minor modifications to the original solution. In LAA, three operators, namely move, swap, and replace, are used. The move operator is used to adjust the value of a decision variable. In the context of single-satellite scheduling, this operator represents adjusting the virtual time window selection and start time of a specific observation task. The swap operator exchanges the numerical values of decision variables, symbolizing the exchange of the start times of two observation tasks. Additionally, the replace operator optimizes the solution based on the existing optimal solution. In other words, the original solution will be modified by incorporating a portion of the decision variables of the current optimal solution to improve the objective. In LAA, these three operators are selected according to their respective probabilities. Adaptive Large Neighborhood Search Algorithm (ALNS): ALNS searches the solution space through destruction and repair operations. In each iteration, ALNS dynamically adjusts the selection probability of the operators to adapt to the current search state.

[0233] (II) Adaptive Large Neighborhood Search Algorithm

[0234] The Adaptive Large Neighborhood Search algorithm (ALNS) is a well-established method widely used to address combinatorial optimization challenges and has achieved successful applications in various fields. Additionally, ALNS has been applied to solve the AEOS scheduling problem by skillfully allocating tasks, scheduling tasks, and considering constraints and objectives. The adaptability of this algorithm enables it to modify search strategies and parameters based on the information obtained, thereby efficiently traversing the solution space. By leveraging local search, metaheuristics, and intelligent learning mechanisms, ALNS guides the search towards potential solutions and converges to near-optimal or optimal solutions that satisfy the constraints. In summary, ALNS provides a powerful and adaptable strategy for optimizing satellite scheduling problems in dynamic environments. Although ALNS has been successfully applied to various combinatorial optimization problems including the AEOS scheduling problem, there seems to be a lack of research integrating it with BLO to fully utilize its potential for enhancing robustness.

[0235] The ALNS method is used to solve the lower-level answering questions, which consists of two basic components:

[0236] 1) Local search layer: It includes destruction and repair operations. These operations are alternately implemented to traverse the solution space and promote the transformation of solutions by modifying the neighborhood.

[0237] 2) Adaptive operator selection layer: In each iteration, the operator dynamically adjusts its score according to predefined criteria. More specifically, the criteria for updating the score are heuristically defined as follows:

[0238]

[0239] Next, the probability p i , that is, the weight, of selecting operator i in the next iteration is updated according to the score s i of operator i at the end of the current iteration through the following formula:

[0240]

[0241] The above specific implementation manners do not constitute a limitation on the protection scope of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of this application shall be included within the protection scope of this application.

[0242] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device. Thus, they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0243] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that various modifications or deformations that can be made without creative efforts on the basis of the technical solution of the present invention are still within the protection scope of the present invention.

[0244] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A dual-level dual-objective optimization scheduling method based on a large language model, characterized in that: Follow these steps: S1. Establish a two-layer dual-objective optimization model in the server; The dual-layer dual-objective optimization model is divided into two sub-models, an outer layer and an inner layer; The outer sub-model is used for prompt word optimization and algorithm evolution, and the inner sub-model is used for solving specific task scheduling problems; S2. Establish upper and lower models for AISTS problems based on multiple different variables, different constraints and different objective functions; S3. Dual-objective optimization model of prompt adjustment and algorithm evolution; S4. Two-stage method based on LLM; S41. Input scene sc, initial prompt set p0, initial algorithm set AL0; S42. Execute Algorithm 2 based on p0, AL0 and sc to obtain the fitness value Fit0 of p0 and the optimal algorithm population The best algorithm in S43. P p ←p0;AL p ←AL0; coming soon Assign value to the optimal algorithm al * , assign p0 to the current prompt set P p , assign AL0 to the current algorithm set AL p ; S44. According to the current fitness Fit p From P p Select the first n prompts in P p The evolution strategy is completed above, and the offspring prompt p0 is generated; S45. If the termination condition is not met, return to step 42; otherwise, output al * , which is the optimal algorithm on scene sc; Among them, let P be the prompt set; p i Represents a single prompt; Indicated by p i Generated algorithm; Fit m Represents a set of fitness values, where m is the number of iterations; fit i is a single fitness value; Prompt initialization: manually design prompts or randomly generate prompts as the initial population: P0 = {p1, p2, ..., p N }, and generate a fitness set: Tip selection: Use roulette wheel selection or tournament selection method to select a tip from the population P according to its fitness value t-1 Select a certain number of prompts p r1 ,p r2 ,...,p rk As a parent prompt; The algorithm 2 is: Step 1: Input scene sc, prompt set p0, algorithm set AL0; Step 2: Initialization, setup Assign the empty set to al * Step 3: For each prompt p∈P, perform the following steps; Step 4: Assign the algorithm population AL to the current prompt P, that is, AL p ←AL; Step 5: Prompt P to use LLM to perform evolutionary operations and obtain the offspring algorithm set AL0; Step 6: For each algorithm al in the descendant algorithm set AL0, evaluate its fitness under the scenario sc to obtain the fitness set Fit0; Step 7: Select the algorithm with the best fitness from the offspring algorithm set AL0 If the fitness function value Update the optimal algorithm Coming soon Assign to al * ; Step 8: If the termination condition is not met, return to step 5; otherwise, output the optimal algorithm al * , fitness set Fit0, and offspring algorithm set AL0.

2. A dual-level dual-objective optimization scheduling method based on a large language model as claimed in claim 1, characterized in that: The outer sub-model generates effective prompt words through the prompt optimization framework PEF to guide the evolution of the inner sub-model algorithm.

3. A dual-layer dual-objective optimization scheduling method based on a large language model as claimed in claim 2, characterized in that: The inner sub-model automatically designs a specific algorithm for solving the AISTS problem through the algorithm evolution framework AEF.

4. A dual-layer dual-objective optimization scheduling method based on a large language model as claimed in claim 3, characterized in that: The upper-level model for the AISTS problem is the task allocation stage, which is to assign appropriate satellites to each task and reasonably distribute tasks among satellites to ensure that the tasks are completed within a given time window.

5. A dual-layer dual-objective optimization scheduling method based on a large language model as claimed in claim 4, characterized in that: The lower-level model for the AISTS problem is the single satellite scheduling stage, which is to determine the task execution order for each satellite within a specified time window and optimize the utilization efficiency of satellite resources.

6. A dual-layer dual-objective optimization scheduling method based on a large language model as claimed in claim 5, characterized in that: The question-answer negotiation algorithm QNA is used in the process of assigning tasks to different satellites; Each task is assigned a priority value; QNA prioritizes high-priority tasks; if two tasks have the same priority, the order of the tasks with the same priority will be random; Ensure that the satellite with greater resource availability is selected; if two satellites have the same amount of available resources, the order of the two satellites with the same amount of available resources will be randomized; Each satellite has scheduled tasks during the negotiation process; QNA prioritizes assigning tasks to satellites with the least accumulated tasks and the most remaining resources.

Images