Method and apparatus for processing multi-target task, electronic device, and storage medium

Abstract

The present disclosure provides a multi-target task processing method and device, electronic equipment and storage medium, relates to the technical field of data processing, and specifically relates to the technical field of intelligent search. The specific implementation scheme is: when the LLM set is used for processing strategy evolution of a multi-target task, the robustness of the evolved processing strategy is evaluated based on a perturbed task cluster, and the real-time updating and iteration of the candidate strategy and prompt word are performed based on the robustness evaluation result of the processing strategy, thereby improving the efficiency and speed of the processing strategy evolution of the target task, and improving the robustness and reliability of the evolved processing strategy.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of artificial intelligence technology such as deep learning, large models, intelligent agents, and intelligent scheduling. It also relates to a method, apparatus, electronic device, and storage medium for processing multi-objective tasks. Background Technology

[0002] Large-scale task combinatorial optimization problems are widespread in industry, such as large equipment scheduling, large-scale vehicle routing, and production workshop scheduling. Algorithm strategies generated by calling Large Language Models (LLMs) in related technologies are often only applicable to solving general mathematical problems, and their stability and robustness are insufficient when used for multi-task combinatorial optimization problems in industry. Summary of the Invention

[0003] This disclosure provides an information processing method, apparatus, electronic device, and storage medium.

[0004] According to one aspect of this disclosure, a method for processing multi-objective tasks is provided, comprising:

[0005] Based on the initial prompt words associated with the multi-objective tasks to be processed and multiple candidate processing strategies, a large language model set (LLM) is invoked to obtain multiple evolved processing strategies, wherein the LLM set contains multiple different LLMs. Based on the perturbation task cluster, each of the evolved processing strategies is evaluated to determine the robustness score corresponding to each evolved processing strategy. The perturbation task cluster is obtained by perturbing the multi-objective task based on the perturbation operator. Based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategy, a new candidate processing strategy is determined. Based on the first association information of the evolved processing strategy and / or the second association information of the new candidate processing strategy, new prompt words are determined; Based on the new prompt word and the new candidate processing strategy, the operation of obtaining multiple evolved processing strategies is returned until a target processing strategy for processing the multi-objective task is obtained.

[0006] According to another aspect of this disclosure, a multi-objective task processing apparatus is provided, comprising: The calling module is used to call a large language model set (LLM) based on the initial prompt words associated with the multi-objective task to be processed and multiple candidate processing strategies, so as to obtain multiple evolved processing strategies. The LLM set contains multiple different LLMs. An evaluation module is used to evaluate each of the evolved processing strategies based on a perturbation task cluster, and determine the robustness score corresponding to each of the evolved processing strategies, wherein the perturbation task cluster is obtained by perturbing the multi-objective task based on a perturbation operator; The processing module is used to determine new candidate processing strategies based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategies. The processing module is further configured to determine new prompt words based on the association information of the evolved processing strategy and the new candidate processing strategy; The return module is used to return to the operation of obtaining multiple evolved processing strategies based on the new prompt words and the new candidate processing strategies, until a target processing strategy for processing the multi-objective task is obtained.

[0007] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect.

[0008] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect.

[0009] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a schematic diagram of a multi-target task processing method provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of a multi-target task processing method provided in an embodiment of this disclosure; Figure 3This is a schematic diagram of a multi-target task processing method provided in an embodiment of this disclosure; Figure 4 This is a structural block diagram of a multi-target task processing device provided in an embodiment of this disclosure; Figure 5 A schematic block diagram of an electronic device used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] Existing technologies primarily employ the following three types of technical solutions to address the widespread large-scale task combination optimization problems in industry, such as: large equipment scheduling, large-scale vehicle routing planning, production workshop operation scheduling, power grid unit combination, logistics network optimization, cloud resource scheduling, coupled scheduling, etc. Option 1: Manual modeling combined with commercial solvers Operations research engineers manually build mixed-integer linear programming, mixed-integer nonlinear programming, constrained programming, or other mathematical programming models based on business scenarios. They explicitly write out the decision variables, objective functions, and constraints, and then call solvers such as Gurobi, CPLEX, SCIP, and OR-Tools to solve the model. Finally, they map the variable values ​​output by the solver back to the business decisions. The complete process of this approach generally includes: data preprocessing → model building → solver invocation → solution post-processing and verification → business output. However, the above approach suffers from high modeling costs, long cycles, and the need for significant time to modify the model in response to changes in business requirements.

[0014] Option 2: Manually design heuristic or metaheuristic algorithms When the problem size exceeds the solver's processing capacity, or when the problem contains black-box objectives / constraints that are difficult to model, engineers manually design metaheuristic algorithms such as greedy algorithms, tabu search, simulated annealing, genetic algorithms, differential evolution, adaptive large neighborhood search, and variable neighborhood search. These algorithms typically require designing neighborhood structures, operator sets, and acceptance criteria for specific problems, and conducting extensive parameter tuning and operator combination experiments. However, this approach suffers from high algorithm design and parameter tuning costs, weak cross-problem transferability, and a lack of end-to-end automation.

[0015] Option 3: Evolution of Traditional LLM-Driven Algorithms In recent years, the evolution of operations research optimization algorithms driven by artificial intelligence (AI) agents has gradually emerged. This approach treats algorithm code as the object of evolution, using Large Language Models (LLMs) to mutate and generate candidate codes, and using evaluators to score the candidate codes, forming a closed loop of "LLM generation → evaluator scoring → storing the top K codes in a database → sampling as the next generation's parent," thereby automatically evolving high-quality algorithms. This paradigm is the first to incorporate "algorithm design" itself into the evolutionary goal, but currently it is mainly used for solving general mathematical problems, such as matrix multiplication and filling problems. When applied to large-scale combinatorial optimization problems at the industrial level, there is significant room for improvement in search efficiency and stability. The generated algorithms may lack robustness, meaning that if the input data is changed or other similar scenarios are used, the generated algorithm may report errors or return infeasible solutions. Furthermore, the above process lacks a knowledge graph feedback mechanism, and the massive number of candidate algorithms generated during the evolution process are not stored and reused in a structured manner, resulting in subsequent generations being unable to effectively inherit failure experiences and success patterns.

[0016] This disclosure relates to the field of computer technology, and in particular to the field of artificial intelligence technologies such as deep learning, large models, and intelligent agents.

[0017] Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence.

[0018] Deep learning (DL) learns the inherent patterns and hierarchical representations of sample data. The information gained during this learning process greatly aids in interpreting data such as text, images, and sound. The ultimate goal of deep learning is to enable machines to possess analytical and learning capabilities similar to humans, allowing them to recognize data such as text, images, and sound.

[0019] Large models, also known as Foundation Models, are models that extract knowledge from hundreds of millions of corpora or images, learn, and then produce large models with hundreds of millions of parameters.

[0020] An intelligent agent, also known as a smart agent, is an intelligent entity that autonomously perceives its environment, makes decisions based on goals, and interacts with the environment to achieve specific functions. Intelligent agents perceive changes in the environment (such as through sensors or data input), make judgments and decisions based on their learned knowledge and algorithms, and then execute actions to influence the environment or achieve predetermined goals.

[0021] Intelligent scheduling involves the identification, analysis, allocation, sorting, conflict resolution, and dynamic adjustment of tasks, resources, time windows, and constraints. Its fundamental purpose is to find the optimal or near-optimal task sequence and resource matching relationship from numerous possible execution schemes, while satisfying given constraints (such as priority, deadline, resource capacity, etc.). This process can adaptively handle uncertainty, random disturbances, or real-time changes, thereby improving the overall efficiency, throughput, response speed, or resource utilization of the system. In this invention, intelligent scheduling encompasses, but is not limited to, reinforcement learning-based scheduling, genetic algorithm-optimized scheduling, multi-agent cooperative scheduling, and real-time rescheduling.

[0022] Figure 1 This is a schematic diagram illustrating a multi-objective task processing method provided in an embodiment of this disclosure. The multi-objective task processing method provided in this embodiment can be executed by an artificial intelligence (AI) agent for processing multi-objective tasks, such as... Figure 1 As shown, the method includes: S101: Based on the initial prompt words associated with the multi-objective task to be processed and multiple candidate processing strategies, call the large language model LLM set to obtain multiple evolved processing strategies, where the LLM set contains multiple different LLMs.

[0023] In this embodiment of the disclosure, the multi-objective task can be a task whose solution is composed of multiple different combinations of variables. For example, equipment scheduling tasks, production workshop job scheduling tasks, large-scale vehicle routing tasks, etc.

[0024] For example, for a device scheduling task, the solution to the task may involve the following combinations of variables: device-task pairing (which task is executed by which device (or devices), involving discrete assignment combinations from task to device), task execution order combination (the order in which multiple tasks are processed on the same device (permutations and combinations)), time resource allocation combination (the selection of the start / end time of each task, combined with the order), and combinations of auxiliary resources such as personnel and / or spare parts. These combination dimensions are mutually constrained (e.g., a device cannot do two things at the same time, and tasks have sequential dependencies), together forming a joint solution space for a device scheduling task.

[0025] In some embodiments, the initial prompt words and multiple candidate processing strategies can be pre-set in the AI ​​agent, or the AI ​​agent can query the associated prompt word library and processing strategy library respectively based on the type of multi-objective task (such as equipment scheduling task or logistics network task, etc.) to obtain the initial prompt words and multiple candidate processing strategies.

[0026] Optionally, the candidate processing strategy can be a parent algorithm selected from a pre-defined algorithm code library (or program library) to perform a specific task or as the basis for generating sub-algorithms. The algorithm code library can store multiple pre-written, independently callable algorithm code modules or program segments, each module corresponding to a basic algorithm logic, such as, but not limited to, sorting algorithms, search algorithms, path planning algorithms, data transformation algorithms, optimization algorithms, or neural network layer construction algorithms. This library can exist in the form of a function library, class library, script collection, or any computer-readable form. The candidate processing strategy can be a complete processing method or an abstract computational pattern (e.g., crossover operators or mutation operators in genetic algorithms, or the basic structure of policy networks in reinforcement learning). The evolved processing strategy can be a new processing strategy (or algorithm) generated by evolving the candidate processing strategy (or parent algorithm) (e.g., difference-based mutation, and / or full rewrite, and / or crossover-by-fragment, etc.).

[0027] In some embodiments, the multiple different LLMs in the LLM set can be multiple LLMs with different structures (such as the number of layers, the type of attention mechanism, etc.); or, they can be multiple LLMs with different algorithms (such as the loss function, optimizer, etc.); or, they can be multiple LLMs with different training data (such as data source, data size, etc.) and / or model size, etc., which are not limited in this disclosure.

[0028] Optionally, the LLM set mentioned above may include one or more of the following three types of evolution operators: The Diff-based mutation operator is used to achieve small-step local improvements, and its mathematical expression is shown below: ,in, This represents the parent algorithm in the current iteration; For the i-th large language model, prompt is the cue word that guides the evolution of the LLM; Indicates LLM according to The generated difference description or mutation instructions; This is a function for applying the difference information generated by the LLM to the parent policy (algorithm). This leads to the generation of the offspring algorithm. This mutation operation, by introducing mutation factors that differ structurally or in parameters from the parent algorithm, effectively increases population diversity and prevents the algorithm from getting trapped in local optima. Full rewrite is used to overcome local minima and achieve large leaps. The crossover-by-fragment operator is used to splice backbone fragments from multiple parent algorithms.

[0029] In this embodiment of the disclosure, the AI ​​agent calls multiple different LLMs to perform evolutionary processing on the initial prompt words associated with the multi-target task to be processed and multiple candidate processing strategies. The resulting multiple evolved processing strategies may be the same or different.

[0030] In this embodiment of the disclosure, multiple parallel threads can simultaneously call multiple LLMs in the LLM set to obtain multiple evolved processing strategies, thereby improving the evolution efficiency of the processing strategies and shortening the evolution time of the processing strategies.

[0031] S102, based on the perturbation task cluster, evaluate each evolved processing strategy and determine the robustness score corresponding to each evolved processing strategy. The perturbation task cluster is obtained by perturbing multi-objective tasks based on perturbation operators.

[0032] In this embodiment of the disclosure, the perturbation operator may include any one or more types of perturbation operators such as constraint reduction, parameter perturbation, scaling, and data translation.

[0033] Optionally, constraint reduction, similar to cutout (or clipping occlusion) regularization, randomly removes non-critical constraints at a certain ratio (e.g., [0, 0.3]); parameter perturbation, used to indicate the application of ±β% (β≤20) uniform perturbation to parameters such as time window, capacity, and demand; scaling, used to indicate the subsampling or synthetic expansion of instance size by γ; and data translation, used to indicate the reasonable translation of translationable data, such as translating the entire date window by Δt for time data to simulate the scenario of "replacing similar new data".

[0034] In this embodiment of the disclosure, a family of perturbation operators is constructed. Then, by using the perturbation algorithm, the original problem can be solved. Mapped as a perturbation problem .

[0035] In this embodiment of the disclosure, to avoid overfitting the evolved processing strategy to the current instance, a family of perturbation operators is constructed, where each operator maps the original problem (original task) to a perturbation problem (perturbation task cluster). The evolved processing strategy is evaluated based on the perturbation task cluster. For example, the robustness score of the evolved processing strategy can be determined based on whether it can run successfully and obtain a feasible solution for the perturbation task cluster.

[0036] S103. Based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategy, determine a new candidate processing strategy.

[0037] In this embodiment of the disclosure, the robustness of the candidate processing strategy can be improved by comparing the robustness scores corresponding to the evolved processing strategy and the candidate processing strategy. The processing strategy with the higher robustness score is determined as the new candidate processing strategy.

[0038] S104, based on the first association information of the evolved processing strategy and the second association information of the new candidate processing strategy, determine the new prompt word.

[0039] The aforementioned first and second association information can be any information that describes the generation process of the evolved processing strategy. For example, it may include prompts used during the evolution of the evolved processing strategy, the source processing strategy corresponding to the evolved strategy (i.e., which strategy the evolved strategy evolved from), the robustness score of the evolved processing strategy, which LLM the evolved strategy evolved from, causal attribution, etc. Causal attribution describes which components of the source (candidate) strategy have a key impact on the performance improvement or decline of the evolved strategy. This disclosure does not limit this aspect.

[0040] In this embodiment, if the new candidate processing strategy is an evolved processing strategy, the first association information and the second association information are the same, so a new prompt word can be determined based on the first association information (or the second association information). Alternatively, if the new candidate processing strategy is different from the evolved processing strategy (that is, the robustness score of the evolved processing strategy may be lower than the robustness score of the original candidate processing strategy), then the AI ​​agent can use the evolution process of the evolved processing strategy as a counterexample to guide the generation of new prompt words. This allows new prompt words to be generated based on the first and second association information, enabling the new prompt words to guide the LLM towards a more accurate evolution direction, thereby maximizing the efficiency of strategy evolution.

[0041] S105, based on the new prompt words and new candidate processing strategies, return to the operation of obtaining multiple evolved processing strategies until the target processing strategy for processing multi-objective tasks is obtained.

[0042] In this embodiment of the disclosure, after determining the new candidate processing strategy and the new prompt word, the LLM set call, robustness evaluation, and determination of the new candidate strategy and the new prompt word can be repeatedly performed until the target processing strategy is obtained.

[0043] Optionally, the AI ​​agent can determine the target policy when the robustness score of the newly obtained evolved treatment policy does not change significantly compared to the robustness score of the previous treatment policy. For example, the robustness score of the evolved treatment policy over K consecutive generations... Robustness score compared to the previous generation's processing strategy The difference is less than a certain threshold (e.g.) ), At that time, the target processing strategy was determined.

[0044] Optionally, the AI ​​agent can reach a set upper limit after the number of evolutions (or iterations) (e.g., When the AI ​​agent exhausts its computational resources, it determines the evolved strategy as the target processing strategy. Alternatively, the AI ​​agent can determine the evolved strategy as the target processing strategy when its computational resources are exhausted, etc., but this disclosure does not limit the specific implementation of the embodiments.

[0045] In this embodiment, the AI ​​agent achieves end-to-end closed-loop processing of multi-objective tasks through natural language problem processing → structured model (processing strategy) evolution → data extraction, transformation, and loading (ETL) → metaheuristic solution → target processing strategy output. This improves the efficiency of multi-objective task processing while ensuring the robustness of the obtained target processing strategy.

[0046] In this embodiment, when using the LLM set to evolve the processing strategy for multiple objectives, the robustness of the evolved processing strategy is evaluated based on the perturbation task cluster, and the candidate strategies and prompts are updated and iterated in real time based on the robustness evaluation results of the processing strategy. This improves the efficiency and speed of the processing strategy evolution for the objective task, and enhances the robustness and reliability of the evolved processing strategy.

[0047] Figure 2 This is a schematic diagram of a multi-objective task processing method provided in an embodiment of this disclosure. Figure 2 As shown, the method includes: S201, Obtain description information related to the multi-objective task, including associated basic processing strategies, declared parameters, and prohibited parameters.

[0048] In some embodiments, the baseline processing strategy can be a pre-defined, unadjusted or unadaptively optimized basic (or coarse) processing strategy for this type of multi-objective task. This baseline processing strategy is used to generate initial feasible solutions or establish a basic solution framework for the problem in the early stages of the solution process, serving as a reference benchmark for subsequent evaluation and improvement of hybrid metaheuristic algorithms. This effectively reduces the search space and avoids the resource waste caused by LLM performing strategy searches in a large-scale search space.

[0049] Optionally, the basic processing strategy can be determined by the AI ​​agent based on the type of multi-objective task by retrieving a basic processing strategy library. The AI ​​agent can maintain strategies categorized by task type. Category baseline library The task types can include the following: Traveling Salesman Problem (TSP), Vehicle Routing Problem (VRP), Job Shop Scheduling Problem (JSSP), Mixed Integer Programming (MIP), Quadratic Assignment Problem (QAP), etc. After obtaining the description information of each multi-objective task to be processed, the type of multi-objective task to be processed can be determined based on the description information. Then, the corresponding basic processing strategy (rough approach) can be retrieved. For example, the basic processing approach for VRP might be based on the Clarke-Wright saving algorithm and 2-opt, thereby compressing the LLM code search space by two to three orders of magnitude.

[0050] Optionally, parameters are declared to the LLM to specify the operations that must be performed during policy evolution, such as the algorithms that must be used. For example, parameters can be used to declare that the LLM must use a hybrid metaheuristic algorithm as the backbone for iteration or solution during policy evolution. Optionally, the parameters can also include the type or name of the metaheuristic algorithm that can be used. For example, if the parameters include Genetic Algorithm (GA), Simulated Annealing (SA), Tabu Search (TS), Adaptive Large Neighborhood Search (ALNS), etc., then it can be determined that at least one hybrid metaheuristic algorithm needs to be used as the core solution framework when evolving the target task, and the hybrid objects must be two or more algorithms (or their variants) from (GA, SA, TS, ALNS). Purely monomorphic heuristics (e.g., using genetic algorithms alone or simulated annealing alone) are not allowed, nor are fully deterministic methods (e.g., exact algorithms) or non-metamorphic heuristics (e.g., greedy constructs) from GA, SA, TS, ALNS, etc.

[0051] Optionally, the prohibition parameter is used to declare to the LLM operations that are prohibited during policy evolution. For example, the prohibition parameter may include prohibited functions, prohibited output content, etc. In this embodiment of the disclosure, in order to avoid the problem of memory overflow caused by the LLM directly calling the solver to solve the task during policy iteration, the type and / or name of the solver to be prohibited can be determined first based on the description information of the multi-objective task (such as the type of the corresponding problem).

[0052] In some embodiments, the description information of the multi-objective task may be input by the user or automatically generated by the AI ​​agent based on the multi-objective task to be processed; this disclosure does not limit this.

[0053] Optionally, the descriptive information of the multi-objective task may include the problem P0 of the original combinatorial optimization problem corresponding to the multi-objective task, the original data D0, the objective function and constraints, and the evaluation function of the combinatorial optimization problem.

[0054] In this embodiment of the disclosure, in order to enable the AI ​​agent to solve the combinatorial optimization problem P0 using a hybrid metaheuristic algorithm, the iterative object in the combinatorial optimization solution process can first be converted from the traditional decision variable x into a processing strategy (or algorithm code) vector a, and then the algorithm code vector a can be iteratively optimized.

[0055] Optionally, for the objective combinatorial optimization problem The traditional solution paradigm is:

[0056] in As decision variables, For the original problem The objective function, For the original problem The constraint set, st, is used to indicate the constraints in minimizing the objective function. When x is subject to the set (subject to, st), the variable x must be bound to the set (subject to, st). .

[0057] The iterative format of traditional optimization algorithms is as follows: ,or ,in Decision variables The iterative operator. In this embodiment of the disclosure, the iterative object is determined by the decision variable. Upgrade to a processing strategy vector ,make This indicates the processing strategy. The solution obtained by running on the input data Then the original problem It is restructured into the following "meta-optimization problem":

[0058] in Indicates processing strategy The validity of the syntax and runtime (e.g., whether it can run normally, whether it complies with constraints such as "disable solver"); To solve A 0-1 discriminant function to determine whether the original problem's constraints (conditions) are satisfied. If the constraints of the original problem are satisfied, then The value is 1 if it is set to 1, and 0 otherwise. The positive scores obtained by transforming the values ​​of the objective function ( (where the value is a constant). Finally, divide the score by... To keep the score below 1, which facilitates the evolution of LLM.

[0059] S202 generates the initial prompt word based on the basic processing strategy, declared parameters, and prohibited parameters.

[0060] In this embodiment, by first determining the associated basic processing strategy, declaration parameters, and prohibition parameters based on the description information of the multi-objective task to be processed, and then generating initial prompt words, the AI ​​agent can avoid evolving the strategy from scratch. Instead, the search space of LLM is significantly compressed by using the basic processing strategy, declaration parameters, and prohibition parameters, compressing the search space of LLM by several orders of magnitude. Under the same computing resource budget, the system can cross the local optimum earlier and enter the fine evolution stage of high-quality algorithms, which greatly improves the search efficiency and evolution speed of the AI ​​agent and increases the speed of processing multi-objective tasks.

[0061] S203, based on the first sampling probability, the second sampling probability and the third sampling probability, sample multiple candidate processing strategies from the elite strategy set, the optimal compromise frontier strategy set and the failed strategy set, respectively.

[0062] The elite strategy set refers to the set of top-ranked, best-performing processing strategies determined based on a pre-defined function or metric. The optimal trade-off front strategy set, in a multi-objective optimization problem, is the set of all non-dominated processing strategies, where each strategy represents the best trade-off between different objectives, and the decision-maker cannot improve other objectives without sacrificing the performance of one objective. The failed strategy set refers to the set of strategies that, during the solution process, are deemed infeasible, fail to meet key constraints, fall into an abnormal state, or fail to produce any improvement in consecutive iterations.

[0063] Optionally, based on the first sampling probability, samples can be taken from the elite policy set in descending order of the original performance score of the policy to maximize the number of high-performance processing policies among the sampled candidate processing policies.

[0064] In this embodiment of the disclosure, the elite strategy set, the optimal compromise frontier strategy set, and the failure strategy set can be preset; or they can be generated iteratively by the AI ​​agent during the processing of other related tasks. This embodiment of the disclosure does not limit this.

[0065] Optionally, the optimal compromise frontier strategy set can be a three-dimensional target space composed of robustness score, constraint score, and original performance score. The constraint score can be determined based on the effectiveness of the processing strategy and the processing results when processing the perturbed task cluster; this disclosure does not impose any limitations on this.

[0066] In this embodiment of the disclosure, the candidate processing strategies sampled from the failure strategy set can serve as a reverse warning to help LLM avoid repeating errors.

[0067] The first sampling probability, the second sampling probability, and the third sampling probability may be the same or different. Optionally, for different multi-target tasks, the first sampling probability, the second sampling probability, and the third sampling probability may be different, partially the same, or all the same; this disclosure does not limit this.

[0068] S204, based on the initial prompt words and multiple candidate processing strategies, calls the large language model LLM set to obtain multiple evolved processing strategies.

[0069] In this embodiment of the disclosure, when calling the LLM set, the sys.modules dictionary (which caches all imported modules) can be intercepted, monitored, or dynamically modified at runtime through the Prompt instruction to force LLM to only generate hybrid metaheuristic backbones (such as combined variants of genetic algorithm GA, simulated annealing SA, tabu search TS, adaptive large neighborhood search ALNS, large neighborhood search LNS, variable neighborhood search VNS, GRASP, etc.), while prohibiting the calling of external solvers, thereby avoiding memory overflow during the evolution of LLM.

[0070] S205, based on the perturbation task cluster, evaluates each evolved processing strategy and determines the robustness score corresponding to each evolved processing strategy.

[0071] The perturbation task cluster is obtained by perturbing the multi-objective task based on the perturbation operator.

[0072] S206. Based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategy, determine a new candidate processing strategy.

[0073] The implementation methods of S204-S206 can be referred to in detail in the implementation methods of S101-S103 in the above embodiments, and will not be repeated here.

[0074] In some embodiments, the AI ​​agent may determine any evolved processing strategy as a new candidate processing strategy if the robustness score corresponding to any evolved processing strategy is greater than the robustness score of the candidate processing strategy.

[0075] In some embodiments, after determining the robustness score of the evolved processing strategy, the AI ​​agent can update the elite strategy set, the optimal compromise frontier strategy set, and / or the failed strategy set based on the robustness score of the evolved processing strategy. For example, the evolved processing strategy with a robustness score greater than the original processing strategy can be added to the elite strategy set, or the evolved processing strategy with a robustness score less than the original processing strategy can be added to the failed strategy set, and so on. Then, a new candidate processing strategy can be selected from the updated elite strategy set, the optimal compromise frontier strategy set, and / or the failed strategy set.

[0076] Optionally, the AI ​​agent can update the first sampling probability, the second sampling probability, and / or the third sampling probability based on the robustness score of the evolved processing strategy, the robustness score of the source processing strategy, and the strategy set to which the source processing strategy belongs. For example, if the robustness score of the evolved processing strategy corresponding to a certain strategy set is greater than the robustness score of the associated candidate processing strategies, then it can be determined that the processing strategy obtained after the evolution of the strategies within that strategy set is more robust when solving multi-objective tasks. In order to maximize the convergence speed of the target strategy, the AI ​​agent can increase the sampling probability corresponding to any strategy set to sample more candidate processing strategies from that strategy set for processing strategy evolution.

[0077] Optionally, if the robustness score of the evolved processing strategy corresponding to any policy set is less than the robustness score of the associated candidate processing strategy, the AI ​​agent can reduce the sampling probability corresponding to any policy set to reduce the amount of invalid evolved processing strategies executed by the AI ​​agent.

[0078] Optionally, the AI ​​agent can increase or decrease the sampling probability corresponding to a certain policy set based on a preset step size. Alternatively, the increase or decrease in the sampling probability corresponding to a policy set can be determined based on the trend of the robustness score of the evolved processing policy corresponding to a certain policy set compared to the robustness score of the original processing policy. For example, if 80% of the evolved processing policies corresponding to a certain policy set have a robustness score greater than that of the original processing policy, while 70% of the evolved processing policies corresponding to another policy set have a robustness score less than that of the original processing policy, then the increase in the sampling probability corresponding to the first policy set can be equal to or slightly greater than the decrease in the sampling probability corresponding to the second policy set.

[0079] It should be noted that when the AI ​​agent adjusts the first sampling probability, the second sampling probability, and / or the third sampling probability, it needs to ensure that the sum of the three sampling probabilities is 1.

[0080] S207, based on the first association information of the evolved processing strategy, constructs a knowledge graph for multi-objective tasks.

[0081] The first associated information includes one or more of the following: source processing strategy, evolved processing strategy, prompt words used in the evolution process, and robustness score (robustness score of the source processing strategy and / or robustness score of the evolved processing strategy).

[0082] In this embodiment of the disclosure, in order to ensure that the AI ​​agent can continuously evolve the processing strategy based on the accurate evolution direction, the AI ​​agent can generate and update the knowledge graph of the strategy evolution process for the multi-objective task in real time during the processing strategy evolution process, so that the historical evolution process can provide feedback to guide the execution of subsequent evolution processes.

[0083] Optionally, nodes in the knowledge graph can represent source processing strategies and evolved processing strategies. Edges between nodes in the knowledge graph can be used to indicate the evolutionary relationship between processing strategies (the source processing strategy points to the evolved processing strategy, and the prompt words used during the evolution process), the similarity between processing strategies (edges within the tree edit distance threshold; for example, if the tree edit distance between two processing strategies is less than the threshold, then an edge is established between them), and causal attribution ("which type of prompt change brings which type of performance improvement"). Furthermore, since the performance of the evolved processing strategy (e.g., accuracy, robustness score, etc.) may be lower than that of the original processing strategy during the evolution of the processing strategy, this embodiment of the disclosure may also record information such as robustness score, whether the evolution process is a positive process (e.g., if the performance of the evolved processing strategy is better than that of the original processing strategy, then the evolution process is a positive process, otherwise it can be a negative process, etc.), and the reasons for the positive (or negative) evolution process (e.g., the robustness score increases (or decreases), etc.) in the knowledge graph. This embodiment of the disclosure does not limit this.

[0084] In some embodiments of this application, the nodes, node attributes (such as the policy set to which the processing policy node belongs, the comprehensive evaluation value (change) of the processing policy node, etc.), or the connection edges between nodes (such as the existence of evolutionary relationships between nodes, or causal attribution between nodes, etc.) of the pre-constructed multi-temporal prediction task knowledge graph can be updated based on the first association information. For example, if the nodes in the knowledge graph include processing policy nodes and performance indicator nodes, then if the evolved processing policy is inconsistent with all the nodes in the original knowledge graph, the evolved processing policy can be added to the knowledge graph as a node.

[0085] Optionally, the attributes (such as changes in policy set affiliation and trends in performance evaluation values) and / or edges of existing nodes can be updated based on the policy set to which the evolved processing strategy belongs and the corresponding causal attribution information. Through this update process, the knowledge graph gradually accumulates the history of policy evolution, success patterns, and lessons learned from failures in each iteration, providing structured knowledge support for subsequent policy retrieval and prompt word generation.

[0086] S208, based on the second association information of the new candidate processing strategy, query the knowledge graph to obtain the third association information of other processing strategies similar to the new candidate processing strategy.

[0087] S209, Generate new prompt words based on the third related information.

[0088] In this embodiment, after determining a new candidate processing strategy, the third association information of other similar processing strategies is queried from the knowledge graph based on the second association information of the new candidate processing strategy (e.g., the strategy set to which it belongs, the source processing strategy, the prompt words used in the evolution process, etc.). Then, based on the third association information (e.g., the prompt words used in the process of evolving other processing strategies as source strategies to obtain new evolved processing strategies, the robustness score of the new evolved processing strategy, the causal attribution path, and the final evolution effect, etc.), new prompt words can be generated. This realizes a text gradient-driven, highly flexible search process. By introducing structured prior ideas in the initial stage of evolution and the Prompt sampling process, the model is guided to perform targeted searches, which improves the probability and efficiency of the model generating more accurate processing strategies that are more in line with knowledge constraints.

[0089] S210, based on the new prompt words and new candidate processing strategies, return to the operation of obtaining multiple evolved processing strategies until the target processing strategy for processing multi-objective tasks is obtained.

[0090] The implementation of S210 can be described in detail in the implementation of S105 in the above embodiments, and will not be repeated here.

[0091] For example, This represents the partial prompts used in step k, such as information about the multi-objective task to be processed and constraints, etc., used to declare to the LLM. The resulting evolved processing strategy after step k is then... It can be represented as:

[0092] in, This is based on knowledge graphs and provides conditional information that can be input into an LLM (part of the new prompt words).

[0093] In this embodiment, by introducing a differentiated sampling mechanism of elite policy set, optimal compromise frontier policy set, and failed policy set, and combining it with the generation and evolution capabilities of Large Language Model (LLM), high-quality processing policies can be adaptively explored and generated in the solution space. Simultaneously, the robustness score of the policy is evaluated using perturbation task clusters, and candidate policies are dynamically updated based on the robustness score, significantly improving the stability and generalization ability of the policy under different task scenarios. Furthermore, by constructing a knowledge graph for multi-objective tasks and performing association queries and prompt word enhancement on similar policies, the reuse of historical experience and precise directional guidance for policy iteration are achieved, effectively reducing invalid searches, accelerating the convergence process, and ultimately obtaining efficient, robust, and interpretable target processing policies for multi-objective tasks.

[0094] Figure 3 This is a schematic diagram of a multi-objective task processing method provided in an embodiment of this disclosure. Figure 3 As shown, the method includes: S301, determine the initial prompt words and multiple candidate processing strategies for the multi-target task association to be processed.

[0095] S302, based on the initial prompt words and multiple candidate processing strategies, calls the large language model LLM set to obtain multiple evolved processing strategies.

[0096] The specific implementation of steps S301-302 can be found in the detailed description of steps S201-S204, and will not be repeated here.

[0097] S303 performs AST analysis on each evolved processing strategy to determine the effectiveness of the evolved processing strategy and whether there are any unauthorized calls.

[0098] Optionally, if any evolved processing strategy is determined to be invalid (e.g., due to syntax errors or semantic consistency issues) after AST analysis, it can be discarded directly. Alternatively, if AST analysis determines that the evolved processing strategy calls a module or algorithm declared in the prohibited parameters, it can also be discarded directly.

[0099] In the embodiments of the present disclosure, the program code or structured representation corresponding to the evolved processing strategy may be parsed into the form of an Abstract Syntax Tree (AST), and each node and sub-node in the tree is traversed to detect syntax correctness, semantic consistency, and the presence of explicitly prohibited external calls or module references. Through this analysis, it can be determined whether the currently evolved processing strategy meets the preset validity conditions, and whether it contains references to illegal calls (such as prohibited solver interfaces or system functions). In the present invention, the AST analysis result is first used as the basis for screening qualified evolved processing strategies and triggering a penalty or rollback mechanism, so as to ensure that the finally generated target processing strategy has legality and executability.

[0100] In S304, under the condition that any evolved processing strategy is valid and no illegal call exists, n perturbation tasks randomly selected from the perturbation task cluster are processed based on the any evolved processing strategy, so as to obtain n processing results.

[0101] Wherein, n is an integer less than N / 2, and N is the number of perturbation tasks contained in the perturbation task cluster.

[0102] Optionally, the processing result obtained by processing the perturbation task based on any evolved processing strategy may include any information indicating the processing procedure for processing the perturbation task and the finally obtained result. For example, it may include whether any evolved processing strategy can process the perturbation task (for example, whether the execution process is abnormally terminated), whether the obtained processing result is valid (for example, the processing result is available, or the processing result is unavailable, etc.), which is not limited in the embodiments of the present disclosure.

[0103] In S305, n reference scores of any evolved processing strategy are determined according to the n processing results.

[0104] In the embodiments of the present disclosure, when it is determined that any evolved processing strategy is valid and no illegal call exists (that is, any evolved processing strategy meets the primary screening condition), a small sample (n<N / 2) of perturbation tasks is randomly selected from the perturbation task cluster to further evaluate and verify the any evolved processing strategy.

[0105] Optionally, the validity score of any evolved processing strategy may be first determined according to whether any evolved processing strategy can run and whether it meets the prohibition parameter Then, based on the processing results, determine whether the constraints of the target task are met (that is, whether the solution obtained after processing the perturbation task based on any evolved processing strategy is a feasible solution, such as a feasible scheduling scheme for equipment or vehicles), and determine the constraint score of any evolved processing strategy. If the constraints are met, then it can be determined. otherwise Then, based on the effectiveness score and constraint score of any evolved processing strategy, a reference score for that evolved processing strategy is determined. In perturbation instances, scores are awarded as long as the constraints are met. This approach is both fast and avoids the optimal value of the perturbation causing it to "eat up" the score of the original problem.

[0106] In this embodiment, any evolved processing strategy can be further screened based on the ratio of the number of scores of 0 in the n reference scores to n. Optionally, if the ratio of the number of scores of 0 in the n reference scores of any evolved processing strategy to n is greater than a first threshold, it can be determined that the robustness of any evolved processing strategy is poor, and thus any evolved processing strategy can be discarded. The first threshold can be pre-configured or determined according to the type of multi-objective task; for example, the first threshold may be 0.2, 0.3, 0.40, etc., and this embodiment does not limit this.

[0107] S306, if the ratio of the number of scores of 0 to n in any evolved processing strategy is less than a threshold, then based on any evolved processing strategy, process the N perturbation tasks in the perturbation task cluster to obtain N processing results.

[0108] S307, based on N processing results, determine the robustness score of any evolved processing strategy.

[0109] Optionally, the robustness score of an evolved processing strategy can be determined based on the N processing results of N perturbed tasks.

[0110] For example, in this embodiment of the disclosure, a reference score for each perturbation task can be determined based on the processing result of a certain evolved processing strategy for each perturbation task. Then, based on N reference scores, the average reference score is determined. For example, if N=20, then the average reference score is... Then, based on the original performance score (Score(a)) of the evolved processing strategy (which can be determined according to the accuracy, efficiency, etc. of the processing strategy), and preset weight values, the robustness score of the evolved processing strategy is determined. Specifically, the robustness score of a given evolved processing strategy... It can be determined based on the following formula: , where λ is a regularization parameter greater than or equal to 0.

[0111] In this embodiment of the disclosure, Let be the robust regularization function, then the robust training objective of this disclosure is: In other words, in code space A, we are looking for an optimal solution a. This makes Score(a) and The weighted sum of the two parts reaches its maximum, where Score(a) is the original performance score (such as accuracy, efficiency, etc.), and the higher the value, the better the basic performance. It is a robust regularization function (e.g., measuring the stability of code against input perturbations or attacks), where λ≥0 is the regularization parameter used to control the weight of robustness in the overall objective. Thus, this function balances basic performance and robustness; by adjusting λ, the robustness of the model or code can be improved while maintaining a certain level of performance.

[0112] In this embodiment of the disclosure, by constructing a cluster of perturbation operators and then generating perturbation task instances, it is possible to prevent the evolved processing strategy from "overfitting" the input data instead of solving a whole class of problems. This evaluation rule can very effectively induce the AI ​​Agent to generate high-quality and robust algorithms, so that the generated algorithms avoid the problem of "performing well on the current input data but performing poorly on other data".

[0113] In this embodiment, the robustness score of the evolved processing strategy is determined by performing a three-stage evaluation: static analysis based on AST, small-sample perturbation evaluation, and full robustness evaluation. Compared with directly performing a full robustness evaluation on the evolved processing strategy, this method greatly reduces the amount of data processed and resources used for robustness evaluation of the evolved processing strategy, thereby improving the efficiency and speed of robustness evaluation.

[0114] In some possible implementations, after determining the reference score and robustness score of the evolved processing strategy, the AI ​​agent can update the optimal compromise frontier strategy set, the elite strategy set, and the failed strategy set based on the reference score and robustness score of the evolved processing strategy. For example, the AI ​​agent can add any evolved processing strategy to the elite strategy set if the reference score and robustness score of any evolved processing strategy in the optimal compromise frontier strategy set all meet the conditions; or, if the reference score and robustness score of any evolved processing strategy do not meet the conditions, it can add any evolved processing strategy to the failed strategy set; or, if the reference score and robustness score of any evolved processing strategy all meet the conditions, it can add any evolved processing strategy to the optimal compromise frontier strategy set, etc. The conditions for meeting the reference score and robustness score can be, for example, a score threshold, or the amount of score increase compared to the original processing strategy's reference score and robustness score, or a score ratio, etc., which are not limited in this embodiment.

[0115] In this embodiment of the disclosure, by adding the better-performing processing strategies from the optimal compromise frontier strategy set to the elite set, or by adding newly generated better-performing processing strategies to the optimal compromise frontier strategy set, or by adding newly generated poor-performing processing strategies to the failure strategy set, dynamic updates of the optimal compromise frontier strategy set, the elite strategy set, and the failure strategy set, and closed-loop optimization of the candidate processing strategy set are achieved, thereby further improving the efficiency and speed of the processing strategy generation process.

[0116] S308, based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategy, a new candidate processing strategy is determined.

[0117] S309, based on the first association information of the evolved processing strategy and / or the second association information of the new candidate processing strategy, determine the new prompt word.

[0118] S310, based on the new prompt word and the new candidate processing strategy, return to the operation of obtaining multiple evolved processing strategies until a target processing strategy for processing the multi-objective task is obtained.

[0119] The specific implementation of steps S308-S310 can be found in the descriptions in S103-105 and S206-S210, and will not be repeated here.

[0120] The multi-objective task processing method provided in this disclosure realizes end-to-end autonomous completion of the entire process of modeling, solving, data processing and result output, and has robustness to changes in data distribution.

[0121] For example, for variants of the Vehicle Routing Problem (VRP) with time windows, capacity, heterogeneous fleets, and multiple warehouses, the system can automatically evolve a customized hybrid meta-heuristic based on the "Clarke-Wright algorithm + large neighborhood search" baseline in the Prompt. Due to the introduction of operator perturbation-enhanced regularization, the algorithm evolved by the same system can adapt to inputs from different cities, different fleet sizes, and different time periods, avoiding frequent retraining.

[0122] Alternatively, for problems involving in-the-loop constraints and mixed differential equations, since calling the solver is prohibited and the iteration object is algorithm code rather than explicit variables, LLM can directly call the black-box / simulation interface within the algorithm and drive the evolution by returning the results through the evaluator, without the need for explicit modeling.

[0123] For example, for problems with tens of millions or more variables, the solution provided in this disclosure can be adopted. Since heuristic algorithms naturally support sparse computation, sliding windows, and hierarchical solutions, and with the "memory sensitive" prompt in the Prompt, heuristic algorithms with sparse computation mechanisms can be evolved (similar to the evolution from standard attention mechanisms to sparse attention mechanisms), thereby completing the solution within controlled memory.

[0124] In this embodiment, AST analysis is introduced to verify the effectiveness of the evolved processing strategies and detect illegal calls, ensuring the legitimacy and executability of the strategies participating in subsequent evaluations, and avoiding resource waste and system anomalies caused by invalid or illegal strategies. Simultaneously, a two-stage robustness evaluation mechanism is adopted: first, a small number (n) of perturbation tasks are used to quickly screen the evolved processing strategies; only after passing the initial screening (i.e., the proportion of non-zero scores meets the threshold requirement) are all (N) perturbation tasks comprehensively evaluated and robustness scores calculated, thus significantly reducing computational overhead while ensuring evaluation reliability. Furthermore, prompt words are dynamically updated based on the association information between the evolved strategies and candidate strategies, and the evolution process of the processing strategies is repeatedly executed in conjunction with iterative candidate strategies. This adaptively guides the search process towards a highly robust and effective strategy space, improving the stability, security, and reliability of the obtained target processing strategies for multi-objective tasks.

[0125] Figure 4 This is a structural block diagram of a multi-target task processing device provided in an embodiment of this disclosure. Figure 4 As shown, the multi-objective task processing device 400 includes: The calling module 401 is used to call the large language model LLM set based on the initial prompt words associated with the multi-target task to be processed and multiple candidate processing strategies, so as to obtain multiple evolved processing strategies. The LLM set contains multiple different LLMs. Evaluation module 402 is used to evaluate each of the evolved processing strategies based on the perturbation task cluster, and determine the robustness score corresponding to each of the evolved processing strategies, wherein the perturbation task cluster is obtained by perturbing the multi-objective task based on the perturbation operator; The processing module 403 is used to determine a new candidate processing strategy based on the robustness score corresponding to each of the evolved processing strategies and the robustness score corresponding to the candidate processing strategies. The processing module 403 is further configured to determine new prompt words based on the association information of the evolved processing strategy and the new candidate processing strategy; The return module 404 is used to return to the operation of obtaining multiple evolved processing strategies based on the new prompt word and the new candidate processing strategy, until a target processing strategy for processing the multi-objective task is obtained.

[0126] In some embodiments, the processing module is further configured to: Obtain the description information associated with the multi-objective task, the basic processing strategy, the declared parameters, and the prohibited parameters, wherein the declared parameters are used to declare to the LLM the operations that must be performed during the policy evolution process, and the prohibited parameters are used to declare to the LLM the operations that are prohibited from being performed during the policy evolution process; Based on the aforementioned basic processing strategy, declaration parameters, and prohibition parameters, the initial prompt words are generated; Based on the first sampling probability, the second sampling probability, and the third sampling probability, multiple candidate processing strategies are sampled from the elite strategy set, the optimal compromise frontier strategy set, and the failed strategy set, respectively.

[0127] In some embodiments, the processing module 403 is further configured to: An Abstract Syntax Tree (AST) analysis is performed on each of the evolved processing strategies to determine the effectiveness of the evolved processing strategies and whether there are any illegal calls. If any evolved processing strategy is effective and there is no illegal invocation, based on the evolved processing strategy, n randomly selected perturbation tasks in the perturbation task cluster are processed to obtain n processing results, where n is less than an integer less than N / 2, and N is the number of perturbation tasks contained in the perturbation task cluster. Based on the n processing results, determine n reference scores for any evolved processing strategy; If the ratio of the number of scores of 0 in the n reference scores of any evolved processing strategy to n is greater than a first threshold, then the evolved processing strategy is discarded.

[0128] In some embodiments, the processing module 403 is further configured to: The effectiveness score of any evolved processing strategy is determined based on whether it can be run and whether it meets the prohibition parameters. Based on the processing result, determine the constraint score of any evolved processing strategy to see if the constraints of the target task are met. Based on the effectiveness score and constraint score of any evolved processing strategy, a reference score is determined for each evolved processing strategy.

[0129] In some embodiments, the processing module 403 is further configured to: If the ratio of the number of scores of 0 in the n reference scores of any evolved processing strategy to n is less than a threshold, then based on any evolved processing strategy, the N perturbation tasks in the perturbation task cluster are processed respectively to obtain N processing results. Based on the N processing results, determine the robustness score of any evolved processing strategy.

[0130] In some embodiments, the processing module 403 is further configured to perform one or more of the following: If the reference score and robustness score of any evolved processing strategy in the optimal compromise frontier strategy set both meet the conditions, then the evolved processing strategy is added to the elite strategy set. If the reference score and robustness score of any evolved processing strategy do not meet the conditions, the evolved processing strategy will be added to the set of failed strategies. If the reference score and robustness score of any evolved processing strategy both meet the conditions, then the evolved processing strategy is added to the optimal compromise frontier strategy set.

[0131] In some embodiments, the processing module 403 is further configured to perform one or more of the following: If the robustness score of the evolved processing strategy corresponding to any strategy set is greater than the robustness score of the associated candidate processing strategy, the sampling probability corresponding to any strategy set is increased. If the robustness score of the evolved processing strategy corresponding to any policy set is lower than the robustness score of the associated candidate processing strategy, the sampling probability corresponding to any policy set shall be reduced.

[0132] In some embodiments, the processing module 403 is further configured to: If the robustness score of any evolved processing strategy is greater than the robustness score of the candidate processing strategy, then the evolved processing strategy is determined as a new candidate processing strategy.

[0133] In some embodiments, the processing module 403 is further configured to: Based on the first association information, a knowledge graph of the multi-objective task is constructed, wherein the first association information includes one or more of the following: source processing strategy, evolved processing strategy, prompt words used in the evolution process, and robustness score; Based on the second association information, the knowledge graph is queried to obtain third association information of other processing strategies similar to the new candidate processing strategy; The new prompt word is generated based on the third association information.

[0134] The multi-objective task processing device in this embodiment is used to execute the multi-objective task processing method to realize the processing of multi-objective tasks. The specific execution process of the multi-objective task processing device can be referred to the detailed description of the multi-objective task processing method above, which will not be repeated here.

[0135] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0136] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0137] Figure 5 A schematic block diagram of an electronic device for implementing embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0138] like Figure 5As shown, device 500 includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 502 or a computer program loaded from storage unit 508 into random access memory (RAM) 503. RAM 503 may also store various programs and data required for the operation of device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0139] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0140] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model processing strategies, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as information processing methods. For example, in some embodiments, the information processing method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the information processing method described above may be performed. Alternatively, in other embodiments, the computing unit 501 may be configured to perform information processing methods by any other suitable means (e.g., by means of firmware).

[0141] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0142] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0143] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0144] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0145] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0146] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0147] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0148] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. 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 principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for processing multi-objective tasks, wherein, The method includes: Based on the initial prompt words associated with the multi-objective tasks to be processed and multiple candidate processing strategies, a large language model set (LLM) is invoked to obtain multiple evolved processing strategies, wherein the LLM set contains multiple different LLMs. Based on the perturbation task cluster, each of the evolved processing strategies is evaluated to determine the robustness score corresponding to each evolved processing strategy. The perturbation task cluster is obtained by perturbing the multi-objective task based on the perturbation operator. Based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategy, a new candidate processing strategy is determined. Based on the first association information of the evolved processing strategy and / or the second association information of the new candidate processing strategy, new prompt words are determined; Based on the new prompt word and the new candidate processing strategy, the operation of obtaining multiple evolved processing strategies is executed until a target processing strategy for processing the multi-objective task is obtained.

2. The method as described in claim 1, wherein, Prior to invoking the Large Language Model (LLM) set, the method further includes: Obtain the description information associated with the multi-objective task, the basic processing strategy, the declared parameters, and the prohibited parameters, wherein the declared parameters are used to declare to the LLM the operations that must be performed during the policy evolution process, and the prohibited parameters are used to declare to the LLM the operations that are prohibited from being performed during the policy evolution process; Based on the aforementioned basic processing strategy, declaration parameters, and prohibition parameters, the initial prompt words are generated; Based on the first sampling probability, the second sampling probability, and the third sampling probability, multiple candidate processing strategies are sampled from the elite strategy set, the optimal compromise frontier strategy set, and the failed strategy set, respectively.

3. The method as described in claim 1, wherein, The method based on the perturbation task cluster evaluates each evolved processing strategy and determines the robustness score corresponding to each evolved processing strategy, including: An Abstract Syntax Tree (AST) analysis is performed on each of the evolved processing strategies to determine the effectiveness of the evolved processing strategies and whether there are any illegal calls. If any evolved processing strategy is effective and there is no illegal invocation, based on the evolved processing strategy, n randomly selected perturbation tasks in the perturbation task cluster are processed to obtain n processing results, where n is less than an integer less than N / 2, and N is the number of perturbation tasks contained in the perturbation task cluster. Based on the n processing results, determine n reference scores for any evolved processing strategy; If the ratio of the number of scores of 0 in the n reference scores of any evolved processing strategy to n is greater than a first threshold, then the evolved processing strategy is discarded.

4. The method of claim 3, wherein, The step of determining n reference scores for any evolved processing strategy based on the n processing results includes: The effectiveness score of any evolved processing strategy is determined based on whether it can be run and whether it meets the prohibition parameters. Based on the processing result, determine the constraint score of any evolved processing strategy to see if the constraints of the target task are met. Based on the effectiveness score and constraint score of any evolved processing strategy, a reference score is determined for each evolved processing strategy.

5. The method of claim 3, wherein, After determining the n reference scores for any of the evolved processing strategies, the method further includes: If the ratio of the number of scores of 0 in the n reference scores of any evolved processing strategy to n is less than a threshold, then based on any evolved processing strategy, the N perturbation tasks in the perturbation task cluster are processed respectively to obtain N processing results. Based on the N processing results, determine the robustness score of any evolved processing strategy.

6. The method of claim 5, wherein, After determining the robustness score of any of the evolved processing strategies, one or more of the following are also included: If the reference score and robustness score of any evolved processing strategy in the optimal compromise frontier strategy set both meet the conditions, then the evolved processing strategy is added to the elite strategy set. If the reference score and robustness score of any evolved processing strategy do not meet the conditions, the evolved processing strategy will be added to the set of failed strategies. If the reference score and robustness score of any evolved processing strategy both meet the conditions, then the evolved processing strategy is added to the optimal compromise frontier strategy set.

7. The method of claim 6, wherein, After determining the robustness score corresponding to each of the evolved processing strategies, the method further includes one or more of the following: If the robustness score of the evolved processing strategy corresponding to any strategy set is greater than the robustness score of the associated candidate processing strategy, the sampling probability corresponding to any strategy set is increased. If the robustness score of the evolved processing strategy corresponding to any policy set is lower than the robustness score of the associated candidate processing strategy, the sampling probability corresponding to any policy set shall be reduced.

8. The method as described in any one of claims 1-7, wherein, The step of determining new candidate processing strategies based on the robustness scores corresponding to each evolved processing strategy and the robustness scores corresponding to the candidate processing strategies includes: If the robustness score of any evolved processing strategy is greater than the robustness score of the candidate processing strategy, then the evolved processing strategy is determined as a new candidate processing strategy.

9. The method as described in any one of claims 1-7, wherein, The determination of new prompt words based on the first association information of the evolved processing strategy and / or the second association information of the new candidate processing strategy includes: Based on the first association information, a knowledge graph of the multi-objective task is constructed, wherein the first association information includes one or more of the following: source processing strategy, evolved processing strategy, prompt words used in the evolution process, and robustness score; Based on the second association information, the knowledge graph is queried to obtain third association information of other processing strategies similar to the new candidate processing strategy; The new prompt word is generated based on the third association information.

10. A multi-objective task processing apparatus, comprising: The calling module is used to call a large language model set (LLM) based on the initial prompt words associated with the multi-objective task to be processed and multiple candidate processing strategies, so as to obtain multiple evolved processing strategies. The LLM set contains multiple different LLMs. An evaluation module is used to evaluate each of the evolved processing strategies based on a perturbation task cluster, and determine the robustness score corresponding to each of the evolved processing strategies, wherein the perturbation task cluster is obtained by perturbing the multi-objective task based on a perturbation operator; The processing module is used to determine new candidate processing strategies based on the robustness score corresponding to each evolved processing strategy and the robustness score corresponding to the candidate processing strategies. The processing module is further configured to determine new prompt words based on the association information of the evolved processing strategy and the new candidate processing strategy; The return module is used to return to the operation of obtaining multiple evolved processing strategies based on the new prompt words and the new candidate processing strategies, until a target processing strategy for processing the multi-objective task is obtained.

11. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-9.

12. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-9.

13. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-9.

Images