Automated design optimization method for llm-based assembly line workstation decoding algorithm

Abstract

The application discloses an automatic design and optimization method of an LLM-based assembly line workstation decoding algorithm, belongs to the technical field of intelligent manufacturing system optimization and computer intelligence, and is characterized in that a self-evolution algorithm design framework is constructed, a high-performance decoding algorithm for solving a parallel mixed-flow assembly line balancing problem is automatically generated, and the decoding algorithm is integrated into an upper genetic algorithm of a double-layer optimization algorithm based on reinforcement learning assistance, so that a traditional paradigm depending on manual design is completely changed, and the solving efficiency and quality of the parallel assembly line balancing problem are significantly improved. Meanwhile, the self-evolution mechanism enables the decoding algorithm to have the capability of continuous improvement, can automatically adjust a design strategy according to different problem characteristics, and improves the adaptability and robustness of the algorithm, so that the efficiency and quality of production line planning are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing system optimization and computer intelligence cross-technology, and in particular relates to an automatic design optimization method for assembly line workstation decoding algorithm based on LLM. Background Technology

[0002] In modern manufacturing, asynchronous mixed-flow assembly lines are core equipment for meeting the demands of flexible production with diverse product types and small batches. To increase the capacity of critical processes, production lines are typically designed with multiple parallel assembly workstations within an assembly center. While this parallel workstation configuration improves the production line's throughput, it also significantly increases the complexity of production line balancing. For example, on an assembly line with parallel workstations, each assembly center can have multiple assembly workstations, each completing its task independently. However, equipment duplication leads to increased costs, and task allocation must satisfy cycle time constraints and priority relationships.

[0003] Currently, solving the parallel assembly line balancing problem mainly relies on evolutionary algorithms, and the performance of these algorithms largely depends on their "decoding algorithm," which transforms the encoded solution into a specific task allocation scheme. Existing decoding algorithms suffer from the following technical limitations in design and application: First, existing decoding algorithms heavily rely on domain expert knowledge, resulting in an inefficient and subjective design process. Currently used decoding rules in industry, such as the "minimize workstation idle time" rule, are entirely designed manually by domain experts based on their personal experience. This is a typical labor-intensive, trial-and-error-driven process, not only time-consuming and labor-intensive, but also making it difficult to guarantee the quality and stability of the design. When production line configurations or products change, experts need to reinvest significant effort in adjustments, and the system lacks adaptability.

[0004] Secondly, existing decoding algorithms typically use single, fixed rules, failing to effectively balance multiple conflicting optimization objectives. The parallel assembly line balancing problem is essentially a multi-objective optimization problem, usually requiring the simultaneous minimization of the total number of workstations, minimization of loop time, and maximization of the smoothness exponent. These objectives inherently conflict. However, once the existing decoding rules are determined, they remain fixed during algorithm execution, unable to dynamically adjust the emphasis on different objectives according to the search progress. This results in insufficient exploration capability on complex Pareto fronts, making it difficult to obtain high-quality, diverse solution sets.

[0005] Furthermore, existing decoding algorithms are rigidly coupled with the solution framework, lacking the ability for co-evolution. In traditional metaheuristic assembly line optimization algorithms, the decoding algorithm is typically embedded as a static module. It cannot learn from the overall search experience of the algorithm, nor can it generate effective feedback and interaction with the population evolution process. This "design-as-fixed" approach results in poor robustness and generalization ability of the solver when facing different problem instances or dynamic production environments.

[0006] Furthermore, in reality, the production line balancing problem often requires joint optimization with other decisions such as buffer configuration, forming a complex two-layer optimization problem. Existing technology includes a "two-layer optimization algorithm based on reinforcement learning," which uses a genetic algorithm in the upper layer to handle the production line balancing problem and deep reinforcement learning in the lower layer to handle the buffer configuration problem. While this two-layer framework can handle joint optimization, its upper-layer genetic algorithm still relies on pre-set, fixed decoding rules, which becomes a bottleneck restricting its overall performance. Simultaneously, solving this two-layer optimization problem is computationally expensive, placing extremely high demands on the efficiency of the upper-layer algorithm, which traditional fixed decoding algorithms struggle to meet.

[0007] In summary, existing technologies lack a systematic solution for the efficient and adaptive solution of parallel assembly line balancing problems that can automatically and efficiently design high-performance decoding algorithms and integrate them into complex optimization frameworks (such as bi-layer optimization). Therefore, there is an urgent need in this field for an innovative technical approach to overcome these bottlenecks. Summary of the Invention

[0008] To address the aforementioned issues, this application provides an automatic design and optimization method for decoding algorithms at assembly line workstations based on LLM (Large Language Model). By constructing a self-evolving algorithm design framework, a high-performance decoding algorithm for solving the balancing problem of parallel mixed-flow assembly lines is automatically generated and integrated into an upper-level genetic algorithm based on a reinforcement learning-assisted two-layer optimization algorithm, thereby significantly improving the efficiency and quality of production line planning.

[0009] To achieve the above objectives, this application adopts the following technical solution: An automatic design optimization method for assembly line workstation decoding algorithms based on LLM includes the following steps: S101. Input the parameters for the parallel assembly line balancing problem, including the standard time matrix of all tasks, the priority relationship diagram between tasks, the production cycle time, the cost of setting up parallel workstations, and the set of optimization objectives. S102. Execute the self-evolutionary method to design the SE-EoH process. This process takes the large language model as the core and automatically executes the design and optimization loop of the workstation decoding algorithm. S103. Based on the previous steps, iterate and output the parallel workstation decoding rule with the highest overall fitness score on all test problem sets after performance verification. S104. The parallel workstation decoding rules are integrated as a decoding module into the upper-level genetic algorithm of the reinforcement learning-assisted two-layer optimization system to complete the solver integration. S105. Input the parameters of the parallel assembly line balancing problem into the integrated solver, run the solver, and use the automatically designed decoding rules to decode the sequence encoding and output the Pareto optimal solution set.

[0010] Furthermore, the design and optimization loop of the automatic decoding algorithm includes the following steps: Initialization phase: Define the instruction prompts for the decoding algorithm design task and set the initial evolution parameters; Initial population generation: The large language model generates multiple decoding rule design schemes according to the design task requirements, forming an initial population; Fitness evaluation: Each decoding rule is temporarily integrated into a pre-defined test environment containing the upper-level genetic algorithm, and its performance is evaluated on a benchmark problem set and the overall fitness value is calculated; Self-evolutionary suggestion generation: Analyze the characteristic differences between the best and worst performing individuals in the current population, generate self-evolutionary suggestions, and integrate them into the design instructions; Offspring generation: Based on self-evolutionary hints and defined evolutionary methods, new decoding rule design schemes are generated; Environmental selection: The individuals with the highest fitness are selected from the parent and offspring to form a new generation of population, and the process is repeated until the termination condition is met.

[0011] Furthermore, in the initialization phase, the instructions for defining the design task include: Define a function named "parallel_station_decode", where "parallel_station_decode" is the "parallel workstation decoding rule". Its input parameters include the task sequence, the task processing time list and the production cycle time, and the output is the line balancing scheme. Design requirements: intelligently determine parallel workstation settings, support workstation merging optimization, achieve multi-objective trade-offs, meet constraints, and ensure algorithm efficiency.

[0012] Furthermore, in the initial population generation step, each individual generated by the large language model is a binary tuple containing: Heuristic description: Natural language explanation of the decoding rule thinking, including task allocation strategy, parallelism calculation method, and merging decision logic; Code implementation: The executable function code corresponding to the description can convert the task sequence into a specific line balancing scheme.

[0013] Furthermore, the fitness assessment step includes: Temporary integration: Integrating the decoding rule code into the upper-level genetic algorithm of the test environment; Benchmarking: The ensemble algorithm is run on a benchmark set containing multiple standard test problems, with each test problem run independently multiple times to eliminate randomness; Performance metrics: Record the total number of workstations, total cost, load balancing index, and computing efficiency metrics; Calculate fitness score: Use a weighted comprehensive scoring method to assign appropriate weights to different performance indicators and calculate the comprehensive fitness score.

[0014] Furthermore, the self-evolutionary suggestion generation step includes: Performance analysis: Based on the overall fitness value, identify the best and worst individuals in the current population; Feature extraction and comparison: Analyze the language descriptions and code structures of the two to extract the key design features that lead to their performance differences; Experience summary and suggestion generation: Invoke the large language model to generate self-evolving suggestions containing successful strategies and improvement recommendations; Hint consolidation: The self-evolution hints and initial instructions have been merged into an enhanced design instruction.

[0015] Furthermore, the offspring generation step employs the following four evolutionary settings: The first approach, E1: Exploratory creation, requires that the large language model not rely entirely on any single existing design, but rather conceive entirely new solutions from scratch based on the nature of the problem and evolutionary experience; The second approach, E2: Integrative Creation, requires a large language model to comprehensively analyze the heuristic descriptions and code implementations of multiple excellent individuals, extract and integrate their advantages to create new decoding rules; The third approach, M1, involves structural modifications that require the large language model to make key changes to the algorithm structure or logical flow of well-performing individuals. The fourth approach, M2, involves parametric tuning, which requires large language models to optimize algorithm behavior by adjusting constants, coefficients, or thresholds in the code.

[0016] Furthermore, the environment selection step includes: The previous generation population is merged with the offspring population generated in this round to form a temporary mixed population; Re-perform fitness assessment on each individual in the temporary mixed population; The elite selection strategy in multi-objective optimization is used to select the best individuals to form the next generation of the population. Check the termination conditions; if they are not met, start a new evolutionary cycle.

[0017] Furthermore, the solver integration step includes: The parallel workstation decoding rules are deeply embedded into the upper-level genetic algorithm as a core component for genotype-to-phenotype conversion; Design an encoding scheme based on task sequences to ensure full compatibility with decoding rules; Reconstruct the crossover and mutation operators to fully utilize the intelligent features of the decoding rules; Achieve deep coupling between fitness calculation and decoding rules.

[0018] Furthermore, the self-evolutionary method designs a self-evolutionary mechanism for the SE-EoH process, which, through deep collaboration across the following three dimensions, constructs a self-evolutionary system to achieve continuous optimization: By analyzing historical design experience, a dynamic optimization system for design strategies is established; By employing a synergistic combination of four evolutionary strategies, an organic balance between exploration and utilization can be achieved; The population diversity parameters, strategy allocation ratios, and evaluation criterion weights are automatically adjusted based on the evolutionary process.

[0019] Compared with the prior art, this application has the following advantages: This application leverages the powerful code generation and reasoning capabilities of large language models to fundamentally change the traditional paradigm of relying on manual design by constructing a self-evolving algorithm design framework. This significantly improves the efficiency and quality of algorithm design and integrates it into the upper-level genetic algorithm of a reinforcement learning-assisted two-layer optimization algorithm. Through seamless integration with the two-layer optimization framework, a high-performance decoding algorithm for solving the parallel mixed-flow assembly line balancing problem is automatically generated. This enables the automated design of parallel workstation decoding algorithms, forming a powerful end-to-end optimization solver that significantly improves the efficiency and quality of solving the parallel assembly line balancing problem. At the same time, the self-evolution mechanism enables the decoding algorithm to continuously improve, automatically adjusting the design strategy according to different problem characteristics, thereby enhancing the algorithm's adaptability and robustness. Attached Figure Description

[0020] Figure 1 A flowchart of the automatic design and optimization method for workstation decoding algorithms provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the working principle of the automatic design and optimization method for workstation decoding algorithms provided in this application embodiment. Figure 3 This is a flowchart illustrating the design of the self-evolutionary algorithm provided in the embodiments of this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0022] The overall technical solution of this application is to construct a self-evolutionary algorithm design framework called "SE-EoH (Self-Evolutionary Evolution of Heuristics)". This framework uses a large language model as the core engine, automatically executes the design and optimization loop of the decoding algorithm, and integrates the optimized decoding rules into a two-layer optimization solver with a genetic algorithm as the upper layer, which is used to efficiently solve the parallel assembly line balancing problem.

[0023] The following is combined with Figures 1-3 This application provides a detailed description of a collaborative optimization process for mixed-flow assembly lines based on deep reinforcement learning, as provided in the embodiments of this application.

[0024] First, combine Figure 1 This paper details the overall process of the automatic design and optimization method for the LLM-based assembly line workstation decoding algorithm in the embodiments of this application. Figure 1 As shown, the automatic design and optimization method for decoding algorithms provided in this application includes the following steps: S101. Problem Input and Definition: Input parameters for the parallel assembly line balancing problem, including: the standard time matrix of all tasks, the priority relationship diagram between tasks, the production cycle time, the cost of setting up parallel workstations, and the set of optimization objectives.

[0025] Specifically, the set of optimization objectives may include minimizing the number of workstations, minimizing total cost, maximizing the equilibrium rate, etc.

[0026] S102. Start the self-evolving method design framework SE-EoH and execute the self-evolving algorithm. This framework takes a large language model as its core and automatically executes the design and optimization loop of the decoding algorithm.

[0027] Based on design requirements, the system invokes the corresponding SE-EoH process to automatically generate and optimize parallel workstation decoding rules. The SE-EoH framework is an innovative architecture combining a large language model with evolutionary computation. Its workflow includes six core steps: initialization, initial population generation, fitness evaluation, self-evolutionary hint generation, offspring generation, and environment selection. Within this framework, the large language model continuously generates and improves parallel workstation decoding rules based on design task requirements and evolutionary feedback.

[0028] Specifically, it includes: Initialization phase: Define the instruction prompts for the decoding algorithm design task, set evolutionary parameters, including population size, maximum number of iterations, evolutionary strategy configuration, and evaluation benchmark problem set; Initial population generation: The large language model generates multiple decoding rule design schemes according to the design task requirements, forming an initial population; each scheme includes a natural language description and executable code; Fitness evaluation: Each decoding rule is integrated into the basic optimization framework, and its performance is evaluated on a benchmark problem set to calculate the overall fitness value; Self-evolutionary suggestion generation: Analyze the characteristic differences between the best and worst individuals in the current population to generate self-evolutionary suggestions; Offspring generation: Based on self-evolutionary hints and four evolutionary strategies, a new decoding rule design scheme is generated; Environmental selection: The individuals with the highest fitness are selected from the parent and offspring to form a new generation of population, and the process is repeated until the termination condition is met.

[0029] S103, Parallel workstation decoding rule generation: Through multiple rounds of iterative loops, output the parallel workstation decoding rule code that has been validated on all test problem sets and has the highest overall fitness score.

[0030] Within the SE-EoH framework, LLM, guided by an evolutionary strategy, undergoes multiple iterations, ultimately outputting one or more validated high-performance parallel workstation decoding rule codes. These high-performance parallel workstation decoding rules are the decoding rules with the highest overall fitness scores. A high overall fitness score means that, during the environment selection steps of multiple iterations, the individual selected to enter the next generation or as the final output, based on its overall fitness score, is typically the individual with the highest score within the comparison range.

[0031] S104 Solver Integration: Integrate the decoding rules into the upper-level genetic algorithm of the two-layer optimization algorithm based on reinforcement learning assistance.

[0032] The decoding rules of the parallel workstations obtained in S103 are integrated, like a "software chip," into a pre-built upper-level genetic algorithm based on a two-layer optimization algorithm framework assisted by reinforcement learning.

[0033] S105, Problem Solving and Solution Output: Input the defined problem into the integrated solver, run the evolutionary algorithm and use automatically designed decoding rules to decode the sequence code, and output the Pareto optimal solution set.

[0034] The problem defined in S101 is input into the integrated solver. The solver runs its internal evolutionary algorithm and uses the decoding rules of automatically designed parallel workstations to decode the sequence encodings in the upper-level optimization algorithm, ultimately outputting a Pareto optimal solution set. Each solution represents a feasible line-balanced scheme and is accompanied by its performance metrics on various optimization objectives, allowing decision-makers to make a final selection based on their preferences.

[0035] This embodiment leverages the powerful code generation and reasoning capabilities of a large language model to achieve complete automation in the design of decoding algorithms for parallel assembly line workstations. By constructing a self-evolving algorithm design framework, the decoding algorithm possesses deep adaptability, dynamically adjusting its internal decision logic and control parameters based on the characteristics of specific problems and the algorithm's search progress. This enhances the algorithm's robustness and performance across different scenarios, establishing a self-evolutionary mechanism for algorithm design. Through historical performance analysis and self-evolutionary suggestion generation, the algorithm design framework possesses continuous learning and improvement capabilities, automatically adjusting design strategies according to different production scenarios and problem characteristics. This represents a shift from single-algorithm design to an adaptive algorithm design paradigm, significantly improving design efficiency and reducing reliance on domain experts. Simultaneously, the automatically designed high-performance decoding algorithm is integrated into a reinforcement learning-assisted two-layer optimization algorithm framework, forming a powerful, end-to-end optimization solver specifically designed for efficiently solving the costly combinatorial optimization problem of parallel assembly line balancing. Through the collaborative optimization of the upper-layer genetic algorithm and the lower-layer deep reinforcement learning, joint optimization of assembly line balancing and buffer configuration is achieved, significantly improving overall solution efficiency and solution quality.

[0036] The above combination Figure 1 This application introduces an automatic design and optimization method for decoding algorithms of parallel assembly line workstations based on LLM drivers, as provided in the embodiments of this application. The following section combines... Figure 2-3 This paper describes the specific process of constructing the SE-EoH framework in the embodiments of this application.

[0037] The SE-EoH framework is the core innovation of this application; it is an automatic algorithm design framework that combines large language models with evolutionary computation. This framework is used to automatically design and optimize decoding rules for parallel assembly line workstations. The overall workflow of the SE-EoH framework is as follows: Figure 2 As shown, its workflow includes the following six core steps: S201: Initialization Phase: Initialization of the design task.

[0038] S2011, Define the instruction prompts for the design task, including function interfaces, input parameters, design requirements, and constraints.

[0039] The initialization phase is the starting point of the entire SE-EoH framework, mainly completing the preparatory work for the algorithm design task. This phase requires clearly defining the design goals, configuring algorithm parameters, and laying the foundation for the subsequent evolutionary process.

[0040] Define the design task (instruction prompting project): This is the "design blueprint" that guides the LLM process and must be clear and unambiguous. For example, when designing parallel workstation decoding rules, the prompt might look like this: "Task: Design an efficient decoding rule algorithm for parallel assembly line balancing."

[0041] Function interface: def parallel_station_decode(task_sequence, task_times, cycle_time), where "parallel_station_decode" is a function defined for the parallel workstation decoding rules. enter: task_sequence: Task sequence ⟨x_1,x_2,...,x_n> task_times: A list of task processing times [t_1, t_2, ..., t_n], cycle_time: Production cycle time (ct). Output: Line balancing scheme Design requirements: Parallel workstation decision-making: It is necessary to intelligently determine when to set up parallel workstations. When the total processing time of the workstations exceeds the cycle time, the optimal degree of parallelism is automatically calculated.

[0042] Workstation merging optimization: It can identify the synergistic effect of adjacent workstations and intelligently merge them while meeting the cycle time constraints.

[0043] Multi-objective trade-offs: achieving a balance between minimizing the number of workstations, minimizing total cost, and maximizing load balancing.

[0044] Constraints must be satisfied: Strictly satisfy task priority relationships and cycle time constraints.

[0045] Algorithm efficiency: Ensures computational efficiency in the decoding process and supports real-time applications.

[0046] Please first generate a natural language description explaining your design concept, and then provide complete, runnable code. S2012. Set the evolution parameters, including population size N=10 and maximum number of iterations I_max=50.

[0047] Population size N: The population size N determines the number of candidate decoding rules retained in each generation. It directly affects the algorithm's exploration ability and computational cost. For example, N=10 (10 optimal individuals are retained in each generation).

[0048] Maximum number of iterations: The maximum number of iterations I_max sets the upper limit of the entire evolution process and is one of the main conditions for the algorithm to terminate (e.g., I_max=50).

[0049] S2013: Configure the calling ratio of four evolution strategies E1, E2, M1, and M2 during offspring generation.

[0050] This configuration defines the frequency of use of the four core evolutionary strategies when generating offspring (i.e., the ratio of E1, E2, M1, and M2 calls), aiming to balance the relationship between "exploration" and "improvement." In the "offspring generation" step, offspring generation tasks will be allocated according to this ratio. (For example, if 10 offspring are needed per generation, then 3 will be generated through E1 and 3 through E2, and 2 through M1 and 2 through M2.) S2014. Establish an evaluation benchmark question set, which includes standard test questions of different sizes.

[0051] Evaluation Benchmark Problem Set: The evaluation benchmark problem set serves as an objective standard and testing environment for measuring the performance of decoding rules. It provides a unified and fair platform for quantitatively evaluating the performance of each decoding rule generated by LLM. Its comprehensiveness directly determines the generalization ability of the designed algorithm; a well-designed decoding rule must perform excellently in various production scenarios. In the "fitness evaluation" step, each candidate decoding rule will be run on this complete benchmark problem set. Its overall fitness is the weighted average of the performance of all test cases. This method forces the evolutionary process to find a general, robust, and excellent decoding rule, rather than an "overfitting" solution that only performs well on specific problems but poorly in other scenarios. Decoding rules designed in this way have the potential to be directly applied to various complex and variable scenarios in actual production.

[0052] For the selection of a large language model, DeepSeek-V3.2-Exp can be selected as the large language model engine. In steps S202 and S203, the system will call the DeepSeek-V3.2-Exp model through the API, send the carefully constructed prompt words (including task instructions, context, self-evolutionary prompts, etc.) to the model, and receive the generated code and description text.

[0053] S202: Initial Population Generation The initial population generation phase leverages the reasoning and creative capabilities of the large language model to generate diverse decoding rule design schemes. Each scheme includes a complete algorithm description and implementation code. The generated schemes undergo initial screening to ensure basic usability. The goal of this phase is to establish an initial population with sufficient diversity, providing rich material for subsequent evolutionary optimization. For example, based on the above instructions, LLM independently generates N different decoding rule design schemes, forming the initial population P_0={h_1,h_2,...,h_10}. Each individual h_i (i.e., each scheme) is a tuple: Heuristic description: Natural language explanation of the decoding rule thinking, including task allocation strategy, parallelism calculation method, merging decision logic, etc.

[0054] Code implementation: The executable function code corresponding to this description is capable of converting the task sequence into a specific line balancing scheme.

[0055] S203: Fitness Assessment: The fitness evaluation phase involves comprehensive performance testing of each individual in the population, using quantified evaluation metrics to reflect the quality of the decoding rules. This phase provides the basis for evolutionary selection, driving the population towards a better direction. It mainly includes the following sub-steps: Temporary integration: Integrating the decoding rule code of h_i into the upper-level genetic algorithm of a basic two-layer optimization algorithm framework.

[0056] Benchmarking: The ensemble algorithm is run on a benchmark set set up during the initialization phase, which includes multiple standard test problems (e.g., 6-task, 25-task, 75-task, etc.). Each test problem is run independently multiple times to eliminate randomness.

[0057] Performance metrics: Record key performance indicators for each test case, including: total number of workstations (minimize), total cost including workstation and equipment costs (minimize), load balancing index (maximize), and computational efficiency (minimize decoding time). Fitness score calculation: A weighted comprehensive scoring method is used to assign appropriate weights to different performance indicators and calculate the comprehensive fitness score. Individuals that perform well and consistently across all test questions receive a higher fitness score.

[0058] S204: Self-evolutionary hint generation: This step is the core component of the SE-EoH framework's "self-learning" and "intelligent evolution." Through a pre-defined self-evolutionary mechanism, it systematically analyzes the successes and failures of the current evolutionary cycle, transforming raw, implicit performance data into structured, actionable evolutionary knowledge. This process simulates the ability of human experts to summarize patterns from historical designs, generating "self-evolutionary hints" as dynamic instructions. These hints provide clear optimization directions for subsequent evolutionary cycles, significantly improving search efficiency and preventing large language models from repeatedly and blindly exploring in an ineffective design space. The sub-steps for generating self-evolutionary hints are as follows: Performance Analysis: The system sorts all individuals in the current population based on the comprehensive fitness score calculated in the "fitness evaluation" step. Subsequently, the system will clearly identify two key individuals: (1) Best individual (h_best): The individual with the highest fitness score in the current population, representing the most successful decoding rule design in this generation of evolution. (2) Worst individual (h_worst): The individual with the lowest fitness score in the current population, whose design strategy has been proven to be ineffective or seriously flawed.

[0059] Feature extraction and comparison: The system performs in-depth analysis of the "heuristic description" and "code implementation" of h_best and h_worst to extract the key design features that lead to their performance differences. (1) Analysis of h_best: The system parses its natural language description and statically analyzes its code structure to identify successful and reusable design patterns. For example, it may find that its code contains logic for dynamically adjusting the allocation threshold based on the remaining load, or that it calculates the correlation between tasks when merging workstations. (2) Analysis of h_worst: Similarly, the system analyzes its description and code to find the root cause of the poor performance. For example, it may find that it uses a fixed, unadjusted threshold, or that it does not consider the workstation merging process at all.

[0060] Experience Summary and Hint Generation: This sub-step invokes the large language model, acting as an "experience summarizer." The system organizes the structured feature comparison information extracted in the previous step, along with the fitness background of h_best and h_worst, into a clear analysis report, which is then used as input to the LLM. The generated self-evolutionary hints include, "Historical experience shows that successful decoding rules all employ dynamic threshold adjustment strategies and consider task relevance during workstation merging. Failed designs often use fixed rules and ignore multi-objective optimization. In subsequent designs, please prioritize dynamic adjustment strategies and multi-objective trade-off mechanisms." Hint integration: The newly generated "self-evolution hint" is integrated with the original "initialization instruction" to form an enhanced composite instruction for the next round of "offspring generation".

[0061] S205: Offspring generation (crossover and mutation): This step is the core driving force behind the population evolution within the SE-EoH framework. It simulates the "reproduction" process in biological evolution, utilizing four carefully designed evolutionary strategies to create new offspring individuals based on the parent population and self-evolutionary cues. These four strategies form a complete process from exploration to improvement: E1 and E2 strategies focus on exploring new design ideas and combinatorial innovations, belonging to global search; M1 and M2 strategies focus on deep optimization of existing superior individuals, belonging to local search. This multi-strategy parallel approach effectively balances development and exploration, ensuring that the population can continuously discover novel solutions while refining existing potential solutions, thereby continuously driving improvements in the performance of decoding rules. The four strategies for offspring generation are as follows: 1. Strategy E1 (Exploratory Creativity): This strategy aims to break free from the current design paradigm of the population, prevent the evolutionary process from prematurely converging to local optima, preserve the long-term evolutionary potential of the population, and encourage LLM to make groundbreaking innovations. It requires LLM not to rely entirely on any single existing design, but to conceive entirely new solutions from scratch based on the nature of the problem and evolutionary experience.

[0062] The core instruction is: "Completely ignore all existing design implementations and, based solely on the original 'initialization instructions' and the latest 'self-evolutionary hints,' conceive a completely new parallel workstation decoding rule from scratch." 2. Strategy E2 (Integrative Creation): This strategy aims to generate innovation through "knowledge recombination." It requires LLM to analyze the strengths of multiple top performers and quickly combine proven and effective design patterns, potentially producing offspring that outperform any parent. This is a highly efficient, recombination-based exploration.

[0063] The core instruction is: "Please comprehensively analyze the 'heuristic descriptions' and 'code implementations' of the following three excellent individuals [h_2, h_5, h_8]. Please extract and integrate their most prominent advantages to create a new decoding rule with stronger overall performance." 3. Strategy M1 (Structural Modification): This strategy focuses on deeply optimizing existing excellent designs to address potential structural defects or bottlenecks, thereby achieving further performance breakthroughs.

[0064] The core instruction is: "Here is the code for a well-performing decoding rule h2. Please make a few key modifications to its algorithm structure or logic flow to improve its performance. For example, you could improve the termination condition of its task allocation loop, refactor the calculation logic for the number of parallel workstations, or optimize the judgment criteria for workstation merging." 4. Strategy M2 (parametric adjustment): This strategy focuses on rapid and fine-tuning of high-performing parent algorithms. It requires the LLM to optimize algorithm behavior by modifying constants, coefficients, or thresholds in the code. This is an efficient local search strategy that can quickly uncover the performance potential of well-structured individuals with untuned parameters, or achieve performance improvements on specific problem types, at a relatively low computational cost.

[0065] The core instruction is: "Here is the code for decoding rule h7. Try adjusting its internal constants, coefficients, or thresholds (such as load rate threshold, similarity merging threshold, scaling factor in parallel calculation, etc.) to see if you can optimize the parameters to make it perform better." Through the parallel execution of the above four strategies, a total of 10 offspring were generated in this cycle according to the preset ratio (e.g., E1:3, E2:3, M1:2, M2:2), forming offspring population C. This new population includes both bold innovative explorations and meticulous refinement of existing results, providing rich and high-quality candidate solutions for the next round of environmental selection.

[0066] S206: Environmental Selection Environmental selection is the "survival of the fittest" stage in the SE-EoH framework's evolutionary cycle. Its core function is to select the best individuals from parents and offspring to form a new generation. This step, by applying selection pressure, ensures that the overall quality of the population improves generation by generation, ultimately converging to a high-performance decoding rule. It simulates the natural selection process, using multi-objective evaluation and elite preservation mechanisms to drive evolution towards the optimal overall performance while maintaining population diversity. The specific steps are as follows: The system merges the previous generation population P (10 individuals) with the offspring population C (10 individuals) generated in this round through four evolutionary strategies, forming a temporary hybrid population of 20 individuals. For each individual in this temporary hybrid population, the complete process of step 3 (fitness assessment) is re-executed. Based on the fitness data obtained from the re-assessment, the system employs an elite selection strategy from multi-objective optimization to precisely select the 10 best individuals from the 20 temporary individuals, forming the next generation population P_next. After completing the environmental selection, the system checks the termination conditions. If these conditions are not met, a new evolutionary cycle is started; if they are met, the final result is output.

[0067] Through these six iterative steps, the SE-EoH framework can systematically generate and optimize parallel workstation decoding rules. Each step has a clear functional positioning and operation specifications, forming a complete automated algorithm design system.

[0068] Furthermore, the automatic design process of the parallel workstation decoding rules in this embodiment is as follows: The automatic design of decoding rules for parallel workstations is a core output goal of the SE-EoH framework. Currently, traditional decoding rules typically involve the following manual design steps: Step 1: Initialize the index of currently pending tasks Assembly Center Serial Number Number of assembly stations .

[0069] Step 2: Create a new assembly center Based on the order in which tasks are assigned, select the [number] task in that order. Tasks at each location Assigned to assembly center At the same time, an assembly workstation was set up. Total processing time: , .

[0070] Step 3: Based on the order of task assignment, select the [number] task in that order. Tasks at each location .

[0071] Step 4: If satisfied Then the task Assign to the current assembly center k, and simultaneously let , Otherwise, let Return to step 2.

[0072] Step 5: Check if there are any unassigned tasks. If so, return to step 3. Otherwise, proceed to step 6.

[0073] Step 6: Complete the task allocation operation and perform parallel workstation allocation. Iterate through the total working time of all assembly centers, and if the conditions are met... This will make the current work center The number of parallel assembly workstations increased , are non-negative integers and satisfy .

[0074] Step 7: Iterate through the total working time of all assembly centers. If the conditions are met... ( To satisfy the smallest positive integer in the above formula, the assembly center arrive Merge into one assembly center, and set the number of assembly workstations to be [number missing]. .

[0075] Step 8: Complete the decoding operation to obtain the line balanced scheme.

[0076] While traditional decoding rules provide basic functionality, their rigid structure and single decision-making logic severely limit their optimization potential in dynamic production environments.

[0077] The decoding rules generated through the SE-EoH framework can autonomously adjust their internal decision-making mechanisms based on specific production scenarios and optimization goals, thereby finding a better balance among multiple conflicting objectives. The specific steps are as follows: First, the core problems that the decoding rules need to solve and the various requirements they need to meet are clearly defined, providing clear goals and constraints for subsequent algorithm design. Second, existing manually designed decoding rules are analyzed to establish performance benchmarks and identify their limitations, providing directions for improvement in automated design. Third, the six core steps constructed using the SE-EoH framework are utilized to generate and optimize decoding rules through multiple rounds of evolutionary iteration. The best-performing decoding rule is selected from the evolutionary results and prepared for integration into the optimization system.

[0078] The above comparison clearly demonstrates that the decoding rules automatically designed by the SE-EoH framework are no longer a simple, step-by-step procedure, but an adaptive decision-making system embedded with domain knowledge and optimization intelligence. It can handle complex situations and conflicting objectives that traditional rules cannot address, highlighting the technical advantages of this invention in solving complex assembly line balancing problems, thereby generating significantly higher-quality solutions for parallel assembly line balancing problems.

[0079] Furthermore, in this embodiment, integrating the automatically designed parallel workstation decoding rules into a reinforcement learning-assisted two-layer optimization framework is a key application aspect of this application. This integration forms a complete collaborative optimization system, where the upper layer focuses on generating line balancing schemes, and the lower layer is responsible for optimizing buffer configurations. The two layers work closely together through automatically designed intelligent decoding rules and a deep reinforcement learning agent, achieving a significant improvement in the overall performance of assembly line design. The specific integration steps are as follows: Step 1: Establish a two-layer optimized system architecture and define the data flow and interface specifications between each component.

[0080] First, a clear two-layer optimization system architecture needs to be established, clearly defining the data flow and interface specifications between each component. This step is fundamental to the entire integration process, ensuring that all components can work together effectively. Regarding data flow, the upper-layer optimizer uses a genetic algorithm framework to generate and optimize line balancing schemes, while the lower-layer optimizer focuses on buffer capacity configuration based on deep reinforcement learning. The assembly line simulator handles scheme performance evaluation and verification, and the control coordinator manages the collaborative work and data exchange between the two optimizers. At the data interface design level, the system defines the standard format for upper-layer output data (covering detailed information such as workstation configuration, task allocation, and parallel settings), establishes specifications for lower-layer input data (ensuring that line balancing scheme features can be accurately extracted and represented), and establishes a performance feedback data transmission protocol that includes multi-dimensional indicators such as throughput, cost, and balance. By clearly defining the data flow and interface specifications, an efficient and stable two-layer optimization integration framework is constructed.

[0081] Step 2: Integrate the decoding rules automatically designed by the SE-EoH framework into the upper-layer optimizer.

[0082] The high-performance decoding rules automatically designed by the SE-EoH framework in this embodiment are deeply embedded into the genetic algorithm of the upper-level optimizer, making them the core component for genotype-to-phenotype transformation. At the algorithm framework level, the traditional genetic algorithm is systematically modified. A task sequence-based encoding scheme is designed to ensure full compatibility with the decoding rules. Crossover and mutation operators are reconstructed to fully utilize the intelligent characteristics of the decoding rules, and fitness calculation is deeply coupled with the decoding rules to ensure the uniformity of evaluation criteria. This deep integration not only makes the decoding rules the core component of gene expression in the genetic algorithm, but more importantly, it integrates the intelligent decision-making capabilities of the decoding rules into the entire optimization search process, significantly improving the search efficiency and solution quality of the genetic algorithm when solving complex assembly line balancing problems.

[0083] Step 3: Train the reinforcement learning agent and integrate it into the lower-level optimizer.

[0084] The system trains a deep reinforcement learning agent to quickly generate optimal buffer configurations based on the line balancing scheme provided by the upper layer. Specific implementations include: In state space design, the system encodes key indicators such as workstation load rate and processing time variation coefficient, extracts key features such as task allocation patterns from the line balancing scheme output by the upper-layer decoding rules, and integrates the current buffer configuration state and real-time performance indicators to form a complete system state representation; in action space definition, a discrete action space is constructed with multiple capacity levels preset for each buffer for the agent to choose from, while a continuous action space is designed to support precise adjustment and fine-tuning of buffer capacity; in reward function construction, a positive performance reward mechanism based on throughput improvement is designed, cost penalty constraints are set for over-investment and resource waste, a balanced reward strategy to encourage load balancing configuration schemes is formulated, and additional stability incentives to improve system robustness are established. Through this multi-dimensional, multi-objective reinforcement learning framework, the agent can quickly generate high-quality buffer configuration schemes based on the line balancing scheme, comprehensively considering multiple optimization objectives such as production efficiency, cost control, and system stability, and form an effective collaborative optimization mechanism with the upper-layer optimizer.

[0085] Step 4: Establish a collaborative working mechanism between the upper and lower level optimizers to jointly optimize overall performance.

[0086] A deep collaborative working mechanism is established between upper and lower level optimizers to achieve joint optimization of overall performance. This is achieved through a two-level collaborative mechanism: at the process level, a working mode combining alternating optimization and parallel execution is designed. After the upper-level genetic algorithm generates a line balancing scheme, the lower-level reinforcement learning agent immediately performs buffer configuration optimization, forming a closed-loop feedback. At the information level, a bidirectional data channel is established, with the upper layer transmitting workstation load distribution and task allocation characteristics to the lower layer, and the lower layer feeding back the impact analysis of buffer configuration on system performance to the upper layer. This deep collaborative mechanism organically integrates line balancing optimization and buffer configuration, significantly improving the overall system performance while ensuring the optimization quality of each optimizer through mutual promotion and collaborative search between the upper and lower level optimizers.

[0087] The systematic implementation of the above four steps ensures deep integration between the decoding rules and the two-layer optimization framework. This integration not only guarantees that the intelligent characteristics of the decoding rules are fully utilized, but also achieves a significant improvement in overall optimization performance through the synergistic effect of the two-layer optimizers. In practical applications, the system demonstrates excellent solution quality, optimization efficiency, and adaptability, providing a reliable solution for complex assembly line design problems.

[0088] In this embodiment, the self-evolution mechanism of the SE-EoH framework is the core driving force for its continuous optimization and performance improvement. This mechanism constructs a complete self-learning and evolution system through the deep cooperation of the following three dimensions: (1) Performance-driven cue evolution mechanism: This mechanism establishes a dynamic optimization system for design strategies by systematically analyzing historical design experience. Specifically, the system continuously tracks the performance of each decoding rule on a benchmark set, identifying common characteristics of high-performance rules and typical defects of low-performance rules. Based on these analysis results, the system automatically generates targeted evolutionary hints, such as: "Historical experience shows that decoding rules using dynamic threshold adjustment strategies perform well in load balancing metrics, while rules with fixed thresholds are prone to uneven workstation load. Adaptive parameter adjustment mechanisms should be given priority in subsequent designs." This data-driven evolutionary hint ensures that the design process always progresses in the right direction.

[0089] (2) Multi-strategy exploration mechanism: The system employs a synergistic approach of four evolutionary strategies (i.e., four evolutionary strategies for offspring generation) to achieve an organic balance between exploration and utilization. These four strategies are executed in parallel according to a preset ratio, ensuring that the potential of existing excellent designs is continuously deepened while maintaining innovative vitality.

[0090] (3) Dynamic parameter adjustment mechanism: The system automatically adjusts key operating parameters based on the evolutionary process, including: population diversity parameters: dynamically adjusting selection pressure based on convergence to prevent premature convergence; strategy allocation ratios: dynamically adjusting the weight of each strategy in offspring generation based on its historical performance; and evaluation criterion weights: adjusting the relative importance of each performance indicator based on the current optimization focus. This adaptive parameter adjustment ensures that the evolutionary process maintains optimal search efficiency at different stages.

[0091] Through the close collaboration of these three dimensions, the SE-EoH framework has achieved a qualitative leap from simple algorithm generation to intelligent self-evolution, ensuring continuous improvement in the quality of decoding rule design and rapid convergence.

[0092] The following is a specific application example of using the SE-EoH framework to automatically design and integrate decoding rules for parallel workstations. This example details the complete implementation process of automatically designing parallel workstation decoding rules using the SE-EoH framework and integrating them into a reinforcement learning-assisted two-layer optimization algorithm. The specific steps are as follows: (1) Problem input and definition: First, the specific parameters for the parallel assembly line balancing problem are defined, including: the standard time matrix for all tasks, the priority relationship diagram between tasks, the production cycle time, the cost of setting up parallel workstations, and the set of objectives to be optimized. The specific parameter settings are shown in Table 1 below: Table 1: Specific parameters for the parallel assembly line balancing problem:

[0093] (2) Start the SE-EoH framework: like Figure 2 The SE-EoH framework overall workflow diagram shown below, upon starting the automated algorithm design engine, will execute the following steps: Initialization phase: The configuration parameters for the SE-EoH framework are shown in Table 2 below: Table 2: SE-EoH Framework Parameters

[0094] Define design task instruction prompts: Task: Design an efficient decoding rule algorithm for parallel assembly line balancing.

[0095] Function interface: defparallel_station_decode(task_sequence, task_times, cycle_time): Input: Task sequence, task time list, production cycle time Design requirements: The algorithm must intelligently determine parallel workstation settings, support workstation merging optimization, implement multi-objective tradeoffs, satisfy constraints, and guarantee algorithm efficiency.

[0096] Initial population generation: The initialization prompts are input into the large language model, generating 10 different decoding rule designs to form the initial population P0. Each individual contains a heuristic description and a corresponding executable code implementation.

[0097] Fitness assessment: Each generated decoding rule undergoes a fitness evaluation, with evaluation metrics including: total number of workstations (weight: 0.4), total cost (weight: 0.4), load balancing index (weight: 0.2), and computational efficiency (as an auxiliary metric). Finally, the fitness value for each individual rule is calculated using a weighted comprehensive score based on the fitness weights of each evaluation metric, using the following formula: Fitness = 0.4 × (1 - standardized value of number of workstations) + 0.4 × (1 - standardized value of cost) + 0.2 × load balancing index.

[0098] (3) Evolutionary optimization process: Based on the evaluation results of the initial population, multiple rounds of evolutionary optimization are performed: Self-evolutionary hint generation: Analyze the characteristic differences between the best-performing and worst-performing individuals to generate the following self-evolutionary hints: Based on the design experience of the previous round, successful decoding rules typically have the following characteristics: When setting up parallel workstations, a gradual increase strategy should be adopted. When merging work centers, consider the synergistic effects of adjacent work centers. Introducing a multi-objective trade-off mechanism and dynamically adjusting optimization priorities: In the new round of design, please prioritize these success factors and avoid the following problems: excessive use of parallel workstations leading to increased costs; ignoring opportunities for work center merging; and a single optimization objective causing a decline in other performance indicators. Offspring generation: Four evolutionary strategies are used to generate new decoding rule design schemes: E1 Strategy: Exploring New Decoding Logic and Allocation Strategies E2 Strategy: Combining the shared ideas of existing outstanding individuals M1 Strategy: Optimize the structure of high-performing individuals. M2 strategy: Adjust key parameters in the decoding rules.

[0099] Environmental selection: The 10 individuals with the highest fitness from the parent and offspring generations are selected to form a new generation of the population.

[0100] (4) Final decoding rule output: After 50 rounds of iteration, the decoding rule with the highest fitness is selected as the final output. The final line balancing scheme is shown in Table 3 below: Table 3: Final Line Balancing Scheme Automatically Decoded:

[0101] (5) Solver integration and problem solving: The obtained decoding rules are integrated into a two-layer optimization algorithm framework based on reinforcement learning assistance: The decoding rules are embedded into the upper-level genetic algorithm as a core component; Design compatible encoding schemes and genetic operators; Achieve deep coupling between fitness calculation and decoding output.

[0102] The defined problem is input into the integrated solver, the evolutionary algorithm is run and the solution is decoded using automatically designed decoding rules, and the Pareto optimal solution set is output.

[0103] To illustrate the technical advantages of this invention, the traditional manual decoding process and results are shown in Table 4 below: Table 4: Traditional Manual Decoding Process and Results

[0104] Compared to the manual design process, the automated design solution, after 50 rounds of iterative optimization, demonstrates significantly superior overall system performance. Specifically, the automated design utilizes only 3 assembly centers, while manual design requires 4. Although the automated design uses one more machine, it achieves significant overall cost savings by replacing the higher fixed cost of the assembly centers (30,000 yuan) with the lower cost of the machine itself (3,000 yuan). Furthermore, the number of parallel workstations is proportionally matched to the workstation load, showcasing higher resource utilization efficiency and intelligent trade-off capabilities, thus confirming its superiority.

[0105] In practical applications, the powerful reasoning and code generation capabilities of large language models break through the limitations of traditional algorithm design, creating a new paradigm for automatic algorithm design. This invention enables the automatically designed decoding rules to have good generalization ability, adapt to assembly line balancing problems of different scales and production requirements, significantly reduce production line design costs, improve production efficiency, and provide reliable technical support for intelligent manufacturing systems.

[0106] All parts not described in detail in this application are existing conventional technologies and will not be elaborated here.

[0107] It is understood that the above specific description of the present invention is only for illustrating the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention to achieve the same technical effect; as long as the use needs are met, they are all within the protection scope of the present invention.

Claims

1. An automatic design optimization method for assembly line workstation decoding algorithms based on LLM, characterized in that, Includes the following steps: S101. Input the parameters for the parallel assembly line balancing problem, including the standard time matrix of all tasks, the priority relationship diagram between tasks, the production cycle time, the cost of setting up parallel workstations, and the set of optimization objectives. S102. Execute the self-evolutionary method to design the SE-EoH process. This process takes the large language model as the core and automatically executes the design and optimization loop of the workstation decoding algorithm. S103. Based on the previous steps, iterate and output the parallel workstation decoding rule with the highest overall fitness score on all test problem sets after performance verification. S104. The parallel workstation decoding rules are integrated as a decoding module into the upper-level genetic algorithm of the reinforcement learning-assisted two-layer optimization system to complete the solver integration. S105. Input the parameters of the parallel assembly line balancing problem into the integrated solver, run the solver, and use the automatically designed decoding rules to decode the sequence code and output the Pareto optimal solution set. The design and optimization loop of the automatic decoding algorithm includes the following steps: Initialization phase: Define the instruction prompts for the decoding algorithm design task and set the initial evolution parameters; Initial population generation: The large language model generates multiple decoding rule design schemes according to the design task requirements, forming an initial population; Fitness evaluation: Each decoding rule is temporarily integrated into a pre-defined test environment containing the upper-level genetic algorithm, and its performance is evaluated on a benchmark problem set and the overall fitness value is calculated; Self-evolutionary suggestion generation: Analyze the characteristic differences between the best and worst performing individuals in the current population, generate self-evolutionary suggestions, and integrate them into the design instructions; Offspring generation: Based on self-evolutionary hints and defined evolutionary methods, new decoding rule design schemes are generated; Environmental selection: The individuals with the highest fitness are selected from the parent and offspring to form a new generation of population, and the process is repeated until the termination condition is met.

2. The method according to claim 1, characterized in that, During the initialization phase, the instructions for defining the design task include: Define a function named "parallel_station_decode", where "parallel_station_decode" is the "parallel workstation decoding rule". Its input parameters include the task sequence, the task processing time list and the production cycle time, and the output is the line balancing scheme. Design requirements: intelligently determine parallel workstation settings, support workstation merging optimization, achieve multi-objective trade-offs, meet constraints, and ensure algorithm efficiency.

3. The method according to claim 1, characterized in that, In the initial population generation step, each individual generated by the large language model is a binary tuple containing: Heuristic description: Natural language explanation of the decoding rule thinking, including task allocation strategy, parallelism calculation method, and merging decision logic; Code implementation: The executable function code corresponding to the description can convert the task sequence into a specific line balancing scheme.

4. The method according to claim 1, characterized in that, The fitness assessment steps include: Temporary integration: Integrating the decoding rule code into the upper-level genetic algorithm of the test environment; Benchmarking: The ensemble algorithm is run on a benchmark set containing multiple standard test problems, with each test problem run independently multiple times to eliminate randomness; Performance metrics: Record the total number of workstations, total cost, load balancing index, and computing efficiency metrics; Calculate fitness score: Use a weighted comprehensive scoring method to assign appropriate weights to different performance indicators and calculate the comprehensive fitness score.

5. The method according to claim 1, characterized in that, The self-evolutionary suggestion generation step includes: Performance analysis: Based on the overall fitness value, identify the best and worst individuals in the current population; Feature extraction and comparison: Analyze the language descriptions and code structures of the two to extract the key design features that lead to their performance differences; Experience summary and suggestion generation: Invoke the large language model to generate self-evolving suggestions containing successful strategies and improvement recommendations; Hint consolidation: The self-evolution hints and initial instructions have been merged into an enhanced design instruction.

6. The method according to claim 1, characterized in that, The offspring generation step employs the following four evolutionary settings: The first approach, E1: Exploratory creation, requires that the large language model not rely entirely on any single existing design, but rather conceive entirely new solutions from scratch based on the nature of the problem and evolutionary experience; The second approach, E2: Integrative Creation, requires a large language model to comprehensively analyze the heuristic descriptions and code implementations of multiple excellent individuals, extract and integrate their advantages to create new decoding rules; The third approach, M1, involves structural modifications that require the large language model to make key changes to the algorithm structure or logical flow of well-performing individuals. The fourth approach, M2, involves parametric tuning, which requires large language models to optimize algorithm behavior by adjusting constants, coefficients, or thresholds in the code.

7. The method according to claim 1, characterized in that, The environment selection step includes: The previous generation population is merged with the offspring population generated in this round to form a temporary mixed population; Re-perform fitness assessment on each individual in the temporary mixed population; The elite selection strategy in multi-objective optimization is used to select the best individuals to form the next generation of the population. Check the termination conditions; if they are not met, start a new evolutionary cycle.

8. The method according to claim 1, characterized in that, The solver integration step includes: The parallel workstation decoding rules are deeply embedded into the upper-level genetic algorithm as a core component for genotype-to-phenotype conversion; Design an encoding scheme based on task sequences to ensure full compatibility with decoding rules; Reconstruct the crossover and mutation operators to fully utilize the intelligent features of the decoding rules; Achieve deep coupling between fitness calculation and decoding rules.

9. The method according to claim 1, characterized in that, The self-evolutionary method designs a self-evolutionary mechanism for the SE-EoH process, which is constructed through deep collaboration across the following three dimensions to achieve continuous optimization: By analyzing historical design experience, a dynamic optimization system for design strategies is established; Four evolutionary strategies are employed in synergy; The population diversity parameters, strategy allocation ratios, and evaluation criterion weights are automatically adjusted based on the evolutionary process.

Images