Dynamic multi-modal multi-objective optimization method based on large language model and dual-archiving strategy

Through a dynamic multi-modal multi-objective optimization method based on large language model and dual archive strategy, the multi-modal complexity in the multi-objective optimization problem is solved, and effective solution set generation under the mutual constraints of multiple goals is achieved to ensure understanding The convergence and diversity of the optimization problem improve the efficiency of solving optimization problems.

CN119066947APending Publication Date: 2024-12-03XIAN UNIV OF TECH
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411004494.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art is difficult to effectively solve the multimodal situation in the multi-objective optimization problem, especially when multiple goals are restricted and conflicted, it is difficult to find the optimal trade-off solution set, and there are multimodal situations in the multi-objective optimization problem. The situation leads to the complexity of multimodal multi-objective optimization problem.

Method used

Adopting a dynamic multimodal multi-objective optimization method based on large language model and dual archive strategy, combining convergence and diversity through initialization parameters, decision variable analysis, reference vector set update, environmental change monitoring and prediction mechanism of large language model The archive updates generate the final solution set, ensuring good convergence and diversity in the target space and decision space.

Benefits of technology

It realizes rapid analysis and prediction under environmental changes, generates high-quality final solution sets, maintains good convergence and diversity of solutions, and improves the solution ability and efficiency of multi-objective optimization problems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066947A_ABST
    Figure CN119066947A_ABST
Patent Text Reader

Abstract

The invention discloses a dynamic multi-modal multi-objective optimization method based on a large language model and a double-archiving strategy. The method is specifically implemented according to the following steps: initializing parameters; detecting an independent convergence related decision variable set by adopting a decision variable analysis technology; a group of reference vector sets R which are well distributed in a target space is utilized to guide updating of the diversity archives; updating the convergence archive AC and the diversity archive AD; when the circulation condition is met, an environment change monitoring strategy is implemented; generating corresponding individuals based on a prediction mechanism strategy of a large language model, and updating the convergence archive and the diversity archive respectively; from the first generation to the MaxGen generation, individuals are selected from the convergence archive AC and the diversity archive AD to serve as parent generations, and offspring individuals are generated through crossover and variation; and recombining individuals in the convergence archive AC and the diversity archive AD to obtain a final solution FS. According to the method, a high-quality migration solution is reserved by selecting parameters, and the coevolution efficiency and the solving capability of the algorithm are dynamically improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent computing methods, and in particular relates to a dynamic multi-modal multi-objective optimization method based on a large language model and a dual archiving strategy. Background Art

[0002] We often encounter problems requiring multiple optimal solutions. These problems are often referred to as multimodal optimization problems. Research on this topic has been ongoing for quite some time. However, most current multimodal problems are limited to a single objective, or multimodal problems. These problems are typically addressed using various niche strategies, which divide the population into clusters and allow each cluster to evolve independently. This allows for finding multiple optimal solutions and prevents premature convergence. However, in reality, many problems involve multiple objectives, which are mutually constrained and conflicting. Optimizing one objective may degrade another. These problems are called multi-objective optimization problems. The key to solving these problems lies in identifying the optimal trade-off between these objectives and providing compromise solutions so that decision makers can select the optimal solution based on their specific circumstances. However, in earlier research, multi-objective optimization problems were often considered unimodal, meaning that a single optimal Pareto solution exists. However, in real life, multi-objective optimization problems also have multimodal situations, that is, there are multiple different global or local Pareto solution sets, and these optimal solutions may correspond to different objectives or objective combinations, thus leading to the emergence of multimodal multi-objective optimization problems. Summary of the Invention

[0003] The purpose of the present invention is to provide a dynamic multimodal multi-objective optimization method based on a large language model and a dual-archiving strategy, which solves the multimodal multi-objective optimization problem in complex environmental changes.

[0004] The technical solution adopted by the present invention is a dynamic multimodal multi-objective optimization method based on a large language model and a dual-archiving strategy, which is specifically implemented in the following steps:

[0005] Step 1: Initialize parameters;

[0006] Step 2: Detect independent convergence-related decision variable sets using decision variable analysis techniques;

[0007] Step 3: Use a set of reference vectors R that are well distributed in the target space to guide the update of the diversity profile;

[0008] Step 4: Update convergence archive A C ;

[0009] Step 5. Update Diversity Archive A D ;

[0010] Step 6: When the cycle conditions are met, implement the environmental change monitoring strategy;

[0011] Step 7: Based on the prediction mechanism strategy of the large language model, generate corresponding individuals and update the convergence archive and diversity archive respectively;

[0012] Step 8: From the first generation to the MaxGen generation, archive A from convergence C and Diversity Archive A D Select individuals as parents and generate offspring individuals through crossover and mutation;

[0013] Step 9: Archive the convergence A C and Diversity Archive A D The individuals in are reorganized to obtain the final solution FS.

[0014] The present invention is also characterized in that:

[0015] In step 1, the initialization parameters include population size N, convergence archive A C 、Diversity Archive A D 、Archive A C The scale N C 、Archive A D The scale N D .

[0016] Step 2 is implemented as follows:

[0017] Step 2.1. Randomly select nSel individuals from the current population and perturb each decision variable in each individual nPer times to generate a series of new sample solutions SP. Normalize the sample solutions SP in the target space and then fit the normalized sample solutions generated by each variable into a straight line L.

[0018] Step 2.2, calculate the relationship between each fitting line L and the hyperplane f1+…+f M = 1, where the normal represents the convergence direction, M represents the number of target problems, and each decision variable is associated with nSel angles. A smaller angle indicates that the variable contributes more to convergence, while a larger angle indicates that the variable contributes more to diversity.

[0019] Step 2.3: Use the k-mean clustering method to divide the decision variables into two groups according to the angle characteristics. The decision variable group with the smaller average angle is identified as the variable related to convergence and put into X IC The other set of variables identified as being related to diversity is placed in the X re In the collection;

[0020] Step 2.4, determine X IC Are the decision variables in X re If any decision variable in X interacts with each other, then change this decision variable from X IC Delete and add to X re middle.

[0021] Step 3 is implemented as follows:

[0022] Step 3.1. Use the UniformPoint function to generate a set of reference vectors R that are well distributed in the target space:

[0023] [R,N] = UniformPoint(N,M) (1)

[0024] Where N is the number of reference points; M is the number of objective functions that need to be considered simultaneously in the dynamic multimodal multi-objective optimization problem; the function returns a set of N uniformly distributed reference vectors R with M objectives, which is used to guide the update of the diversity archive.

[0025] Step 3.2: Generate the initial population P:

[0026] x=repmat(lu(1,∶),N,1)+rand(N,D).*(repmat(lu(2,∶)-lu(1,∶),N,1)) (2)

[0028] Where x represents the initialization population; lu represents the upper and lower bounds of the population; N represents the population size; and D represents the dimension.

[0029] Step 4 is implemented as follows:

[0030] Step 4.1: Merge the solution set and save the current population P and convergence archive A C The individuals in are merged together to form a candidate solution set Q, and the convergence index of each individual x in Q is calculated as follows:

[0031]

[0032] In the formula, the convergence index is obtained by averaging the function values ​​of individual x on all objective problems, where M represents the number of objective functions that need to be considered simultaneously in the dynamic multimodal multi-objective optimization problem; f m (x) represents the function value of individual x on the mth target problem;

[0033] Step 4.2: For every two individuals x in Q i ,x j , calculate their distance in the independently converged decision subspace:

[0034] d(x i ,x j )=Variable_Distance(x i ,x j ,X IC ) (4)

[0035] Step 4.3: Initially create an empty temporary archive Q′ and a new ranking counter nrank=1, and save the convergence archive A C Set to an empty set;

[0036] Step 4.4: Select the optimal solution, and select the solution x with the minimum convergence index from Q min , remove it from Q and add it to the convergence archive A C In

[15] , a convergence metric is used to provide strong selection pressure for the pareto front:

[0037] x min =argminfs(x) (5)

[0038] Step 4.5: For each solution x in Q i , if it is the same as the optimal solution x min The distance d(x min ,x i ) is smaller than the radius of the preset microhabitat σ niche , then delete it from Q and add it to the temporary archive Q′;

[0039] Step 4.6: Update the candidate solution set Q. If Q is empty, reload it as the solution set in Q′ and increase the ranking counter nrank.

[0040] Step 4.7. Repeat steps 4.4-4.6 until A C The number of solutions reaches the predetermined convergence archive size N C , returns the updated convergence archive A C .

[0041] Step 5 is implemented as follows:

[0042] Step 5.1: Archive the current diversity A D The merged set of the current population P is sorted non-dominated to form a non-dominated hierarchy (Fronts), and the solution set F1, F2, ..., F at each level is determined. i :

[0043] F=F1∪F2∪…F i =Nondominated_rank(A D ∪P) (6)

[0044] Step 5.2: Initialize the diversity archive, merge all layers except the last non-dominated layer, and put them into the diversity archive A. D In the last non-dominated layer F i Put it into the candidate solution set Q;

[0045] Step 5.3: Calculate each solution x in the candidate solution set Q i The angle between it and a set of reference vectors R that are well redistributed in the target space:

[0046] θ(x i ,r i )=Angle((x i ,r i ) (7)

[0047] Step 5.4: For every two individuals x in Q i ,x j , calculate their distance in the remaining relevant decision subspace, and assign each solution to the cluster C corresponding to the nearest reference vector. In each cluster, if there is a solution whose distance to the existing solution is less than the radius σ of the niche niche , then the solution is placed into the corresponding secondary cluster C′ to reduce the crowding of the solution space;

[0048] Step 5.5: If the number of solutions in all clusters exceeds the pre-set diversity archive size N D , then delete the solution farthest from the cluster center in the cluster with the largest number of solutions; if it is less than the size N of the diverse archive D , then the solution closest to the cluster center in the secondary cluster is added to the cluster with the least number of solutions;

[0049] Step 5.6: Merge all clusters to form and return the final diversity archive A D .

[0050] In step 6, the proposed method is used to detect whether the problem has changed, and the problem is detected from the diversity archive A. DA certain number of individuals are selected for re-evaluation. If the value of the objective function or constraint of any individual changes during the re-evaluation process, the problem is considered to have changed.

[0051] Step 7 is implemented as follows:

[0052] Step 7.1. Collect and organize all historical data related to the problem, including the population data of the convergence archive, as data for training the large model;

[0053] Step 7.2: Normalize the data to make it suitable for input into the LLM and determine the features that are important for predicting the population.

[0054] Step 7.3: Design clear instructions or prompts to guide the LLM in making the forecast. The instructions include the forecast objectives, environmental factors to be considered, and any specific requirements.

[0055] Step 7.4: In the diversity archive, select the best half of the individuals and keep them in the population associated with the new diversity archive. The remaining half of the individuals are reinitialized to increase the diversity of the population.

[0056] Step 8 is implemented as follows:

[0057] Step 8.1: Select the parent from the convergence archive and the diversity archive using the binary tournament algorithm, and set the probability P of selecting the parent from the convergence archive con , the probability of selecting a parent from the diversity archive is 1-P con , P con The larger the value, the more important the role of convergence archiving in population evolution;

[0058] Step 8.2: For each parent individual x i , using the mutation operator DE / current-to-rand / 1 based on the differential evolution algorithm to generate the offspring mutant individual v i :

[0059] v i =x i +rand*(x r1 -x i )+F*(x r2 -x r3 ) (8)

[0060] Where r1, r2, and r3 are three different integers randomly selected from [1, N], rand is a uniformly distributed random number selected from [0, 1], and F is a scaling factor;

[0061] Step 8.3, each mutant individual vi Individual x of the same parent i Cross-generate N experimental individuals u i,j That is, the offspring individuals:

[0062]

[0063] Among them, i∈{1,2,…,N}, j∈{1,2,…,D}, cross control parameter CR∈[0,1], rand is a random number uniformly distributed between [0,1], j rand is an integer randomly selected from [1,D];

[0064] Step 8.4: Update the convergence archive and diversity archive again using steps 4 and 5.

[0065] Step 9 is implemented as follows:

[0066] Step 9.1: If the set of independent convergence-related decision variables is empty, the final solution set is equal to the diverse archive. If it is not empty, perform the following steps:

[0067] Step 9.2: Select the solution x with the smallest convergence index in the convergence archive. min and use it as a reference point;

[0068] Step 9.3. Set a precision level ∈ peak , save the convergence to x min The convergence index difference is less than ∈ peak The solution is added to the peak solution set A peak , thereby finding the Pareto optimal solution or a solution very close to the Pareto optimal solution with different independent convergence-related decision variable values;

[0069] Step 9.4: For A peak Each peak solution x i and Diversity Archive A D The solution x in j , create a new solution x by l :

[0070] For the set of independent convergent decision variables X IC The variables k, x l The corresponding variable x in k,l Taken from the peak solution x k,i ;

[0071] For the set of independent convergence-related decision variables X IC The variable x l The corresponding variable x in k,l Solution x from the Diversity Archive k,j;

[0072] Step 9.5: Add the newly created solution to the final solution set FS and repeat step 9.4. The final solution set FS will then contain the Pareto optimal solution generated by the reorganization strategy.

[0073] The beneficial effects of the present invention are:

[0074] This dynamic multimodal, multi-objective optimization method, based on a large language model and a dual-archive strategy, analyzes the properties of decision variables and their relationships to guide population evolution. Using a framework of convergence archives and diversity archives, it addresses the challenge of maintaining good convergence and diversity of solutions in both the objective and decision spaces. When the environment changes, the large language model's prediction mechanism and historical information are leveraged to generate an initial population that approximates the new PS in the new environment. LLM, through contextual learning and complex reasoning capabilities, can rapidly process and analyze new environmental data, providing immediate information for predicting the formation of new populations and enabling more accurate predictions of environmental changes. A new initial population is generated from the diversity archive using a strategy of reinitializing half of the individuals to increase population diversity. During population evolution, individuals are selected from the convergence archive and the diversity archive as parents to generate offspring, and the convergence archive and diversity archive are updated separately. At the end of the evolution process, the solutions in the convergence and diversity archives are recombined to produce a large set of final solutions with good overall performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a flow chart of the dynamic multimodal multi-objective optimization method based on a large language model and a dual-archiving strategy of the present invention;

[0076] Figure 2 It is a detailed relationship diagram of population evolution, convergence archiving, diverse archiving and final solution set in the dynamic multimodal multi-objective optimization method based on the large language model and dual archiving strategy of the present invention;

[0077] Figure 3 It is a schematic diagram of the interaction of a large language model based on a large language model and a dynamic multimodal multi-objective optimization method of a dual-archiving strategy of the present invention. DETAILED DESCRIPTION

[0078] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0079] Example 1

[0080] The present invention is based on a dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy, such as Figure 1 As shown, please follow the steps below:

[0081] Step 1: Initialize parameters, including population size N and convergence archive A C 、Diversity Archive A D 、Archive A C The scale N C 、Archive A D The scale N D ;

[0082] Step 1.1, set the number of iterations t = 1, the population size N = 50 / 100, the population dimension D = 25 / 50, and the number of tasks K; generate K*N individuals as the initial population P k , k=1,2,…,K, and for P k Perform fitness evaluation; set the maximum function evaluation times FESMAX = 100000 / 200000, the initial evaluation times FES = 0; set the adaptive selection parameters Unify the multi-task search space and make dimension D=max j (D j ), j is the jth task;

[0083] Step 1.2, set CR i =0.5,F i =0.5,μ CR =0.5,μ F =0.5, Among them, each individual is assigned a crossover probability CR i and variable asynchronous length F i , μ CR ,μ F The crossover probability CR is i and variable asynchronous length F i The average value, S CR ,S F are the CRs corresponding to the successful individuals i ,F i The set of , B is the inferior solution set;

[0084] Step 2: Use decision variable analysis technology to detect independent convergence-related decision variable sets and put them into X IC In the set, other decision variables are placed in X re In the collection;

[0085] Step 3: Use a set of reference vectors R that are well distributed in the target space to guide the update of the diversity profile, and create an initial population P by randomly generating N individuals;

[0086] Step 4: Update convergence archive A C ;

[0087] Step 5. Update Diversity Archive AD ;

[0088] Step 6: When the cycle conditions are met, implement the environmental change monitoring strategy;

[0089] Step 7: If the environment changes, then based on the prediction mechanism strategy of the large language model, corresponding individuals are generated according to the independent convergence-related decision variables; based on the strategy of reinitializing half of the individuals, corresponding individuals are generated according to the other decision variables, and the convergence archive and diversity archive are updated respectively;

[0090] Step 8: From the first generation to the MaxGen generation, from the convergence archive A2 and the diversity archive A D Individuals are selected as parents, and offspring individuals are generated through crossover and mutation, and the convergence archive and diversity archive are updated again;

[0091] Step 9: Archive the convergence A C and Diversity Archive A D The individuals in the are reorganized to obtain the final solution FS under the current environment. The relationship between the convergence archive, diversity archive and the final solution is as follows: Figure 2 shown.

[0092] Example 2

[0093] The present invention is based on a dynamic multi-modal multi-objective optimization method using a large language model and a dual-archiving strategy, wherein step 2 is specifically implemented according to the following steps:

[0094] Step 2.1. Randomly select nSel individuals from the current population and perturb each decision variable in each individual nPer times to generate a series of new sample solutions SP. Normalize the sample solutions SP in the target space and then fit the normalized sample solutions generated by each variable into a straight line L.

[0095] Step 2.2, calculate the relationship between each fitting line L and the hyperplane f1+…+f M = 1. The normal line represents the convergence direction, M represents the number of target problems, and each decision variable is associated with nSel angles. A smaller angle indicates that the variable contributes more to convergence, while a larger angle indicates that the variable contributes more to diversity.

[0096] Step 2.3: Use the k-mean clustering method to divide the decision variables into two groups based on the angle characteristics. The decision variable group with the smaller average angle is identified as the variable related to convergence and is placed in X IC The other set of variables identified as being related to diversity is placed in the X re In the collection;

[0097] Step 2.4, determine X IC Are the decision variables in X re If any decision variable in X interacts with each other, then change this decision variable from X IC Delete and add to X re middle.

[0098] Step 3 is implemented as follows:

[0099] Step 3.1. Use the UniformPoint function to generate a set of reference vectors R that are well distributed in the target space:

[0100] [R,N] = UniformPoint(N,M) (1)

[0101] Where N is the number of reference points; M is the number of objective functions that need to be considered simultaneously in the dynamic multimodal multi-objective optimization problem. The function returns a set of N uniformly distributed reference vectors R with M objectives, which is used to guide the update of the diversity archive.

[0102] Step 3.2: Generate the initial population P:

[0103] x=repmat(lu(1,∶),N,1)+rand(N,D).*(repmat(lu(2,∶)-lu(1,∶),N,1)) (2)

[0105] Where x represents the initialization population; lu represents the upper and lower bounds of the population; N represents the population size; and D represents the dimension.

[0106] Step 4 is implemented as follows:

[0107] Step 4.1: Merge the solution set and save the current population P and convergence archive A C The individuals in are merged together to form a candidate solution set Q, and the convergence index of each individual x in Q is calculated as follows:

[0108]

[0109] In the formula, the convergence index is obtained by averaging the function values ​​of individual x on all objective problems, where M represents the number of objective functions that need to be considered simultaneously in the dynamic multimodal multi-objective optimization problem; f m (x) represents the function value of individual x on the mth target problem;

[0110] Step 4.2: For every two individuals x in Q i ,x j , calculate their distance in the independently converged decision subspace:

[0111] d(x i ,x j )=Variable_Distance(x i ,x j ,X IC ) (4) Any method for calculating distance can be used, here Chebychev distance is used;

[0112] Step 4.3: Initially create an empty temporary archive Q′ and a new ranking counter nrank=1, and save the convergence archive A C Set to an empty set;

[0113] Step 4.4: Select the optimal solution, and select the solution x with the minimum convergence index from Q min , remove it from Q and add it to the convergence archive A C In this step, the goal is to use the convergence indicator to provide strong selection pressure for the pareto front;

[0114] x min =argminfs(x) (5)

[0115] Step 4.5: For each solution x in Q i , if it is the same as the optimal solution x min The distance d(x min ,x i ) is smaller than the radius of the preset microhabitat σ niche , then delete it from Q and add it to the temporary archive Q′. This step aims to improve the diversity in the subspace of independent convergence-related decision variables;

[0116] Step 4.6: Update the candidate solution set Q. If Q is empty, reload it as the solution set in Q′ and increase the ranking counter nrank.

[0117] Step 4.7. Repeat steps 4.4-4.6 until A C The number of solutions reaches the predetermined convergence archive size N C , returns the updated convergence archive A C This allows the algorithm to maintain convergence while also enabling the convergence archive to find different Pareto optimal solutions (or solutions very close to them) in independent subspaces of convergence-related decision variables. These solutions are defined as peak solutions and are used to restructure the strategy.

[0118] Step 5 is implemented as follows:

[0119] Step 5.1: Archive the current diversity AD The merged set of the current population P is sorted non-dominated to form a non-dominated hierarchy (Fronts), and the solution set F1, F2, ..., F at each level is determined. i :

[0120] F=F1∪F2∪…F i =Nondominated_rank(A D ∪P) (6)

[0121] Step 5.2: Initialize the diversity archive, merge all layers except the last non-dominated layer, and put them into the diversity archive A. D In the last non-dominated layer F i Put it into the candidate solution set Q;

[0122] Step 5.3: Calculate each solution x in the candidate solution set Q i The angle between it and a set of reference vectors R that are well redistributed in the target space:

[0123] θ(x i ,r i )=Angle((x i ,r i ) (7)

[0124] Step 5.4: For every two individuals x in Q i ,x j , calculate their distance in the remaining relevant decision subspace, and assign each solution to the cluster C corresponding to the nearest reference vector. In each cluster, if there is a solution whose distance to the existing solution is less than the radius σ of the niche niche , then the liberation is placed into the corresponding secondary cluster C′ to reduce the crowding of the solution in the space and improve the diversity of the remaining decision subspace;

[0125] Step 5.5: If the number of solutions in all clusters exceeds the pre-set diversity archive size N D , then delete the solution farthest from the cluster center in the cluster with the largest number of solutions (i.e. the solution with the largest angle); if it is less than the size N of the diverse archive D , then the solution closest to the cluster center in the secondary cluster (i.e., the solution with the smallest angle) is added to the cluster with the least number of solutions; this step is intended to further improve the diversity performance of the target space;

[0126] Step 5.6: Merge all clusters to form and return the final diversity archive A D .

[0127] Example 3

[0128] The present invention is based on a dynamic multimodal multi-objective optimization method based on a large language model and a dual archive strategy, wherein in step 6, the method proposed by Deb et al. is used to detect whether the problem has changed. D A certain number of individuals are selected for re-evaluation. If the value of the objective function or constraint of any individual changes during the re-evaluation process, the problem is considered to have changed.

[0129] Since the goal of solving multimodal problems is to find multiple Pareto optimal solution sets for such problems, and a large number of such Pareto optimal solution sets are retained in the convergence preservation set, the objective function values ​​of these solution sets may be repeated. Therefore, in order to avoid repeated calculations and affect the accuracy of environmental change detection, only the individuals from the diversity archive A are selected. D Select from .

[0130] Step 7 is implemented as follows:

[0131] Step 7.1, collect and organize all historical data related to the problem, including the population data of the convergence archive, that is, the individual decision value, fitness value and environmental parameters as data training large model, LLM interactive diagram is as follows Figure 3 As shown;

[0132] Step 7.2: Normalize the data to make it suitable for input into the LLM. At the same time, determine which features are important for predicting the population, which may include population diversity, fitness distribution, and historical optimal solutions.

[0133] Step 7.3: Design clear instructions or prompts to guide the LLM in making predictions. The instructions should include the prediction goal, the environmental factors to be considered, and any specific requirements. The following is the interaction strategy between the population and the large model:

[0134]

[0135] Step 7.4: In the diversity archive, select the best half of the individuals and keep them in the population associated with the new diversity archive. The remaining half of the individuals are reinitialized to increase the diversity of the population.

[0136] Step 8 is implemented as follows:

[0137] Step 8.1: Select the parent from the convergence archive and the diversity archive using the binary tournament algorithm, and set the probability P of selecting the parent from the convergence archive com , the probability of selecting a parent from the diversity archive is 1-P con , P con The larger the value, the more important the role of convergence archiving in population evolution;

[0138] Step 8.2: For each parent individual x i , using the mutation operator DE / current-to-rand / 1 based on the differential evolution algorithm (DE) to generate the offspring mutant individual v i :

[0139] v i =x i +rand*(x r1 -x i )+F*(x r2 -x r3 ) (8)

[0140] Where r1, r2, and r3 are three different integers randomly selected from [1, N], rand is a uniformly distributed random number selected from [0, 1], and F is a scaling factor;

[0141] Step 8.3, each mutant individual v i Individual x of the same parent i Cross-generate N experimental individuals u i,j That is, the offspring individuals:

[0142]

[0143] Among them, i∈{1,2,…,N}, j∈{1,2,…,D}, cross control parameter CR∈[0,1], rand is a random number uniformly distributed between [0,1], j rand is an integer randomly selected from [1,D];

[0144] Step 8.4: Update the convergence archive and diversity archive again using steps 4 and 5.

[0145] Step 9 is implemented as follows:

[0146] Step 9.1: If the set of independent convergence-related decision variables is empty, the final solution set is equal to the diverse archive. If it is not empty, perform the following steps:

[0147] Step 9.2: Select the solution x with the smallest convergence index in the convergence archive. min and use it as a reference point;

[0148] Step 9.3. Set a precision level ∈ peak , save the convergence to x min The convergence index difference is less than ∈ peak The solution is added to the peak solution set A peak, thereby finding the Pareto optimal solution or a solution very close to the Pareto optimal solution with different independent convergence-related decision variable values;

[0149] Step 9.4: For A peak Each peak solution x i and Diversity Archive A D The solution x in j , create a new solution x by l :

[0150] For the set of independent convergent decision variables X IC The variables k, x l The corresponding variable x in k,l Taken from the peak solution x k,i ;

[0151] For the set of independent convergence-related decision variables X IC The variable x l The corresponding variable x in k,l Solution x from the Diversity Archive k,j ;

[0152] Step 9.5: Add the newly created solutions to the final solution set FS and repeat step 9.4. The final solution set FS then contains the Pareto optimal solutions generated by the reorganization strategy. These solutions have good convergence and diversity in both the objective space and the decision space.

[0153] Through the above approach, the present invention's two-stage adaptive knowledge transfer learning evolutionary multi-task optimization algorithm, based on the concept of adaptively controlling population evolution, adopts a dual strategy combining affine transformation and local outlier detection model to construct a two-stage knowledge transfer mechanism. This adaptively solves MTO problems of different properties under the black-box problem, while simultaneously utilizing adaptive parameter selection to retain high-quality transfer solutions, dynamically improving the algorithm's co-evolutionary efficiency and solution-solving capabilities. In summary, the present invention has universal applicability and superior performance.

Claims

1. A dynamic multi-modal multi-objective optimization method based on a large language model and dual-archiving strategy, characterized in that: Follow the steps below to implement it: Step 1, initialization parameters; Step 2: Detect independent convergence-related decision variable sets using decision variable analysis techniques; Step 3: Use a set of reference vectors R that are well distributed in the target space to guide the update of the diversity profile; Step 4: Update convergence archive A C ; Step 5. Update Diversity Archive A D ; Step 6: When the cycle conditions are met, implement the environmental change monitoring strategy; Step 7: Based on the prediction mechanism strategy of the large language model, generate corresponding individuals and update the convergence archive and the diversity archive respectively; Step 8. From the first generation to the MaxGen generation, from the convergence archive A C and Diversity Archive A D Select individuals as parents and generate offspring individuals through crossover and mutation; Step 9: Archive Convergence A C and Diversity Archive A D The individuals in are reorganized to obtain the final solution FS.

2. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: In step 1, the initialization parameters include the population size N, the convergence archive A C 、Diversity Archive A D 、Archive A C The size N C 、Archive A D The size N D .

3. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, randomly select nSel individuals from the current population, and perform nPer perturbations on each decision variable in each individual to generate a series of new sample solutions SP, normalize the sample solutions SP in the target space, and then fit the normalized sample solution sets generated by each variable into a straight line L; Step 2.2, calculate the relationship between each fitting line L and the hyperplane f1+…+f M = 1, where the normal represents the convergence direction, M represents the number of target problems, and each decision variable is associated with nSel angles. A smaller angle means that the variable contributes more to convergence, while a larger angle means that the variable contributes more to diversity. Step 2.3: Use the k-mean clustering method to divide the decision variables into two groups according to the angle characteristics. The decision variable group with a smaller average angle is identified as a variable related to convergence and put into X IC The other group of variables identified as being related to diversity is placed in the X re In the collection; Step 2.4: Determine X IC Are the decision variables in X re If any decision variable in X interacts with each other, then change this decision variable from X IC Delete and add to X re middle.

4. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1, use the UniformPoint function to generate a set of reference vectors R that are well distributed in the target space: [R,N]=UniformPoint(N,M) (1) Where N represents the number of reference points; M represents the number of objective functions that need to be considered simultaneously in the dynamic multimodal multi-objective optimization problem; the set of approximately N uniformly distributed reference vectors R with M objectives returned by the function is used to guide the update of the diversity archive; Step 3.2, generate the initial population P: x=repmat(lu(1,∶),N,1)+rand(N,D).*(repmat(lu(2,∶)-lu(1,∶),N,1)) (2) Where x represents the initialization population; lu represents the upper and lower bounds of the population; N represents the population size; and D represents the dimension.

5. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: The step 4 is specifically implemented according to the following steps: Step 4.1: Merge the solution set and save the current population P and convergence archive A C The individuals in are merged together to form a candidate solution set Q, and the convergence index of each individual x in Q is calculated as shown in the following formula: In the formula, the convergence index is obtained by averaging the function values ​​of individual x on all objective problems, where M represents the number of objective functions that need to be considered simultaneously in the dynamic multi-modal multi-objective optimization problem; f m (x) represents the function value of individual x on the mth target problem; Step 4.2: For every two individuals x in Q i ,x j , calculate their distance in the independently convergent decision subspace: d(x i ,x j )=Variable_Distance(x i ,x j ,X IC ) (4) Step 4.3: Initially create an empty temporary archive Q′ and a new ranking counter nrank=1, and save the convergence archive A C Set to an empty set; Step 4.4: Select the optimal solution, and select the solution x with the minimum convergence index from Q min , remove it from Q and add it to the convergence archive A C In , a convergence metric is used to provide strong selection pressure for the pareto front: x min =argminfs(x) (5) Step 4.5: For each solution x in Q i , if it is the same as the optimal solution x min The distance d(x) in the subspace of independent convergent decision variables min ,x i ) is smaller than the radius σ of the preset microhabitat niche , then delete it from Q and add it to the temporary archive Q′; Step 4.6: Update the candidate solution set Q. If Q is empty, reload it as the solution set in Q′ and increase the ranking counter nrank; Step 4.

7. Repeat steps 4.4-4.6 until A C The number of solutions reaches the predetermined convergence archive size N C , returns the updated convergence archive A C .

6. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: The step 5 is specifically implemented according to the following steps: Step 5.1: Archive the current diversity D The merged set of the current population P is sorted non-dominatedly to form a non-dominated hierarchy (Fronts), and the solution set F1, F2, ..., F at each level is determined i : F=F1∪F2∪…F i =Nondominated_rank(A D ∪P) (6) Step 5.2: Initialize the diversity archive, merge all levels except the last non-dominated level, and put them into diversity archive A. D In the last non-dominated layer F i Put it into the candidate solution set Q; Step 5.3: Calculate each solution x in the candidate solution set Q i The angle between it and a set of reference vectors R that are well redistributed in the target space: θ(x i ,r i )=Angle((x i ,r i ) (7) Step 5.4: For every two individuals x in Q i ,x j , calculate their distance in the remaining relevant decision subspace, and assign each solution to the cluster C corresponding to the nearest reference vector. In each cluster, if there is a solution whose distance to the existing solution is less than the radius σ of the niche niche , then the solution is liberated into the corresponding secondary cluster C′ to reduce the crowding of solutions in the space; Step 5.5: If the number of solutions in all clusters exceeds the pre-set diversity archive size N D , then delete the solution farthest from the cluster center in the cluster with the largest number of solutions; If less than the size N of the multi-archive D , then the solution closest to the cluster center in the secondary cluster is added to the cluster with the least number of solutions; Step 5.6: Merge all clusters to form and return the final diversity archive A D .

7. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: In step 6, the proposed method is used to detect whether the problem has changed, and the diversity archive A D A certain number of individuals are selected for re-evaluation. If the value of the objective function or constraint of any individual changes during the re-evaluation process, the problem is considered to have changed.

8. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1 is characterized in that: The step 7 is specifically implemented according to the following steps: Step 7.1, collect and organize all historical data related to the problem, including the population data of convergence archive, as data to train the large model; Step 7.2: Normalize the data to make it suitable for input into the LLM and determine the features that are important for predicting the population. Step 7.3: Design clear instructions or prompts to guide the LLM in making the forecast, including the forecast objectives, environmental factors to be considered, and any specific requirements; Step 7.4: In the diversity archive, select the best half of the individuals and keep them in the population associated with the new diversity archive. The remaining half of the individuals are reinitialized to increase the diversity of the population.

9. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1, characterized in that: The step 8 is specifically implemented according to the following steps: Step 8.1: Select the parent from the convergence archive and the diversity archive using the binary tournament algorithm, and set the probability P of selecting the parent from the convergence archive con , the probability of selecting a parent from the diversity archive is 1-P con , P con The larger the value, the more important the role of convergence archiving in population evolution; Step 8.2: For each parent individual x i , using the mutation operator DE / current-to-rand / 1 based on the differential evolution algorithm to generate the offspring mutant individual v i : v i =x i +rand*(x r1 -x i )+F*(x r2 -x r3 ) (8) Where r1, r2, and r3 are three different integers randomly selected from [1, N], rand is a uniformly distributed random number selected from [0, 1], and F is a scaling factor; Step 8.3: Each mutant individual v i Individual x of the same parent i Cross-generate N experimental individuals u i,j That is, the offspring individuals: Where i∈{1,2,…,N}, j∈{1,2,…,D}, cross control parameter CR∈[0,1], rand is a random number uniformly distributed between [0,1], j rand is an integer randomly selected from [1,D]; Step 8.4: Use steps 4 and 5 to update the convergence archive and diversity archive again.

10. The dynamic multi-modal multi-objective optimization method based on a large language model and a dual-archiving strategy according to claim 1, characterized in that: The step 9 is specifically implemented according to the following steps: Step 9.1: If the set of independent convergence-related decision variables is empty, the final solution set is equal to the diverse archive. If it is not empty, perform the following steps; Step 9.2: Select the solution x with the smallest convergence index in the convergence archive min , and use it as a reference point; Step 9.

3. Set a precision level∈ peak , save the convergence to x min The convergence index difference is less than ∈ peak The solution is added to the peak solution set A peak In order to find the Pareto optimal solution or a solution very close to the Pareto optimal solution with different independent convergence related decision variable values; Step 9.4: For A peak Each peak solution x i and Diversity Archive A D The solution x in j , create a new solution x by l : For the set of independent convergent decision variables X IC The variables k, x l The corresponding variable x in k,l Taken from the peak solution x k,i ; For the set of independent convergence-related decision variables X IC The variable, x l The corresponding variable x in k,l Solution x taken from the Diversity Archive k,j ; Step 9.5: Add the newly created solution to the final solution set FS, and repeat step 9.

4. Then the final solution set FS contains the Pareto optimal solution generated by the reorganization strategy.

Citation Information

Cited By

  • Indoor access point optimized placement method and system based on multi-modal large language model

    CN120857136A

  • Indoor access point optimized placement method and system based on multi-modal large language model

    CN120857136B

  • Double-file super-multi-objective optimization method and system based on enhanced knowledge sharing

    CN121809101A