CN117787331A - Agent-assisted co-evolution method based on improved grouping strategy - Google Patents
Agent-assisted co-evolution method based on improved grouping strategy Download PDFInfo
- Publication number
- CN117787331A CN117787331A CN202311842226.1A CN202311842226A CN117787331A CN 117787331 A CN117787331 A CN 117787331A CN 202311842226 A CN202311842226 A CN 202311842226A CN 117787331 A CN117787331 A CN 117787331A
- Authority
- CN
- China
- Prior art keywords
- sub
- population
- individual
- value
- optimal solution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 239000002245 particle Substances 0.000 claims abstract description 29
- 230000008569 process Effects 0.000 claims abstract description 22
- 238000003860 storage Methods 0.000 claims abstract description 11
- 238000005516 engineering process Methods 0.000 claims abstract description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 67
- 230000006870 function Effects 0.000 claims description 43
- 238000011156 evaluation Methods 0.000 claims description 38
- 239000003795 chemical substances by application Substances 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 19
- 239000013598 vector Substances 0.000 claims description 18
- 238000004590 computer program Methods 0.000 claims description 16
- 230000002860 competitive effect Effects 0.000 claims description 13
- 238000009826 distribution Methods 0.000 claims description 6
- 230000001174 ascending effect Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 abstract description 37
- 238000002474 experimental method Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 8
- 238000012545 processing Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 5
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000000354 decomposition reaction Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000000585 Mann–Whitney U test Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000000758 substrate Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to the technical field of large-scale expensive optimization, in particular to an agent-assisted co-evolution method based on an improved grouping strategy, which comprises the following specific steps: firstly, on the basis of the existing random grouping technology, the effectiveness of grouping is improved by adding a shared variable into each sub-problem; secondly, the competition particle swarm CSO method is improved, and EI values are used as standards for judging winners and losers in the population updating process, so that the searching process is diversified, and different areas in the solution space can be explored; finally, a two-stage filling criterion is provided, and by selecting different candidate solutions in different stages, not only can a wide search space be explored in the overall optimization process, but also convergence can be accelerated by local search when the optimal solution is approached. Furthermore, the invention also provides a computer readable storage medium and a computer device.
Description
Technical Field
The invention relates to the technical field of large-scale expensive optimization, in particular to an agent-assisted co-evolution method based on an improved grouping strategy.
Background
In practice, many engineering problems involve hundreds or even thousands of decision variables, often referred to as this type of problem being a high-dimensional or large-scale optimization problem. Taking the minimization problem as an example, the single objective optimization problem can be expressed by the following formula:
min F(x),subject to x＝(x 1 ,x 2 ,...,x n )∈Ω
in the method, in the process of the invention,decision space representing a question->a i And b i Respectively representing the upper and lower bounds of the i-th dimension, x represents an n-dimensional decision vector, and F (x) represents the objective function. When n is larger than or equal to 1000, the problem at this time is called a large-scale single-target optimization problem.
As the number of decision variables increases, the search space increases, the number of locally optimal solutions increases exponentially, and the complexity of the above problem increases, so it is difficult to find a good solution. To solve such problems, meta heuristic algorithms are widely used, such as evolutionary algorithm EA and swarm intelligence algorithm SIA, because they are simple to implement and have no differential requirements on the objective function. Currently, the problem of large-scale optimization can be largely divided into two types: based on a decomposition method, the method utilizes the concept of 'divide-and-conquer', divides a large-scale optimization problem into a plurality of sub-problems, each sub-problem only comprises a small amount of decision variables, divides the sub-problem into a plurality of sub-problems with a small number of decision variables, and searches the optimal solution of the whole problem by cooperatively solving the optimal solution of all the sub-problems; non-decomposition-based methods, which typically use new evolutionary techniques to address the problem of large-scale optimization.
Both types of algorithms exhibit good performance in solving the large-scale optimization problem. However, some large-scale optimization problems involve computationally expensive objective evaluation procedures. Taking complex aerodynamic design optimization as an example, a large number of simulations and wind tunnel tests are required to evaluate potential design configurations each time. It is clear that many existing large-scale optimization algorithms are limited in solving this computationally expensive problem, because they require a large number of real objective function evaluations to search for globally optimal or near optimal solutions, which consume a large amount of computational cost each time. Proxy models, such as gaussian process models, radial basis function models, support vector machine models, artificial neural network models, polynomial regression models, etc., have been widely used to solve the computationally expensive optimization problem. However, as the dimension of the optimization problem increases, the proxy model requires more sample training models with real objective function evaluations, which is impractical for computationally expensive problems. Thus, proxy models are often used in collaborative co-evolution methods to overcome the limitations imposed by high dimensionality and sample scarcity. While proxy models exhibit good performance in assisting large-scale optimization algorithms in solving expensive large-scale optimization problems, there are still some challenges that need to be addressed.
First, random grouping techniques, while having low computational costs, do not guarantee that related variables are grouped into the same sub-problem, potentially resulting in searches that deviate from the correct direction. Second, the choice of filling criteria is very important for expensive large-scale optimization problems, which can find a good solution within a limited computational budget. Since large-scale optimization problems typically have more locally converging regions, how to avoid these regions when selecting solutions is also a matter of concern.
Disclosure of Invention
Aiming at the defects existing in the prior art, the invention provides a proxy assisted co-evolution method based on an improved grouping strategy, provides a new technical scheme for the problem of expensive large-scale optimization, and balances the convergence and diversity of populations while improving the grouping effectiveness.
Aiming at the defect that the current random grouping technology cannot ensure that related variables are grouped into the same sub-problem, the shared variable is selected on the basis of the current random grouping technology so as to improve the probability of dividing the related variables into the same sub-problem, so that a population finds a better solution; meanwhile, the method for preparing the competitive particle swarm CSO is improved, and a group of samples with good fitness values are used for replacing the global optimal solution, so that the population can be quickly converged, meanwhile, the diversity of the population is ensured, and the population is not easy to fall into local optimal; in addition, a two-stage filling criterion is provided, the population P is used as a candidate solution in the early stage of the experiment, the corresponding dimension of the current global optimal solution is replaced by the sub solution with the smallest fitness value on each updated sub problem in the later stage of the experiment, the obtained k synthesized solutions are used as candidate solutions, and different candidate solutions are selected in different stages, so that a wide search space can be explored in the whole optimization process, convergence can be accelerated through local search when the optimal solution is approached, and meanwhile, the global search capability of the algorithm can be prevented from being trapped in local optimal and enhanced.
The invention provides a proxy assisted co-evolution method based on an improved grouping strategy, which comprises the following specific steps:
s1, parameter setting: setting the population scale as N, the number of individuals in an initial database as DBsize, the dimension of a decision space as D and the maximum number of true evaluation as FE max The current times are truly evaluated as FE, the number of sub-questions or sub-populations is K, and the maximum dimension of the sub-questions is D sub The maximum number of shared variables is Ds;
s2, randomly generating a database by using Latin hypercube, carrying out real evaluation on all individuals in the database, storing the evaluation result into the database, and defining the individual with the smallest fitness value in the database as a global optimal solution gbest; in addition, a Latin hypercube is used for random initialization to generate a population P with a scale of N, and the speed of all individuals in the initial population is set to be 0;
s3, carrying out ascending order sorting on individuals in the database according to the real fitness value, and selecting individuals with the ranking names of the first M to form a better solution set gbests according to the sorting result;
s4, dividing the original large-scale problem into K sub-problems by selecting a shared variable in the prior random grouping technology;
s5, training a corresponding trained agent model for each sub-problem, and calculating the average value of fitness values of each individual on all the sub-problems
S6, sequentially updating each sub-problem by applying an improved competitive particle swarm CSO algorithm until all the sub-problems are updated, and outputting a new population P;
s7, determining candidate solutions by adopting a double-stage strategy, calculating the fitness value of each individual in the candidate solutions by utilizing the K trained agent models obtained in the step S5, calculating the evaluation index SD value of each individual according to the fitness value, and taking the individual with the minimum SD value as a new global optimal solution gbest;
s8, carrying out real evaluation on the new global optimal solution gbest, comparing the current global optimal solution gbest with the new global optimal solution gbest, and updating the global optimal solution gbest and the database according to a comparison result;
s9, judging whether a termination condition is met, if yes, evaluating termination, and outputting a global optimal solution gbest; otherwise, the steps S3-S9 are continued.
Preferably, the step S4 specifically includes the steps of:
s41, in the range [1, D s ]Internally generating a random number d s Randomly selecting d among all decision variables of the current decision space s Individual variables as shared variables X sv Put into each sub-problem;
s42, in the range [1, D sub -d s ]Internally generating a random number d r Removing shared variable X from the current decision space sv The remaining variable X other D is selected randomly r Putting the variable into the first sub-problem;
s43, repeating the step S42 until d which is randomly selected is put into the K sub-questions r Stopping after the residual variables;
s44, outputting the K sub-questions after grouping.
Preferably, the step S5 specifically includes:
s51, generating a random number n as an interval, and selecting a plurality of individuals with equal intervals as a training set from the individuals with the smallest fitness value according to the sequencing result in the step S3;
s52, respectively extracting a corresponding sub-training set for each sub-problem from the training set according to the dimension corresponding to each sub-problem;
s53, training a proxy model for each corresponding sub-problem by using each sub-training set, so that each sub-problem respectively obtains a corresponding trained proxy model;
s54, extracting respective corresponding sub-populations from the current population P according to the dimension corresponding to each sub-problem, and calculating the fitness value of each individual in each sub-population by using K trained agent models, wherein each individual obtains K fitness values;
s55, solving the average value of the fitness values of each individual on all the sub-problems according to the K fitness values obtained by the individual
Preferably, the specific update steps of each sub-problem are as follows:
Step1, extracting a corresponding kth sub-population from the current population P according to the dimension corresponding to the kth sub-problem;
step2, evaluating the fitness value of the individual in the current kth sub-population by using a trained proxy model corresponding to the kth sub-problemBased on the fitness value of the individuals in the k sub-population evaluated +.>Calculating the expected lifting value EI of the individuals in the kth sub-population, wherein the expected lifting value EI is calculated according to the following formula:
wherein i represents the ith individual, k represents the kth sub-population, C i，k Representing a minimum fitness value f in a database min Estimated fitness value with the ith individual in the kth sub-populationDifference between S i，k Mean value representing fitness value of the ith individual over all subproblems +.>Estimated fitness value +.>The absolute value of the difference value between the two groups is K, the number of the sub-groups, and phi respectively represent the standard distribution and the probability density distribution of the normal function;
step3, randomly selecting two individuals which do not participate in updating from the kth sub-population, comparing the EI values of the two individuals, taking the individual with large EI value as a winner and the individual with small EI value as a loser, and updating the speed and the position of the individuals on the current sub-population by applying an improved competitive particle swarm CSO algorithm based on the EI value comparison result of the two individuals, wherein the speed updating formula of the loser and the winner is as follows:
The loser and winning position update formulas are as follows:
X l，k (t+1)＝X l，k (t)+V l，k (t)+R 7 (k，t)ΔV l，k (t+1)
ΔV l，k (t+1)＝V l，k (t+1)-V l，k (t)
X w，k (t+1)＝X w，k (t)+V w，k (t)+R 8 (k，t)ΔV w，k (t+1)
ΔV w，k (t+1)＝V w，k (t+1)-V w，k (t)
wherein l represents a loser, w represents a winner, X l，k (t) and X w，k (t) position vectors representing the failure and winning of the kth sub-population at the t th generation, X l，k (t+1) and ζ w，k (t+1) represents the position vectors of the loser and winner of the kth sub-population at the t+1 th generation, V l，k (t) and V w，k (t) velocity vectors representing the failure and winning of the kth sub-population at the t th generation, V l，k (t+1) and V w，k (t+1) represents the velocity vector of the loser and winner, deltaV, of the kth sub-population at the t+1 th generation, respectively l，k (t+1) and DeltaV w，k (t+1) represents the variation of the speeds of the losing particles and the winning particles of the kth sub-population at the t+1 generation, respectively; r is R 1 (k，t)、R 2 (k，t)、R 3 (k，t)、R 4 (k，t)、R 5 (k，t)、R 6 (k，t)、R 7 (k, t) and R 8 (k, t) is the random coefficient of the kth sub-population at the t generation, at [0,1]Randomly generating in a range; and->Randomly selecting the three preferred solutions at the t generation from the preferred solution set gbests at the t generation;Control parameters of loser for control parameters +.>Control parameter of winner +.0.5>0.1;
step4, repeatedly executing the Step3 until all individuals in the current sub population participate in comparison and updating and then stop;
step5, judging whether the maximum iteration number of the current sub-population update is reached, if so, stopping the current sub-population update to obtain a new kth sub-population; otherwise, continuing to execute the steps Step2-Step5;
Step6, replacing the corresponding dimension on the population P by the new kth sub-population to form a new population P.
Preferably, the step S7 specifically includes:
s71, when FE/FE max When < b, taking the new population P as a candidate solution; otherwise, evaluating the fitness value of the individual in each sub-problem updated in the evaluation step S6 by using the trained agent model corresponding to each sub-problem, further selecting the individual with the smallest fitness value in each sub-problem, taking out the corresponding dimension of each sub-problem as the sub-solution of the sub-problem, and then respectively replacing the corresponding dimension of the global optimal solution gbest by the sub-solutions of the sub-problems to obtain k synthetic solutions as candidate solutions;
s72, evaluating the fitness value of each individual in the candidate solution by sequentially using the k trained agent models obtained in the step S5;
s73, evaluating an evaluation index SD value of each individual in the candidate solutions based on an average deviation strategy, and taking the individual with the minimum SD value as a new global optimal solution gbest, wherein the evaluation index SD value has the following calculation formula;
where K represents the number of trained proxy models,representing the estimated fitness value of the kth trained surrogate model for individual i, +.>Mean value representing estimated fitness values of all individuals in K trained surrogate models, +. >Andmaximum estimated fitness value and minimum estimated fitness value of all individuals in K trained agent models are divided by +.>Is a normalization process.
Preferably, the step S8 specifically includes: firstly, carrying out real evaluation on a new global optimal solution gbest, and storing the new global optimal solution gbest and a fitness value thereof in a database for updating the database; and secondly, comparing the fitness value of the current global optimal solution gbest with the fitness value of the new global optimal solution gbest, if the fitness value of the new global optimal solution gbest is better than the fitness value of the current global optimal solution gbest, updating the current global optimal solution gbest, otherwise, not updating the current most global optimal solution gbest.
Preferably, when Step3 is performed and the number of individuals in any one sub-population is singular, the last individual does not perform Step3, i.e. the individual does not update the speed and position, and participates in the next execution Step directly.
Preferably, the proxy model used in the step S5 is a radial basis function RBF model, and the termination condition in the step S9 is that the execution is terminated when the maximum number of evaluations is performed.
The present invention provides a computer readable storage medium having a computer program stored thereon, which when executed by a processor, enables a proxy assisted co-evolution method based on an improved grouping strategy.
The invention also provides a computer device comprising a memory, a processor and a computer program stored on the memory and operable on the processor, which when executed by the processor, implements a proxy assisted co-evolution method based on an improved grouping strategy.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention selects the shared variable based on the prior random grouping technology, and adds the shared variable for each sub-problem to improve the probability of dividing the related variable into the same sub-problem and the effectiveness of grouping.
2. In the invention, in the process of updating the population by using the competitive particle swarm CSO method, the EI value is used for replacing the fitness value to judge the winner and the loser, and meanwhile, the potential expansibility of the convergence degree of the particles is considered.
3. The invention also provides a two-stage filling criterion, wherein the population P is used as a candidate solution in the early stage of the experiment, the updated sub solution with the smallest fitness value on each sub problem in the later stage of the experiment is used for respectively replacing the corresponding dimension of the current global optimal solution, the obtained k synthetic solutions are used as candidate solutions, and different candidate solutions are selected in different stages, so that not only can a wide search space be explored in the whole optimization process, but also convergence can be accelerated through local search when the optimal solution is approached, and meanwhile, the global search capability of the algorithm can be prevented from being trapped in local optimal and enhanced.
Drawings
In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a proxy assisted co-evolution method based on an improved grouping strategy in an embodiment of the invention;
FIG. 2 is a schematic diagram of one embodiment of grouping decision variables in an embodiment of the invention;
FIG. 3 is a schematic diagram of one embodiment of an updated population P in an embodiment of the invention;
FIG. 4 is a flowchart of a specific update procedure for each sub-problem in the embodiment of the present invention;
fig. 5 is a flowchart of obtaining a new global optimal solution using a two-stage strategy in an embodiment of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
In the description of the present invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. The specific meaning of the above terms in the present invention will be understood in specific cases by those of ordinary skill in the art.
When solving a large-scale optimization problem by using a decomposition-based method, the complete problem needs to be divided into a plurality of sub-problems with smaller dimensions, and random grouping is a relatively common method for dividing decision variable groups. However, in a random grouping strategy, decision variables are typically divided randomly, without regard to the relationship between the decision variables. This approach, while relatively simple and easy to implement and computationally inexpensive, does not guarantee that the relevant variables are grouped into the same sub-problem, possibly resulting in a search that deviates from the correct direction. Furthermore, since large-scale optimization problems typically have more locally converging regions, how to avoid these regions in the selection of the filling criteria is also a matter of concern.
The invention provides a proxy assisted co-evolution method based on an improved grouping strategy, which aims at improving the problem that the current random grouping technology does not consider the relation among decision variables so as to not ensure that related variables are grouped into the same sub-problem. Firstly, the invention selects the shared variable based on the prior random grouping technology, and adds the shared variable for each sub-problem to improve the probability of dividing the related variable into the same sub-problem and the grouping effectiveness; secondly, in the population updating process by using the competitive particle swarm CSO method, the EI value is used for replacing the fitness value to judge winners and losers, meanwhile, the potential expansibility of the convergence degree of particles is considered, and by improving the competitive particle swarm CSO method, the searching process is diversified, and the searching process is also beneficial to exploring different areas in the solution space; in addition, a two-stage filling criterion is provided, the population P is used as a candidate solution in the early stage of the experiment, the corresponding dimension of the current global optimal solution is replaced by the sub solution with the smallest fitness value on each updated sub problem in the later stage of the experiment, the obtained k synthetic solutions are used as candidate solutions, and different candidate solutions are selected in different stages, so that a wide search space can be explored in the whole optimization process, convergence can be accelerated through local search when the optimal solution is approached, and meanwhile, the global search capability of the algorithm can be prevented from being trapped in local optimal and enhanced; and finally, selecting a new global optimal solution from the determined candidate solutions, and updating the global optimal solution and the database by comparing the current global optimal solution with the real evaluation result of the new global optimal solution.
Referring to fig. 1, an agent assisted co-evolution algorithm based on an improved grouping strategy, english is abbreviated as SACC-NGs, specifically comprises the following steps:
s1, parameter setting: setting the population scale as N, the number of individuals in an initial database as DBsize, the dimension of a decision space as D and the maximum number of true evaluation as FE max True evaluation of the current times being FE, the number of sub-questions or sub-populations being K, the maximum of the sub-questionsDimension D sub The maximum number of shared variables is Ds.
In the embodiment of the application, the English language of the agent assisted co-evolution algorithm based on the improved grouping strategy is called Surrogate-Assisted Cooperative Coevolutionary Optimization algorithm with a New Grouping Strategy, and therefore the algorithm is simply called SACC-NGs in the application.
S2, randomly generating a database by using Latin hypercube, carrying out real evaluation on all individuals in the database, storing the evaluation result into the database, and defining the individual with the smallest fitness value in the database as a global optimal solution gbest; in addition, a Latin hypercube is used to randomly initialize a population P of size N, and the speed of all individuals in the initial population is set to 0.
S3, carrying out ascending sort on individuals in the database according to the real fitness value, and selecting individuals with the first M sorting names to form a better solution set gbests according to the sorting result.
Referring to fig. 2, S4, in the existing random grouping technology, the original large-scale problem is divided into K sub-problems by selecting a shared variable, and the specific steps are as follows:
s41, in the range [1, D s ]Internally generating a random number d s Randomly selecting d among all decision variables of the current decision space s Individual variables as shared variables X sv Put into each sub-problem;
s42, in the range [1, D sub -d s ]Internally generating a random number d r Removing shared variable X from the current decision space sv The remaining variable X other D is selected randomly r Putting the variable into the first sub-problem;
s43, repeating the step S42 until d which is randomly selected is put into the K sub-questions r Stopping after the residual variables;
s44, outputting the K sub-questions after grouping.
Note that, in the embodiment of the present application, all decision variables=shared variable X in the decision space sv +residual variable X other And share variablesMaximum number D of (2) s < maximum dimension D of sub-problem sub 。
In order to further promote interaction between different sub-questions, in each evaluation, the shared variable is selected and put into each sub-question, so that in the case that each sub-question has the shared variable, as long as the variable randomly selected from the remaining variables is related to the shared variable, the variable randomly selected from the remaining variables is in the same sub-question as the shared variable no matter which sub-question is divided into, and the probability of dividing the related variable into the same sub-question and the effectiveness of grouping are improved. In addition, the influence degree of different dimensions on the final result is different, if the selected shared variable is the dimension with larger influence on the overall adaptability evaluation result, the population is more likely to find a better result after multiple updates.
S5, training a corresponding trained agent model for each sub-problem, and calculating the average value of fitness values of each individual on all the sub-problemsThe method comprises the following specific steps:
s51, generating a random number n as an interval, and selecting a plurality of individuals with equal intervals as a training set from the individuals with the smallest fitness value according to the sequencing result in the step S3.
In the embodiment of the present application, the range of values of the random number n is [1,5].
S52, respectively extracting a corresponding sub-training set for each sub-problem from the training set according to the dimension corresponding to each sub-problem.
S53, training the proxy model for each corresponding sub-problem by using each sub-training set, so that each sub-problem respectively obtains a corresponding trained proxy model.
S54, extracting respective corresponding sub-populations from the current population P according to the dimension corresponding to each sub-problem, and calculating the fitness value of each individual in each sub-population by using K trained agent models, wherein each individual obtains K fitness values.
S55, according to eachSolving the mean value of the fitness values of the individual on all the sub-problems by the K fitness values obtained by the individual
In this embodiment of the present application, the proxy model used in step S5 is a radial basis function RBF model.
It should be noted that, since the radial basis function RBF model is suitable for handling expensive optimization problems within 100 dimensions, the maximum dimension D of the sub-problem in the embodiment of the present application sub 100. In addition, the radial basis function RBF model is constructed by directly calling pre-written codes through matlab software, wherein the codes for constructing the radial basis function RBF model are in the prior art, and therefore the codes for constructing the radial basis function RBF model are not specifically described in the application.
The conventional formula for updating the population position by using the competitive particle swarm CSO algorithm is as follows:
X l (t+1)＝X l (t)+V l (t+1)＝X l (t)+V l (t)+ΔV l (t+1)
in the formula DeltaV l (t+1) represents the variation of the failed particle velocity, l represents the failure party, w represents the winner, X l (t) and X w (t) position vectors representing the loser and the winner at the t-th generation, V l (t) represents the velocity vector of the failure party at the t th generation, V l (t+1) represents the velocity vector of the failure party at the t+1st generation, R 1 (t)、R 2 (t) and R 3 (t) is a random coefficient, in [0,1 ]]Randomly generating in a range;for controlling parameters, control->Influence on the position update of the failure location;Indicating the average position of all particles at the t-th generation.
The traditional CSO algorithm using the competitive particle swarm only updates the speed and the position of the failure party at a time, so that the population convergence is slow. In addition, the exploration of the population in the global space only depends on the learning process of the winner of the loser item, so that the population is not developed enough.
In order to balance the exploration and development of the population and enhance the diversity of the population, the competitive particle swarm CSO algorithm is improved, and the global optimal solution is replaced by a group of samples with good fitness value by improving the update formula of a loser and increasing the update formula of a winner, so that the population can be quickly converged, the diversity of the population is ensured, and the population is not easy to fall into local optimal.
Referring to fig. 3, S6, an improved competitive particle swarm CSO algorithm is applied to update each sub-problem in turn, until a new population P is output after all the sub-problems are updated.
Referring to fig. 4, in the embodiment of the present application, each sub-problem specifically updates the following steps:
step1, extracting a corresponding kth sub-population from the current population P according to the dimension corresponding to the kth sub-problem.
It should be noted that, in the embodiment of the present application, the first sub-problem corresponds to the first sub-population, the second sub-problem corresponds to the second sub-population, …, and the kth sub-problem corresponds to the kth sub-population.
Step2, evaluating the fitness value of the individual in the current kth sub-population by using a trained proxy model corresponding to the kth sub-problemBased on the fitness value of the individuals in the k sub-population evaluated +. >Calculating the period of individuals in the kth sub-populationThe expected boost value EI is calculated as follows:
wherein i represents the ith individual, k represents the kth sub-population, C i，k Representing a minimum fitness value f in a database min Estimated fitness value with the ith individual in the kth sub-populationDifference between S i，k Mean value representing fitness value of the ith individual over all subproblems +.>Estimated fitness value +.>The absolute value of the difference value between the two groups is K, the number of the sub-groups, and phi respectively represent the standard distribution and the probability density distribution of the normal function;
it should be noted that since the sub-population is extracted from the population P according to the dimension corresponding to each sub-problem, there is a corresponding sub-population for each sub-problem and a corresponding trained surrogate model for each sub-problem, the embodiments of the present application use K to represent the number of sub-populations, sub-problems or trained surrogate models simultaneously. In addition, in each sub-problem specific updating step, K is uniformly described as the number of sub-populations, and K is described as the kth sub-population in the embodiment of the present application.
In the application, the fitness value of each individual in the sub-population is an estimated value, and the estimated value can only represent the fitness of the individual in the sub-population, but cannot represent the fitness of the individual in other sub-populations, so that the fitness estimated value is used as an index for judging the winner and the loser, which is not beneficial to the development and exploration of the sub-problem and is easy to deviate from the correct searching direction. The EI value is used for judging indexes of winners and losers by replacing the fitness value, so that development and exploration of sub-problems are considered, and an attempt is made to keep a correct search direction for large-scale problems.
Step3, randomly selecting two individuals which do not participate in updating from the kth sub-population, comparing the EI values of the two individuals, taking the individual with large EI value as a winner and the individual with small EI value as a loser, and updating the speed and the position of the individuals on the current sub-population by applying an improved competitive particle swarm CSO algorithm based on the EI value comparison result of the two individuals.
In the embodiment of the present application, the speed update formulas of the loser and the winner are as follows:
wherein l represents a loser, w represents a winner, X l，k (t) and X w，k (t) position vectors representing the failure and winning of the kth sub-population at the t th generation, V l，k (t) and V w，k (t) velocity vectors representing the failure and winning of the kth sub-population at the t th generation, V l，k (t+1) andV w，k (t+1) represents the speed vector of the loser and winner of the kth sub-population at the t+1 th generation, respectively; r is R 1 (k，t)、R 2 (k，t)、R 3 (k，t)、R 4 (k，t)、R 5 (k, t) and R 6 (k, t) is the random coefficient of the kth sub-population at the t generation, at [0,1]Randomly generating in a range;andrandomly selecting the three preferred solutions at the t generation from the preferred solution set gbests at the t generation;Control parameters of loser for control parameters +.>Control parameter of winner +.0.5>0.1.
It should be noted that the number of the substrates, And->Is randomly chosen from the current preferred solution set gbests. Wherein the losing party updates its own speed and position according to random one individual in the gbests set and the winning party, and the winning party updates its own speed and position according to random two individuals in the gbests set. In the method, the set gbests of better solutions with higher rank is used for replacing the mean value of the population position in the original formula, so that failed particles can learn from the better solutions in the population, and the convergence of a CSO algorithm is facilitated to be accelerated. In addition, for the winning particle, the formula is updated to follow the loser to make itThey can learn from two best solutions gbests, helping to enhance the exploratory capacity of winning particles. To ensure diversity, all particles are let learn from random solutions in the highest-ranked best solution set gbest, instead of the traditional global optimal solution gbest. This approach diversifies the search process and helps explore different regions in the solution space.
In the embodiment of the application, the position update formulas of the loser and the winning position are as follows:
X l，k (t+1)＝X l，k (t)+V l，k (t)+R 7 (k，t)ΔV l，k (t+1)
ΔV l，k (t+1)＝V l，k (t+1)-V l，k (t)
X w，k (t+1)＝X w，k (t)+V w，k (t)+R 8 (k，t)ΔV w，k (t+1)
ΔV w，k (t+1)＝V w，k (t+1)-V w，k (t)
wherein l represents a loser, w represents a winner, X l，k (t) and X w，k (t) position vectors representing the failure and winning of the kth sub-population at the t th generation, V l，k (t) and V w，k (t) velocity vectors representing the failure and winning of the kth sub-population at the t th generation, V l，k (t+1) and V w，k (t+1) represents the velocity vector of the loser and winner, deltaV, of the kth sub-population at the t+1 th generation, respectively l，k (t+1) and DeltaV w，k (t+1) represents the variation of the speeds of the losing particles and the winning particles of the kth sub-population at the t+1 generation, respectively; r is R 7 (k, t) and R 8 (k, t) is the random coefficient of the kth sub-population at the t generation, at [0,1]Randomly generated in the range.
It should be noted that, in order to suppress the step size of the speed change from being too large, it may prevent the overall search and avoid sinking into the locally optimal solution, in this embodiment of the present application, the random number R is used 7 (k, t) and R 8 (k, t) to change the step size of the particle learning.
Step4, repeating the Step3 until all individuals in the current sub-population participate in comparison updating and stop.
In this embodiment, when Step3 is executed, if the number of individuals in any one of the sub-populations is singular, the last individual does not execute Step3, i.e. the individual does not update the speed and the position, and directly participates in the next execution Step.
Step5, judging whether the maximum iteration number of the current sub-population update is reached, if so, stopping the current sub-population update to obtain a new kth sub-population; otherwise, go on to Step2-Step5.
Step6, replacing the corresponding dimension on the population P by the new kth sub-population to form a new population P.
And S7, determining a candidate solution by adopting a double-stage strategy, calculating the fitness value of each individual in the candidate solution by utilizing the K trained agent models obtained in the step S5, calculating the evaluation index SD value of each individual according to the fitness value, and taking the individual with the minimum SD value as a new global optimal solution gbest.
Referring to fig. 5, in the embodiment of the present application, the specific steps of step S7 are:
s71, when FE/FE max When < b, taking the new population P as a candidate solution; otherwise, the trained agent model corresponding to each sub-problem is used for evaluating the fitness value of the individual in each sub-problem updated in the evaluation step S6, the individual with the smallest fitness value in each sub-problem is further selected, the corresponding dimension of each sub-problem is taken out as the sub-solution of the sub-problem, and then the sub-solutions of the sub-problems are used for replacing the corresponding dimension of the global optimal solution gbest respectively, so that k synthetic solutions are obtained as candidate solutions.
In this application, b is a control two-stage conversion parameter, randomly generated within the range of [0,1 ]. In the experimental process, the values of the parameter b are respectively 0, 0.2, 0.5 and 1, and according to the experimental result, the optimization performance is better when the value of the parameter b is 0.2. Furthermore, in the examples of the present application, each synthesis solution is an individual.
And S72, evaluating the fitness value of each individual in the candidate solution by sequentially using the k trained agent models obtained in the step S5.
S73, evaluating an evaluation index SD value of each individual in the candidate solutions based on an average deviation strategy, and taking the individual with the minimum SD value as a new global optimal solution gbest, wherein the evaluation index SD value has the following calculation formula;
where K represents the number of trained proxy models,representing the estimated fitness value of the kth trained surrogate model for individual i, +.>Mean value representing estimated fitness values of all individuals in K trained surrogate models, +.>Andmaximum estimated fitness value and minimum estimated fitness value of all individuals in K trained agent models are divided by +.>Is a normalization process.
In the present application, according to FE/FE max And selecting different candidate solutions in different stages according to the comparison result of the control two-stage conversion parameter b. In the early stage of the experiment, a new population P is used as a candidate solution, and a wider area can be explored by evaluating the whole solution, so that the diversity of the algorithm is enhanced, the algorithm can perform more comprehensive searching, and the opportunity of finding the global optimal solution is increased; in the later period of the experiment, the corresponding dimension of the current global optimal solution is respectively replaced by the updated sub solution with the minimum fitness value on each sub problem, the obtained k synthesized solutions are used as candidate solutions, and only the global optimal solution is changed A portion of the dimensions of the solution corresponds to performing a local search, thereby speeding up the convergence process. Through the strategy, the method and the device can explore a wide search space in the whole optimization process, can accelerate convergence through local search when approaching to the optimal solution, and can avoid sinking into local optimal and enhance the global search capability of an algorithm.
S8, carrying out real evaluation on the new global optimal solution gbest, comparing the current global optimal solution gbest with the new global optimal solution gbest, and updating the global optimal solution gbest and the database through a comparison result.
In this embodiment, the step S8 specifically includes: firstly, carrying out real evaluation on a new global optimal solution gbest, and storing the new global optimal solution gbest and a fitness value thereof in a database for updating the database; and secondly, comparing the fitness value of the current global optimal solution gbest with the fitness value of the new global optimal solution gbest, if the fitness value of the new global optimal solution gbest is better than the fitness value of the current global optimal solution gbest, updating the current global optimal solution gbest, otherwise, not updating the current most global optimal solution gbest.
It should be noted that, in this application, for the new global optimal solution gbest, whether or not its fitness value is better than that of the current global optimal solution gbest, the new global optimal solution gbest and its fitness value are stored in the database for updating the database.
S9, judging whether a termination condition is met, if yes, evaluating termination, and outputting a global optimal solution gbest; otherwise, the steps S3-S9 are continued.
In this embodiment of the present application, the termination condition in step S9 is that the execution is terminated when the maximum number of evaluations is reached.
It should be noted that the agent-assisted co-evolution method based on the improved grouping strategy provided by the embodiments of the present application is suitable for solving the single-objective large-scale expensive optimization problem.
The effectiveness of the algorithm is verified through 15 test functions of CEC2013, and compared with some most advanced large-scale expensive algorithms, the algorithm comprises a proxy assisted simplified algorithm with random grouping for processing large-scale expensive optimization problem SEA-LEEA-RG, a proxy assisted evolutionary algorithm SAEA-RFS with random feature selection for the large-scale expensive problem, a proxy assisted differential evolutionary algorithm SADE-AMSS with adaptive subspace search for the large-scale expensive optimization problem and a proxy model assisted co-evolution algorithm SACC-RBFN for processing large-scale optimization problem SACC.
It should be noted that, there are various algorithms in the large-scale optimization problem of Investigating suggorate model assisted cooperative coevolution for large scale optimization using the proxy model to assist the co-evolution, and the comparison algorithm SACC-RBFN used in the present application is only one of the algorithms mentioned in the method.
Based on the separability of the variables, the 15 test functions of CEC2013 can be classified into 5 classes, including fully resolvable functions F1-F3, partially resolvable functions F4-F7 with separable variables, partially resolvable functions F8-F11 without separable variables, overlapping functions F12-F14, and fully non-resolvable function F15. Wherein the number of decision variables of the functions F1-F12 and F15 is 1000, and the number of decision variables of the functions F13 and F14 is 905.
In the experiment, the initial population scale N is 10, the number of the sub-questions is 20, the number of shared variables among the sub-questions is 10, the optimization iteration number of the sub-questions is 5, and the maximum dimension of each sub-question is 100. The initial database individual number is 200, which is twice the largest dimension of the sub-problem. The termination conditions for all algorithms in this experiment were set to reach the maximum number of evaluations, i.e. 11 x d, d represents the dimension of the decision space. Furthermore, each algorithm was run independently 25 times.
Further, in the optimization process, the number of the preferable solution sets gbests is 20, and the parameter b of the handover selection candidate solution is set to 0.2.
In the experiments of the application, the significance of the results of the SACC-NGs and other algorithms of the algorithm is evaluated at a significance level of 0.05 by adopting a Wilcoxon rank sum test, and signs "+", "-" and "=" indicate that the SACC-NGs of the algorithm have obviously better performance, obviously worse performance or no significant difference compared with the other algorithms, and are specifically shown in table 1.
Table 1, comparison results of the present algorithm and other 4 algorithms on CEC 2013 test function
As can be seen from Table 1, the present algorithm SACC-NGs obtained 12 "good", 2 "bad" and 1 "similar" results, respectively, as compared to SEA-LEEA-RG; compared with SAEA-RFS, the algorithm SACC-NGs respectively obtain 12 'good', 0 'bad' and 3 'similar' results; compared with SACC-RBFN, the algorithm SACC-NGs respectively obtain 11 'good', 4 'bad' and 0 'similar' results; compared with SADE-AMSS, the algorithm SACC-NGs respectively obtain 14 'good', 0 'bad' and 1 'similar' results.
Compared with other 4 algorithms, the SACC-NGs of the algorithm shows excellent performance on 9 test functions. Among them, the present algorithm SACC-NGs is particularly excellent in solving the partial resolvable functions F4-F5, F7-F9 and F11. In addition, the present algorithm SACC-NGs also exhibits good performance on the overlapping functions F13-F14 and the completely non-resolvable function F15. Notably, the present algorithm SACC-NGs achieves results that are better than the existing methods by an order of magnitude or even more for functions F5, F8, F9, F14 and F15.
In the partial decomposable function F6 containing separable variables, although the present algorithm SACC-NGs may not obtain the absolute best results, its performance is comparable to the other 4 algorithms.
The algorithm SACC-NGs is superior to only one or two other algorithms for the fully resolvable functions F1-F3, the partially resolvable function F10 without separable variables, and the overlap function F12.
In summary, according to the above comparison results, the present algorithm SACC-NGs performs excellently in solving partially resolvable functions, overlapping functions, and completely non-resolvable functions, but encounters challenges in solving completely resolvable functions, and the present algorithm SACC-NGs exhibits poor performance in terms of convergence for challenging functions F3, F6, and F10.
In addition, the invention provides a computer readable storage medium, wherein the storage medium is stored with a computer program, and when the computer program is executed by a processor, the agent assisted co-evolution method based on the improved grouping strategy can be realized.
In addition, the invention also provides a computer device, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein when the processor executes, the agent assisted co-evolution method based on the improved grouping strategy can be realized.
It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product on one or more computer-usable storage media having computer-usable program code embodied therein, including but not limited to disk storage and optical storage devices, and the like.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus/systems, and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
Those skilled in the art will appreciate that implementing all or part of the above-described methods in accordance with the embodiments may be accomplished by way of a computer program stored on a computer readable storage medium, which when executed may comprise the steps of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a read-only memory ROM or a random access memory RAM.
Finally, it should be noted that: the foregoing description is only illustrative of the preferred embodiments of the present invention, and although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments described, or equivalents may be substituted for elements thereof, and any modifications, equivalents, improvements or changes may be made without departing from the spirit and principles of the present invention.
Claims (10)
1. The agent assisted co-evolution algorithm based on the improved grouping strategy is characterized by comprising the following specific steps:
s1, parameter setting: setting the population scale as N, the number of individuals in an initial database as DBsize, the dimension of a decision space as D and the maximum number of true evaluation as FE max The current times are truly evaluated as FE, the number of sub-questions or sub-populations is K, and the maximum dimension of the sub-questions is D sub The maximum number of shared variables is Ds;
s2, randomly generating a database by using Latin hypercube, carrying out real evaluation on all individuals in the database, storing the evaluation result into the database, and defining the individual with the smallest fitness value in the database as a global optimal solution gbest; in addition, a Latin hypercube is used for random initialization to generate a population P with a scale of N, and the speed of all individuals in the initial population is set to be 0;
s3, carrying out ascending order sorting on individuals in the database according to the real fitness value, and selecting individuals with the ranking names of the first M to form a better solution set gbests according to the sorting result;
s4, dividing the original large-scale problem into K sub-problems by selecting a shared variable in the prior random grouping technology;
s5, training a corresponding trained agent model for each sub-problem, and calculating the average value of fitness values of each individual on all the sub-problems
S6, sequentially updating each sub-problem by applying an improved competitive particle swarm CSO algorithm until all the sub-problems are updated, and outputting a new population P;
s7, determining candidate solutions by adopting a double-stage strategy, calculating the fitness value of each individual in the candidate solutions by utilizing the K trained agent models obtained in the step S5, calculating the evaluation index SD value of each individual according to the fitness value, and taking the individual with the minimum SD value as a new global optimal solution gbest;
s8, carrying out real evaluation on the new global optimal solution gbest, comparing the current global optimal solution gbest with the new global optimal solution gbest, and updating the global optimal solution gbest and the database according to a comparison result;
s9, judging whether a termination condition is met, if yes, evaluating termination, and outputting a global optimal solution gbest; otherwise, the steps S3-S9 are continued.
2. The agent assisted co-evolution algorithm according to claim 1, wherein the step S4 comprises the following specific steps:
s41, in the range [1, D s ]Internally generating a random number d s Randomly selecting d among all decision variables of the current decision space s Individual variables as shared variables X sv Put into each sub-problem;
S42, in the range [1, D sub -d s ]Internally generating a random number d r Removing shared variable X from the current decision space sv The remaining variable X other D is selected randomly r Putting the variable into the first sub-problem;
s43, repeating the step S42 until d which is randomly selected is put into the K sub-questions r Stopping after the residual variables;
s44, outputting the K sub-questions after grouping.
3. The agent assisted co-evolution algorithm according to claim 2, wherein the step S5 comprises the following specific steps:
s51, generating a random number n as an interval, and selecting a plurality of individuals with equal intervals as a training set from the individuals with the smallest fitness value according to the sequencing result in the step S3;
s52, respectively extracting a corresponding sub-training set for each sub-problem from the training set according to the dimension corresponding to each sub-problem;
s53, training a proxy model for each corresponding sub-problem by using each sub-training set, so that each sub-problem respectively obtains a corresponding trained proxy model;
s54, extracting respective corresponding sub-populations from the current population P according to the dimension corresponding to each sub-problem, and calculating the fitness value of each individual in each sub-population by using K trained agent models, wherein each individual obtains K fitness values;
S55, solving the average value of the fitness values of each individual on all the sub-problems according to the K fitness values obtained by the individual
4. A proxy assisted co-evolution algorithm based on an improved grouping strategy as claimed in claim 3, wherein each sub-problem specific update step is:
step1, extracting a corresponding kth sub-population from the current population P according to the dimension corresponding to the k sub-problems;
step2, evaluating the fitness value of the individual in the current kth sub-population by using a trained proxy model corresponding to the kth sub-problemBased on the fitness value of the individuals in the k sub-population evaluated +.>Calculating the expected lifting value EI of the individuals in the kth sub-population, wherein the expected lifting value EI is calculated according to the following formula:
wherein i represents the ith individual, k represents the kth sub-population, C i，k Representing a minimum fitness value f in a database min Estimated fitness value with the ith individual in the kth sub-populationDifference between S i，k Mean value representing fitness value of the ith individual over all subproblems +.>Estimated fitness value +.>The absolute value of the difference value between the two groups is K, the number of the sub-groups, and phi respectively represent the standard distribution and the probability density distribution of the normal function;
Step3, randomly selecting two individuals which do not participate in updating from the kth sub-population, comparing the EI values of the two individuals, taking the individual with large EI value as a winner and the individual with small EI value as a loser, and updating the speed and the position of the individuals on the current sub-population by applying an improved competitive particle swarm CSO algorithm based on the EI value comparison result of the two individuals, wherein the speed updating formula of the loser and the winner is as follows:
the loser and winning position update formulas are as follows:
X l，k (t+1)＝X l，k (t)+V l，k (t)+R 7 (k，t)ΔV l，k (t+1)
ΔV l，k (t+1)＝V l，k (t+1)-V l，k (t)
X w，k (t+1)＝X w，k (t)+V w，k (t)+R 8 (k，t)ΔV w，k (t+1)
ΔV w，k (t+1)＝V w，k (t+1)-V w，k (t)
wherein l represents a loser, w represents a winner, X l，k (t) and X w，k (t) position vectors representing the failure and winning of the kth sub-population at the t th generation, X l，k (t+1) and X w，k (t+1) represents the position vectors of the loser and winner of the kth sub-population at the t+1 th generation, V l，k (t) and V w，k (t) velocity vectors representing the failure and winning of the kth sub-population at the t th generation, V l，k (t+1) and V w，k (t+1) represents the velocity vector of the loser and winner, deltaV, of the kth sub-population at the t+1 th generation, respectively l，k (t+1) and DeltaV w，k (t+1) represents the variation of the speeds of the losing particles and the winning particles of the kth sub-population at the t+1 generation, respectively; r is R 1 (k，t)、R 2 (k，t)、R 3 (k，t)、R 4 (k，t)、R 5 (k，t)、R 6 (k，t)、R 7 (k, t) and R 8 (k, t) is the random coefficient of the kth sub-population at the t generation, at [0,1 ]Randomly generating in a range; and->Randomly selecting the three preferred solutions at the t generation from the preferred solution set gbests at the t generation;Control parameters of loser for control parameters +.>Control parameter of winner +.0.5>0.1;
step4, repeatedly executing the Step3 until all individuals in the current sub population participate in comparison and updating and then stop;
step5, judging whether the maximum iteration number of the current sub-population update is reached, if so, stopping the current sub-population update to obtain a new kth sub-population; otherwise, continuing to execute the steps Step2-Step5;
step6, replacing the corresponding dimension on the population P by the new kth sub-population to form a new population P.
5. The agent assisted co-evolution algorithm based on the improved grouping strategy according to claim 4, wherein the step S7 comprises the following specific steps:
s71, when FE/FE max When < b, taking the new population P as a candidate solution; otherwise, evaluating the fitness value of the individual in each sub-problem updated in the evaluation step S6 by using the trained agent model corresponding to each sub-problem, further selecting the individual with the smallest fitness value in each sub-problem, taking out the corresponding dimension of each sub-problem as the sub-solution of the sub-problem, and then respectively replacing the corresponding dimension of the global optimal solution gbest by the sub-solutions of the sub-problems to obtain k synthetic solutions as candidate solutions;
S72, evaluating the fitness value of each individual in the candidate solution by sequentially using the k trained agent models obtained in the step S5;
s73, evaluating an evaluation index SD value of each individual in the candidate solutions based on an average deviation strategy, and taking the individual with the minimum SD value as a new global optimal solution gbest, wherein the evaluation index SD value has the following calculation formula;
where K represents the number of trained proxy models,representing the estimated fitness value of the kth trained surrogate model for individual i, +.>Mean value representing estimated fitness values of all individuals in K trained surrogate models, +.>And->Maximum estimated fitness value and minimum estimated fitness value of all individuals in K trained agent models are divided by +.>Is a normalization process.
6. The agent assisted co-evolution algorithm according to claim 5, wherein the step S8 is specifically: firstly, carrying out real evaluation on a new global optimal solution gbest, and storing the new global optimal solution gbest and a fitness value thereof in a database for updating the database; and secondly, comparing the fitness value of the current global optimal solution gbest with the fitness value of the new global optimal solution gbest, if the fitness value of the new global optimal solution gbest is better than the fitness value of the current global optimal solution gbest, updating the current global optimal solution gbest, otherwise, not updating the current most global optimal solution gbest.
7. The agent assisted co-evolution algorithm according to claim 6, wherein when Step3 is performed, and the number of individuals in any sub-population is singular, the last individual does not perform Step3, i.e. the individual does not update the speed and location, and participates in the next execution Step.
8. The agent assisted co-evolution algorithm according to claim 6, wherein the agent model used in the step S5 is a radial basis function RBF model, and the termination condition in the step S9 is terminated when the maximum number of evaluations is performed.
9. A computer readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, a proxy assisted co-evolution method based on an improved grouping strategy according to any one of claims 1-8 is implemented.
10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, which when executed by the processor, implements an improved grouping strategy-based proxy assisted co-evolution method according to any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311842226.1A CN117787331A (en) | 2023-12-28 | 2023-12-28 | Agent-assisted co-evolution method based on improved grouping strategy |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311842226.1A CN117787331A (en) | 2023-12-28 | 2023-12-28 | Agent-assisted co-evolution method based on improved grouping strategy |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117787331A true CN117787331A (en) | 2024-03-29 |
Family
ID=90394147
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311842226.1A Pending CN117787331A (en) | 2023-12-28 | 2023-12-28 | Agent-assisted co-evolution method based on improved grouping strategy |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117787331A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118333161A (en) * | 2024-06-12 | 2024-07-12 | 杭州电子科技大学 | Optimization method based on multi-group competition multi-objective optimization confidence rule base |
-
2023
- 2023-12-28 CN CN202311842226.1A patent/CN117787331A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118333161A (en) * | 2024-06-12 | 2024-07-12 | 杭州电子科技大学 | Optimization method based on multi-group competition multi-objective optimization confidence rule base |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103049792B (en) | Deep-neural-network distinguish pre-training | |
| CN109325516B (en) | Image classification-oriented ensemble learning method and device | |
| CN107506590A (en) | A kind of angiocardiopathy forecast model based on improvement depth belief network | |
| CN114386694A (en) | Drug molecule property prediction method, device and equipment based on comparative learning | |
| CN117787331A (en) | Agent-assisted co-evolution method based on improved grouping strategy | |
| CN102314614A (en) | Image semantics classification method based on class-shared multiple kernel learning (MKL) | |
| CN114548591A (en) | Time sequence data prediction method and system based on hybrid deep learning model and Stacking | |
| CN112465226B (en) | User behavior prediction method based on feature interaction and graph neural network | |
| CN115101145B (en) | Medicine virtual screening method based on self-adaptive meta-learning | |
| CN105260746A (en) | Expandable multilayer integrated multi-mark learning system | |
| Demirel et al. | Meta-tuning loss functions and data augmentation for few-shot object detection | |
| CN114971066A (en) | Knowledge tracking method and system integrating forgetting factor and learning ability | |
| CN113191450B (en) | Weak supervision target detection algorithm based on dynamic label adjustment | |
| Mao et al. | An XGBoost-assisted evolutionary algorithm for expensive multiobjective optimization problems | |
| CN113706285A (en) | Credit card fraud detection method | |
| CN117349362A (en) | Dynamic knowledge cognitive hierarchy mining method, system, equipment and terminal | |
| Rahman et al. | Implementation of artificial neural network on regression analysis | |
| Jing et al. | NASABN: A neural architecture search framework for attention-based networks | |
| Kavipriya et al. | Adaptive weight deep convolutional neural network (AWDCNN) classifier for predicting student’s performance in job placement process | |
| CN115661546A (en) | Multi-objective optimization classification method based on feature selection and classifier joint design | |
| Jouyban et al. | Using the modified k-mean algorithm with an improved teaching-learning-based optimization algorithm for feedforward neural network training | |
| CN115345106A (en) | Verilog-A model construction method, system and equipment of electronic device | |
| CN115083511A (en) | Peripheral gene regulation and control feature extraction method based on graph representation learning and attention | |
| CN114818859A (en) | Method and device for diagnosing condition of heat distribution pipe network, terminal equipment and storage medium | |
| CN112529179A (en) | Genetic algorithm-based confrontation training method and device and computer storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |