Novel co-evolution multi-objective optimization method

Abstract

The invention provides a novel co-evolution multi-objective optimization method. According to the method, a cooperative coevolution algorithm (CCEA) is adopted to simulate the cooperative influence between individuals; a distribution estimation strategy is adopted to simulate and solve the direction selection problem that majority decides minority in transaction cooperation, and efficient global search is carried out; simulating a leading effect of an expert group by adopting a Pareto dominating concept and a domain elite strategy; a'large blood exchange 'strategy is adopted to simulate a reform mode when an individual or a tissue pauses when facing a major problem, more effective local search is carried out, and the algorithm is prevented from entering local optimum.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a collaborative evolutionary multi-objective optimization method. Background Art

[0002] Many complex problems in life, such as vehicle path planning, medical problems, power dispatching problems, etc., are often multi-objective optimization problems, and the optimization objectives are mutually conflicting, that is, only one set of Pareto optimal solutions can be obtained. The current multi-objective evolutionary algorithms (MOEA), such as NSGA-II, ATHE, etc., all use the Pareto optimal solution method to compare the advantages and disadvantages of the objectives, and select and evolve individuals accordingly. Compared with the previous non-Pareto methods, they can converge to a more realistic Pareto frontier solution. However, the individual relationships in these algorithms are still too simple compared with the real society. They only perform hybridization and mutation based on the advantages and disadvantages of the individuals themselves, lacking the various mutual influences between individuals in the real society, which is obviously inconsistent with the situation in the real society.

[0003] The co-evolutionary algorithm is such an evolutionary algorithm that takes into account the mutual influence between individuals, and is particularly suitable for situations where there are multiple influencing factors in the problem itself. For problems with multiple influencing factors, each influencing factor can be represented by a population, and the correlation between multiple influencing factors can be reflected through cooperation between populations. The method of representing different individuals through random combinations between populations also increases the diversity of individuals, helps the algorithm to search globally, and avoids the algorithm from falling into a local optimal solution too early. However, this evolutionary method may also be somewhat blind.

[0004] Therefore, how to design an effective technical solution to achieve efficient global search and avoid local optimal solutions for multi-objective problem optimization to suit practical applications is a problem that needs to be solved urgently. Summary of the invention

[0005] The purpose of the present invention is to provide a novel collaborative evolutionary multi-objective optimization method, which can achieve efficient global search and avoid local optimal solutions for multi-objective problem optimization.

[0006] To achieve the above object, the technical solution of the present invention is:

[0007] 1. A novel collaborative evolutionary multi-objective optimization method, characterized by comprising:

[0008] S1. Adopting the domain elite set strategy: In NCMOA, a fixed-size domain elite set A is set. The individuals entering this set are the best individuals in evolution so far. As evolution progresses, the elites or experts in the set will also face renewal. The leader-follower clustering algorithm is used to perform clustering analysis on the excellent individual set, so that the frontier represented by each subclass found in the excellent set can be preserved and the frontier distribution of the focused search guarantee algorithm can be obtained;

[0009] S2. Use cooperative co-evolution to evaluate individual cooperation: randomly combine the individuals to be evaluated in the sub-population with the individuals in the remaining sub-populations to form complete individuals, and then judge the quality based on the Pareto concept. If the external individual is better, update the corresponding individual in set A, otherwise discard it; repeat this step until all individuals in the sub-population are evaluated;

[0010] S3, randomly select N excellent individuals from the external elite set to generate the parent population: sort the non-dominated solution set found by the algorithm based on Pareto's non-dominated sorting and crowding sorting, so as to judge the quality of the solution, and then randomly select N individuals from the external elite set, and these N individuals are used as the parent population;

[0011] S4. Evolving the population using the EDA evolutionary method: the parent population is used as an excellent set of individuals for EDA to establish a probability model, and the sampling opportunities of the stable sub-population are randomly transferred to other sub-populations that have not stabilized;

[0012] S5. Adopt the "big blood transfusion" strategy: When individuals or entire populations are facing stagnation, adopt the "big blood transfusion" strategy to make major replacements and adjustments.

[0013] The present invention proposes a novel collaborative evolution multi-objective optimization method. The method uses a cooperative collaborative evolution algorithm (CCEA) to simulate the cooperative influence between individuals; uses a distribution estimation strategy to simulate the direction selection problem of majority decision of minority in transaction cooperation, and conducts efficient global search; uses the Pareto dominance concept and domain elite strategy to simulate the leading role of expert groups; uses a "big blood transfusion" strategy to simulate the reform mode when individuals or organizations face major problems and stagnate, and conducts more effective local search to prevent the algorithm from entering the local optimum. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Population clustering effect diagram

[0015] Figure 2 The first generation of individual cooperation diagram

[0016] Figure 3 Parent population generation method

[0017] Figure 4 Individual cooperation mode diagram

[0018] Figure 5 Population replacement diagram. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings. The following embodiments do not constitute a limitation of the present invention.

[0020] This embodiment provides a novel collaborative evolution multi-objective optimization method, comprising the following steps:

[0021] S1. Establish a set of domain elites and initialize parameters.

[0022] In NCMOA, a fixed-size domain elite set A is set, the number of individuals in the set is set to M, the total number of sub-populations is set to C, and the size of each sub-population is set to N; the domain elite set A is set to empty, the maximum capacity of A is set to Mmax, the minimum capacity is set to Mmin, Mmin ≥ N; the evolutionary generation number T = 1, and each sub-population is randomly initialized to generate;

[0023] The individuals that enter this group are the most outstanding individuals or "experts" in their fields that have evolved so far;

[0024] As evolution progresses, the elites or experts in the collection will also face renewal and replacement. The renewal method is as follows:

[0025] S11. When the number of individuals M in the set is less than Mmin, random sampling is used to combine individuals, and the Pareto dominance concept is used to select excellent individuals to supplement the set;

[0026] S12, when the number of individuals M in the set exceeds Mmin, the parent population is selected by random sampling;

[0027] S13. When the number of individuals M in the set exceeds Mmax, the leader-follower clustering algorithm is used to classify the individuals in the set. After obtaining the classified individuals, the components of the individuals are placed in various groups respectively, so that each group will also have its own subclasses after clustering.

[0028] This implementation uses a clustering algorithm to perform cluster analysis on a collection of outstanding individuals, such as Figure 1 As shown, each rectangular shaded part represents a corresponding subclass, and each subpopulation is autonomously divided into multiple classes. By establishing a probability model for each subclass in the population, the advantage of this is that the frontier represented by each subclass found in the excellent set can be preserved and searched in focus, thus ensuring the frontier distribution of the algorithm.

[0029] S2. Use the cooperative co-evolution method to evaluate individual cooperation.

[0030] In this implementation, the cooperative co-evolution method is used to evaluate individuals. As Figure 2 shown, each sub-population represents an independent variable. The individuals in the sub-population are randomly combined with the individuals in the remaining sub-populations to form complete individuals. When the sub-population (variable) tends to be stable, a cooperation method based on the "total replacement" strategy is used to generate complete individuals, and the objective vector is calculated. According to the Pareto concept, the superiority and inferiority of the objective vector and the individuals inside set A are judged. If the external individual is more excellent, the corresponding individual in set A is updated; otherwise, it is discarded. Repeat this step until all individuals in the sub-population are evaluated.

[0031] The details of the "total replacement" strategy are as follows:

[0032] S21. The "total replacement" strategy for variables:

[0033] By calculating the variance Vi (1 ≤ i ≤ C) of the parental sub-population, it is used to judge whether each sub-population in the parental population has tended to be consistent. Then, each individual in the corresponding sub-population of the offspring population is evaluated in a loop. As Figure 4 shown, V1 and Vc-1 are less than Vmin, that is, sub-population 1 and sub-population C-1 tend to be consistent. While V2 and Vc are greater than Vmin, that is, sub-population 2 and sub-population C do not tend to be consistent. Then, the individuals in sub-populations 2 and C of the offspring sub-population need to be evaluated in a loop. For example, when evaluating individual individual ji (the ith individual in sub-population j, j = 2), a complete individual i' needs to be combined. Individual i' is composed of genes from the parental population and the offspring population. For the gene at the kth position, when k ≠ j, if Vk < Vmin, the corresponding gene comes from the corresponding sub-population of the parental population; otherwise, the corresponding gene comes from the corresponding sub-population of the offspring population. When k = j, the corresponding gene position is individual ji .

[0034] S22. The "total replacement" strategy for the population:

[0035] In the later stage of evolution, there must be many non-dominated solutions in the offspring population. These non-dominated solutions are added to the external elite set. If, after crowding distance sorting, none of these non-dominated solutions in the offspring population can be retained in the external elite set, it is considered that this generation of population fails to update the external elite population. As Figure 5 shown, the population enters a stagnant state, and the individuals in sub-population 2 tend to be consistent, that is, V2 < Vmin. Then, sub-population 2 needs to be replaced as a whole.

[0036] Replacement strategy: Use uniform distribution to randomly initialize and generate subpopulations. Cycle through the evaluation of individuals in the replaced subpopulations, and combine them into complete individuals, such as Figure 5 As shown, for individuals in subpopulation 2, it is necessary to select the remaining C-1 individuals from the parent population to form a complete individual.

[0037] S3. Calculate the size of the domain elite set A and determine whether A needs to be supplemented.

[0038] This implementation uses the size of the elite set A in the calculation field to determine whether A needs to be supplemented. If the number of elements in the set A is less than Mmin, A needs to be supplemented. According to the method in step S2, the individuals of each subpopulation are randomly combined again to generate the target vector, and the non-dominated individuals are added to A (because of the random combination method, it is easy to generate a new solution, which is different from the solution in the previous step), which can easily supplement A to Mmin; otherwise, if |A|≥Mmin, go to step S4;

[0039] S4. Generate the parent population by randomly selecting N excellent individuals from the external elite set.

[0040] This implementation uses Pareto-based non-dominated sorting and crowding sorting to sort the non-dominated solution set found by the algorithm, thereby optimizing the elite set, and then randomly selects N individuals from the external elite set. These N individuals can serve as the parent population;

[0041] like Figure 3 As shown, the same position elements of N individuals selected from the elite set form each subpopulation, using the same Figure 1 The opposite operation is to put the C elements of the complete individuals into C sub-populations respectively, so as to obtain C next-generation parent groups of size N.

[0042] The advantage of using this method to generate the parent population is that the parent population can always be generated on the basis of an excellent gene, and the algorithm's evolutionary search can always proceed in the direction of the Pareto frontier without the situation of gene regression. In addition, the new excellent individuals generated on the basis of this parent population can also supplement and optimize the external elite set; in this way, the elite set and the parent population always promote each other, accelerating the efficiency of evolutionary search.

[0043] S5. Evolve using EDA.

[0044] This implementation uses leader-follower to perform cluster analysis on set A, obtains different classes, and applies EDA to each class for mathematical modeling, and samples each class separately according to the mathematical model, and finally generates N new individuals. Each individual has C elements, and the elements at the corresponding positions of N individuals are combined together to form C sub-populations, and the size of each sub-population is N.

[0045] S6. Calculate the size of the domain elite set A and determine whether it needs to be pruned.

[0046] If the number of elements in set A is greater than Mmax, A needs to be pruned. According to the leader-follower classification, the crowding distance of individuals in the same classification is calculated, and individuals with small distances are removed to make the size of A equal to Mmax; otherwise, if |A|≤Mmax, proceed to step S7.

[0047] S7. Evolutionary algebra T=T+1.

[0048] S8. If the maximum number of evolutionary generations is reached, the algorithm terminates and set A is output as the non-dominated solution set for multi-objective calculation; otherwise, go to step S2.

[0049] The present invention proposes a new type of collaborative evolutionary multi-objective optimization method. This method is a multi-objective evolutionary algorithm that simulates the development model of human society. NCMOA uses a cooperative collaborative evolutionary algorithm (CCEA) to simulate the influence of cooperation between individuals; uses a distribution estimation strategy to simulate the direction selection problem of majority decision minority in transaction cooperation, and conducts efficient global search; uses the Pareto dominance concept and domain elite strategy to simulate the leading role of expert groups; uses a "big blood transfusion" strategy to simulate the way individuals or organizations change when they face major problems and stagnate, and conducts more effective local search to prevent the algorithm from entering the local optimum.

[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A novel collaborative evolutionary multi-objective optimization method, characterized in that: include: S1. Adopting the domain elite set strategy: In NCMOA, a fixed-size domain elite set A is set. The individuals entering this set are the best individuals in evolution so far. As evolution progresses, the elites or experts in the set will also face renewal. The leader-follower clustering algorithm is used to perform clustering analysis on the excellent individual set, so that the frontier represented by each subclass found in the excellent set can be preserved and the frontier distribution of the focused search guarantee algorithm can be obtained; S2. Use cooperative co-evolution to evaluate individual cooperation: randomly combine the individuals to be evaluated in the sub-population with the individuals in the remaining sub-populations to form complete individuals, and then judge the quality based on the Pareto concept. If the external individual is better, update the corresponding individual in set A, otherwise discard it; repeat this step until all individuals in the sub-population are evaluated; S3, randomly select N excellent individuals from the external elite set to generate the parent population: sort the non-dominated solution set found by the algorithm based on Pareto's non-dominated sorting and crowding sorting, so as to judge the quality of the solution, and then randomly select N individuals from the external elite set, and these N individuals are used as the parent population; S4. Evolving the population using the EDA evolutionary method: the parent population is used as an excellent set of individuals for EDA to establish a probability model, and the sampling opportunities of the stable sub-population are randomly transferred to other sub-populations that have not stabilized; S5. Adopt the "big blood transfusion" strategy: When individuals or entire populations are facing stagnation, adopt the "big blood transfusion" strategy to make major replacements and adjustments.

2. A novel collaborative evolutionary multi-objective optimization method as claimed in claim 1, characterized in that: The above-mentioned method is to establish a fixed-size domain elite set A, and the individuals entering this set are the best individuals or domain "experts" evolved so far; As evolution progresses, the elites or experts in the set will also face replacement. Considering the leading role of group evolution and the need for optimal efficiency, the number of individuals M in the set will have a minimum Mmin and a maximum Mmax requirement. Clustering algorithms are used to cluster the excellent individual set. After obtaining the classified individuals, the components of the individuals are placed in various groups. In this way, there will be subclasses after clustering in various groups. By establishing a probability model for each subclass in the population, the frontier represented by each subclass found in the excellent set can be preserved and searched in focus. The specific method is as follows: 2)1. When the number of individuals M in the set is less than Mmin, random sampling is used to combine individuals, and the Pareto dominance concept is used to select excellent individuals to supplement the set; 2)2. When the number of individuals M in the set exceeds Mmin, random sampling is used to select the parent population; 2)3. When the number of individuals M in the set exceeds Mmax, a clustering algorithm is used to classify the individuals in the set. After obtaining the classified individuals, the components of the individuals are placed in various groups respectively. In this way, each group will also have its own subclasses after clustering. Where M is the number of individuals in the set, Mmin is the minimum number of individuals in the set, Mmax is the maximum number of individuals in the set, Mmin and Mmax are fixed values, Mmin is used to satisfy the generation of the parent population, and Mmax is used to maintain the efficiency of evolution.

3. A novel collaborative evolutionary multi-objective optimization method as claimed in claim 1, characterized in that: The cooperative co-evolution method is adopted to evaluate individual cooperation, specifically as follows: The individuals to be evaluated currently in the sub-population are randomly combined with the individuals in the remaining sub-populations to form complete individuals. When the sub-population tends to be stable, a cooperation method based on the "big replacement" strategy is adopted to generate complete individuals, and the objective vector is calculated; the superiority and inferiority of the objective vector and the individuals inside the set A are judged according to the Pareto concept; if the external individual is more excellent, the corresponding individual in the set A is updated, otherwise it is discarded; this step is repeated until all individuals in the sub-population are evaluated.

4. A novel collaborative evolutionary multi-objective optimization method as claimed in claim 1, characterized in that: The method of randomly selecting N excellent individuals from the external elite set is adopted to generate the parent population, specifically as follows: The non-dominated solution set found by the algorithm is sorted according to the non-dominated sorting and crowding degree sorting of Pareto to judge the superiority and inferiority of the solutions, and then N individuals are randomly selected from the external elite set, N < Mmax, and these N individuals are used as the parent population. In this way, the parent population can always be generated on the basis of an excellent gene, and the evolutionary search of the algorithm is always carried out in the direction of the Pareto front. Moreover, the new excellent individuals generated on the basis of this parent population can also supplement and optimize the external elite set. In this way, the elite set and the parent population always promote each other to accelerate the efficiency of evolutionary search; it can also be used as the modeling object of the EDA evolutionary operation.

5. A novel collaborative evolutionary multi-objective optimization method as claimed in claim 1, characterized in that: The method of evolving the population by adopting the EDA (Estimation of Distribution Algorithm) evolutionary method is to establish an excellent individual set of the probability model and randomly transfer the sampling opportunity of the sub-population that tends to be stable to other sub-populations that do not tend to be stable, specifically including: Use EDA to statistically analyze the regular model of the most excellent individual group in each evolutionary process, and sample and generate new individuals according to this excellent model. Analyze and classify the variable value distributions of these excellent individuals to establish a probability model according to their corresponding distributions. As the evolution progresses, some sub-populations (variables) will tend to be consistent, and the sub-populations generated by establishing probability models for them will also tend to be the same; and the sampling opportunity of the sub-population that tends to be stable is randomly transferred to other sub-populations that do not tend to be stable.

6. A novel collaborative evolutionary multi-objective optimization method as claimed in claim 1, characterized in that: When the individuals and the entire population are facing stagnation, the individuals or the population in the set are replaced and adjusted, specifically as follows: 6)1. For individuals: During the evolution process, when it is found that a certain individual is very excellent, this individual is used with a high probability to replace the blood of the population (only replace the positions of the variables that tend to be stable. In this way, the algorithm can explore and fine-tune around the excellent individual to accelerate the convergence of the algorithm); Cooperation method: First, calculate the variance Vi (1 ≤ i ≤ C) of the parent sub-population to judge whether each sub-population in the parent population has tended to be consistent. If there are sub-populations that have not tended to be consistent, then loop to evaluate each individual in the corresponding sub-population that has not tended to be consistent in the offspring population; 6)2. For the entire population: When the population shows evolutionary stagnation, the entire population will be replaced; Stagnation criterion: When the algorithm fails to update the external elite population after continuous evolution of Gen generations, it can be determined that the population has entered an evolutionary stagnation state, where Gen is a fixed value; Replacement strategy: Use uniform distribution to randomly initialize and generate sub-populations; Cooperation method: Circularly evaluate the individuals in the replaced sub-population, and combine them into complete individuals.

Images