Large-scale problem optimization method based on random interpolation
Through the parallel EMT framework and multi-layer knowledge migration module mapped to low-dimensional space in high-dimensional space for optimization, the problem of high-dimensional large-scale optimization is solved, and efficient optimization is achieved in unknown effective low-dimensionality. Improved the scope of application of running speed and random interpolation.
Patent Information
- Application Number
- CN202410963817.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-22
AI Technical Summary
The high-dimensional large-scale optimization problem is computationally expensive during the iteration process, resulting in a decrease in search efficiency. The existing random interpolation dimensionality reduction method requires prior knowledge and cannot be effectively applied to unknown effective low-dimensionality situations.
Using a parallel EMT framework based on island model, the multi-layer knowledge migration module exchanges information between different tasks and islands, and uses a random interpolation matrix to map high-dimensional space to low-dimensional space, and optimizes it in low-dimensional space, avoiding the first The need to verify knowledge has improved the optimization speed and scope of application.
It realizes efficient optimization of large-scale problems without knowing effective low-dimensionality in advance, improves the scope of application of operating speed and random interpolation, and ensures optimization quality and efficiency.
Smart Images

Figure CN119004949A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to high-dimensional large-scale problem optimization, and specifically to a large-scale problem optimization method based on random interpolation. Background Art
[0002] Under the objective environment of explosive data growth, more complex and higher-dimensional large-scale optimization problems have emerged. Solving such problems is of great significance in fields such as scientific computing and actual production. Large-scale optimization problems have the characteristics of large dimensions of solution space. In the optimization process of such problems, the algorithm will encounter the following problems: 1) High-dimensional space causes the population to complete a large number of calculations during the iteration process; 2) High-dimensional solution space greatly delays the search process and reduces search efficiency.
[0003] In order to solve the problem of large amount of computation in population iteration, since population iteration involves a large number of repeated operations on individuals, existing studies have verified that the algorithm can be accelerated by parallel processing of a large number of ALU units in GPU. As an emerging optimization strategy, the multi-task optimization algorithm EMT can play its evolutionary advantages when there are multiple problems to be solved. It can not only handle different problems at the same time, but also deal with the situation where a single problem is re-described to obtain multiple problems to be solved. Moreover, the self-evolution process and knowledge transfer process within multiple tasks can be executed in parallel. Therefore, EMT is particularly suitable for GPU parallel computing.
[0004] In order to solve the problem of reduced search efficiency due to high-dimensional solution space, a large number of studies have proposed general solutions to large-scale optimization problems. Among them, random interpolation is a dimensionality reduction method based on effective low dimensionality. It is based on a large number of matrix operations and can be accelerated through parallel operations. Many scholars have focused on the research of its effectiveness and feasibility, such as: [Yinglan Feng, Liang Feng, Yaqing Hou, Kay Chen Tan, Sam Kwong. EMT-ReMO: Evolutionary Multitasking for High-Dimensional Multi-Objective Optimization via Random Embedding [C]. IEEE Congress on Evolutionary Computation, 2021: 1672-1679] and [Yinglan Feng, Liang Feng, Yaqing Hou, Kay Chen Tan. Large-Scale optimization via Evolutionary Multitasking assisted Random Embedding [C]. IEEE Congress on Evolutionary Computation, 2020: 1-8]. However, the dimensionality reduction method of random interpolation requires prior knowledge. For example, in the literature [Hong Qian, Yi-Qi Hu, Yang Yu. Derivative-free optimization of high-dimensional non-convex functions by sequential random embeddings [C]. The Proceedings of the 25th International Joint Conference on Artificial Intelligence, 2016: 1946-1952], the authors found that reconstructing the original solution space through a single interpolation matrix would produce interpolation gaps, so they proposed a method of using multiple interpolation matrices to reduce the gaps and thus improve the accuracy of random interpolation. However, experimental results show that only under the specific condition of known effective low dimensions, the optimal value can be obtained through evolution by reducing the original space to the vicinity of the corresponding space of the effective low dimension. In actual optimization problems, the real effective low dimension of the problem cannot be known in advance, nor can it be guaranteed to be optimized under specific conditions. As a result, the dimensionality reduction method of random interpolation cannot be effectively applied to theoretical research or practical applications when the effective low dimension of the problem is unknown. Summary of the invention
[0005] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a large-scale problem optimization method based on random interpolation, which alleviates the limitations of random interpolation in large-scale problem optimization and improves the running speed.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] A large-scale problem optimization method based on random interpolation includes the following steps:
[0008] Step 1: Construct a large-scale multi-task optimization algorithm model LSEMT-Re based on random interpolation
[0009] The parallel EMT framework based on the island model is adopted to divide the populations in all tasks into multiple highly parallel islands. Evaluation, sorting and optimization are performed on an island-by-island basis. A multi-layer knowledge transfer module is introduced into the parallel EMT framework based on the island model to realize conditionally triggered knowledge transfer between individuals belonging to different tasks and between individuals belonging to different islands of the same task.
[0010] Step 2: Use the LSEMT-Re algorithm to optimize large-scale problems
[0011] Step 2.1: Task initialization
[0012] Each task divides the population into four independent islands, and the individuals in each island evolve independently. Based on the given high-dimensional problem, the dimension of the original solution space is D, and the dimension of the low-dimensional evolution space for different tasks is d. i Initialize, d i Satisfy the principle of gradual increase, according to d i Generate random interpolation matrix A corresponding to different tasks i , used to map the low-dimensional evolutionary space population to the original solution space, where A i ~N(0,1 / d i ); For the random interpolation matrix A of different tasks i , the inverse mapping matrix pinvA is obtained by formula (1) i , which is used to map the population from the original solution space back to the low-dimensional space, and then in the low-dimensional space d corresponding to each task i Generate population Subsequent populations evolve independently on islands, through Mapping the population to the original solution space Complete the unification of populations in different dimensions, evaluate and sort the populations according to the objective function, and formula (1) is expressed as:
[0013]
[0014] Step 2.2: Generate offspring based on selective mating
[0015] After the LSEMT-Re algorithm enters the iteration, when specific conditions are met, the multi-layer knowledge transfer module performs cross-island individual migration within the task and implicit knowledge migration across tasks, extracts effective information from tasks in different low-dimensional evolutionary spaces, and migrates it to other low-dimensional tasks, which accelerates convergence. Then, the parent generation is determined by random selection, and the offspring evolution is completed in the low-dimensional evolutionary space.
[0016] Step 2.3: Population update based on the principle of survival of the fittest
[0017] The offspring are mapped to the original solution space through the inverse mapping matrix for evaluation and sorting, and compared with the parent generation to eliminate inferior individuals and select the offspring for the next round of evolution;
[0018] Repeat steps 2.1 to 2.3 until the iteration ends, complete the self-evolution of all tasks, and achieve large-scale problem optimization.
[0019] Furthermore, the step 2.1 also includes: adaptively adjusting the normal distribution obeyed by the random interpolation matrix according to the dimension of the low-dimensional evolution space of the task in question, so as to reduce the variance of the data in the matrix.
[0020] Furthermore, the process of cross-island individual migration within the task of step 2.2 is as follows: Replace the good and bad individuals and synchronize the population changes to the original space. The evolution process is divided into T tasks, each task is divided into K islands, each island takes the adjacent island as the target, and the K islands form a ring topology, and select the target migration island in turn. If the best individual on the target island is better than the worst individual on the island, it will be replaced, and then the second-best individual will be compared with the second-worst individual on the island, and so on, to complete the cross-island individual migration within the task.
[0021] Furthermore, the process of implicit knowledge transfer across tasks in step 2.2 is as follows: select the task whose low-dimensional space dimension is closest to the target task as the source task, and utilize the population in the original solution space by simulating binary crossover The offspring are generated and compared with the worst individual on the island where the parent of the target task is located. If it is better, it is replaced and the changes are synchronized to the low-dimensional space.
[0022] Compared with the prior art, the present invention has the following technical effects:
[0023] The present invention can optimize multiple tasks at the same time by means of EMT, and reduces a large-scale problem to different dimensions through random interpolation method, places it in different tasks, and completes a large number of parallel operations through GPU, thereby alleviating the limitation of random interpolation in processing large-scale optimization problems that requires prior knowledge; in order to ensure that the optimization direction of the low-dimensional evolutionary space can represent the better optimization direction of the original solution space, that is, to ensure the continuity of the evolution of high-dimensional and low-dimensional spaces, through knowledge transfer, the favorable information in the low-dimensional evolutionary space is exchanged to other tasks, the overall optimization speed is accelerated, and the favorable information in the high-dimensional evolutionary space is exchanged to other tasks, and the optimization potential of other tasks is improved, and finally the purpose of solving large-scale optimization problems is achieved. In short, the present invention not only greatly improves the running speed while ensuring good optimization quality, but also improves the scope of application of the dimensionality reduction method of random interpolation. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 : Parallel EMT framework diagram based on island model;
[0025] Figure 2 : Operation flow chart of LSEMT-Re algorithm;
[0026] Figure 3 : When the effective low dimension T = 30, the impact of the number of tasks T on the optimization performance of different problems;
[0027] Figure 4 : When the effective low dimension T = 100, the impact of the number of tasks T on the optimization performance of different problems;
[0028] Figure 5 :When the effective low dimension T = 30, the inter-island migration period η m Impact on optimization performance of different problems;
[0029] Figure 6 :When the effective low dimension T = 100, the inter-island migration period η m Impact on optimization performance of different problems;
[0030] Figure 7 : When the effective low dimension T = 30, the cross-task migration period Δ m Impact on optimization performance of different problems;
[0031] Figure 8 : When the effective low dimension T = 100, the cross-task migration period Δ m Impact on optimization performance of different problems;
[0032] Fig. 9 :When the effective low dimension T=30, the cross-task migration period η t Impact on optimization performance of different problems;
[0033] Fig.10 :When the effective low dimension T = 100, the cross-task migration period η t Impact on optimization performance of different problems;
[0034] Fig.11 : When the effective low dimension T = 30, the number of cross-task migrations Δ t Impact on optimization performance of different problems;
[0035] Fig.12 : When the effective low dimension T = 100, the number of cross-task migrations Δ t Impact on optimization performance of different problems;
[0036] Fig.13 : Comparative experiment process of various algorithms running time bar chart. DETAILED DESCRIPTION
[0037] The specific contents of the present invention are further explained in detail below in conjunction with embodiments.
[0038] A large-scale problem optimization method based on random interpolation includes the following steps:
[0039] Step 1: Build a large-scale multi-task optimization algorithm model based on random interpolation
[0040] Use Figure 1 The parallel EMT framework based on the island model shown in the figure divides the populations in all tasks into multiple islands, and performs evaluation sorting and subsequent optimization on an island-by-island basis, so that multiple islands have high parallelism, thereby giving full play to the parallel capability of the GPU.
[0041] In order to further exert the parallel capability of GPU and obtain higher optimization efficiency, a multi-layer knowledge transfer module is introduced to realize conditionally triggered knowledge transfer between individuals belonging to different tasks and between individuals belonging to different islands of the same task. That is, knowledge transfer includes cross-island individual transfer within the task as shown in Table 2 and implicit knowledge transfer across tasks as shown in Table 3, where: cross-island individual transfer within the task does not involve changes in population information, and only replaces inferior individuals, so that GPU can give full play to its parallel capability and obtain higher optimization efficiency; implicit knowledge transfer across tasks can make up for the disadvantageous characteristics of the evolution of each task and expand the advantageous characteristics. This is because by using the EMT optimization principle, under the premise that there is no similarity between tasks, knowledge transfer can still accelerate convergence. If there is similarity between tasks, the probability of positive transfer is greater. Multiple tasks that solve the same problem have certain similarities, and different tasks are in different low-dimensional evolutionary spaces and have differentiated evolutionary characteristics.
[0042] The dimensionality reduction strategy of random interpolation is adopted. When dealing with a single large-scale optimization problem, the original solution space is reconstructed into multiple different low-dimensional evolutionary spaces through random interpolation matrices of different dimensions and placed in different tasks. The optimization speed and quality of tasks reduced to different dimensions also vary due to the different dimensions. Through knowledge transfer, the tasks optimized in low-dimensional space help other tasks to accelerate convergence, and the tasks in the low-dimensional evolutionary space of higher dimensions help other tasks to ensure the optimization potential, so there is no need to know the prior knowledge of the effective low dimension in advance, thereby alleviating the limitations of random interpolation methods in dealing with large-scale optimization problems.
[0043] The Large Scale Multi-task Optimization Algorithm based on Random Embeddin (LSEMT-Re) proposed in this embodiment can improve the bottleneck problem of the random interpolation method that is effectively low-dimensional and unknowable through the framework of multi-task parallel optimization.
[0044] like Figure 2 As shown in the figure, the operation process of the LSEMT-Re algorithm is as follows: first, perform task initialization on the Host side of the GPU and write the initialized task information into the global memory of the GPU; then, call each kernel function on the Host side to activate the population initialization, interpolation matrix initialization, population evaluation and sorting on the Device side of the GPU. After the data is calculated, it is written into the global memory of the GPU. Among them, population evaluation and sorting use shared memory to accelerate the thread parallelism within the thread block, and only share the necessary information of the corresponding individuals within the thread block. After completing all initializations, the Host side control algorithm enters the iterative process, and controls the number of evaluations by presetting the number of evaluations in advance. The length of the optimization process, the preset number of evaluations is the same as the number of evaluations of the existing comparison algorithm; when specific conditions are met, the corresponding kernel function of the migration is called, and knowledge migration is performed on the Device side, including the kernel (M) that completes the exchange of individuals across islands within the task, and the kernel (IT) that completes the external knowledge migration of the task, and then writes the data to the global memory. Finally, the Host side controls and calls kernel (G) and kernel (S), performs low-dimensional space evolution on the Device side and returns to the original space evaluation sorting, completes the population update in island units, writes the data to the global memory, and then the program control returns to the Host side;
[0045] When the LSEMT-Re algorithm optimizes large-scale optimization problems, it forms multiple different low-dimensional evolutionary spaces by reconstructing a single large-scale optimization problem. Each low-dimensional evolutionary space is optimized within a separate task. From the beginning of the iteration process, all tasks execute the process in parallel to complete self-evolution. When the migration cycle is met, knowledge migration is performed on all tasks at the same time.
[0046] Step 2: As shown in Table 1, the LSEMT-Re algorithm is used to optimize the large-scale problem, including the following steps:
[0047] Step 2.1: Task initialization
[0048] Each task divides the population into four independent islands, and the individuals in each island evolve independently. Based on the given high-dimensional problem, the dimension of the original solution space is D, and the dimension of the low-dimensional evolution space for different tasks is d. i Initialize, d i Satisfy the principle of gradual increase, according to d i Generate random interpolation matrix A corresponding to different tasks i , used to map the low-dimensional evolutionary space population to the original solution space, where A i ~N(0,1 / d i );
[0049] When A i When the matrix obeys the standard normal distribution, after the matrix is multiplied by the large population, the probability of data crossing the boundary is high, resulting in reduced performance of random interpolation. Therefore, the normal distribution obeyed by the random interpolation matrix is processed, and adaptive adjustments are made according to the size of the low-dimensional evolutionary space of the task in question to reduce the variance of the data in the matrix, thereby reducing the possibility of crossing the boundary after matrix multiplication and improving optimization efficiency.
[0050] Interpolation matrix A for different tasks i , the inverse mapping matrix pinvA is obtained by formula (1) i , which is used to map the population from the original solution space back to the low-dimensional space, and then in the low-dimensional space d corresponding to each task i Generate population The subsequent populations will undergo independent self-evolution and other operations on islands. Mapping the population to the original solution space Complete the unification of populations in different dimensions, evaluate and sort the populations according to the objective function, and formula (1) is expressed as:
[0051]
[0052] Step 2.2: Generate offspring based on selective mating
[0053] After entering the iteration process, when specific conditions are met, the multi-layer knowledge transfer module performs cross-island individual migration within the task and implicit knowledge migration across tasks, extracts effective information from tasks in different low-dimensional evolutionary spaces, and migrates it to other low-dimensional tasks, which accelerates convergence. Then, the parent generation is determined by random selection, and the offspring evolution is completed in the low-dimensional evolutionary space.
[0054] Step 2.3: Population update based on the principle of survival of the fittest
[0055] The offspring are mapped to the original solution space through the inverse mapping matrix for evaluation and sorting. After comparison with the parent generation, inferior individuals are eliminated to select the offspring for the next round of evolution.
[0056] Repeat steps 2.1 to 2.3 until the iteration ends and the self-evolution of all tasks is completed. This part involves frequent spatial mapping processes, processing the population information of the low-dimensional evolution space or the original solution space in different tasks, making the populations in the two spaces asynchronous. During the implementation process, special attention should be paid to the synchronization after the population changes.
[0057] Table 1 Operation process of LSEMT-Re algorithm
[0058]
[0059]
[0060] Table 2. Cross-island individual migration within the task
[0061]
[0062]
[0063] As shown in Table 2, individuals within the same task are all in the same low-dimensional space. The process of cross-island individual migration within the task does not involve individual evaluation and does not require the use of the population of the original solution space. Therefore, it is carried out in the low-dimensional space, that is, using the low-dimensional space population After performing corresponding operations and completing the replacement of the good and bad individuals in the low-dimensional space, the population changes are synchronized to the original space. The evolution process is divided into T tasks, each task is divided into K islands, each island takes the adjacent island as the target, and the K islands form a ring topology. The target migration island is selected in turn. If the best individual on the target island is better than the worst individual on the island, it is replaced, and then the second-best individual is compared with the second-worst individual on the island, and so on. A total of Δ m In the second comparison, the cross-island individual migration algorithm within the task uses a triple for loop. In the actual operation of the algorithm, based on the convenient parallel processing mechanism provided by CUDA, the outer double loop is placed in T*K thread blocks, and each thread block calls Δ mThreads can complete the migration of all individuals in all tasks within one operation time;
[0064] Table 3 Implicit knowledge transfer across tasks
[0065]
[0066]
[0067] As shown in Table 3, in the process of implicit knowledge transfer across tasks, the populations within different tasks are in different low-dimensional evolutionary spaces, but the populations of all tasks are in the same original solution space, so the implicit knowledge transfer process across tasks is carried out in the original space, that is, using the population in the original solution space. Corresponding operations are performed. After completing the knowledge transfer, the changed population information needs to be synchronized to the low-dimensional space. The lower the similarity between the two tasks, the easier it is to generate negative migration information, thereby wasting evolutionary resources. Therefore, cross-task knowledge transfer is performed between tasks with the highest similarity. Since the same evaluation function is used, all T tasks have similarity. When the dimensions of the low-dimensional evolutionary space of the two tasks are closer, the information lost after mapping from the original solution space to the low-dimensional space is also closer, and thus has a higher similarity. Therefore, the task with the closest low-dimensional space dimension to the target task is selected as the source task, and the offspring is generated by simulated binary crossover (SBX), which is compared with the worst individual on the island where the parent of the target task is located. If it is better, it is replaced and the changes are synchronized to the low-dimensional space. The triple cycle of implicit knowledge transfer across tasks can be completed in O(1) time complexity.
[0068] In order to verify the performance of the large-scale problem optimization algorithm LSEMT-Re based on random interpolation proposed in this embodiment, it is compared with several existing single-objective large-scale optimization algorithms:
[0069] 1. Design test function
[0070] The dimensionality reduction strategy of random interpolation is based on the concept of effective low dimension, that is, some dimensions of a problem to be optimized have little effect on the evaluation of the objective function, while other dimensions have a decisive influence on the evaluation. Most practical problems have such characteristics. In order to meet this characteristic, the existing large-scale optimization problems are artificially modified to make some dimensions more important. The embodiment of this article assumes that the effective low dimensions of the objective function corresponding to the large-scale optimization problem are 30 and 100 respectively. However, during the experiment, it only simulates the real problem solving process and does not know the situation of this parameter. The dimension selection of the low-dimensional evolution space under the multi-task framework is randomly generated and increases successively. If better results are obtained in solving large-scale optimization problems, it means that the limitation of random interpolation that requires the prior knowledge of effective low dimensions in advance can be alleviated.
[0071] See Table 4. Six standard functions with different characteristics, Sphere, Ackley, RosenBrock, Rastrgin, Griewank and Schwefel, are selected as test cases, where f(x) is the large-scale optimization problem to be solved; D is the number of decision variables, that is, the dimension of the original solution space; d is the effective low dimension of the problem, that is, the first d decision variables have a greater impact on the function evaluation, and the last Dd decision variables have a smaller impact on the function evaluation;
[0072] In all problems, except for the RosenBrock function whose optimal value is (1, 1, ..., 1), the optimal values of the other problems are all (0, 0, ..., 0). In order to avoid the code design being biased towards taking the minimum value as 0 and to ensure the accuracy of the experiment and the fairness of the comparison of algorithms, the objective functions of these problems were shifted to the right by 0.2.
[0073] Table 4 Six test functions
[0074]
[0075]
[0076] 2. Experimental environment and parameter settings
[0077] The LSEMT-Re algorithm uses GPU as the computing platform, uses NVIDIA3070ti 8GB graphics card, the programming language is CUDAC, the operating system is Windows 10, and EMT uses GA as the solver, with a crossover probability of 1.0 and a mutation probability of 1 / dim, where dim is the dimension of the population when performing the mutation operation;
[0078] The comparison algorithms all use CPU as the computing platform, using Intel Core i9-10900K@3.70GHz processor, 64GB RAM, MATLAB as the programming language, and Windows 10 as the operating system;
[0079] The experimental parameters are shown in Table 5, including:
[0080] 2.1) The dimension D of the original solution space should meet the dimension setting requirements for solving large-scale problems, and make the comparison algorithm run on the CPU without causing the algorithm to crash due to hardware computing bottlenecks. Under the premise of meeting the above requirements, the dimension should be as high as possible. Through testing, D is more suitable to be 2000;
[0081] 2.2) The number of islands K. The algorithm is based on the island model. The task population needs to be divided into multiple islands. In GPU computing, the data in the island will be processed inside the thread block. According to the GPU hardware architecture, the number of thread blocks is set to 2. n The best value is when n=1, the population will be divided into two islands, and the cross-island individual migration within the task will be carried out between the two islands. There will be more identical individuals in the two islands, which will greatly reduce the richness of the population; when n=3, the population will be divided into 8 islands, and there will be too few individuals in each island, which is not conducive to self-evolution based on the island; therefore, taking n=2 and dividing the population into 4 islands can not only meet the requirements of forming an appropriate number of topological rings, but also ensure that the number of individuals in the island is relatively reasonable;
[0082] 2.3) The population size S of each island. In the CUDA architecture design, the thread declaration uses the thread bundle as the minimum resource allocation unit. Each thread bundle includes 32 threads. During the optimization process, each individual is calculated in a separate thread. In order to prevent the GPU from wasting resources during resource allocation and avoid unnecessary resource overhead, the number of individuals in the island is set to 32n for the best result. When n=1, there are too few individuals in the island, which affects the evolution efficiency. When n=3, the overall population of a single task is too large. Therefore, n=2 is taken, and a single island contains 64 individuals.
[0083] 2.4) Single-task population size N, a single task contains 4 islands, and the population size of each island is 64. Therefore, the population size in a single task is 4*64=256;
[0084] 2.5) Total number of evaluations G. In order to ensure the principle of fairness when comparing algorithms, all algorithms should use the same population size to optimize the same problem under the same total number of evaluations. The LSEMT-Re algorithm uses T tasks to solve a large-scale optimization problem at the same time. During the experiment, when T>10 and the population iterates 100 times, some of the compared algorithms encounter hardware bottlenecks due to insufficient resources under the same number of evaluations, resulting in program crashes. Therefore, when T=10 and the population iterates 100 times, the total number of evaluations used is 260,000 as the standard for comparison with other algorithms. During the optimization process, knowledge transfer is performed once every 5 generations, and the corresponding evaluation cycle is about 13,000 times, that is, knowledge transfer is performed once every 13,000 evaluations.
[0085] Table 5 Experimental parameters
[0086]
[0087]
[0088] 3. Ablation experiment
[0089] Since some parameters cannot be optimally selected through theoretical analysis, the number of tasks T, the island migration period η, and m , the number of inter-island migrations Δ m , cross-task migration cycle η t , the number of cross-task migrations Δ t A total of 5 parameters to be determined were tested to determine the best parameter settings. Three problems with different characteristics, Sphere, Ackley, and RosenBrock, were selected from the existing test set and tested at 30 and 100 effective low dimensions respectively. The experimental results were averaged over 30 independent experiments to more accurately count the impact of parameter values on experimental performance.
[0090] 3.1) Number of tasks T
[0091] In order to study the impact of the number of tasks T on the optimization performance, this part of the parameter setting experiment is carried out. When a single large-scale optimization problem is solved by EMT, it is necessary to include a task in the original solution space to ensure the continuity of the optimization process, and also to include a task in the low-dimensional evolution space to accelerate the convergence of the problem. Therefore, T should be ≥ 3. The final test value selected is T = {3, 4, 5, 6, 7, 8, 9, 10}. During the experiment, other parameters are set as follows: η m =26000,η t =78000,Δ m =10,Δ t =10;
[0092] Due to the large difference in the objective function values, in order to more directly observe the impact of parameters on the solution performance, the results of some objective functions are magnified. If the same situation occurs, subsequent experiments will also use this processing method. The experimental results are the average of 30 independent experiments, and are tested under problems with different characteristics. Although different tasks evolve in different low-dimensional spaces, the more tasks there are, the easier it is to disperse computing resources, thereby reducing the effect of task evolution. Figure 3 and Figure 4 From the ablation experiment, we can see that T=3 is a more appropriate number of tasks, so this parameter value will continue to be used in subsequent experiments;
[0093] 3.2) Inter-island migration period η m
[0094] The essence of inter-island migration is the replacement of individuals, and no new individuals are generated. Therefore, the occurrence cycle should not be too fast. The strategy of performing intra-island migration once every {5, 10, 15, 20, 25} generations is adopted, corresponding to the evaluation cycle η m={13000,26000,39000,52000,65000}, other parameters during the experiment were set as, T=3,η t =78000,Δ m =10,Δ t = 10, and tested at 30 and 100 effective low dimensions respectively. The results are as follows Figure 5 and Figure 6 As shown, when η m =13000, the function values of the four curves reach the optimal value, and the performance of other curves is also good, so η is finally determined. m =13000, and this value was used in subsequent experiments;
[0095] 3.3) The number of inter-island migrations Δ m
[0096] If the number of inter-island migration is too large, the islands within the mission will become homogenous, which will greatly reduce the richness of the population and thus slow down the evolution process. If the number of inter-island migration is too small, it will not accelerate convergence. Therefore, the number of migration is determined to be between 1 / 6 and 1 / 4 of the size of a single island population, that is, Δ m ={10,11,12,13,14,15}, other parameters are set as T=3,η m =13000,η t =78000,Δ t = 10, and tested at 30 and 100 effective low dimensions respectively. The results are as follows Figure 7 and Figure 8 As shown, when Δ m =11, it has good results in all three test functions, so Δ m =11, and this value was used in subsequent experiments;
[0097] 3.4) Cross-task migration cycle η t
[0098] Cross-task knowledge transfer can accelerate convergence and in some cases help specific tasks escape local optimality. However, due to the differences in the processing tasks themselves, it also generates negative migration information, wastes evaluation resources, and has a negative impact on the evolution process. The cross-task migration cycle η t The setting principle is slower than the inter-island migration period η m , take η t ={65000,78000,91000,104000}, the rest of the parameters are set as T=3,η m =13000,Δ m =11,Δ t=10; the test was performed under 30 and 100 effective low dimensions respectively, and the Sphere function was expanded by 10 times under 30 effective low dimensions, and the Sphere function was expanded by 2 times under 100 effective low dimensions. The results are as follows Fig. 9 and Fig.10 As shown, when η t When =65000, more curves can obtain the optimal value, and this value will be used in subsequent experiments;
[0099] 3.5) The number of cross-task migrations Δ t
[0100] New individuals will be generated during the cross-task migration process, and evaluation resources will be used to evaluate these individuals, thereby reducing the overall evolution process of the task. Therefore, the number should not be set too much. The number of cross-task migrations is set with reference to the number of inter-island migrations, Δ t ={10,11,12,13,14,15}, other parameters are set as T=3,η m =13000,Δ m =11,η t = 65000, tested at 30 and 100 effective low dimensions respectively, and at 30 effective low dimensions, the Sphere function was expanded by 10 times, and at 100 effective low dimensions, the Sphere function was expanded by 2 times. The results are as follows Fig.11 and Fig.12 As shown, Δ t There is no significant difference in the values, so we randomly select Δ t =10, and continued to use it in subsequent experiments;
[0101] 4. Comparative experiment
[0102] The idea of using multiple tasks to solve the same problem is that when T=3 and G=260000, the population size is T*N(3*256)=768 when solving a single large-scale problem, and the total number of evaluations is 260000. The classic large-scale optimization problem algorithms ABC, DE, OFA, SHADE, GA and IMODE that process single objective functions are selected as comparison algorithms. The performance of the comparison algorithms and LSEMT-Re is tested using 12 objective functions expanded from 6 large-scale optimization problems.
[0103] 4.1) Comparative experimental results
[0104] Table 6 shows the function value comparison results of the comparison algorithm and LSEMT-Re. The black bold indicates the optimal result. In order to intuitively show the statistical difference, "+", "-", "=" are used to indicate that the performance of each existing algorithm in the corresponding experiment is "better", "worse" or "equivalent" to the LSEMT-Re algorithm. The statistical method uses the rank sum test to compare the significant differences of the algorithms.
[0105] Table 6 Comparison results of function values between the comparison algorithm and LSEMT-Re
[0106]
[0107] As can be seen from Table 6, the LSEMT-Re algorithm of this embodiment obtains the optimal value in 5 objective functions including RosenBrock-30, RosenBrock-100, Rastrigin-30, Griewank-30 and Griewank-100, and GA obtains the optimal value in 4 test functions including Ackley-30, Ackley-100, Schwefel-30 and Schwefel-100. It can be seen that the optimization performance of the LSEMT-Re algorithm of this embodiment for large-scale problems is significantly better than that of ABC, DE, OFA, and SHADE, and is comparable to that of GA, and slightly inferior to that of IMODE. Although it failed to obtain the optimal value in more than half of the test problems, it can obtain the optimal value on more problems than the comparison algorithm, and the solution quality ranks among the top three of all algorithms;
[0108] 4.2) Comparative experimental results analysis
[0109] On the one hand, the large randomness of random interpolation has an impact on the performance of the LSEMT-Re algorithm, which is mainly reflected in the degree to which the matrix obeys the normal distribution. If the matrix is closer to the normal distribution, it is less likely to cross the boundary during the spatial reconstruction process, and the global optimal solution or local optimal value is more likely to appear in the low-dimensional evolution space. Therefore, the randomness of random interpolation has a negative impact on the experimental results;
[0110] On the other hand, the random interpolation process is based on matrix operations. After mapping the original space to the low-dimensional space, it will cause information loss, resulting in random interpolation gaps. The smaller the dimension of the low-dimensional evolution space, the larger the interpolation gap, and the more difficult it is to obtain a better solution in the low-dimensional space, which has a negative impact on the performance of the algorithm and limits the upper limit that the algorithm can reach. If the final optimal solution of the original space is (0, 0, ..., 0) D , then through the random interpolation matrix, the optimal solution is mapped to the low-dimensional space as (0, 0, ..., 0) d, still within the definition of the function, when designing the test function, in order to prevent the algorithm from evolving towards 0, all test functions are translated, and the optimal solution is also translated accordingly, becoming (a, a, ..., a) D , then mapping this optimal value to a low-dimensional space cannot guarantee the optimal solution (a1, a2, ..., a d ) d Still within the given range, that is, in the low-dimensional space, there may not be an optimal solution for the function or there may be fewer high-quality solutions, which also has a negative impact on the performance of random interpolation;
[0111] 4.3) Comparison of the running time of various algorithms in the experiment
[0112] Fig.13 The total time of the comparison algorithms ABC, DE, OFA, SHADE, GA and IMODE and the LSEMT-Re algorithm of this embodiment after optimizing 12 large-scale problems is shown in seconds. It can be seen that the LSEMT-Re algorithm has the shortest time of 1968s. Compared with ABC, DE, OFA and SHADE, it can obtain the optimal solution in a shorter time; although GA and IMODE have better optimization effects, they also consume the most time in the serial operation process. That is to say, although LSEMT-Re has no obvious advantage over GA and IMODE in solution quality, its solution speed is 7.8 times and 8.9 times that of these two algorithms respectively. It can be seen that through GPU parallel computing, the algorithm greatly reduces the time cost and significantly improves the running speed in the optimization process of large-scale problems;
[0113] Since the comparison algorithm and the LSEMT-Re algorithm proposed in this embodiment use different computing platforms, in order to compare all algorithms as fairly as possible, the experimental environment selected Intel i9 flagship series CPU and NVIDIA 30 series mid-range GPU. Although it is impossible to make a more scientific comparison of algorithms on different platforms under completely equal conditions, the experimental conclusions are of reference value, verifying that the GPU-based LSEMT-Re algorithm has advantages in time and performance when solving large-scale optimization problems.
Claims
1. A large-scale problem optimization method based on random interpolation, characterized in that: The steps include: Step 1: Construct a large-scale multi-task optimization algorithm model LSEMT-Re based on random interpolation The parallel EMT framework based on the island model is adopted to divide the populations in all tasks into multiple highly parallel islands. Evaluation, sorting and optimization are performed on an island-by-island basis. A multi-layer knowledge transfer module is introduced into the parallel EMT framework based on the island model to realize conditionally triggered knowledge transfer between individuals belonging to different tasks and between individuals belonging to different islands of the same task. Step 2: Use the LSEMT-Re algorithm to optimize large-scale problems Step 2.1: Task initialization Each task divides the population into four independent islands, and the individuals in each island evolve independently. Based on the given high-dimensional problem, the dimension of the original solution space is D, and the dimension of the low-dimensional evolution space for different tasks is d. i Initialize, d i Satisfy the principle of gradual increase, according to d i Generate random interpolation matrix A corresponding to different tasks i , used to map the low-dimensional evolutionary space population to the original solution space, where A i ~N(0,1 / d i ); For the random interpolation matrix A of different tasks i , the inverse mapping matrix pinvA is obtained by formula (1) i , which is used to map the population from the original solution space back to the low-dimensional space, and then in the low-dimensional space d corresponding to each task i Generate population Subsequent populations evolve independently on islands, through Mapping the population to the original solution space Complete the unification of populations in different dimensions, evaluate and sort the populations according to the objective function, and formula (1) is expressed as: Step 2.2: Generate offspring based on selective mating After the LSEMT-Re algorithm enters the iteration, when specific conditions are met, the multi-layer knowledge transfer module performs cross-island individual migration within the task and implicit knowledge migration across tasks, extracts effective information from tasks in different low-dimensional evolutionary spaces, and migrates it to other low-dimensional tasks, which accelerates convergence. Then, the parent generation is determined by random selection, and the offspring evolution is completed in the low-dimensional evolutionary space. Step 2.3: Population update based on the principle of survival of the fittest The offspring are mapped to the original solution space through the inverse mapping matrix for evaluation and sorting, and compared with the parent generation to eliminate inferior individuals and select the offspring for the next round of evolution; Repeat steps 2.1 to 2.3 until the iteration ends, complete the self-evolution of all tasks, and achieve large-scale problem optimization.
2. The large-scale problem optimization method based on random interpolation according to claim 1 is characterized in that: The step 2.1 also includes: adaptively adjusting the normal distribution obeyed by the random interpolation matrix according to the dimension size of the low-dimensional evolution space of the task in question, so as to reduce the variance of the data in the matrix.
3. The large-scale problem optimization method based on random interpolation according to claim 1 is characterized in that: The process of cross-island individual migration within the task of step 2.2 is as follows: Replace the good and bad individuals and synchronize the population changes to the original space. The evolution process is divided into T tasks. Each task is divided into K islands. Each island takes the adjacent island as its target. The K islands form a ring topology. The target migration island is selected in turn. If the best individual on the target island is better than the worst individual on the island, it is replaced. Then the second-best individual is compared with the second-worst individual on the island, and so on, to complete the cross-island individual migration within the task.
4. The large-scale problem optimization method based on random interpolation according to claim 1 is characterized in that: The process of implicit knowledge transfer across tasks in step 2.2 is as follows: select the task with the closest dimension in the low-dimensional space to the target task as the source task, and utilize the population in the original solution space by simulating binary crossover. The offspring are generated and compared with the worst individual on the island where the parent of the target task is located. If it is better, it is replaced and the changes are synchronized to the low-dimensional space.
