A multi-objective sparse competition decomposition method and system for automobile exterior design and a computer storage device thereof

Abstract

The application provides a multi-objective sparse competition decomposition method, system and computer storage device for automobile shape design. The method sets initialization processing, adaptive parameter updating, competition decomposition mechanism setting, set optimization and loop judgment execution steps, adopts a limited number of weight vectors to decompose the multi-objective problem, significantly reduces the size of the sub-problem under the premise of ensuring the distribution quality of the solution set, thereby improving the overall efficiency of large-scale expensive multi-objective optimization; and the search state of the sub-problem is comprehensively evaluated, and the sub-problems are dynamically divided into different competition sets according to the flexible sorting, so that different sub-problems play different roles in the search process to execute different evolution processes, and the adaptive allocation of search resources is realized.

Description

Technical Field

[0001] This invention relates to the field of optimization computing technology, and in particular to a multi-objective sparse competitive decomposition method, system, and computer storage device for automobile exterior design. Background Technology

[0002] In practical applications such as aerospace design, complex engineering system optimization, neural network structure search, and photonic structure design, it is often necessary to simultaneously optimize multiple conflicting performance indicators, such as efficiency, stability, and structural performance. The evaluation of these performance indicators typically relies on high-precision numerical simulations or physical experiments, resulting in high costs for each evaluation and thus forming an expensive multi-objective optimization problem. As the complexity of engineering systems increases, the dimensionality of decision variables in related problems continues to rise, reaching hundreds or even thousands of dimensions in some applications, further evolving the problem into a large-scale, expensive multi-objective optimization problem, significantly increasing its solution difficulty.

[0003] For multi-objective optimization problems, decompositional multi-objective evolutionary algorithms improve scalability to some extent by decomposing the original multi-objective problem into a set of scalar subproblems and solving them collaboratively. However, existing decompositional methods still have shortcomings in large-scale and expensive scenarios. On the one hand, to obtain a good Pareto front approximation, a large number of weight vectors are usually introduced, leading to a surge in the number of subproblems, significantly increasing computational overhead and the difficulty of training and updating the surrogate model. On the other hand, existing methods often use a uniform search strategy for all subproblems, failing to distinguish the differences in convergence state and diversity contribution of different subproblems, which can easily lead to unreasonable allocation of search resources, thereby reducing overall optimization efficiency, especially under expensive evaluation conditions. To reduce the cost of objective function evaluation, some studies have introduced evolutionary optimization methods assisted by surrogate models. Among them, classification models guide the search direction by distinguishing between potentially good and bad solutions, and have a certain robustness in cases with limited sample size or noise. However, in existing technologies, the integration of classification models and decompositional multi-objective optimization frameworks is still not tight enough, making it difficult to achieve dynamic search resource allocation for different subproblems in large-scale and expensive multi-objective problems.

[0004] In practical engineering applications of multi-objective optimization, automotive exterior design optimization is a representative and typical problem, with similar problems also widely existing in fields such as airfoil design. Automotive exterior design improves multiple performance indicators, including aerodynamics, vehicle stability, and occupant comfort, by adjusting the geometric parameters of the body, auxiliary components, and local curved surfaces. As the level of design refinement increases, these problems typically exhibit high-dimensional, multi-objective characteristics, with the number of decision variables reaching tens to hundreds. During the conceptual design and simulation analysis stages, designers often aim to fully explore the design space without introducing manufacturing or safety constraints.

[0005] For example, car exterior design is one parameterized variable The description defines each variable as a geometric feature of the vehicle's overall outline, local surface shapes, or auxiliary components. These include, but are not limited to, front and rear curvature parameters, roof height, side profile control point positions, spoiler angles, and local proportion parameters. These parameters collectively constitute the high-dimensional design space of the vehicle's exterior. The objective function vector is defined as follows: ,in It can represent the air drag coefficient. It can be expressed as the lift coefficient. The objective function can represent the passenger compartment volume or space utilization rate, and other objective functions can be extended according to specific design requirements. The above performance indicators are often conflicting, requiring a comprehensive trade-off among multiple objectives. During the conceptual design and simulation analysis phases, designers often aim to fully explore the automotive exterior design space without introducing manufacturing constraints, safety regulations, or regulatory limitations to obtain potentially superior exterior design solutions. Therefore, this type of problem can be formally represented as an unconstrained multi-objective optimization problem: , In practical applications, the objective function of automotive exterior design typically relies on computational fluid dynamics simulations or other numerical analysis methods for evaluation. Each evaluation process is computationally complex and time-consuming, resulting in high evaluation costs. Therefore, this problem is an unconstrained, large-scale, and expensive multi-objective optimization problem. Under limited evaluation budgets, it places high demands on the search efficiency and resource allocation capabilities of optimization algorithms. Existing technologies still struggle to balance solution quality and computational efficiency when solving such problems.

[0006] Therefore, in view of the problems existing in the prior art, it is of great importance to provide a multi-objective sparse competitive decomposition technique that can significantly reduce the size of subproblems and improve the overall efficiency of large-scale and expensive multi-objective optimization while ensuring the quality of solution set distribution. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-objective sparse competitive decomposition method, system, and computer storage device for automotive exterior design, applicable to unconstrained multi-objective high-dimensional automotive exterior design optimization scenarios. By introducing an intelligent search scheduling mechanism within a decompositional multi-objective optimization framework, it can dynamically allocate evolutionary resources based on the search status and potential contribution of different sub-problems, thereby effectively improving optimization efficiency and solution set quality under a finite objective function evaluation budget. While ensuring the diversity and convergence of multi-objective solution sets, it reduces the computational burden caused by high-dimensional design spaces and expensive evaluations, making the proposed multi-objective sparse competitive decomposition method applicable to unconstrained large-scale expensive multi-objective optimization scenarios such as automotive exterior design and airfoil design. This provides an efficient and reliable optimization method for the conceptual design and simulation analysis of complex engineering systems, thereby solving the problems of excessive number of sub-problems, unreasonable allocation of search resources, and low optimization efficiency in existing technologies for large-scale expensive multi-objective optimization problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A multi-objective sparse competitive decomposition method for automobile exterior design, characterized in that the multi-objective sparse competitive decomposition method includes: Step S1. Initialization process, which is performed according to the following steps: Step S1-1: Using the simplex lattice point design method in 3D target space generation Uniform points Generate weight vector ; Step S1-2: Calculate the weight vector Calculate the Euclidean distance between every two points, sort the distances of each vector in ascending order, and select the nearest one. The set of neighbors forms an adjacency index. Obtain the neighbor matrix This is used for subsequent sub-problem grouping or neighborhood search; Step S1-3: Generate the initial population using a Latin hypercube The initial population and its target values ​​are then evaluated and stored in an external archive. The ideal point is initialized based on the objective function value of the current population. The population size is set to N, and the objective number is set to M. Then the ideal point... It consists of the minimum values ​​of each objective dimension, i.e. ( ),in , Let i be the i-th individual in the population. Let it be the function value of the j-th objective; Steps S1-4: Initialize a local support vector machine (SVM) surrogate model for each sub-problem to assist in sub-problem search, and train the SVM surrogate model to obtain a model for predicting arbitrary input points. An SVM proxy model that categorizes and sets the values ​​of their decision functions; Step S2. Adaptive parameter update: Adaptively update local crossover probabilities and maximum number of offspring Among them, the maximum number of offspring generation Based on the number of assessments of the current population and the maximum number of assessments of the population Decide; Step S3. Set up a competitive decomposition mechanism: flexibly sort the subproblems, and then divide the population into a winning set and a losing set according to the flexible sorting strategy. ; Step S4. Set optimization: Optimize both the failure set and the winning set simultaneously. The winning set is configured to explore the potential of the neighborhood, while the failure set is configured to explore and try to find a better solution in the neighborhood. Step S5. Loop judgment execution: Repeat steps S2-S4 until the maximum number of evaluations for the population is reached. .

[0009] Specifically, the above technical solution effectively reduces the number of sub-problems through a sparse decomposition strategy, lowering the computational burden of decomposition-based multi-objective optimization in large-scale scenarios and providing feasibility assurance for optimization under expensive evaluation conditions. Secondly, by introducing a dynamic sub-problem partitioning mechanism based on competitive ranking, it can adaptively allocate different optimization strategies according to the search state of the sub-problems, avoiding the resource waste caused by using a uniform search method for all sub-problems. Furthermore, this method uses a classification surrogate model to pre-screen candidate solutions, which can stably distinguish potentially excellent solutions from inferior solutions even with a limited number of samples, significantly reducing unnecessary and expensive evaluation times of the objective function. In addition, this method effectively balances convergence performance and solution set diversity through differentiated exploration and development strategies, making it particularly suitable for high-dimensional, multi-objective optimization problems with high evaluation costs. Finally, this method has no specific requirements for the form of the objective function, possessing good versatility and engineering applicability, and can be widely applied in the design of complex engineering systems and intelligent optimization fields.

[0010] In steps S1-4 above, the training method for the SVM proxy model is as follows: training samples are generated from the current archive. The solution is constructed from the given information, where the decision vector for each sample is... Corresponding to input The initial label is set to negative. For each subproblem adjacency index set Calculate the scalarized objective function value for each sample in the archive. And select the best sample for each neighbor subproblem, and label it as the positive class. To avoid duplicate selections; after the training samples are constructed, the decision vector of each sample is normalized to the unit hypercubic space. ,in and These are the lower and upper bounds of the decision variables, respectively; then, the SVM model is trained using radial basis function kernels.

[0011] Above, in steps S1-4, the kernel function of the radial basis function kernel is defined as... Its penalty coefficient is ,in The kernel parameters are used; the trained SVM model is used to predict arbitrary input points. The categories and their decision function values ,in, is the Lagrange multiplier obtained during support vector machine training, used to characterize the contribution of samples to the classification hyperplane, and is calculated by constraining the relationship between sample class and kernel function during the support vector machine training phase; b is the bias term used to determine the position of the classification boundary, which is determined by the decision function value combined with the support vectors.

[0012] In step S2 above, the local crossover probability is adaptively updated. The calculation method is as follows: Maximum number of offspring generated The calculation method is as follows: ,in and It is a constant. This represents the number of times the current population has been assessed. This represents the maximum number of assessments for the population.

[0013] In step S4 above, the optimization process for the failure set is as follows: based on archives Construct the first Local SVM proxy model for each sub-problem; based on local crossover probability. Parents are selected from the neighborhood or the winner set; differential evolution is performed under the guidance of the SVM model to generate candidate solutions. ; Evaluate candidate solutions And update the archive. With the ideal point Then, the population is updated through a stable matching population selection mechanism. The optimization process for the winning set is as follows: based on archives Construct the first Local SVM proxy model for each sub-problem; partitioning decision variables into highly sensitive variables. and low-sensitivity variables Then one of its solutions Guided by the SVM proxy model, highly sensitive variables were sequentially processed. and low-sensitivity variables Continue the differential evolution search; for the generated candidate solutions Conduct a realistic assessment and update the archive. With the ideal point Then, the population is updated through a stable matching population selection mechanism. .

[0014] Specifically, in the above technical solution, a dynamic variable segmentation strategy is adopted for variable partitioning. This strategy divides the decision space into highly sensitive variables based on variable sensitivity. and low-sensitivity variables To improve the efficiency of high-dimensional optimization, the algorithm uses the variance of each variable in the population as a sensitivity proxy, assuming that variables with larger variances have a more significant impact on the objective function. The algorithm calculates the variance of all variables, sorts them in descending order, and then divides the variables with the largest variances into the top 50% according to a preset ratio (default is 50%). The rest are This strategy can dynamically focus on key variables, reduce the effective search dimensionality, and is suitable for high-dimensional problems where the importance of variables differs significantly. It is often used in co-evolutionary or hierarchical optimization frameworks. Its advantages lie in its simple computation and strong adaptability, but relying solely on variance may lead to misjudgments of importance, and the fixed split ratio lacks flexibility.

[0015] In step S4 above, a stability ranking mechanism is also included, which is used to select candidate solutions from the candidate solution set. Choose the optimal solution to the subproblem.

[0016] Preferably, for each candidate solution and the weight vector of each subproblem Calculate its modified Tchebycheff value. ,in To solve In the The value on each target, For the ideal point, The weight vector components are then calculated; finally, the weight vector of each solution relative to the subproblem is calculated. vertical distance ,in For the target vector, The angle between the candidate solution and the weight vector of the subproblem is denoted as .

[0017] Specifically, the stability ranking mechanism described above is used to assign stable representative solutions to each subproblem from the candidate solution set. For each subproblem, it selects the solution with the smallest Tchebycheff value from the unassigned candidate solutions as the potential optimal solution. If this solution has not been assigned, it is assigned directly; otherwise, by comparing vertical distances, the solution with the smaller distance is selected to ensure the representativeness of the subproblem's direction. This process is iterated until all subproblems obtain a unique representative solution, thereby ensuring the stability of each subproblem while avoiding redundant allocation of solutions and improving the uniformity and distribution quality of solutions in the target space. Finally, the solution set output by the stable matching mechanism serves as the stable representative for each subproblem, used for subsequent evolution or archive updates.

[0018] Without departing from the core idea of ​​this invention, the aforementioned technical solutions can be equivalently replaced as follows: First, in this invention, the multi-objective optimization problem is decomposed by a finite number of weight vectors. As an alternative implementation, the weight vectors can be generated using different uniform distribution strategies or adaptive adjustment strategies, such as weight vector adjustment methods driven by adaptive updates based on reference points and historical solution distribution, or hierarchical, non-uniform weight vector distribution methods to adapt to different forms of Pareto fronts. Second, the competition evaluation criteria for subproblems are not limited to non-dominated ranking, diversity indicators, or local density indicators. They can also be based on convergence speed, historical improvement magnitude, subproblem stability, or comprehensive evaluation results of multiple indicators. The number of subproblem sets is not limited to two categories and can be extended to multi-level competition sets to adapt to optimization needs of different scales or stages. Finally, the evaluation indicators used for subproblem competition ranking in this invention can take various forms. As an alternative, evaluation indicators based on target spatial distance, solution set coverage, improvement rate, or historical performance trends can be introduced, or weighted combination and multi-criteria decision-making methods can be used to comprehensively rank multiple indicators.

[0019] The present invention also provides a multi-objective sparse competitive decomposition system for automobile exterior design, which applies the aforementioned multi-objective sparse competitive decomposition method for automobile exterior design.

[0020] The present invention also provides a computer storage device that stores a plurality of instructions, which are adapted to be loaded by a processor and executed in steps S1-S5.

[0021] The beneficial effects of this invention are: This invention provides a multi-objective sparse competitive decomposition method and system for automobile exterior design. The method includes initialization processing, adaptive parameter updating, setting up a competitive decomposition mechanism, set optimization, and iterative judgment execution. It decomposes the multi-objective problem using a limited number of weight vectors, significantly reducing the size of sub-problems while ensuring the quality of solution set distribution, thereby improving the overall efficiency of large-scale and expensive multi-objective optimization. Furthermore, by comprehensively evaluating the search status of sub-problems and using flexible sorting to dynamically divide sub-problems into different competitive sets, different sub-problems play different roles in the search process and perform different evolutionary processes, achieving adaptive allocation of search resources. Attached Figure Description

[0022] Figure 1 The flowchart shows the multi-objective sparse competitive decomposition method for automobile exterior design provided by this invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments. However, the scope of protection of this invention is not limited to the specific embodiments described below.

[0025] like Figure 1 As shown, this embodiment provides a multi-objective sparse competitive decomposition method for automobile exterior design, the method including: Step S1. Initialization process, which is performed according to the following steps: Step S1-1: Using the simplex lattice point design method in 3D target space generation Uniform points Generate weight vector ; Step S1-2: Calculate the weight vector Calculate the Euclidean distance between every two points, sort the distances of each vector in ascending order, and select the nearest one. The set of neighbors forms an adjacency index. Obtain the neighbor matrix This is used for subsequent sub-problem grouping or neighborhood search; Step S1-3: Generate the initial population using a Latin hypercube The initial population and its target values ​​are then evaluated and stored in an external archive. The ideal point is initialized based on the objective function value of the current population. The population size is set to N, and the objective number is set to M. Then the ideal point... It consists of the minimum values ​​of each objective dimension, i.e. ( ),in , Let i be the i-th individual in the population. Let it be the function value of the j-th objective; Steps S1-4: Initialize a local support vector machine (SVM) surrogate model for each sub-problem to assist in sub-problem search, and train the SVM surrogate model to obtain a model for predicting arbitrary input points. An SVM proxy model for the categories and their decision function values; the training method of the SVM proxy model is as follows: training samples are generated from the current archive. The solution is constructed from the given information, where the decision vector for each sample is... Corresponding to input The initial label is set to negative. For each subproblem adjacency index set Calculate the scalarized objective function value for each sample in the archive. And select the best sample for each neighbor subproblem, and label it as the positive class. After the training samples are constructed, the decision vector of each sample is normalized to the unit hypercubic space. ,in and These are the lower and upper bounds of the decision variables, respectively; then, the SVM model is trained using a radial basis function kernel; the kernel function of the radial basis function kernel is defined as... Its penalty coefficient is ,in The kernel parameters are used; the trained SVM model is used to predict arbitrary input points. The categories and their decision function values ,in, is the Lagrange multiplier obtained during support vector machine training, used to characterize the contribution of a sample to the classification hyperplane, and is calculated by constraining the relationship between the sample class and the kernel function during the support vector machine training phase; b is the bias term used to determine the position of the classification boundary, which is determined by the decision function value combined with the support vectors. Step S2. Adaptive parameter update: Adaptively update local crossover probabilities and maximum number of offspring Among them, the maximum number of offspring generation Based on the number of assessments of the current population and the maximum number of assessments of the population Decision; Adaptive update of local crossover probabilities The calculation method is as follows: Maximum number of offspring generated The calculation method is as follows: ,in and It is a constant. This represents the number of times the current population has been assessed. The maximum number of assessments for the population; Step S3. Set up a competitive decomposition mechanism: flexibly sort the subproblems, and then divide the population into a winning set and a losing set according to the flexible sorting strategy. ; Step S4. Set optimization: Optimize both the failure set and the winning set simultaneously. The winning set is configured to explore the potential of the neighborhood, while the failure set is configured to explore and try to find a better solution in the neighborhood. The optimization process for the failure set is as follows: based on archives Construct the first Local SVM proxy model for each sub-problem; based on local crossover probability. Parents are selected from the neighborhood or the winner set; differential evolution is performed under the guidance of the SVM model to generate candidate solutions. ; Evaluate candidate solutions And update the archive. With the ideal point Then, the population is updated through a stable matching population selection mechanism. ; The optimization process for the winning set is as follows: based on archives Construct the first Local SVM proxy model for each sub-problem; partitioning decision variables into highly sensitive variables. and low-sensitivity variables Then one of its solutions Guided by the SVM proxy model, highly sensitive variables were sequentially processed. and low-sensitivity variables Continue the differential evolution search; for the generated candidate solutions Conduct a realistic assessment and update the archive. With the ideal point Then, the population is updated through a stable matching population selection mechanism. ; A stability ranking mechanism is set up to select candidate solutions. Choose the optimal solution to the subproblem; for each candidate solution and the weight vector of each subproblem Calculate its modified Tchebycheff value. ,in To solve In the The value on each target, For the ideal point, The weight vector components are then calculated; finally, the weight vector of each solution relative to the subproblem is calculated. vertical distance ,in For the target vector, The angle between the candidate solution and the weight vector of the subproblem is denoted as .

[0026] Step S5. Loop judgment execution: Repeat steps S2-S4 until the maximum number of evaluations for the population is reached. .

[0027] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A multi-objective sparse competitive decomposition method for automobile exterior design, characterized in that, The multi-objective sparse competitive decomposition method includes: Step S1. Initialization process, which is performed according to the following steps: Step S1-1: Using the simplex lattice point design method in 3D target space generation Uniform points Generate weight vector ; Step S1-2: Calculate the weight vector Calculate the Euclidean distance between every two points, sort the distances of each vector in ascending order, and select the nearest one. The set of neighbors forms an adjacency index. Obtain the neighbor matrix This is used for subsequent sub-problem grouping or neighborhood search; Step S1-3: Generate the initial population using a Latin hypercube The initial population and its target values ​​are then evaluated and stored in an external archive. The ideal point is initialized based on the objective function value of the current population. The population size is set to N, and the target number is set to M. Then the ideal point... It consists of the minimum values ​​of each objective dimension, i.e. ( ),in , Let i be the i-th individual in the population. Let it be the function value of the j-th objective; Steps S1-4: Initialize a local support vector machine (SVM) surrogate model for each sub-problem to assist in sub-problem search, and train the SVM surrogate model to obtain a model for predicting arbitrary input points. An SVM proxy model that categorizes and sets the values ​​of their decision functions; Step S2. Adaptive parameter update: Adaptively update local crossover probabilities and maximum number of offspring Among them, the maximum number of offspring generation Based on the number of assessments of the current population and the maximum number of assessments of the population Decide; Step S3. Set up a competitive decomposition mechanism: flexibly sort the subproblems, and then divide the population into a winning set and a losing set according to the flexible sorting strategy. ; Step S4. Set optimization: Optimize both the failure set and the winning set simultaneously. The winning set is configured to explore the potential of the neighborhood, while the failure set is configured to explore and try to find a better solution in the neighborhood. Step S5. Loop judgment execution: Repeat steps S2-S4 until the maximum number of evaluations for the population is reached. .

2. The multi-objective sparse competitive decomposition method according to claim 1, characterized in that, In steps S1-4, the SVM proxy model is trained as follows: training samples are generated from the current archive. The solution is constructed from the given information, where the decision vector for each sample is... Corresponding to input The initial label is set to negative. For each subproblem adjacency index set Calculate the scalarized objective function value for each sample in the archive. And select the best sample for each neighbor subproblem, and label it as the positive class. After the training samples are constructed, the decision vector of each sample is normalized to the unit hypercubic space. ,in and These are the lower and upper bounds of the decision variables, respectively; then, the SVM model is trained using radial basis function kernels.

3. The multi-objective sparse competitive decomposition method according to claim 2, characterized in that, In steps S1-4, the kernel function of the radial basis function kernel is defined as follows: Its penalty coefficient is ,in For kernel parameters; The trained SVM model is used to predict any input point. The categories and their decision function values ,in, is the Lagrange multiplier obtained during support vector machine training, used to characterize the contribution of samples to the classification hyperplane, and is calculated by constraining the relationship between sample class and kernel function during the support vector machine training phase; b is the bias term used to determine the position of the classification boundary, which is determined by the decision function value combined with the support vectors.

4. The multi-objective sparse competitive decomposition method according to claim 1, characterized in that, In step S2, the local crossover probability is adaptively updated. The calculation method is as follows: Maximum number of offspring generated The calculation method is as follows: ,in and It is a constant. This represents the number of times the current population has been assessed. This represents the maximum number of assessments for the population.

5. The multi-objective sparse competitive decomposition method according to claim 1, characterized in that, In step S4, the optimization process for the failure set is as follows: based on archives Construct the first Local SVM proxy model for each sub-problem; based on local crossover probability. Parents are selected from the neighborhood or the winner set; differential evolution is performed under the guidance of the SVM model to generate candidate solutions. ; Evaluate candidate solutions And update the archive. With the ideal point Then, the population is updated through a stable matching population selection mechanism. .

6. The multi-objective sparse competitive decomposition method according to claim 1, characterized in that, In step S4, the optimization process for the winning set is as follows: based on archives Construct the first Local SVM proxy model for each sub-problem; Decision variables are divided into highly sensitive variables and low-sensitivity variables Then one of its solutions Guided by the SVM proxy model, highly sensitive variables were sequentially processed. and low-sensitivity variables Continue the differential evolution search; for the generated candidate solutions Conduct a realistic assessment and update the archive. With the ideal point Then, the population is updated through a stable matching population selection mechanism. .

7. The multi-objective sparse competitive decomposition method according to claim 5 or 6, characterized in that, Step S4 further includes setting a stability ranking mechanism, which is used to select candidate solutions from the candidate solution set. Choose the optimal solution to the subproblem.

8. The multi-objective sparse competitive decomposition method according to claim 7, characterized in that, For each candidate solution and the weight vector of each subproblem Calculate its modified Tchebycheff value. ,in To solve In the The value on each target, For the ideal point, The weight vector components are then calculated; finally, the weight vector of each solution relative to the subproblem is calculated. vertical distance ,in For the target vector, The angle between the candidate solution and the weight vector of the subproblem is denoted as .

9. A multi-objective sparse competitive decomposition system for automobile exterior design, characterized in that, The multi-objective sparse competitive decomposition method for automobile exterior design as described in any one of claims 1 to 8 is applied.

10. A computer storage device storing a plurality of instructions, characterized in that, The instructions are applicable to being loaded by a processor and executed as described in any one of claims 1-8.

Images