In space engineering, the electronic component layout has a very important impact on the centroid stability and heat dissipation of devices. However, the diversity of components, a variety of spatial location constraints, and thermal performance constraints have brought great challenges to the layout design. Moreover, the thermodynamic simulation is very expensive and time-consuming. In a research paper recently published in Space: Science & Technology, Handing Wang from School of Artificial Intelligence, Xidian University, proposed a surrogate-assisted evolutionary algorithm with restart strategy to reduce the cost.
First of all, the author introduced some preliminaries to establish a basis for the research. The component layout optimization problems (CLOPs) was formulated as the expensive constrained optimization problems (ECOPs). The thermal layout simulation of electronic components was time-consuming, so it was expensive to evaluate the performance of a solution. Therefore, it was unrealistic to use amount of real function evaluations in the optimization process. One of the core issues in ECOPs was how to handle constraints. There were mainly three kinds of constraint handling methods. The first was the penalty function method to impose penalties on the constraints. The second method was the feasibility rule, which provided a comparison rule among the feasible solutions and the infeasible solutions. The stochastic ranking was also a kind of feasibility rule method. The third was the multi-objective method which regarded the constraints as the objective to transform the origin constrained optimization problems into the multi-objective optimization problems. In addition, radial basis function network (RBFN) and differential evolution (DA) were also introduced by the author.
Then, the author proposed surrogate-assisted evolutionary algorithm with restart strategy (SAEA-RS) for solving the ECOPs. The author first introduced the framework. The algorithm had two coordinate optimization processes. One was a global search process, and another one was a local search process. The local search can converge rapidly but easily fall into the local optimum, while the global search was on the contrary. To combine the advantages of both strategies, a restart strategy was introduced to restart the population when the search fell into the local optimum or fell behind the global search. A local search method was proposed to exploit the local region efficiently, and a new constraint violation calculation method was developed for handling the constraints. For adapting to the different characteristics of the two optimization processes, the feasibility ranking was applied to the local search, and the stochastic ranking was applied to the global search. Moreover, to reduce the number of the real function evaluations, the RBFN models were established for each real objective and constraint functions to evaluate the newly generated individuals. For improving the models’ prediction ability, a model management strategy was proposed by the author. The RBFN models were updated during the optimization process by the model management strategy.
Afterwards, ablation experiments and comparative experiments with other algorithms were conducted before applying the proposed algorithm to solve a CLOP on a space circuit board. Firstly, to verify the performance of each part of the proposed algorithm, the ablation experiments were carried out on the restart strategy and the local search. Therefore, there were three different algorithms which are a basic SAEA without the restart strategy and the local search and a SAEA with the local search (SAEA-LS) and the SAEA-RS. In this experiment, the surrogate model of all algorithms adopted the RBFN model, and the kernel function adopted cubic kernel function. The model management strategy and the constraint violation calculation method adopted the methods proposed by the author. According to the results, the SAEA-RS had the faster convergence speed than the SAEA and SAEA-LS due to the restart strategy. The constraint violation calculation method was benefit for improving the performance of the proposed algorithm, which can increase the convergence speed on some problems. Secondly, the author carried out the comparative experiment to compare the performance of the proposed algorithm with other existing algorithms. The SAEA-RS is compared with two state-of-the-art algorithms which are the global and local surrogate-assisted differential evolution algorithm (GloSADE) and the evolutionary algorithm with multiple penalty functions and multiple local surrogates (MPMLS). The comparative experiments were, respectively, conducted on three sets of benchmark problems, which are the CEC2006, the CEC2010, and the CEC2017. The experimental results showed that SAEA-RS had the highest convergence speed and found the best solutions on most benchmark problems. Besides, when the number of real evaluations was small, the advantages of SAEA-RS were more prominent and more suitable for expensive problems than GLoSADE and MPMLS. Finally, the proposed algorithm was applied to solve a CLOP on a space circuit board. The simulation results showed that the problem was well solved. The overall centroid of the components was centered which guaranteed the stability of the overall board. The components were all in contact with heat pipe so that the components can dissipate the heat through the heat pipes. In conclusion, the proposed algorithm converged under the small number of thermodynamic simulations, which greatly sped up the optimization process, which indicated the great significance of SAEA-RS to solve real-world engineering problems.
Author: Lei Han, Handing Wang, and Shuo Wang
Title of original paper: A Surrogate-Assisted Evolutionary Algorithm for Space Component Thermal Layout Optimization
Article link: https://doi.org/10.34133/2022/9856362
Journal: Space: Science & Technology
Affiliations: School of Artificial Intelligence, Xidian University, Xi’an, China
School of Computer Science, The University of Birmingham, Birmingham, UK