A data-driven self-adaptive parameter tuning framework for composite automobile part optimization design

Abstract

Population-based heuristic optimization algorithms are wildly used in the automobile optimization design. However, the hyper-parameter tuning has a significant effect on the performance of the most of the heuristic algorithms. In order to take full advantages of the heuristic optimization algorithms, this article proposes a data-driven framework for self-adaptive parameters tuning, which named DSPT. The DSPT framework divides the optimization process into two phases. In the learning phase, the knowledge is learned from abundant benchmark functions. The specifically designed performance metrics are used to relate the characteristics of different problems and algorithm performances. In the optimizing phase, the characteristics of a new problem are firstly extracted. According to the knowledge gained from the learning phase and the problem characteristics gained in this phase, rather than predetermined parameters based on experience, the key parameters are tuned automatically. Therefore, the optimization can continue more efficiently. Based on the newly proposed social spider inspired particle swarm optimization algorithm, the proposed framework is successfully applied to the multi-scale lightweight design of four different composite automobile parts.