A Roulette Wheel Selection-based Data-driven Evolutionary Algorithm for Data-driven Optimization Problems

초록

Data-driven optimization problems (DDOPs) have become increasingly prevalent in real-world applications where analytical models are unavailable or expensive to evaluate. Data-driven evolutionary algorithms (DDEAs) are potentially useful in solving DDOPs by leveraging surrogate models to estimate the fitness of candidate solutions, thus reducing the number of expensive evaluations. However, surrogate models trained on limited data are prone to overfitting, which reduces their generalization performance and hinders the search process. To address this limitation, this paper proposes a novel roulette wheel selection-based DDEA, referred to as RWS-DDEA. The core idea is to construct an ensemble of surrogate models and apply a roulette wheel selection (RWS) strategy to select models for fitness prediction based on both accuracy and diversity. This approach enhances the robustness of fitness approximation by reducing the probability of overfitting. The proposed RWS-DDEA is evaluated on a set of DDOPs against several state-of-the-art DDEAs. Experimental results show that RWS-DDEA achieves competitive or superior performance across various metrics. Furthermore, ablation studies confirm the effectiveness of the RWS method, highlighting its role as a simple yet powerful enhancement to ensemble-based surrogate modeling. Overall, this work presents an effective model management strategy for the field of data-driven optimization and provides insights for future research on enhancing data efficiency and generalization in DDEAs.

키워드