Dynamic Adaptive Deep Learning approach for Algorithm Selection in BBOP

Abstract

Choosing the most suitable algorithm for tackling black-box continuous optimization problem (BBOP) is a difficult task. As numerous optimization algorithms are evaluated annually, there is an urgent need for automated methods for algorithm selection for single-objective black-box optimization. This technique is better known as the Algorithm Selection Problem, where an effective algorithm tailored to each particular problem can be chosen for that particular problem. It gained much importance in the last decades to find a means by which researchers could identify some existing algorithms best suited for working on particular problems, rather than inventing new ones. We will be proposing, in this paper, a dynamic adaptive model that shall make use of Deep Learning (DL) technology coupled with Exploratory Landscape Analysis techniques (ELA) for predicting the single best solver for any given problem set. In particular, we combine the Exploratory landscape analysis techniques to visualize the landscape of problem features with a number of high-performance algorithm portfolios to create deep learning models which predict the optimal solver for any given problem set. It outperforms the performance of conventional optimization algorithms on continuous black-box optimization problems. The proposed DL model will adapt dynamically to changing characteristics, concerning problem instances, thus being even more effective and efficient.