Automated algorithm selection for black-box optimization using light gradient boosting machine

Many evolutionary algorithms have been designed to address industrial black-box optimization problems in the real world. No single algorithm can outperform others across all problem instances. Algorithm selection methods aim to help users to automatically choose the best algorithm for new problems without expertise in evolutionary algorithm. However, the existing methods are implemented on a limited number of handcrafted benchmarks which lack practicality, and there is no general metric for measuring the best algorithm for black-box problems with unknown optimum. To tackle these issues, we propose an algorithm selection method for black-box optimization using light gradient boosting machine, where a tree-based random instance generation method is introduced to create diverse problem instances simulating real-world cases, and a metric is proposed to evaluate the performance of evolutionary algorithms on real-world black-box optimization considering both convergence speed and value. Experimental results show that our method achieves an accuracy of 72.23% on our generated dataset, and has lower computational cost compared to existing methods.