Automated Algorithm Selection: from Feature-Based to Feature-Free Approaches

Abstract We propose a novel technique for algorithm-selection, applicable to optimisation domains in which there is implicit sequential information encapsulated the data, e.g., online bin-packing. Specifically we train two types of recurrent neural networks predict packing heuristic bin-packing, selecting from four well-known heuristics. As input, RNN methods only use sequence item-sizes. This contrasts typical approaches algorithm-selection require model be trained using domain-specific instance features that need first derived input data. The are shown capable achieving within 5% oracle performance on between 80.88 and 97.63% instances, depending dataset. They also outperform classical machine learning models features. Finally, hypothesise proposed perform well when instances exhibit some structure results discriminatory with respect set test this hypothesis by generating fourteen new datasets increasing levels structure, show critical threshold required before delivers benefit.