Learning Individualized Hyperparameter Settings

Abstract

The performance of optimization algorithms, and consequently of AI/machine learning solutions, is strongly influenced by the setting of their hyperparameters. Over the last decades, a rich literature has developed proposing methods to automatically determine the parameter setting for a problem of interest, aiming at either robust or instance-specific settings. Robust setting optimization is already a mature area of research, while instance-level setting is still in its infancy, with contributions mainly dealing with algorithm selection. The work reported in this paper belongs to the latter category, exploiting the learning and generalization capabilities of artificial neural networks to adapt a general setting generated by state-of-the-art automatic configurators. Our approach differs significantly from analogous ones in the literature, both because we rely on neural systems to suggest the settings, and because we propose a novel learning scheme in which different outputs are proposed for each input, in order to support generalization from examples. The approach was validated on two different algorithms that optimized instances of two different problems. We used an algorithm that is very sensitive to parameter settings, applied to generalized assignment problem instances, and a robust tabu search that is purportedly little sensitive to its settings, applied to quadratic assignment problem instances. The computational results in both cases attest to the effectiveness of the approach, especially when applied to instances that are structurally very different from those previously encountered.