Learning Individualized Hyperparameter Settings

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