Hyper-parameter Optimisation by Restrained Stochastic Hill Climbing.

Abstract

Abstract. Machine learning practitioners often refer to hyper-parameter optimisation (HPO) as an art form and a skill that requires intuition and experience; Neuroevolution (NE) typically employs a combination of manual and evolutionary approaches for HPO. This paper explores the integration of a stochastic hill climbing approach for HPO within a NE algorithm. We empirically show that HPO by restrained stochastic hill climbing (HORSHC) is more effective than manual and pure evolutionary HPO. Empirical evidence is derived from a comparison of: (1) a NE algorithm that solely optimises hyper-parameters through evolution and (2) a number of derived algorithms with random search optimisation integration for optimising the hyper-parameters of a Neural Network. Through statistical analysis of the experimental results it has been revealed that random initialisation of hyper-parameters does not significantly affect the final performance of the Neural Networks evolved. However, HORSHC, a novel optimisation approach proposed in this paper has been proven to significantly out-perform the NE control algorithm. HORSHC presents itself as a solution that is computationally comparable in terms of both time and complexity as well as outperforming the control algorithm.