Initializing hyper-parameter tuning with a metaheuristic-ensemble method: a case study using time-series weather data


ÖZTÜRK M. M.

EVOLUTIONARY INTELLIGENCE, 2022 (Journal Indexed in ESCI) identifier identifier

  • Publication Type: Article / Article
  • Publication Date: 2022
  • Doi Number: 10.1007/s12065-022-00717-y
  • Title of Journal : EVOLUTIONARY INTELLIGENCE
  • Keywords: Echo state network, Hyper-parameter optimization, Metaheuristic initializer, Ensemble algorithm, Weather forecasting, ECHO STATE NETWORK, OPTIMIZATION ALGORITHM, SEARCH

Abstract

Hyper-parameter optimization (HO), regardless of the type of optimization, inherently not only increases the completion time of the algorithm to be optimized but also creates a remarkable computational burden. However, employing the most suitable HO technique for a specific problem may not be sufficient to improve the performance of the selected machine learning algorithm. In such cases, it is common to deploy default values of the initialization hyper-parameters of HO. Instead, a configured set of initialization hyper-parameters of HO is significantly more impactful than a default mode of HO. In this study, a metaheuristic ensemble technique is proposed to configure the initialization hyper-parameters of HO. The proposed method is devised after an extensive time analysis of metaheuristics and applied to Echo State Network (ESN). The experiment performed with weather forecast data shows that metaheuristic initialization methods are quite compatible with evolutionary algorithms. In the benchmark, the proposed method outperformed two alternatives. Probabilistic methods such as Bayesian optimization are not preferable for metaheuristic initialization methods, according to the results of the experiment. Metaheuristic hyper-parameter initialization methods can be performed by utilizing Random search that provides a moderate performance in which there are hardware-restricted sources. Last, the hyper-parameter called leakingrate of ESN is the most sensitive one and creates the largest churns in the prediction performance.