An evolutionary-based hyper-heuristic approach for optimal construction of group method of data handling networks

In this paper we present a novel method aiming at constructing Group Method of Data Handling networks (GMDH), assisted by hyper-heuristics algorithms. The proposed approach is based on an evolutionary hyper-heuristic, which completely automates the GMDH construction, by evolving the number of layers, the polynomial type and the number of selected nodes in each layer of the network. It results in a completely self-organized algorithm called Hyper Heuristic-GMDH (HH-GMDH). In the paper we focus on the definition of the hyper-heuristic approach proposed, including the basic heuristics to be evolved, the evolutionary algorithm encoding, and a comprehensive description of its evolutionary operators. We explore two versions of the HH-GMDH approach, depending on how a regularization parameter (λ) is determined in the algorithm. We have tested the proposed HH-GMDH algorithm in problems from UCI public repository and in two real problems: (1) temperature prediction in Barcelona’s airport and (2) total ozone content prediction at the Iberian Peninsula. In these problems, we show that the proposed HH-GMDH outperforms the classical GMDH network.