Automated Portfolio Generation for Selection Hyper-heuristics: an Application to Protein Structure Prediction on 2D HP Model

Abstract

The present study introduces two methods for constructing portfolios of low-level heuristics, to be used by Selection Hyper-heuristics to solve Protein Structure Prediction with the 2D HP model. Protein Structure Prediction is useful and practical for various application domains, particularly in medicine. As the protein structures are linked to the proteins' functionalities, it is critical to have efficient strategies to identify the proteins' structures. The traditional methods happening in the lab environments are impractical and impossible to go with the pace of the requirements for determining certain protein structures. Algorithmic approaches come into play to fill this gap by utilizing computational resources. Despite the relevant algorithm development efforts, there is no single algorithm that can truly address this structure prediction problem. Different algorithms take the lead on distinct problem scenarios. Selection Hyper-heuristics provide a way of benefiting from multiple algorithms, i.e. low-level heuristics, on-the-fly. However, the choice of those algorithms largely affect the performance. Thus, it is critical to come up with the low-level heuristic sets or portfolios that can be effectively used by Selection Hyper-heuristics for solving a specific target problem. This paper investigates two clustering-based methods for addressing the portfolio design task. The reported experimental analysis shows that the proposed methods are able to deliver competitive portfolios with significantly less number of low-level heuristics than the human-designed complete low-level heuristic set.