Generalized Automated Energy Function Selection for Protein Structure Prediction on 2D and 3D HP Models
The present work applies Algorithm Selection for automatically determining the best energy functions for search algorithms on Protein Structure Prediction. Protein Structure Prediction is a critical problem concerned with exploring the structure of a protein, given an amino acid sequence. For the sake of improved search performance, various models have been introduced such as the Hydrophobic Polar (HP) models. These models make PSP computationally approachable. There exists a large suite of algorithms introduces to solve PSP. As in all other problem domains, it is possible to see that the PSP algorithms considered to be strong can fail to address some problem instances. One way of providing further performance improvements without devising a new algorithm is to automatically determine the best possible algorithm for a given PSP problem instance. Algorithm Selection focuses on automatically choosing algorithms for any given problem solving scenario. Unlike the traditional use of AS, this study accommodates Algorithm Selection for specifying the best suited energy function used for search. The process is designed as a rank prediction task on 7 energy functions for Iterated Local Search as the optimization algorithm. The idea is to assess the best energy function that can guide the search process of Iterated Local Search on a per-instance basis. An experimental analysis is reported on 30 PSP instances, each half belongs to the 2D and 3D HP models, respectively. The results show that Algorithm Selection is capable to offer proper selection of the energy functions with robustness. Additionally, a brief algorithm and instance analysis is reported on the instance hardness and matching algorithmic capabilities. This analysis further provides insights on the instance hardness besides instance and algorithm similarities.