This collaborative project proposes the development of new hybrid approaches (based on metaheuristics and machine learning) to tackle a variety of real-world problems. Specifically, we combine recent advances in these types of techniques to design nominal and ordinal classification models (i.e., multiclass classification when orders can be defined for different categories), regression models, and time series segmentation/prediction models. From a methodological point of view, one of the subprojects (University of Córdoba, UCO) focuses on machine learning aspects, tackling the singular problem of ordinal classification (unbalanced classification, in the context of weekly supervised classification) and new deep learning models for ordinal classification. The second subproject (University of Alcalá de Henares, UAH) will focus on the development of advanced nature-inspired algorithms, specifically coevolution in one population, portfolio search algorithms, automatic estimation of parameters, and fractal-inspired algorithms. The coordination of both proponent teams aims to obtain a new hybrid optimization method for solving problems in two different fields. The proponents have contrasting experience in 1) renewable energy resource evaluation, climatology and meteorology, and 2) biomedical issues. The first area addresses the effects of climate change on the spatio-temporal distribution of renewable energy resources, together with the prediction of extreme events in meteorology (e.g. heat waves, cold waves, wind lamps, low visibility events) and climatology (e.g. spatio-temporal distribution of severe droughts and their precursors). Regarding biomedicine, we propose to address issues such as donor and recipient matching in liver transplants, medical image analysis for melanoma detection and Parkinson’s disease detection, including varying degrees of severity, and treatment allocation for HIV/HCV patients. To address the diversity of real-world problems being considered, the research team includes subject matter experts. These real-world problems often generate time-series data whose complexity makes traditional prediction problems (ARIMA models, recurrent neural networks, etc.) suboptimal. We believe that the developed hybrid metaheuristic can segment and make predictions on this kind of data. The use of advanced metaheuristics is also explored to tune the network architecture of deep learning models (nominal or ordinal) related to image analysis. The proposed project will be managed based on a specific communication and results dissemination plan designed and coordinated by a FECYT certified entity with dedicated personnel. The purpose of this communication plan is to disseminate the results of the project to professional and general audiences. This includes publication of results in selected high-impact journals, communication of results in mass media, and participation in activities promoting scientific culture.
