Scalability of Population-based Stochastic Metaheuristics
Abstract:
The population based stochastic metaheuristics (e.g. evolutionary algorithms, differential evolution, swarm intelligence etc.) have been successfully applied in solving various optimization problems arising in data mining like, for instance, feature and instance selection, classification rules mining, reduction of prototype based models, clustering etc. In the context of big data analysis these metaheuristics face computational challenges related to large and high-dimensional data sets. Once the problem size increases, not only the search space increases significantly but the optimization problem landscape becomes more difficult to explore. Thus natural questions arise such as “are population-based metaheuristics scalable enough to address such challenges?” and “there are some mechanisms which can be used to improve the metaheuristics scalability?”.
This talk will address the scalability issue by discussing first the impact of the problem size on the balance between exploration and exploitation components of the search process and then by analyzing the effectiveness of cooperative coevolution in enhancing the metaheuristics scalability. The analysis mainly relies on the properties of difference-based metaheuristics with a focus on Differential Evolution (DE). The usage of DE as a case study is mainly motivated by the fact that it is one of the simplest but in the same time the most effective metaheuristics for continuous optimization problems. In the context of image processing, variants of DE have been used to estimate the parameters of the affine transformations arising in image registration or of deformable models used for image segmentation and also for endmembers extraction from hyperspectral images.
This talk will address the scalability issue by discussing first the impact of the problem size on the balance between exploration and exploitation components of the search process and then by analyzing the effectiveness of cooperative coevolution in enhancing the metaheuristics scalability. The analysis mainly relies on the properties of difference-based metaheuristics with a focus on Differential Evolution (DE). The usage of DE as a case study is mainly motivated by the fact that it is one of the simplest but in the same time the most effective metaheuristics for continuous optimization problems. In the context of image processing, variants of DE have been used to estimate the parameters of the affine transformations arising in image registration or of deformable models used for image segmentation and also for endmembers extraction from hyperspectral images.
Short Bio:
Daniela Zaharie obtained in 1997 a PhD degree in Mathematics (Probability and Statistics) at the West University of Timisoara, Romania, with a thesis on stochastic modelling of recurrent neural networks. Currently she is professor of Computer Science at the same university. Her main research topics are evolutionary computing, data mining, machine learning, computational biology and image processing. She authored and co-authored over 80 papers in these fields and was recently involved in several national and international projects related to nature-inspired metaheuristics, medical data mining, remote sensing and high performance computing.. She is co-editor of the proceedings of the international Symposium of Symbolic and Numeric Algorithms for Scientific Computing (http://synasc.ro) and member of the editorial board of Soft Computing and of Swarm and Evolutionary Computation journals.
