- Research Grants
|Support type:||Scholarships in Brazil - Scientific Initiation|
|Effective date (Start):||July 01, 2016|
|Effective date (End):||June 30, 2017|
|Field of knowledge:||Physical Sciences and Mathematics - Computer Science|
|Principal Investigator:||André Luis Debiaso Rossi|
|Home Institution:||Universidade Estadual Paulista (UNESP). Campus Experimental de Itapeva. Itapeva , SP, Brazil|
The wood quality classification according its defects is a task performed by industries which supply such raw material for proper use of the product. These procedures are generally performed by human operators trained for this purpose. However, this solution suffers from considerable drawbacks, such as the subjectivity and the tiredness of the operator, after performing the same task for a long period. For this reason, machine learning techniques have been used in quality classification and detection of many products, including wood. Despite the potential, these techniques have free parameters that must be provided by users and that directly influence the predictive performance and robustness of the induced models. Generally, the user sets the values for the parameters by trial and error, which is a unsystematic method and, therefore, often very expensive. Researchers have developed different algorithms and methods to deal with the problem of finding the best values for the free parameters. Recently, meta-heuristics based on evolutionary process, such as genetic algorithms, and swarm intelligence, as the Particle Swarm Optimization, have been used to treat this problem. Meta-heuristics have some basic heuristics to escape from the local minima, which are very common in the parameter tuning problem. This project aims to investigate the use of Particle Swarm Optimization for setting the parameter values for Neural Networks and Support Vector Machines applied to wood quality classification. The classification is made from images of wood boards according to their defects. It is expected that the optimization algorithm will achieve better values for the parameters when compared to ad hoc and trial and error approaches, resulting in models with higher predictive performance.