Por favor, use este identificador para citar o enlazar este ítem:
http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/615| Inverse active machine learning in optimization processes with applications in astronomy | |
| JUAN CARLOS GOMEZ CARRANZA | |
| LUIS OLAC FUENTES CHAVEZ | |
| Acceso Abierto | |
| Atribución-NoComercial-SinDerivadas | |
| Evolutionary computation Learning by example Optimisation | |
| Traditional optimization algorithms have many drawbacks when applied to difficult practical problems. Although the search for a magic optimization algorithm able to solve any kind of problem is far from being achieved, it is possible to establish that problems can be grouped by classes, and in this manner we can propose efficient algorithms to solve problems within a certain class. In this work, we propose some efficient algorithms based on machine learning techniques to smartly solve a set of difficult practical optimization problems that are frequently found in many scientific areas. We are interested in global and constrained non-trivial practical problems with sets of real parameters that have not linear dependency, defined on continuous spaces, and where the target function is highly non-linear and based on a comparison between arrays, and where we want to find a parameterized model for certain object or process. The idea behind this work is to reach an acceptable approximation to global minimum of the target function for the problems defined, trying to use the least number of function evaluations. Trying to reduce the number of function evaluations is important because the function evaluation is generally the most expensive part of the optimization process. This work focuses the research in two kinds of algorithms: modified Evolution Strategies (ES+) and hybrid algorithms. The first one belongs to the class of evolutionary algorithms that have been taking popularity in recent years, and here is presented with a set of modifications done to the basic algorithm to accelerate the convergence. The second one is a hybrid that uses ES+ and either a traditional Hooke-Jeeves (HJ) algorithm or Locally Weighted Linear Regression (LWLR), an instance-based learning algorithm. For testing the efficacy and measure the efficiency of the optimization algorithms we use two schemas: first we try out the algorithms with two known test functions, where we know a priori the analytical minimum of the function; the first is an Euclidean distance from a vector of known parameters, and the second is the classical Rosenbrock’s function. Both of them are used under different conditions, varying initial starting points and number of parameters. In the second schema we apply the algorithms for solving two application problems, both in Astronomy because in this area is easy to find problems that are difficult to solve and that require a lot of computational resources. The propose | |
| Instituto Nacional de Astrofísica, Óptica y Electrónica | |
| 2007-02 | |
| Tesis de doctorado | |
| Inglés | |
| Estudiantes Investigadores Público en general | |
| Gómez-Carranza JC | |
| RADIOTELESCOPIOS | |
| Versión aceptada | |
| acceptedVersion - Versión aceptada | |
| Aparece en las colecciones: | Doctorado en Ciencias Computacionales |
Cargar archivos:
| Fichero | Tamaño | Formato | |
|---|---|---|---|
| GomezCaJC.pdf | 9.85 MB | Adobe PDF | Visualizar/Abrir |