Algoritmos macroevolutivosdesde la simulación hasta el modelo estocástico

  1. Marín Sánchez, Jesús
Dirigée par:
  1. Ricard V. Solé Directeur/trice

Université de défendre: Universitat Politècnica de Catalunya (UPC)

Fecha de defensa: 18 janvier 2005

Jury:
  1. Pere Caminal Magrans President
  2. Jordi Delgado Pin Secrétaire
  3. Richard J. Duro Fernández Rapporteur
  4. Juan Pérez Mercader Rapporteur
  5. Juan Julián Merelo Guervós Rapporteur

Type: Thèses

Teseo: 127694 DIALNET

Résumé

This work includes the analysis of a novel stochastic algorithm named Macroevolutionary Algorithm with a theoretical model to make quantitative predictions about its behavior when is applied to optimize globally a function without constraints, It also includes the design and implementation of a general simulation environment to be easily applied to several problems and allowing to make comparisons with other optimization techniques in order to evaluate the balance between solution quality and efficiency. Through this study, we tried to provide to the algorithm the ability to allow an easier and clearer adjustment of its parameters that control such balance between exploration and exploration than standard genetic algorithms. In this way, though depending upon the problem under consideration, the premature falling into local optima could be reduced, as the performed experiments have demonstrated. Moreover, the algorithm performance ---expressed in terms of balance between solution quality and efficiency--- makes it a very competitive optimization method. The thesis gives especial relevance to the theoretical model that describes the macroscopic behavior of its nonlinear stochastic dynamics along the optimization process. The main idea considers the algorithm dynamics as if it would be a biological process to be modelled, obtaining a coupled system of non-linear equations. The model does use of analytical or statistic information about the function, represented as probability distributions according with fitness values and spatial locations of its points. So, the model is able to make predictions about the algorithm behavior as adjusted as good would be the quality of probability distribution. The model considers all algorithm parameter values and, for each time step, it gives quantitative measures as the found-solution quality or the number of function evaluations. As a consequence, this model would be useful in order to get the opti