Implicit model selection using variable transformation in Estimation of Distribution Algorithms
DEI - Aula 2A
18 luglio 2012
The Estimation of Distribution Algorithms belong to the family of Model Based Search strategies for optimization and they employ a probability model, able to capture the features of high quality solutions, which is iteratively estimated from a set of promising solutions and sampled to generate new candidates for the next iteration. The problem of the choice of the model is crucial, indeed most of EDA literature focuses on employing model selection strategies known from statistics and machine learning. In this work we address the model selection problem from a different perspective: instead of learning a complex, high dimensional probability model we fix a low dimensional model and search for a variable transformation which maximize the likelihood of the transformed individuals with respect to the fixed model chosen a-priori. This approach, along with the particular class of variable transformations we consider, leads to the introduction of a novel class of EDAs, called Function Composition Algorithms, which we test and validate against a set of well known benchmark functions.
Area di ricerca:
Intelligenza artificiale, robotica e computer vision