Adaptive evolutionary algorithms for portfolio selection problems: state of the art and experimental analysis

Abstract

This thesis aims at solving complex portfolio selection problems by introducing an adaptive strategy for parameter control in EAs, with the aim of achieving accurate and robust solutions. In chapter 1 we review a broad set of parameter tuning and parameter control strategies, then we implement an adaptive policy, based on the parameter control technique proposed by Maturana (2010), on a variety of non-convex risk measures, that display many local optima, for which traditional minimization strategies like gradient descent methods are not suitable. The idea behind this method is to solve problems by managing the well-known EvE balance in the context of evolutionary computation, which is widely acknowledged as a key issue in terms of search performance. This approach allows the EA to use an appropriate parameter setting in different stages of the search process, typically by generating large improvements of the solution quality at the beginning and finally by fine-tuning the solution. We apply this method to large scale optimization problems; in particular, we start by considering relatively basic programming problems with easy constraints, then we take into account a set of NP-hard integer programming problems, which display well-known computational issues.