An Exploration of Genetic Algorithms Operators for the Detection of Multiple Change-Points of Exceedances Using Non-homogeneous Poisson Processes and Bayesian Methods

In this paper it is presented an exploration of different strategies to generate solutions in a genetic algorithm for the detection of multiple change-points in univariate time series. The purpose is to find which combination of these is the optimal one while modelling times where there is an exceedance from a given threshold through Non Homogeneous Poisson Processes. Likewise, elements from information theory are taken to define a parsimonious model such that the explained phenomenon has a low memory usage and an optimal quantity of parameters which are estimated through a Bayesian approach. These elements define the objective function. Thus and after evaluating different operators it is found that the optimal strategy to generate and to combine new solutions is through a random keys initialization, selection of the parents through the ranks and Boltzmann tournament method or through a roulette strategy and using a fixed low mutation rate such that the diversity component is supplied through a neighborhood exploration while keeping the fitness of the solutions close to the real value.