Automatic Tuning of a Local Search Algorithm for Estimating Biological Signals JPDs

摘要

The estimation of Joint Probability Distributions (JPD) permeates several bio-signal modeling real world applications. Aiming to provide an efficient solution strategy to a difficult problem that is to estimate the interaction of signals non-simultaneously recorded in biological systems, this paper describes the implementation of local search methods to solve the JPD estimation minimizing the computational time needed. Different strategies were implemented and defined by adjustable parameters in the search algorithms input. With sets of training instances, experiments were executed using Irace, a racing algorithm designed to tune algorithms automatically. The best candidates extracted from those experiments were tested on unseen instances and the results show that Irace were able to find versions of the algorithm with biased strategies that were successful in obtaining the same answers with average fewer iterations than the non-biased one. Moreover, this paper also discusses some peculiarities founded while solving the JPD estimation problem with Gaussian marginal distributions, regarding the robustness of the implemented algorithm according to its parameters and the different instances solved.