Initialization Procedures for Multiobjective Evolutionary Approaches to the Segmentation Issue

Abstract

Abstract. Evolutionary algorithms have been applied to a wide variety of domains with successful results, supported by the increase of computational resources. One of such domains is segmentation, the representation of a given curve by means of a series of linear models minimizing the representation error. This work analyzes the impact of the initialization method on the performance of a multiobjective evolutionary algorithm for this segmentation domain, comparing a random initialization with two different approaches introducing domain knowledge: a hybrid approach based on the application of a local search method and a novel method based on the analysis of the Pareto Front structure.

A review of population initialization techniques for evolutionary algorithms

“…These techniques generally assume that the given optimization problem is a black-box puzzle. Hence, no specific knowledge about the region of interest or building blocks of potential solution is Tometzki two-stage stochastic mixed-integer 2011 [76] Guerrero segmentation 2012 [77] available before running the EA. In the absence of such prior knowledge about the problem, generic population initialization techniques can be used easily and effectively [11].…”

Multiobjective Local Search Techniques for Evolutionary Polygonal Approximation

“…This multi-objective nature is faced with a Multi-objective evolutionary algorithm. Local search algorithms may be introduced to enhance this approach, leading to several issues: the configuration to obtain the different individuals is hard to establish, and each of this individuals requires an independent execution of the local search algorithm, providing disappointing results [4]. This section will present alternative, parameter-free versions of two well known local search algorithms for polygonal approximation which provide a whole Pareto front of solutions: Top-Down and Bottom-up algorithms.…”

“…This algorithm is based on the SPEA2 [14] MOEA, according to the configuration presented in [5]. The default initialization process creates a uniform Pareto Front in terms of coverage of the objectives, as presented in [4]. This section will cover the comparison between the initial and final populations of the two techniques presented and the suggested initialization process.…”

“…In [5] a multi-objective evolutionary algorithm [1] is proposed for the multi-objective solution of the segmentation issue, while in [4] a comparison between different possible initializations was carried, focusing on the different results between a random initialization aiming at the coverage of the obtained Pareto fronts versus the results from different local search techniques. One of the detailed issues is the single-objective nature of the traditional techniques used, which required different executions with different parameters in order to obtain different individuals from the front, also introducing issues regarding the configuration of these techniques to obtain such different individuals.…”

“…This implementation will produce a whole Pareto front from a single execution without parametrization required from the user. These initializations will be later tested as initial populations for a MOEA approach, in order to determine whether they have successfully created better initial populations than the approach presented in [4] and how these initial populations translate into the final results of the algorithm.…”