Emilia Tantar Research Xplore

My view on multi-objective optimization and other related topics


Multi-objective particle methods

Stochastic particle models emerged as one of the most fascinating connection points between applied probability, Bayesian inference and computer science. These particle models can be thought as a new universal and sophisticated Monte Carlo type technique for sampling complex distributions in highly dimensional state spaces. At the same time, interacting particle methods evolved into a revolutionary stochastic simulation technology for solving complex estimation and optimization problems. A unified and simplified perspective is considered in the following, with no further details or implications on branching, adaptive or self-interacting methods. An iterative algorithm can be defined where, at each iteration, all particles forming the population, are subject to independent transitions, followed by a selection step where high-potential particles are duplicated at the expense of the lower ranked ones which are discarded. As independent particles are considered, an Evolutionary Algorithm (EA) – like intrinsic parallel approach results, enabling the exploitation of all data and algorithm centered parallel techniques which exist in the EAs domain. In the multi-objective context, each particle is represented by an archive.

Dynamic multi-objective optimization on online stochastic environments

The focus of this research is set on providing the main structure of dynamic multi-objective optimization algorithms in a learning environment using estimation of distribution (EDA)  like algorithms . The stability and performance guarantees of the evolutionary process are to be controlled by interacting Markov chains techniques.

Landscape analysis in multi-objective combinatorial optimization

My research focuses on landscape analysis for multi-objective problems, that provides solutions for ensuring performance guarantees for approximation algorithms dealing with large scale problem instances. Landscape analysis represents the most natural way of translating visual perception in guiding
tools and techniques.

The large variety of visual perceptions of the solutions landscape provides the aspects of interest in landscape analysis studies. Furthermore, the visual perception of the search space can be integrated in the searching techniques
as to supply tools which adapt themselves to the encountered landscape characteristics. During my thesis, I have firstly identified the complex role played by landscape analysis for the optimization process (in two distinct moments: a priori and online). Furthermore, landscape analysis
has been positioned as regards the different studied characteristics. Among my achievements in this direction, I have integrated landscape analysis as :

  1. a tool for establishing the complexity induced by the topological structure of the landscape of feasible solutions (through a priori studies) ;
  2. a guiding mechanism for interactive evolutionary optimization methods ;
  3. a guiding component for adaptive methods.

eXTReMe Tracker

var EXlogin=’tantare’ // Login
var EXvsrv=’s11′ // VServer


Written by emiliatantar

October 9, 2009 at 3:54 pm