多模态优化论文
A PSO-Based Algorithm With Local Search for M ultimodal Optimization Without Constraints
Omar Andres Carmona CortesI
Andrew Rau-Chaplin2 Rafael Fernandes Lopes I
Departamento Academico de Informatica
Instituto Federal de Educa<;ao, Ciencia e Tecnologia do Maranhao
Av. Getulio Vargas, 04 -Monte Castelo Sao Luis -MA -65025-001 -Brazil Email: {omar.rafaelf}@ifma.edu.br
1
2 Faculty of Computer Science, Dalhousie University
6050 University A venue
Halifax, CA Email: arc@cs.dal.ca
algorithm named LHS, enhancing the exploration and exploitation capabilities of the canonical hybrid
PSO algorithm mixed with a new hybrid local search moderate or low objective variability is observed (pla
PSO, named PSOLHS, is examined against six
PSO and LHS.
Furthermore, a comparison
Abstract-The purpose of this paper is to present a
of local extremes are separated by large areas on which
PSO.
The
known multimodal functions and compared with both canonical
between evolutionary strategies
(ES) and MPSO-LS is
going to show how our approach outperforms these other techniques in almost all benchmark functions. All comparisons are based on a statistical t-test for supporting our results. dal,Optimization.
Index Terms-PSO,
1.
Hybrid Algorithm, Multimo
INTRODUCTION
The term global optimization refers to the process of attempting to find out the solution x* of a set of possible solutions S, which has the optimal value for some fitness function j. In other words, we are trying to find out the solution x* such that x -I-x* :::} j(x*) :::: j(x) [2], in maximization problems, or x -I-x* :::} j(x*) :::; j(x), in minimization problems. In both cases, the best solution is called global optima. Solutions other than global optima are called local optima.
A function to be optimized is monomodal if it has only one global optima and no locals optima. On the other hand, a multimodal function has many locals optima and, normally, one global optima, however, there are multimodal functions with many global optima as well.
Global optimization algorithms, based on mechanisms inherited from biology can be applied effectively when solving continuous problems, specially in difficult cases in which other classical computing strategies, such as gradient, have failed.
Continuous global optimization problems with multimodal objective functions, in which the basins of attraction
teaus), belong to the group of important difficult ones. Another type of difficulty is caused by the low regularity of the objective function when the gradient and the Hessian computation require costly approximative routines or is generally meaningless. [1]
In order to solve multimodal functions, many bioinspired approaches have been proposed. Particularly, different PSO (Particle Swarm Optimization) algorithms have been used for solving multimodal problems. These algorithms are mostly based on existing approaches used in the evolutionary algorithms [3].
Li et al. [4] investigates some mutations operators based on global best particle. Specifically, three mutations operators are used for helping PSO jump out of local optima: Cauchy, Gaussian and Levy. All mutation operators have an equal initial selection ratio with 1/3 and the mutation operators that result in higher fitness values has its ratio increased.
Pat and Hota [6] propose a modified and improved quantum-behaved particle swarm optimization (QPSO) [11] adding fitness weighted recombination, i.e., recombining two particles using arithmetic crossover, which were chosen by the roulette wheel selection.
Chen [7] uses hill climbing to improve the solution of a particle chosen at random. Then, he modified his basic PSO-LS by choosing the best particles (MPSO-LS) as initial solutions for local search, showing that this last approach presents better results.
Ozcan [3] adds a random walk component and a hill climber to enhance the exploration and exploitation capabilities of PSO. Khairy [5] proposes a PSO algorithm with local search algorithm where a small independent swarms is put around a selected particles within the specified
978-1-4673-0793-2/12/$31.00 ©2012 IEEE
热门文档