I am working on a project that implements PSO to solve a nonlinear optimization equation (minimization problem). I want to ensure that PSO is not falling for premature convergence (i.e choosing a solution too early and claiming it to be the best solution).
Hi, this is not a simple question to answer !
But there are some points to focus on to modify the swarm behavior of PSO core algorithm :
- Initialize population using methods to maximize search space coverage (such as latin hypercubes, sequences, ...)
- try different hyper-parameter values (to poderate inertia vs social behavior)
- try different neighborhood topology.
Then there are other evolutions of the core algorithm, but other evolutions as TRIBES, FIPS, ...
you can find more at :
Clerc, M. et J. Kennedy. 2002, "The particle swarm - explosion, stability, and convergence in a multidimensional complex space"
Mendes, R., J. Kennedy et J. Neves. 2004, "The fully informed particle swarm : simpler, maybe better"
Clerc, M. 2003, "TRIBES - Un exemple d’optimisation par essaim particulaire sans paramètre de contrôle"
El Dor, A. 2012, Perfectionnement des algorithmes d’Optimisation par Essaim Particulaire. Application en segmentation d’images et en électronique (study of topologies, in french)
GOOD LUCK !
As pointed out by Peio Loubiere using another topology is useful. For multimodal functions it must be local, either fixed, adaptive or random. Even the old random one used in SPSO 2007 usually produces good results (Particle Swarm Central hppt://particleswarm.info Programs section).
In the ligth of the answers and guiding dirrections provided by Peio Loubiere and Maurice Clerc you may try a two-step-neighborhood search as a replacement for one-step search. Based on
Applied Mathematics Letters, Volume 16, Issue 3, April 2003, Pages 337-343
the 2-step search can help you for escaping from a rugged landscape where the 1-step search may be trapped with the best local solution.