pso-package

A Particle Swarm Optimizer

The package provides an implementation of particle swarm optimization which is consistent with the standard PSO 2007 and 2011 by Maurice Clerc et al. Additionally a number of ancillary routines are provided for easy testing and graphics. package

Details

Package:	pso
Type:	Package
Version:	1.0.4
Date:	2022-04-12
License:	LGPL-3
Depends:	methods

The core function in the package is psoptim which can be used as a drop in replacement for optim. When used without additional control parameters the implementation is intended to be equivalent to SPSO 2007 (by M. Clerc et al.).

Control parameters can be specified for SPSO 2011 (in its basic implementation), to clamp the maximal velocity, provide restarting when the swarm converges to a region as well as using BFGS as a local search strategy. See psoptim for details.

Author(s)

Maintainer: Claus Bendtsen papyrus.bendtsen@gmail.com

See Also

optim.

Examples

## Not run:

## Some examples of using the functions in the package

## Using basic "optim" interface to minimize a function
set.seed(1)
psoptim(rep(NA,2),function(x) 20+sum(x^2-10*cos(2*pi*x)),
        lower=-5,upper=5,control=list(abstol=1e-8))

## Parabola
p <- test.problem("parabola",10) # one local=global minimum
set.seed(1)
o1 <- psoptim(p,control=list(trace=1,REPORT=50))
show(o1)

set.seed(1)
o2 <- psoptim(p,control=list(trace=1,REPORT=50,w=c(.7,.1)))
show(o2)

set.seed(1)
o3 <- psoptim(p,control=list(trace=1,REPORT=1,hybrid=TRUE))
show(o3) ## hybrid much faster

## Griewank
set.seed(2)
p <- test.problem("griewank",10) # lots of local minima
o1 <- psoptim(p,control=list(trace=1,REPORT=50))
show(o1)

## The above sometimes get stuck in a local minima.
## Adding a restart to increase robustness.
set.seed(2)
o2 <- psoptim(p,control=list(trace=1,REPORT=50,reltol=1e-4))
show(o2)

## An then adding the hybrid
set.seed(2)
o3 <- psoptim(p,control=list(trace=1,REPORT=50,reltol=1e-4,
              hybrid=TRUE,hybrid.control=list(maxit=10)))
show(o3)

## Rosenbrock
set.seed(1)
p <- test.problem("rosenbrock",1)
o1 <- psoptim(p,control=list(trace=1,REPORT=50))
show(o1)

## Change to fully informed
set.seed(1)
o2 <- psoptim(p,control=list(trace=1,REPORT=50,p=1))
show(o2)

## Rastrigin
p <- test.problem("rastrigin",10)
set.seed(1)
o1 <- psoptim(p,control=list(trace=1,REPORT=50))
show(o1)

set.seed(1)
o2 <- psoptim(p,control=list(trace=1,REPORT=50,hybrid=TRUE,
              hybrid.control=list(maxit=10)))
show(o2) # better
plot(o1,xlim=c(0,p@maxf),ylim=c(0,100))
lines(o2,col=2) # and much faster convergence

## Ackley
set.seed(1)
p <- test.problem("ackley",10)
o1 <- psoptim(p,control=list(trace=1,REPORT=50))
show(o1)
## End(Not run)