gena (version 1.0.0)

pso: Particle Swarm Optimization

Description

This function allows to use particle swarm algorithm for numeric global optimization of real-valued functions.

Usage

pso(
  fn,
  gr = NULL,
  lower,
  upper,
  pop.n = 40,
  pop.initial = NULL,
  pop.method = "uniform",
  nh.method = "random",
  nh.par = 3,
  nh.adaptive = TRUE,
  velocity.method = "hypersphere",
  velocity.par = list(w = 1/(2 * log(2)), c1 = 0.5 + log(2), c2 = 0.5 + log(2)),
  hybrid.method = "rank",
  hybrid.par = 2,
  hybrid.prob = 0,
  hybrid.opt.par = NULL,
  hybrid.n = 1,
  constr.method = NULL,
  constr.par = NULL,
  random.order = TRUE,
  maxiter = 100,
  is.max = TRUE,
  info = TRUE,
  ...
)

Value

This function returns an object of class pso that is a list containing the following elements:

• par - particle (solution) with the highest fitness (objective function) value.

• value - value of fn at par.

• population - matrix of particles (solutions) of the last iteration of the algorithm.

• counts - a two-element integer vector giving the number of calls to fn and gr respectively.

• is.max - identical to is.max input argument.

• fitness.history - vector which i-th element is fitness of the best particle in i-th iteration.

• iter - last iteration number.

Arguments

fn

function to be maximized i.e. fitness function.

gr

gradient of the fn.

lower

lower bound of the search space.

upper

upper bound of the search space.

pop.n

integer representing the size of the population.

pop.initial

numeric matrix which rows are particles to be included into the initial population. Numeric vector will be coerced to single row matrix.

pop.method

the algorithm to be applied for a creation of the initial population. See 'Details' for additional information.

nh.method

string representing the method (topology) to be used for the creation of neighbourhoods. See 'Details' for additional information.

nh.par

parameters of the topology algorithm.

nh.adaptive

logical; if TRUE (default) then neighbourhoods change every time when the best known (to the swarm) fitnesses value have not increased. Neighbourhoods are updated according to the topology defined via nh.method argument.

velocity.method

string representing the method to be used for the update of velocities.

velocity.par

parameters of the velocity formula.

hybrid.method

hybrids selection algorithm i.e. mechanism determining which particles should be subject to local optimization. See 'Details' for additional information.

hybrid.par

parameters of the hybridization algorithm.

hybrid.prob

probability of generating the hybrids each iteration.

hybrid.opt.par

parameters of the local optimization function to be used for hybridization algorithm (including fn and gr).

hybrid.n

number of hybrids that appear if hybridization should take place during the iteration.

constr.method

the algorithm to be applied for imposing constraints on the particles. See 'Details' for additional information.

constr.par

parameters of the constraint algorithm.

random.order

logical; if TRUE (default) then particles related routine will be implemented in a random order.

maxiter

maximum number of iterations of the algorithm.

is.max

logical; if TRUE (default) then fitness function will be maximized. Otherwise it will be minimized.

info

logical; if TRUE (default) then some optimization related information will be printed each iteration.

...

additional parameters to be passed to fn and gr functions.

Details

Default arguments have been set in accordance with SPSO 2011 algorithm proposed by M. Clerc (2012).

To find information on particular methods available via pop.method, nh.method, velocity.method, hybrid.method and constr.method arguments please see 'Details' section of gena.population, pso.nh, pso.velocity, gena.hybrid and gena.constr correspondingly.

It is possible to provide manually implemented functions for population initialization, neighbourhoods creation, velocity updated, hybridization and constraints in a similar way as for gena.

By default function does not impose any constraints upon the parameters. If constr.method = "bounds" then lower and upper constraints will be imposed. Lower bounds should be strictly smaller then upper bounds.

Currently the only available termination condition is maxiter. We are going to provide some additional termination conditions during future updates.

Infinite values in lower and upper are substituted with -(.Machine$double.xmax * 0.9) and .Machine$double.xmax * 0.9 correspondingly.

By default if gr is provided then BFGS algorithm will be used inside optim during hybridization. Otherwise Nelder-Mead will be used. Manual values for optim arguments may be provided (as a list) through hybrid.opt.par argument.

For more information on particle swarm optimization please see M. Clerc (2012).

References

M. Clerc (2012). Standard Particle Swarm Optimisation. HAL archieve.

Examples

## Consider Ackley function
# \donttest{
fn <- function(par, a = 20, b = 0.2)
{
  val <- a * exp(-b * sqrt(0.5 * (par[1] ^ 2 + par[2] ^ 2))) +
         exp(0.5 * (cos(2 * pi * par[1]) + cos(2 * pi * par[2]))) -
         exp(1) - a
  return(val)
}

# Maximize this function using particle swarm algorithm

set.seed(123)
lower <- c(-5, -100)
upper <- c(100, 5)
opt <- pso(fn = fn, 
           lower = lower, upper = upper,
           a = 20, b = 0.2)
print(opt$par)
# }

## Consider Bukin function number 6

fn <- function(x, a = 20, b = 0.2)
{
  val <- 100 * sqrt(abs(x[2] - 0.01 * x[1] ^ 2)) + 0.01 * abs(x[1] + 10)
  return(val)
}

# Minimize this function using initially provided
# position for one of the particles
set.seed(777)
lower <- c(-15, -3)
upper <- c(-5, 3)
opt <- pso(fn = fn, 
           pop.init = c(8, 2),
           lower = lower, upper = upper,
           is.max = FALSE)
print(opt$par)