source: branches/OKBJavaConnector/ECJClient/src/ec/select/README @ 9449

This package provides various common selection operators.  In ECJ, selection
operators (called SelectionMethods) appear as leaf nodes in breeding pipelines.
If you're having trouble selecting, we suggest you use TournamentSelection,
perhaps with a tournament size of 2.

Some of these methods require that all fitnesses be >= 0.

Other methods, when first requested to select an individual, buffer up a large
number of precomputed selections to provide later, or compute certain
statistics to be used later.  Such selection methods cannot be used efficiently
in steady state algorithms.

ec.select.MultiSelection 
  This isn't a selection method so much as a utility procedure.  The
  algorithm maintains some number of subsidiary selection methods
  (the 'selects').  When asked to produce a child, it picks a subsidiary
  according to its associated probability; it then asks for a child from
  that subsidiary.  The probability is defined in the 'probability'
  parameter in the subsidiary pipeline.  Compare to 
  ec.breed.MultiBreedingPipeline, which does more or less the same
  thing, but as a BreedingPipeline, always copies the children produced
  from its sources.  MultiSelection, like a good selection operator,
  makes no such copies (and so is a bit more efficient in this context).

ec.select.BestSelection
  Always selects uniformly from among the best N individuals in the
  population.  Can also optionally select from among the N worst.

ec.select.TournamentSelection   [acceptable for Steady State]
  When asked to select an individual, first picks N individuals at
  random with replacement, then returns the fittest individual among
  those N.  The size of N is known as the tournament size, and you
  specify this.  Can also optionally select the worst among the N. 

ec.select.FitProportionateSelection
  Selects from among the individuals in the population in proportion
  to their fitness values (which must all be >= 0).

ec.select.SUSSelection
  Performs Stochastic Universal Sampling selection, a variant of
  fitness proportionate selection which guarantees that sufficiently
  fit individuals will always be selected at least once.  When first
  asked to select, SUSSelection computes selection choices for an entire
  populations' worth of individuals beforehand.  It then begins handing
  these choices out in response to successive requests.  When the choices
  have been depleted, SUSSelection computes another populations's worth.
  Like fitness proportionate selection, SUSSelection requires that all
  fitnesses be >= 0.

ec.select.GreedyOverselection
  A version of FitProportionateSelection used in early forms of genetic
  programming.  The individuals are divided by fitness into two groups:
  the good group and the bad group.  To select an individual, the 
  algorithm first picks either the good group or the bad group, then
  performs fitness proportionate selection within that group.  You 
  specify the size of the good group and the probability that it will
  be selected.  Like fitness proportionate selection, SUSSelection
  requires that all fitnesses be >= 0.

ec.select.RandomSelection   [acceptable for Steady State]
  Always provides a random individual.  This is equivalent to a
  tournament selection with a tournament size of 1.

ec.select.FirstSelection    [acceptable for Steady State]
  Always selects the first individual in the population array.  Only
  used for debugging purposes.

ec.select.BoltzmannSelection    [NOTE: UNTESTED]
  Performs Boltzman Selection, variation of fitness proportionate
  selection which makes fitness differences insignificant initially
  and more prominent as time goes on (think Simulated Annealing).
  Each generatiowe compute a temperature, defined as the initial
  temperature (which you define) minus the current generation times
  the cooling rate (which you also define).  If the temperature is
  below 1.0, the fitnesses are not adjusted.  Otherwise each
  fitness is adjusted as E^(fitness / temperature).

ec.select.SigmaScalingSelection   [NOTE: UNTESTED]
  Performs Sigma Scaling, a variation of fitness propoprtionate
  selection which stretches fitnesses out if they have been
  all bunched up together, enabling fitness proportionate 
  selection methods to still distinguish between them.  We begin
  by computing the mean and standard deviation of the current
  fitnesses.  If the standard deviation is not zero, we then
  adjust each fitness as 1 + (fitness - mean)/(2 * standardDeviation).
  If the standard deviation is zero, all fitnesses are set to 1.
  If any fitness is less than a minimal acceptable fitness (the "floor"),
  it is set to the floor.  This prevents fitnesses from getting stretched
  so small that they cannot compete.  You specify the floor.