From Eigenvector Research Documentation Wiki
Revision as of 14:58, 7 November 2016 by imported>Mathias (→‎See Also)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search


Genetic algorithm for variable selection with PLS.


model = gaselctr(x,y,options)
[fit,pop,cavfit,cbfit] = gaselctr(x,y,options)


GASELCTR uses a genetic algorithm optimization to minimize cross validation error for variable selection.


  • x = the predictor block (x-block), and
  • y = the predicted block (y-block) (note that all scaling should be done prior to running GASELCTR).


  • model = a standard GENALG model structure with the following fields:
    • modeltype: 'GENALG' This field will always have this value.
    • datasource: {[1x1 struct] [1x1 struct]}, structures defining where the X- and Y-blocks came from.
    • date: date stamp for when GASELCTR was run.
    • time: time stamp for when GASELCTR was run.
    • info: 'Fit results in "rmsecv", population included variables in "icol"', information field describing where the fitness results for each member of the population are contained.
    • rmsecv: fitness results for each member of the population, for X MxN and Mp unique populations at convergence then rmsecv will be 1xMp.
    • icol: each row of icol corresponds to the variables used for that member of the population (a 1 [one] means that variable was used and a 0 [zero] means that it was not), for X MxN and Mp unique populations at convergence then icol will be MpxN, and
    • detail: [1x1 struct], a structure array containing model details including the following fields:
      • avefit: the average fitness at each generation.
      • bestfit: the best fitness at each generation, and
      • options: a structure corresponding to the options discussed above.

For the second output syntax shown above,

  • fit is the same as model.rmsecv
  • pop is the same as model.icol
  • cavfit is the same as model.detail.avefit
  • cbfit is the same as model.detail.bestfit


options is a structure array with the following fields:

  • plots: ['none' | {'intermediate'} | 'replicates' | 'final' ] Governs plots.
    • 'final' gives only a final summary plot.
    • 'replicates' gives plots at the end of each replicate.
    • 'intermediate' gives plots during analysis.
    • 'none' gives no plots.
  • display: [{'on'}| 'off' ] governs output to the command window.
  • popsize: {64} the population size (16<popsize<256 and popsize must be divisible by 4),
  • maxgenerations: {100} the maximum number of generations (25<mg<500),
  • mutationrate: {0.005} the mutation rate (typically 0.001<mt<0.01),
  • windowwidth: {1} the number of variables in a window (integer window width),
  • convergence: {50} percent of population the same at convergence (typically cn=80),
  • initialterms: {30} percent terms included at initiation (10<it<50),
  • crossover: {2} breeding cross-over rule (cr = 1: single cross-over; cr = 2: double cross-over),
  • algorithm: [ 'mlr' | {'pls'} ] regression algorithm,
  • ncomp: {10} maximum number of latent variables for PLS models,
  • cv: [ 'rnd' | {'con'} ] cross-validation option ('rnd': random subset cross-validation; 'con': contiguous block subset cross-validation),
  • split: {5} number of subsets to divide data into for cross-validation,
  • iter: {1} number of iterations for cross-validation at each generation,
  • preprocessing: {[ ] [ ]} a cell containing standard preprocessing structures for the X- and Y-blocks respectively (see PREPROCESS),
  • preapply: [ {0} | 1 } If 1, preprocessing is applied to data prior to GA. This speeds up the performance of the selection, but may reduce the accuracy of the cross-validation results. Output "fit" values should only be compared to each other. A full cross-validation should be run after analysis to get more accurate RMSECV values.
  • reps: {1} the number of replicate runs to perform,
  • target: a two element vector [target_min target_max] describing the target range for number of variables/terms included in a model n. Outside of this range, the penaltyslope option is applied by multiplying the fitness for each member of the population by:
penaltyslope*(target_min-n) when n<target_min, or
penaltyslope*(n-target_max) when n>target_max.
Field target is used to bias models towards a given range of included variables (see penaltyslope below),
  • targetpct: {1} flag indicating if values in field target are given in percent of variables (1) or in absolute number of variables (0), and
  • penaltyslope: {0} the slope of the penalty function (see target above).


To use mean centering outside the genetic algorithm (no additional centering will be performed within the algorithm) do the following:

  x2 = mncn(x);
  y2 = mncn(y);
  [fit,pop] = gaselctr(x2,y2);

To use mean centering inside the genetic algorithm (centering will be performed for each cross-validation subset) do the following:

  options = gaselctr('options');
  options.preprocessing{1} = preprocess('default', 'mean center');
  options.preprocessing{2} = preprocess('default', 'mean center');
  [fit,pop] = gaselctr(x2,y2,options);

See Also

calibsel, fullsearch, genalg, genalgplot, ipls,Genetic Algorithms for Variable Selection