setsystemconfiguration.m

有监督自组织映射-偏最小二乘算法(A supervised self-organising map–partial least squares algorithm),可以用语多变量数据的回归分析

function [MLKP] = SetSystemConfiguration(MLKP);
%
% !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
% !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
% The following parameters are internal defaults or (experimental) oddities.
% Changing these  might affect the behaviour of the program substantially.
% !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
% !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
%
% for speed-up calculate the totla number of units in the input and output map
MLKP.Nunits=MLKP.Nrow*MLKP.Ncol;
% Maximum output is (consensus) winner, or use principle of Head and Shoulder (only for 'Cls')
MLKP.OutputCriterion='Max';
MLKP.OutputConsensus=1;
% determines whether output is based on Ymap ('Normal') or weighted on Ytr ('Weight')
MLKP.OutputMode='Normal';
% parameters for output criterion (in case output criterion 'HaS' is selected)
MLKP.Threshold = 0.0;
MLKP.Diff = 0.25;
% fraction of value range (per variable) used to initialise maps randomly
% or weight scaling factor if 'KStone' or 'Select' is chosen
MLKP.InitFract = 0.3;
% determines how neighbourhood shrinks during the iterations ('Lin' or 'Exp')
MLKP.Shrink='Exp';
% cut-off determines where the shrinking neighourhood size becomes equal to 0
MLKP.ShrinkCutOff=0.33;
% architecure of maps: 'Hex' (hexagonal configuration) or 'Grd' (rectangular grid)
MLKP.Architecture='Grd';
% shape of the neighbourhood function: 'Lin', 'Blk', 'Gau' or 'Mex'
MLKP.NeighbourFunction='Gau';
% determines whether a fast map update (K-Means clustering like) is invoked if
% the neighbourhood size for updating the XMap and YMap is shrunk to 'FastUpdateSize'
MLKP.FastUpdate='N';
MLKP.FastUpdateSize=0;
% reduce size of X by binning the variables
MLKP.BinnedData = 'N';
MLKP.BinnedSize = 2;
% weight objects proportionally to their class size
% 0: all objects are equally weighted with a value of 1
% 1: the smallest class of objects gets a weight of 1, the rest < 1
%
% !!! only applicable for 'Cls' problems
MLKP.PowerFact=0.125;
% invoke (if set to 'Y') a kernel transformation of Xtr and Xtest
MLKP.KernelTransformation='N';
MLKP.KernelType='rbf';
MLKP.KernelParameter1=10.0;
MLKP.KernelParameter2=3;
% if BDK is selected: update maps object by object ('Y') or batch by batch ('N')
MLKP.MLXModeObjectWise='Y';
% determines whether BDK operates in batch mode in alternating ('Y') or normal XYF mode ('N')
MLKP.MLXModeWeightsReverse='Y';
% determines whether the program operates in silent mode or not
MLKP.SilentMode='N';
% sets colormap mode and axis definition ('normal' or 'square')
MLKP.ColorMap='default';
MLKP.AxisMode='square';
% accept or refuse weight changes according to Boltzmann criterion
% 'Not': no Boltzmann seelection invoked
% 'Exp': exponential decay ('Exp')
% 'Thr': exponential RBF-like decay preceeded by hard threshold ('Thr')
MLKP.BoltzmannMode='Not';
% scaling factor for temperature
MLKP.BoltzmannK=1.0;
% threshold for 'Thr' mode of Boltzmann
MLKP.BoltzmannThreshold=0.95;
% determines whether training is stopped if exit criterion is met
MLKP.CheckConv='N';
% limit of relative absolute differences between old an new X maps before exit
MLKP.ExitLimit=1.0e-8;
% determines whether Y data are permuted for testing overtraining models
MLKP.PermutationTest='N';
% determines whether some noise (a fraction of the full range per variable)
% is added to the map weights
%
% !!! the noise amplitude is weighted by the rescaling factor for the weight decay
MLKP.AddNoise='N';
MLKP.NoiseFraction=0.001;
% number of latent variables in case PLS is selected directly as modeling method
MLKP.LatentVar=1;