Trying FITC and PITC approx. but the toolbox is broken and it won't work.

% MULTIGP: test 1, using SMOS, ASCAT and AMSR2

% Test 1, using SMOS, ASCAT and AMSR2

% Test 2, trying FITC optimization => toolbox won't work with these approx.

clc, clear

rng('default'), rng(0)

% modelType = 'gg';  % Gaussian kernel (for convolution) and Gaussian covariance

% modelType = 'lmc'; % Linear model of corregionalization

% Full (FTC), FITC approx., PITC approx.

% FITC and PITC approx. are broken in the toolbox, they won't work

modelApprox = 'ftc';  % 'ftc', 'fitc', 'pitc'

%% Set needed paths

addpath(genpath('GPmat'))

addpath('netlab')

    datetick, axis tight; grid on

    if n == 1, ax1 = axis; else ax2 = axis; axis([ax1(1:2) ax2(3:4)]); end

    if n == 1, ax1 = axis; else; ax2 = axis; axis([ax1(1:2) ax2(3:4)]); end

    title(ylabels{n})

    hold off

end

%% Options for multiGP model

% Common options

options = multigpOptions('ftc');

% MultiGP options

options = multigpOptions(modelApprox);

switch modelApprox

case {'fitc', 'pitc'}

    % initialInducingPositionMethod options

    % kmeansIsotopic, kmeansHeterotopic, espaced, espacedInRange, fixIndices

    % randomComplete, randomDataIsotopic, nonrandomDataIsotopic, randomDataHeterotopic

    options.initialInducingPositionMethod = 'randomDataIsotopic';

    options.numActive = 30;

    % Choose inducing points as a % of samples

    % options.numActive = fix(length(utc_ref) * 0.02);

    options.beta = 1e-1;

    options.fixInducing = false;

    options.kern.nout = size(Ytrain, 2);

end

% Other options I didn't tried

% options.kern.isArd = 0;

% options.includeNoise = false;

% The optimiser

options.optimiser = 'scg';

% Specific options

mdate = datestr(now, 'yyyymmdd_hhMM');

switch lower(modelType)

    case {'sim', 'gg'}

        fname = sprintf('SMOS_model_%s_site_%s_nlf_%d_sensors_%d_%s.mat', ...

            upper(modelType), site, options.nlf, sensors, mdate);

        fname = sprintf('SMOS_model_%s_%s_site_%s_nlf_%d_sensors_%d_%s.mat', ...

            upper(modelType), upper(modelApprox), site, options.nlf, sensors, mdate);

    case 'lmc'

        fname = sprintf('SMOS_model_%s_site_%s_nlf_%d_rank_%d_sensors_%d_%s.mat', ...

            upper(modelType), site, options.nlf, model.rankCorregMatrix, sensors, mdate);

        fname = sprintf('SMOS_model_%s_%s_site_%s_nlf_%d_rank_%d_sensors_%d_%s.mat', ...

            upper(modelType), upper(modelApprox), site, options.nlf, model.rankCorregMatrix, sensors, mdate);

end

rname = strrep(fname, '_model_', '_results_');

fname = ['models/' fname];

function model = trainLFMGP(Xtrain, Ytrain, options)

q = 1; % Input dimension

d = size(Ytrain, 2) + options.nlf;

X = cell(size(Ytrain, 2) + options.nlf, 1);

Y = cell(size(Ytrain, 2) + options.nlf, 1);

switch options.approx

case {'fitc', 'pitc', 'dtc', 'dtcvar'}

    % Do not include LFs with these approximations, the toolbox already

    % includes them.

    disp('Not including empty LFs')

    X = Xtrain';

    Y = Ytrain';

otherwise

    disp('Including empty LFs')

    % When we want to include the structure of the latent force kernel within

    % the whole kernel structure, and we don't have access to any data from the

    % latent force, we just put zeros in the vector X and empty in the vector y.

    X = cell(size(Ytrain, 2) + options.nlf, 1);

    Y = cell(size(Ytrain, 2) + options.nlf, 1);

    for j = 1:options.nlf

        Y{j} = [];

        X{j} = zeros(1, q);

        Y{i+options.nlf} = Ytrain{i};

        X{i+options.nlf} = Xtrain{i};

    end

end

q = 1; % Input dimension

d = size(Ytrain, 2); % Outputs

% Creates the model

warning('off', 'multiKernParamInit:noCrossKernel')