Daniel@0: function net = gp(nin, covar_fn, prior)
Daniel@0: %GP	Create a Gaussian Process.
Daniel@0: %
Daniel@0: %	Description
Daniel@0: %
Daniel@0: %	NET = GP(NIN, COVARFN) takes the number of inputs NIN  for a Gaussian
Daniel@0: %	Process model with a single output, together with a string COVARFN
Daniel@0: %	which specifies the type of the covariance function, and returns a
Daniel@0: %	data structure NET. The parameters are set to zero.
Daniel@0: %
Daniel@0: %	The fields in NET are
Daniel@0: %	  type = 'gp'
Daniel@0: %	  nin = number of inputs
Daniel@0: %	  nout = number of outputs: always 1
Daniel@0: %	  nwts = total number of weights and covariance function parameters
Daniel@0: %	  bias = logarithm of constant offset in covariance function
Daniel@0: %	  noise = logarithm of output noise variance
Daniel@0: %	  inweights = logarithm of inverse length scale for each input 
Daniel@0: %	  covarfn = string describing the covariance function:
Daniel@0: %	      'sqexp'
Daniel@0: %	      'ratquad'
Daniel@0: %	  fpar = covariance function specific parameters (1 for squared exponential,
Daniel@0: %	   2 for rational quadratic)
Daniel@0: %	  trin = training input data (initially empty)
Daniel@0: %	  trtargets = training target data (initially empty)
Daniel@0: %
Daniel@0: %	NET = GP(NIN, COVARFN, PRIOR) sets a Gaussian prior on the parameters
Daniel@0: %	of the model. PRIOR must contain the fields PR_MEAN and PR_VARIANCE.
Daniel@0: %	If PR_MEAN is a scalar, then the Gaussian is assumed to be isotropic
Daniel@0: %	and the additional fields NET.PR_MEAN and PR_VARIANCE are set.
Daniel@0: %	Otherwise,  the Gaussian prior has a mean defined by a column vector
Daniel@0: %	of parameters PRIOR.PR_MEAN and covariance defined by a column vector
Daniel@0: %	of parameters PRIOR.PR_VARIANCE. Each element of PRMEAN corresponds
Daniel@0: %	to a separate group of parameters, which need not be mutually
Daniel@0: %	exclusive. The membership of the groups is defined by the matrix
Daniel@0: %	PRIOR.INDEX in which the columns correspond to the elements of
Daniel@0: %	PRMEAN. Each column has one element for each weight in the matrix, in
Daniel@0: %	the order defined by the function GPPAK, and each element is 1 or 0
Daniel@0: %	according to whether the parameter is a member of the corresponding
Daniel@0: %	group or not.  The additional field NET.INDEX is set in this case.
Daniel@0: %
Daniel@0: %	See also
Daniel@0: %	GPPAK, GPUNPAK, GPFWD, GPERR, GPCOVAR, GPGRAD
Daniel@0: %
Daniel@0: 
Daniel@0: %	Copyright (c) Ian T Nabney (1996-2001)
Daniel@0: 
Daniel@0: net.type = 'gp';
Daniel@0: net.nin = nin;
Daniel@0: net.nout = 1;  % Only do single output GP
Daniel@0: 
Daniel@0: % Store log parameters
Daniel@0: net.bias = 0;
Daniel@0: net.min_noise = sqrt(eps);  % Prevent output noise collapsing completely
Daniel@0: net.noise = 0;
Daniel@0: net.inweights = zeros(1,nin);  % Weights on inputs in covariance function
Daniel@0: 
Daniel@0: covarfns = {'sqexp', 'ratquad'};
Daniel@0: 
Daniel@0: if sum(strcmp(covar_fn, covarfns)) == 0
Daniel@0:   error('Undefined activation function. Exiting.');
Daniel@0: else
Daniel@0:   net.covar_fn = covar_fn;
Daniel@0: end
Daniel@0: 
Daniel@0: switch covar_fn
Daniel@0: 
Daniel@0:   case 'sqexp'		% Squared exponential
Daniel@0:     net.fpar = zeros(1,1);  % One function specific parameter
Daniel@0:     
Daniel@0:   case 'ratquad' 	% Rational quadratic
Daniel@0:     net.fpar = zeros(1, 2); % Two function specific parameters
Daniel@0: 
Daniel@0:   otherwise
Daniel@0:     error(['Unknown covariance function ', covar_fn]);
Daniel@0: end
Daniel@0: 
Daniel@0: net.nwts = 2 + nin + length(net.fpar);
Daniel@0: 
Daniel@0: if nargin >= 3
Daniel@0:   if size(prior.pr_mean) == [1 1]
Daniel@0:     net.pr_mean = prior.pr_mean;
Daniel@0:     net.pr_var = prior.pr_var;
Daniel@0:   else
Daniel@0:     net.pr_mean = prior.pr_mean;
Daniel@0:     net.pr_var = prior.pr_var;
Daniel@0:     net.index = prior.index;
Daniel@0:   end  
Daniel@0: end
Daniel@0: 
Daniel@0: % Store training data as needed for gpfwd
Daniel@0: net.tr_in = [];
Daniel@0: net.tr_targets = [];