Initialises the Sparse Gaussian Process hyperparameters.
ogphypcovpar(hyplambda, hypmean)
ogphypcovpar(hyplambda) initialises the hyperparameter vector of the
Sparse Gaussian Process structure net, where hyplambda specifies
the precision (inverse variance) of the parameters - where the parameter
order is given by the function ogppak. Since several parameters have to
be positive, ogppak usually contains the log-transformed parameters,
forcing positivity.
If a single parameter is given, zero mean is assumed to the (log)parameters
except for the input scales which are initialised to -4.605-log(2*nin) which
supposes a preference for lengthscales of order 100 increasing with
the number of dimensions.
If unnormalised inputs are used, then it is advised to have the mean values initialised to values proportional to negative log-variance. The inference on the hyper-parameter level is ML2 inference, thus vague priors are beneficial for the stability of the algorithm.
A manual setup of the hyperparameters is recommended only if some specific knowledge about the data is known.
The function sets the fields of the GLOBAL variable net.
ogp, ogppak, ogpcovarp, ogptrain
Copyright (c) Lehel Csató (2001-2004)