matlab - Gaussian Process, negative hyper-parameters? -
matlab - Gaussian Process, negative hyper-parameters? -
i'm doing gaussian processes regression using gpml toolbox. however, after optimization using 'minimize.m' (without mean functions), negative hyper-parameters!
the initial hyper-parameters are:
hyp.cov = [0; 0]; % 2 hyper-parameters in covariance kernel (length-scale & amplitute) hyp.lik = log(0.1); %hyper-parameters of noise
the original training data:
x=[819 1119 1419 1599 1719 1839 1899 2019 2079 2139]; %coordinates y=[105.00 114.33 126.33 130.33 116.33 103.00 103.00 124.67 122.67 109.00]; %training info
in codes, y normalized have 0 mean , unit variance. optimize:
hyp = minimize(hyp, @gp, -100, @infexact, [], {@covseiso}, likfunc, x,y);
after 100 iterations, negative hyper-parameters!!! quite confusing....
however, if don't normalize y, hyper-parameters positive after optimization.
could tell me negative hyper-parameters mean? should normalize data?
the error initial noise-parameter negative:
hyp.lik = log(0.1); %hyper-parameters of noise
so i'd suggset utilize positive noise-variance , seek again.
regarding other question, yes, thought normalize info (i.e. give them zero-mean , unit-variance).
a possible reason why behaviour occurs centralized info might variance becomes smaller variance of original data. remember in bayesian linear regression (--what gaussian processes in principle are), noise variance added covariance matrix. in case, addend negative, , effect of course of study larger smaller variance.
matlab machine-learning hyperparameters
Comments
Post a Comment