Entropy search

Entropy Search

very brief testing

1. download the source code from http://www.probabilistic-optimization.org/Global.html
2. install additional Matlab packages logsumexp and tprod and GPML
3. run, for example, the following code:

% Entropy Search (Henning, 2012) example

clear
f = @(x) sin(x)         % function to optimize
cd /mnt/win/new/doc/skola/phd/2013/03_bbob-2013/EntropySearch/
in.covfunc      = {@covSEard};       % GP kernel
in.covfunc_dx   = {@covSEard_dx_MD}; % derivative of GP kernel. You can use covSEard_dx_MD and covRQard_dx_MD if you use Carl's & Hannes' covSEard, covRQard, respectively.
in.xmin = 0;            % input space bounds
in.xmax = 2*pi;
in.f = f;               % function to optimize
addpath ~/prg/gpeda/src/vendor/gpml-matlab-v3.2/
startup                 % startup of GPML

N = 1;                  % input dimension
hyp.cov         = log([ones(N,1);1]); % hyperparameters for the kernel
hyp.lik         = log([1e-3]); % noise level on signals (log(standard deviation));
in.hyp          = hyp;  % hyperparameters, with fields .lik (noise level) and .cov (kernel hyperparameters), see documentation of the kernel functions for details.
in.LearnHypers  = true; % yes.
in.HyperPrior   = @SEGammaHyperPosterior;
addpath ~/tmp/tprod/    % tprod C++/mex64 library
in.MaxEval = 5;         % 5 evaluations at maximum (it still takes about 2 min)

% EntropySearch itself
result = EntropySearch(in)

% Visualization of results

% just brief intro...
f(result.FunEst)
xx = linspace(0,2*pi,50);
plot(xx,f(xx));
hold on;
scatter(result.FunEst, f(result.FunEst), 'ro');
hold off;

% All 5 models visualization
for i = 1:5
  subplot(1,5,i);
  [my_m my_s2] = gp(result.GPs{i}.hyp, [], [], result.GPs{i}.covfunc, result.GPs{i}.likfunc, result.FunEst(1:i,:), f(result.FunEst(1:i,:)), xx');
  plotErr1(xx', my_m, my_s2, result.FunEst(1:i,:)', f(result.FunEst(1:i,:)'));
  axis([0 2*pi -2 2]);
end;