README file:


May 08, 1999.

1) The software represents an attempt to implement Gaussian Processes (GP) with
a global optimization algorithm. From our experience, we believe our implementation
of Simulated Annealing (SA) is doing a reasonable good job. This is mainly due to the way
we set the standard deviation of the proposed jumps accordingly to each different
component of the parameter vector. This is done by dynamically guessing the "correct"
scale for each component. The present implementation may be further improved.

2) We have tested our implementation against Rasmussen's, this was done in two ways: first to
have an indirect test for the correctness of the implementation, second, to compare
the optimization routines. Originally we implemented priors add did Maximum
A Posteriori, but finally we decided to do Maximum likelihood only.

3) Besides the usual inversion algorithm for inverting a matrix, we coded the innovations
algorithm. This is a neat way of computing the inner products appearing in the likelihood
of GP without explicitly inverting the matrix. See the report for references.

4) We have worked exclusively under Linux, latest versions of the GNU compiler are
needed to handle exceptions. If they are removed, older versions of g++ should 
be able to compile the code. We are using GnuPlot, unfortunately (due to the way we
implemented things), it requires root privileges to run. If you can not be "su"
you should get rid of the lines calling GnuPlot.

5) The foundational C++ classes used as well as SA have been tested thoroughly. 


Feel free to use the code as you need. For questions email to: ferrando@acs.ryerson.ca

Sebastian Ferrando.