From: "Laszlo Sragner" <sragner@mit.bme.hu> Path: news.mathworks.com!newsfeed.mathworks.com!WebX Newsgroups: comp.soft-sys.matlab Subject: Re: Test case rough statistics Message-ID: <eeb4a65.69@WebX.raydaftYaTP> Date: Wed, 13 Nov 2002 06:33:39 -0500 References: <eeb4a65.-1@WebX.raydaftYaTP> <eeb4a65.36@WebX.raydaftYaTP> <eeb4a65.66@WebX.raydaftYaTP> <eeb4a65.68@WebX.raydaftYaTP> Lines: 33 MIME-Version: 1.0 Content-Type: text/plain; charset="ISO-8859-1" Content-Transfer-Encoding: 8bit Xref: news.mathworks.com comp.soft-sys.matlab:142750 Stijn Helsen wrote: > > > Great Rumpuscat wrote: >> >> >> For the three strings that I believe I determined, I was able to > come >> up with better foldings than the front runners. But, somehow I >> wasn't able to take advantage of that information, which I find a >> little mystifying. Anyway, maybe this posting will be of use to >> someone. > Maybe that comes because the random number sequence is changed if you > replace a current solution by a fast one. (Refer to the improvement > by the "holy line" 'rand(63,1)' > > Stijn > I think that it can be useful to omit rand and randn, to make the run and the score for the entries deterministic. It is easy to create pseudo-random numbers if one like to use random search algorithms. It doesn't count if the pseudo random numbers repeat, while they are used only a few times. They can be computed from the given input of the solver function or some random matrices included in the code. If rand omitted even the least improvement in speed or cost can be measured deterministic. It is possible to tune the pseudo-random generator to get better results but if the generator is the same the changes are always measurable Sragner Laszlo