Estevam, As far as I know, P. Larranaga used a GA to search for the best ordering for K2. The fitness he used was the same score function used by K2 on the learned structuce from a particular ordering. We are also doing some research on this direction, but we are trying different fitness functions and GA operators. Our code has been released as open-source software development toolkit for Bayesian network research: BNJ, which stands for Bayesian Network tools in JAVA. Please see the attached references for more information. Hope this helps.
[LPYM+96] P. Larranaga, M. Poza, Y. Yurramendi, R. H. Murga and C.M.H. Kuijpers. Structure Learning of Bayesian Networks by Genetic Algorithms: A Performance Analysis of Control Parameters. IEEE Journal on Pattern Analysis and Machine Intelligence 18(9): 912-926, 1996. http://www.kddresearch.org/Groups/Probabilistic-Reasoning/BNJ/manual/manual.html http://www.kddresearch.org/Groups/Probabilistic-Reasoning/BNJ/manual/ref.html cheers, - -hpguo On Fri, 26 Jul 2002, Estevam Rafael Hruschka Junior wrote: > Dear all, > > > when learning a BN from data, some algorithms (such as K2) are influenced by > the variable order. And if this initial order is not the most appropriated, > the results may be prejudiced. Trying to find this best order may be an > exponential problem (when it's done by exhaustive search). One way to > minimize this problem may be to prune the "less" significant variables. I've > already implemented some initial tests and they showed that it's worthwhile > to look for the best order when trying to improve the learning results. > > > So I was wondering if there is any defined heuristic to deal with this > problem. > > > Thank you so much for your attention. > > > > Estevam. > > - -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_- > Estevam Rafael Hruschka Junior > > [EMAIL PROTECTED] > Curitiba PR. - Brazil. >
