Hello, there - This is a very interesting problem. But I think you don't really want to have a BBN generator which can generate networks really "randomly" because the space of DAGs is infinite and it's impossible(and useless) to deal with networks with extreme parameters. In real world applications we only deal with "nomal" networks. The space of "real world" networks is just a small subset of all networks. So instead of making a random BBN generator, I suggest making a "real world" network generator. You can try to collect all kinds of BBNs you can get, and sample from these real world BBN samples in a random way. Thus you will get a random "real world" network generator. It dosen't sample the space of the DAGs uniformly. Instead it samples the space of "real world DAGs" uniformly.
cheers, - -hpguo On Fri, 26 Oct 2001, Fabio Gagliardi Cozman wrote: > Dear UAIers, > > I would appreciate if anyone could give some pointers > on how to generate Bayesian networks randomly. That is, > how to construct directed acyclic graphs and their > associated parameters in a way that uniformly samples > the space of the DAGs? I have not found a simple explanation > on how to do it, even though many papers refer to networks > that have been generated randomly. I wonder if there is > software available for this. > > Maybe the problem is simple, but it does seem to require > some sophistication. Even to generate the conditional > probabilities, it does not seem that simply generating > tuples and normalizing them will uniformly sample the > space of probability distributions. > > Any help is appreciated. Thanks, > > Fabio Cozman >
