>>>>> "Bo" == Bo Peng <[EMAIL PROTECTED]>
>>>>> on Fri, 24 Jun 2005 10:32:45 -0500 writes:
Bo> On 6/24/05, Prof Brian Ripley <[EMAIL PROTECTED]> wrote:
>> `Research' involves looking at all the competitor methods, devising a
>> near-optimal strategy and selecting amongst methods according to that
>> strategy. It is not a quick fix we are looking for but something that
>> will be good for the long term.
Bo> I am sorry but I am afraid that I do not have enough time and
Bo> background knowledge
Bo> to do a thorough research in this area.
which I think is well understandable.
Bo> I have tried bisection search method and the alias
Bo> method, the latter has greatly improved the performance
Bo> of my bioinformatics application. Since this method is
Bo> the only one mentioned in Knuth's book, I have no idea
Bo> about other alternatives.
I think you've also explored the space of possible inputs a bit
and have suggested that the alias method was "uniformly" better
than the current one, i.e. always better, sometimes only
slightly but sometimes considerably (and never worse).
If this (uniform improvement) can be ``proven'' in some way,
{and that maybe a considerable "if", I haven't started to go in there}
and because the algorithm is relatively simple {i.e., there's
not much code added to the current one},
I'd think that we (R-core) should incorporate the algorithm for
the time being, until someone has time for the ``real research''
and provide even better algorithm(s).
I don't see why the phrase
"the good is the enemy of the better" should apply in this
situation.
Martin Maechler, ETH Zurich
Bo> Attached is a slightly improved version of the alias method.
(deleted for this reply).
Bo> It may be helpful to people having similar problems.
Bo> Thanks.
Bo> --
Bo> Bo Peng
Bo> Department of Statistics
Bo> Rice University.
Bo> http://bp6.stat.rice.edu:8080/
______________________________________________
R-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
