Hi all - there is a project at MIT called FlexGP that has done more work on this.
http://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/index.php?n=Site.FlexGP Unfortunately I can't find a download for the code so I suppose that it's not opensource, however you might like to contact these folks and see if they can help you get a leg up. Best, Simon -----Original Message----- From: Ted Dunning [mailto:ted.dunn...@gmail.com] Sent: 13 January 2014 18:00 To: user@mahout.apache.org Subject: Re: travelling salesman on Mahout On Mon, Jan 13, 2014 at 8:42 AM, Pavan K Narayanan < pavan.naraya...@gmail.com> wrote: > > > Please may I ask why TSP has been removed from Mahout. It was the Genetic Algorithms that were removed. The implementation was unmaintained and not scalable and thus not appropriate for Mahout. > Its just that I > see discussions about distributed Genetic Algorithm and other > evolutionary algorithms being implemented in distributed environment > and feel it would be possible to implement in Mahout as well. > Indeed, it is quite possible. Most evolutionary algorithms are trivially parallel. But somebody has to step up to maintain the code. > > For example, the initial population generation and evaluating the best > initial route could be done in 'n' nodes using 'n' map phases and we > could have a reduce phase where the best of initial route from 'n' > nodes could be taken for further treatment like mutation, etc... What > are your thoughts on this? > I think that it would be better done using something like Giraph or Spark. This isn't a natural fit to map-reduce due to the iteration. Also, TSP itself is not a particularly interesting problem for Mahout. The Concorde system is very mature and already scales to over a hundred nodes using MPI. Mahout is not going touch it and has very little to offer.