On Wed, Jul 23, 2003 at 07:33:41AM -0700, Ian Clarke wrote:
> On Wed, Jul 23, 2003 at 03:26:40PM +0100, Theodore Hong wrote:
> > Is there any way of doing the training incrementally?  That is,
> Hmmm, I didn't realize that it wasn't incremental.  Given that each
> node is likely to have about 50 other nodes for which it must record
> data, and given that several thousand requests per hour would not
> be uncommon - is this going to be practical from a CPU/memory
> requirements perspective?

Incremental training is a topic of research, but it's far too complicated
for us to consider trying in the near future.  Theo's reasoning about
nearness / farness from the decision boundary agrees with my intuition,
but the problem is that adding or removing a point could change
the optimal decision boundary in certain cases because a different
local minimum can be found.  I don't think the CPU load will be a problem,
so long as we spawn off a seperate thread at program startup that just
constantly retrains the system with all of the latest data in a loop.
Then prediction can be done in the seperate main thread rapidly.

Rudi

_______________________________________________
devl mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl

Reply via email to