Ian Clarke <[EMAIL PROTECTED]> wrote:
> On Wed, Jul 23, 2003 at 05:52:17PM +0100, Theodore Hong wrote:
> > Could we perhaps encode all of the samples as binary {0,1} data, but
> > train the classifier to return a continuous result? That is, use
> > samples like:
> >
> > ((time=411,key=0x3a4b,htl=12), target=1) // success
> > ((time=480,key=0x3a4b,htl=15), target=1) // success
> > ((time=485,key=0x3a4c,htl=8), target=1) // success
> > ((time=487,key=0x3a4b,htl=9), target=0) // fail
> >
> > with the regression mode. The predictor will then return values
> > between 0 and 1 which we can interpret as a probability of success.
>
> Perhaps, although it is starting to look really convoluted :-/
Not really, it's just saying that instead of training the SVM to do a
binary prediction:
training: success, fail, fail, success, fail, success
prediction: fail
we do it this way:
training: 1, 0, 0, 1, 0, 1
prediction: 0.47 (probability of success)
This should avoid the scenario you mentioned that all nodes predict
failure and we can't do anything. The code is exactly the same as for
the response time estimator, which looks like this-
training: 1010, 435, 600, 902, 771
prediction: 722 (milliseconds)
> I would be interested to see a critique of what we have now and in
> what ways this approach is likely to be better. The main issue with
> the existing technique that I see is that the "forgetfulness" of the
> continuous averaging algorithms must be decided manually and
> arbitrarily, but if I understand this correctly I don't think this
> is addressing that issue - or is it?
No, it's not addressing the forgetfulness. The hypothesis is just
that an SVM might be more accurate at predicting response times than a
continuous average.
theo
_______________________________________________
devl mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
