Ian wrote:
> This is extremely interesting.  Do you think that these predictions
> (either RTE's or SVM's) are statistically significant (I will need
> to look more closely at the data).

I do believe the results are significant in that they are not just a
lucky run or fluke; I tried many different parameter combinations and
have seen the SVM perform better in most cases.  As to whether or not
the differences matter, I tend to doubt the estimation accuracy improvement
would be much help if the rest of the routing algorithm did not change.
I think that a predictor that improved accuracy only 1%, but was used
as part of the forwarding decision algorithm could still cause a dramatic
improvement in Freenet's performace as a whole, as the final output is
not an estimate but a routing decision.  So really the question to me
is, how do these algorithms perform as part of a larger network?  My
suspicion is that the simple version of routing that I have heard about
involving forwarding to "closest keys" may be creating some congestion
bottlenecks or file "masking" effects, analogous to how a hash table using
linear probing will create more clustered distributions than a hash
table using a secondary hash function.  I think it's clear that there will
be some sort of "self-organizing complexity" as Wolfram would phrase it,
but am still trying to decide how we might characterize and measure
different large-scale network behaviors when under the control of different
routing schemes.  Perhaps it will be useful to add some optional
request-tracking to the protocol so that we can look at actual traces of
a query flowing through the Freenet, and see where the actual holdup
occurs, usually.  I have no idea how feasible this might be; but if we
can't temporarily disable anonymity then I am not sure how else we can
even measure the effects of experimental adjustments.

Rudi

> Matthew is working hard to get the NGR infrastructure in-place -=20
> then we will be able to do a more realistic test that will include
> factors such as CPU and memory requirements which are important
> to consider prior to an actual deployment.
> 
> Ian.
> 
> On Sat, Jul 26, 2003 at 06:08:05AM -0700, Rudi Cilibrasi wrote:
> > Hi again everybody,
> >=20
> > I've just finished running "experiment 2" based on the real-life
> > sampled data Ian recently posted.  Short summary: SVM is about 10%
> > more accurate (in terms of RMS) on this sample, as compared to
> > RTE set on max accuracy.  There's a lot of tweaking and experimentation
> > that we can do, but I think it's wasted until we make a better
> > description of how it is supposed to interface in the real
> > freenet code.  You can see the results and code yourself at:
> >=20
> > http://homepages.cwi.nl/~cilibrar/ngrouting/exp2/
> >=20
> > This one is pretty slow to run, as I'm doing a lot of N^2 type loops
> > for easy implementation -- I think we can probably adjust the
> > time/space/accuracy tradeoff better when we have a clearer idea of
> > what the real interface in Freenet will look like.
> >=20
> > I don't think the 10% number is very significant by itself, but I do
> > expect that if we wind up using the SVM as part of a route
> > choosing scheme, then it may evolve quite different global network
> > behavior than the RTE scheme.  In combination with adding more
> > independent input data, I think we may get very different behavior.
> > Better or worse, I don't know, but I do think it will be interesting
> > to find out. =20
> >=20
> > Rudi
> >=20
> > _______________________________________________
> > devl mailing list
> > [EMAIL PROTECTED]
> > http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
> 
> --=20
> Ian Clarke                                                [EMAIL PROTECTED]
> Coordinator, The Freenet Project            http://freenetproject.org/
> Founder, Locutus                                      http://locut.us/
> Personal Homepage                                 http://locut.us/ian/
> 
> --HcAYCG3uE/tztfnV
> Content-Type: application/pgp-signature
> Content-Disposition: inline
> 
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.2.1 (GNU/Linux)
> 
> iD8DBQE/JWzQQtgxRWSmsqwRAnYAAJ0Usm/J7eMzc5geG8+jPKzJFi5vkwCffNdg
> hVFPDJGqTbz2dFn+gRlEhrE=
> =syXJ
> -----END PGP SIGNATURE-----
> 
> --HcAYCG3uE/tztfnV--
> 
> --__--__--
> 
> Message: 10
> Date: Mon, 28 Jul 2003 20:38:26 +0100
> To: [EMAIL PROTECTED]
> From: Roger Hayter <[EMAIL PROTECTED]>
> Subject: [freenet-dev] Possible security problem with experimental build and/or Sun 
> Java
> Reply-To: [EMAIL PROTECTED]
> 
> I have been running build 6508, snapshot of about 48 hours ago, with Sun 
> 1.4.2.  I noticed today that isproxy would not start.  I found a process 
> which was only identified as "java2", which is the name of the symlink 
> to ..../jre/bin/java, had a tcp connection from a high port on my 
> machine to port 6667 on a host in attbi.com.  I can think of no 
> legitimate explanation of this, but would be glad to hear of one!
> -- 
> Roger Hayter
> 
> 
> --__--__--
> 
> _______________________________________________
> devl mailing list
> [EMAIL PROTECTED]
> http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
> 
> 
> End of devl Digest
_______________________________________________
devl mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl

Reply via email to