Re: [Nmlrg] recent paper on cognition-based networks

Dear all

with respect to the article of Michele and to the comments of Brian, I have
a remark, primarily, on the relation of intent and ML techniques (hope I
got the question correctly).

I will start from the second, that of the relation of intent and
unsupervized machine learning techniques.

In general, my view regarding the relation of intent and ML is that it
depends on the scenarios, on what we are trying to solve. I will take the
example as a basis.

-- The declaration of intent may designate "clusters", and, therefore,
imply that supervized learning techniques are enough. E.g., situations
addressed and network configurations used can be split into clusters that
yield jitter above or below the indicated threshold.

-- The declaration of intent may also leave degrees of freedom. In this
case unsupervized learning may be most appropriate. E.g., we may be
interested in aspects further to the jitter entailed by a situation
addressed and the configuration used. Such aspects may be the energy
consumption, the spectrum needed, etc.

Further views are welcome, and of course further elaboration is possible.

Best regards
Panagiotis

On Tue, Jun 14, 2016 at 5:01 AM, Brian E Carpenter <
brian.e.carpenter@gmail.com> wrote:

> Michele,
>
> I read the paper with interest; it even cites an article by one of
> my colleagues in Auckland (Aniket Mahani).
>
> I have a practical question. You correctly observe that computing
> power is now cheap enough to consider such techniques, but I'm
> wondering what that really means. Can we imagine ML inside
> high-end routers? Line speed firewalls? Or will it be confined
> to specialised nodes that serve up ML results to network devices?
>
> Another question is more abstract. You focus on unsupervised
> learning where (if I understand correctly) the features that are
> learned have no direct relationship to features that humans
> consider important. In that situation, how can we translate what
> we are calling "Intent" in the IETF work on autonomics into
> goals for the ML-based system? Or to give a concrete example,
> suppose there is an Intent statement that jitter (more precisely,
> Packet Delay Variation, RFC3393) for voice streams must be less
> than N milliseconds. How do we relate that to ML-derived features
> in a packet stream?
>
> Regards
>    Brian Carpenter
>
> On 24/05/2016 23:52, Michele Zorzi wrote:
> >
> > Hi all,
> >
> > I would like to point you to a recent paper we published, which tries to
> scope out some vision for cognitive networking beyond
> > the more traditional view of cognitive radio. Please find it on
> IEEExplore (it is open access) at
> > http://ieeexplore.ieee.org/xpl/abstractCitations.jsp?arnumber=7217798
> >
> > I will very much welcome your comments and feeback, especially about the
> practical feasibility of this approach. I will also
> > welcome expressions of interest to collaborate on furthering this into
> something more concrete.
> >
> > Thanks a lot,
> >
> > Michele
> >
> > _______________________________________________
> > nmlrg mailing list
> > nmlrg@irtf.org
> > https://www.irtf.org/mailman/listinfo/nmlrg
> >
>
> _______________________________________________
> nmlrg mailing list
> nmlrg@irtf.org
> https://www.irtf.org/mailman/listinfo/nmlrg
>

-- 
Panagiotis Demestichas
Professor
University of Piraeus
School of Information and Communication Technology
Department of Digital Systems
Telecommunication Networks and integrated Services (TNS) Laboratory
Office:         +302104142758
Laboratory: +302104142749
Cellular:      +306944243856
Skype:        pdemest
Office: 150 Androutsou, 18532, Piraeus, GREECE
Mail :  80 Karaoli and Dimitriou, 18534, Piraeus, GREECE