►
From YouTube: IETF104-NMRG-20190328-1350
Description
NMRG meeting session at IETF104
2019/03/28 1350
https://datatracker.ietf.org/meeting/104/proceedings/
A
This
is
a
RTF
research
group
meeting.
The
iltf
follows
the
ietf
policy
known
as
no
not
well,
and
so
essentially,
this
is
a
reminder
of
IETF
policies
in
effect
on
various
topics
such
as
patents
or
code
of
conduct.
So
you
have
a
list
of
key
elements
reminded
in
this
slide
and
all
details
are
in
the
best
current
practices
document
and
at
the
different
links.
So
please
be
aware
of
that.
A
A
B
A
A
A
Blue
sheets,
so
we
are
circulating
the
blue
sheets
in
some
participant
still
coming
in,
please
remember
to
make
them
circulating
through
the
room.
This
is
very
important
in
order
to
have
a
right
estimation
of
the
number
of
participants
to
get
the
right
types
of
room
for
the
next
meetings,
and
we
are
very
happy
to
be
here
in
this
the
best
meeting
room
of
the
hotel,
where
we
have
actually
first
class
citizen
here
and
I,
don't
know
what
the
people
on
the
left
are
doing
anyway.
A
So
just
keep
recap,
so
we
already
had
one
session
during
this
week
in
in
an
hour
in
the
IETF,
the
session
was
dedicated
to
progress
on
intern
base.
Networking
I
think
it
was
a
pretty
nice
session,
a
lot
of
people
in
the
room,
very
interesting
presentation
and
comments.
The
the
next
step
then,
for
that
will
be
to
really
build
upon
us
feedback
and
the
the
energy
that
we
had
in
order
to
make
it
a
concrete
plan
and
progress
on
the
documents.
A
A
This
is
something
we
have
started
a
couple
of
meetings
ago,
also
on
the
mailing
list
and
through
virtual
meetings,
so
I
asked
some
participants
to
provide
some
input
to
the
discussion
on
various
various
things,
to
wrap
a
bit
to
trigger
discussion
and
feedback,
and
I
will
have
also
some
some
slide
at
the
end,
to
give
some
general
information
on
the
research
group,
but
also
try
to
yes
to
capture
where
we
are
going
and
I
mean,
put
some
some
deadlines
and
some
some
concrete
actions
on
that.
So
this
is
the
main
topic
of
today.
A
So
the
agenda
is
very
simple.
This
is
my
introduction,
which
will
end
in
one
minute.
We
are
faced.
Oh
forgot,
to
mention
very
important
anniversary
energy
is
twenty
years
on
this
year,
this
month,
more
or
less.
What
with
the
official
creation
so
I
think
it's,
the
oldest
research
group
active
in
in
the
IOT
F
as
I,
usually
say,
I,
don't
know
if
it's
a
good
thing
or
a
bad
thing,
but
we
are
here.
We
have
also
the
the
luck
and
the
player
to
have
at
least
one
of
the
previous
energy
chair
in
the
room.
A
Our
our
goal
was
to
to
have
the
two
other
previous
energy
services,
but
unfortunately
they
could
not
attend
and
I'm,
not
even
sure
they
will
be
able
to
attend
remotely
due
to
difference
in
in
timezone
and
other
commitments.
But
still
what
we
have
discussed
together
is
to
put
a
short
with
respective
of
the
different
things
the
research
group
have
studied
in
the
Indus
years
and
that
could
maybe
trigger
a
reflection
on
what
we
have
done.
What
should
we
do
for
the
future?
A
So
don't
look
into
more
than
that
in
this
presentation,
but
I
think
it
was
important
to
to
make
a
point.
Then
we
have
our
second
topic
as
input
for
a
discussion,
the
use
of
AI
techniques
for
network
management.
So
this
is
a
kind
of
collective
input
based
on
several
participants
that
provided
their
views
and
some
some
reflection
on
different
aspect.
That
would
be
useful
for
the
group
to
study
on
AI
for
network
management.
This
is
a
very
open
topic.
A
A
C
Some
twenty
years
ago,
I
I
was
guilty.
Writing
a
chato
proposal
for
a
research
group
which
I
call
the
network
management
research
group
and
it's
kind
of
funny
that
it
still
exists
and
that
I'm
here
and
you
talk
about
the
future
of
it
and
I
see
even
more
funny.
So
the
slide
lists
the
number
of
the
chairs
that
have
been
chairing
the
so
time
and
so
probably
go
to
the
next
slide.
C
So
how
do
you
measure
what
what
the
working
group
or
research
group
does,
while
you
count
as
result
produced
yeah,
so
count
the
number
of
our
C's,
so
there
are
eight
of
C's
published
that
really
came
out
of
the
research
group
over
time.
Now,
if
you
would
say
eight
over
twenty
years,
that
was
probably
not
a
huge
number,
but
at
least
we
managed
aid.
C
We
did
much
better
in
terms
of
number
of
meetings.
So
51
meetings
is
that's
quite
a
lot,
so
there
has
been
so
one
of
the
the
the
properties
of
the
research
group
has
always
been,
and
it
has
been
very
interactive
and
lots
of
discussion
so
stimulating
discussions.
Next
slide
I've
tried
to
put
this
into
a
time
line,
and
so
starting
from
the
beginning
to
the
end,
and
you
see
below
the
main
timeline
you
see,
DRC's
and
how
long
it
did
take
to
produce
those
hour
series.
C
So
if
people
say
that
production
of
RCC
at
slower
and
slower,
this
does
not
prove
this.
It
just
varies
at
the
bottom.
You
see
that
the
timeline
where
the
chairs
have
been
hopping,
the
research
group
to
stay
on
track
and
on
the
top
you
see
working
groups.
There
were
kind
of
spin-offs
coming
out
of
the
an
emoji
next
slide.
Please
so
slide
4.
So
20
years
ago,
when
the
idea
came
along
to
write
a
charter
for
a
research
group
technology,
world
looked
slightly
different.
C
There
was
a
lot
discussion
about
policy
based
management,
I
actually
started
in
the
90s
mid
90s
late
90s.
We
got
cops
and
then
a
little
bit
later
cops
PR,
where's
money.
Well
they're,
not
so
many
young
people
here,
but
if
there
would
be
young
people
here,
they
wouldn't
know
what
it
is
and,
along
with
cops,
PR
came
another
data.
Modeling
language
was
called
s
PPI,
which
was
kind
of
as
my
v2,
which
was
used
for
us
and
P
MIT
modules
to
write
pip
modules
for
cops
PR,
and
so
that
was
kind
of
the
scene.
C
It
was
directory,
enables
networks,
yeah,
so
Jon's
trust,
nodded
directory,
enabled
networks,
and
then
there
was
this
big
policy
information
model
work
done
in
corporation,
where
the
DMT
F,
and
so
several
of
C's,
publish
telling
you
how
to
write
policy
rules,
event,
condition
action
rules.
What
can
you
put
into
an
event
condition
and
so
forth
now?
At
the
same
time,
it
was
clear
that
SNMP
table
isn't
really
doing
enough.
C
It
was
good
for
monitoring,
but
it
wasn't
doing
enough,
but
somehow
the
idea,
the
working
groups
working
on
SNP
got
so
much
stuck
on
security
that
they
kind
of
kept
stuck
focusing
on
security
and
I
didn't
really
want
to
improve
the
the
technology.
Researchers
were
running
prototyping,
you
managed
my
protocols.
They
were
running
over
HTTP,
while
the
SNMP
people
said
everything,
that's
not
running
over
UDP
as
broken
by
design.
C
Nowadays
we
have
rest
comp,
but
that
kind
of
started
some
of
the
ideas
here.
People
started
talking
about
pushing
data
rather
than
pulling
data.
Well,
today
we
call
it
telemetry,
you
know
different
term
for
it
and
there
was
lots
of
Technology
fragmentation
because
everybody
was
coming
up
with
new
technology.
Soap
based
technology
to
shift
management,
data
management,
data
embedded
into
HTML
texts
and
stuff
like
that
beep,
something
in
that
the
IETF
created
at
the
same
time
boxes
were
essentially
very
closed
boxes.
You
had
to
see
a
lie
and
you
could
talk.
C
Snmp
trail
and
I
was
mostly
dirt
and
you
were
totally
depending
on
what
the
vendor
did
put
him
on
the
device
you
couldn't
do
much
more,
so
flexible
flexibility
was
really
limited
by
what
could
be
done
so
one
of
the
first
things
we
try
to
break
this
up
and
try
to
evolve
the
existing
technology.
If
you
go
to
slide
5.
So
that's
where
people
just
a
few
people
were
sitting
together
and
if
you
search
the
internet,
you
find
even
a
photo
from
that
time
we're
sitting
together.
C
C
Okay,
so
I
wrote
a
proposal
after
the
meeting
I
did
hand
it
in
and
I'm
a
big
wine
rip.
He
was
the
IOT
F
chair
that
time
approved
her,
and
so
we
got
going
in
March
1999
slide
6,
so
the
first
phase
was
really
looking
at
management
technology.
And
what
can
we
do
to
improve
it?
So
the
first
one
was
dealing
with
the
question
that
we
now
get
multiple
data
models
written
in
different
languages.
So
do
we
have
to
duplicate
all
our
data
modeling
work
that
seems
very
costly
to
do
so?
C
Could
we
actually
go
for
a
data
modeling
language
that
is
able
to
be
a
protocol
independent,
and
so
we
define
things
once
and
then
we
used
a
lot.
Multiple
protocols
and
the
proposal
that
came
out
of
the
NMR
G
was
called
SMI
and
G
kind
of
trying
to
take.
As
my
what
allow
more
structured
data
alarm,
mappings
to
SNMP
to
cop's,
PR
and
whatever
might
come
along
in
the
future,
there's
actually
turned
into
a
working
group,
but
the
working
group
they
take
a
long
time
to
write
down
a
guidelines
document.
C
C
The
second
goal
was
to
actually
look
at
the
yeah,
the
whole
SNMP
framework.
So
we
wanted
to
have
more
structured
data
in
there.
One
should
be
more
efficient,
have
data
compression
rather
than
although
BER
is
not
that
bad
in
terms
of
encoding
efficiency,
but
you
could
squeeze
out
more.
There
was
another
working
group
coming
along.
C
It
was
called
the
evolution
of
SNMP,
and
that
was
a
really
short
working
group
until
it
found
out
that
it
doesn't
have
any
agreement
on
anything
and
so
later
we
labeled
at
the
end
of
SNMP
working
groups,
because
afterwards,
really
nothing
has
been
done
anymore
in
terms
of
SNP.
That
was
the
end
of
it.
C
A
Phase
I
was
expecting
listen
or
my
previous
code
sure
to
cover
that,
because
at
the
time
of
this
second
phase,
I
was
only
young
participant
to
the
research
group.
So
my
recollection
of
the
fact
and
the
way
I've
maybe
will
express
this
is
more
from
a
participant,
interviewed
and
really
from
a
research
group
point
of
view
anyway.
I
try
to
give
you
some
useful
information
and
put
four
faults
on
this
second
phase.
A
So
this
was
already
an
established
field
in
the
research
community,
but
there
were
some
different
sources
and
different
active
participants
that
converge
on
the
need
to
sit
together
and
try
to
define,
identify
some
gaps,
try
to
clarify
some
concept
and
try
also
an
approach
with
an
eventual
standardization
of
some
of
this,
of
the
elements
required
to
have
an
atomic
network.
So
first
goal.
First
step
was
to
address
some
common
understanding.
As
I
said,
the
app
analysis
also
explores
different
set
of
use
cases.
This
was
supported
by
a
series
of
meeting.
A
The
industry,
adoption
and
the
actual
deployment
of
such
technology
was
quite
low
for
many
years,
and
it
was
one
of
the
challenges
some
of
the
outcome
of
this
activity
was
or
to
heiresses,
75,
75
and
75.
76
one
is
addressing
the
diffusion
of
the
concept
with
a
kind
of
DN,
a
proposal
for
a
high
level
architecture
as
a
guideline,
and
the
second
one
is
covering
the
different
gap
gaps
analysis
in
order
to
to
derive
potential
elements
for
the
research
group,
but
also
what
became
later
on
the
inputs
to
the
animal
working
group.
A
So
another
outcome
was
creation,
I
mean
the
a
you
can
both
effect
in
2014
asking
if
it
was
relevant
to
to
create
a
working
group
and
then,
a
few
months
later,
creation
of
the
animal
working
group,
which
is
still
active,
an
American
group.
The
scope
is
essentially
on
autonomic
networking
protocols,
a
bit
of
architecture
and
sexual
bootstrapping.
So
a
very
specific
scope
also
still
in
this
broad
topic
of
autonomic
network
management,
they
have
been
more
recently
we
targeted
autonomic
3.0
in
just
to
differentiate
with
this
first
goal.
A
First
phase
of
autonomic
network
management
in
energy
not
outside,
but
in
an
energy,
because
when
anima
was
created,
it
somehow
took
with
him
a
lot
of
the
activity,
a
lot
of
also
the
contributors
of
on
this
topic.
But
it
remains
some
gaps.
Some
research
direction
that
were
not
explode
and
also
to
recognize
that
some
elements
in
the
in
our
environment,
where
we're
evolving
so,
for
instance,
the
increasing
importance
and
also
deployment
of
virtualization
approaches,
also
programmable
approaches
so
and
FES
GN,
essentially
also
new
application
areas.
A
We
currently
file
G,
also
the
growing
interest
in
to
IOT
IOT
environment
IOT
networks
and
also
a
bit
more
with
a
different
he
package,
but
all
the
smart
X
example
like
smart
factory,
smart
city,
smart
transport,
whatever
you
can
think
that
was
trying
to
blend
a
kind
of
very
domain-specific
environment,
with
communication
technologies
in
various
in
various
uses,
also
along
those
years,
because
we
are
talking
about
2000
from
2013
to
up
to
now
really
rising
trend
on
what
the
industry
called
Network
automation.
But
so
the
really
industry
call
its
network
automation.
A
A
I
mean
it's
not
you
that
AI
techniques
has
been
used
in
networking
environments,
it's
been
15
years
or
20
years
that
those
techniques
exist
and
has
been
applied
for
solving
different
types
of
problems,
but
with
the
new
wave
I'm,
referring
to
five
years
six
years
ago,
where
this
really
neural
network,
and
especially
deep
neural
networks.
Well,
it's
exploding
and
covering
very
different
field,
and
so
there
is
really
a
new
wave
of
of
techniques,
but
also
of
energy
in
order
to
investigate
any
kind
of
problem
with
those
new
but
new
approaches.
D
Hi
Rick
Taylor,
Airbus
I
need
chair
of
the
DTN
working
group
and
I
was
trying
to
find
a
good
moment
to
to
come
to
the
mic
as
you're
talking
about
autonomics.
There's
DTM
is
working
group
all
about
store-and-forward
systems.
Is
it
comes
out
of
deep
space,
so
NASA
and
the
usual
players,
so
they
independently
or
or
we
know
within
our
working
group,
have
independently
done
a
lot
of
work
over
the
last
five
years
on
what
we're
calling
asynchronous
management
rather
than
autonomic.
But
it
is
autonomic
it's
about.
D
How
do
you
do
event-driven
management
on
a
deeply
disconnected
network?
So
we've
got
a
lot
of
work
going
on
in
the
in
the
group
at
the
moment,
but
we're
really
trying
to
reach
out
to
people
like
the
NMR
G
and
say
we
don't
know
enough
about
management
in
depth.
We
know
enough
about
our
use
cases,
but
we'd
really
like
help
from
you
guys,
because
you
know
management
and
you
know
what
worked
in
the
past
and
where
the
when
the,
where
network
management
is
going
in
the
future.
D
So
if
anyone
is
interested
in
what
we're
doing,
we
have
running
code
on
live
systems,
for
example
the
International
Space
Station,
there's
components
on
there
that
are
managed
using
asynchronous
policy
based
conditionally
event
systems
that
back
down
now
on
to
neck
on
space
systems.
It's
all
really
coming
together
now,
but
we
need
you
plurals
help
to
come
and
make
sure
that
we're
not
making
terrible
mistakes
so
a
bit
of
an
advert
there,
but
it
does
kind
of
fit
in
with
your
goals.
But
this
independently
invented.
Unfortunately,
time.
A
End
I
think
this
is
something
we
should
try
to
follow
up.
Maybe
after
the
meeting
you
and
me
can
have
a
quick
shot,
but
thinking
about
maybe
either
some
joint
description
or
joint
meeting,
maybe
at
some
yeah
or
even
through
a
kind
of
few
scales
up
Church.
If
you
can
express,
as
you
mentioned,
you
are
really
in
your
domain.
D
Mean
we've
got
architectural
drafts,
which
we
can
absolutely
share
and
push
through
onto
your
mailing
list
and
that
kind
of
stuff
what
I
have
with
my
chair
has
on
what
I
don't
want
us
to
do
is
to
go
and
invent
an
entirely
separate
and
standardized
way
of
doing
fifty
percent
of
the
same
things
that
either
is
coming
out
of
the
RTF
or
anima
or
whatever,
and
just
try
and
link
all
this
up
just
because
we
only
have
so
many
cycles,
let's,
let's
not,
learn
them
independently.
Thanks.
A
Also
in
this
autonomic
3.0
at
church
to
address
some
of
the
issue
not
address
in
Quran
was
to
try
to
have
better
link
with
will
word
concerns
and
approbation,
so
try
to
involve
or
to
get
more
aware
of
the
requirement
and
needs
from
the
operator
community,
especially
trying
to
approach
nine
org
and
ripe
communities,
but
also
to
address
or
to
get
more
knowledge
from
other
emerging
disciplines.
If
I
can
call
it
like
them
like
site,
reliability,
engineering
or
network
reliability,
engineering
which
are
approaching
the
way
of
operating
large-scale
infrastructure
in
very
different
way.
A
They
don't
call
network
management
at
all.
They
come
with
very
different
background.
Different
techniques
and
different
I
mean
real
life.
I
mean
we
operationalize
constrain
and
deployments
so
I
think
we
have
to
connect
more
with
those
those
communities
and
those
way
of
doing
things
and
also
effective
deployment.
Try
to
in
this
research
that
could
be
done
in
the
research
group
to
resign
on
the
lack
of
wide
scale
deployment.
A
So,
as
I
said
this
second
phase,
most
of
the
activity
was
drawn
out
with
the
creation
of
an
imam
in
terms
of
fully
coverage
of
topic.
But
still
we
had
some.
We
developed
what
we
call
the
research
agenda
for
for
energy
and
we
wanted
to
address
what
kind
of
new
topics
new
research
direction
will
be
relevant
for
the
photo
research
group.
So
we
will
invest
the
time
into
I
mean
if
you
remember
that
the
previous
two
or
three,
your
sessions,
inviting
new
topics,
discussion
of
inside
meetings
etc.
A
So
one
of
the
key
key
topic
that
is
also
was
the
topic
of
the
previous
session.
This
week
is
on
intent
or
intern
base.
Networking
just
one
way
to
phrase
it
a
bit
differently
of
what
we
see
regularly
to
see
intent
as
a
means
for
better
usability
and
manageability
of
the
networks,
because
I
think
what
I've
seen
especially
in
IETF
and
maybe
some
other
standardization
organization,
is
that
we
are
very
much
concerned
with
the
format
of
the
things
and
some
mechanism
or
some
protocols.
A
But
we
are
in
the
network,
management,
research,
group
and
I
think
we
should
invest
more
also
in
understanding
and
putting
at
the
start
of
the
work,
the
manageability
aspects
of
those
environments
and
usability.
If
this
is
a
bit,
it
may
be
far
fetching
for
our
researcher,
but
all
the
the
the
people
will
actually
use
a
system
is
should
be
a
key
key
concern
and
so
within
10.
This
is
a
one
approach
to
to
try
to
to
have
better
ability.
A
Essentially,
through
use
of
abstractions,
so
how
not
to
go
into
every
little
details
of
how
every
things
are:
encoded
must
be
configured
monitored,
but
to
have
a
useful
abstractions
in
order
to
to
use
the
things
at
the
right
levels
and
also
the
mechanism
in
order.
The
abstraction
is
not
necessarily
sufficient.
You
need
also
to
have
the
mechanism
in
order
to
to
go
from
different
abstractions
level
and
to
bind
to
the
different
functionality,
also
an
important
aspect
in
this
manageability
all
to
achieve
that's
what
I
call
transfer
of
knowledge
and
unreasoning
from
human
to
machines.
A
A
key
element
here
is
that
if
you
want
to
be
successful
within
turn-based
at
working,
it
means
that
the
machine
the
system
needs
to
have
different
set
of
capabilities,
cognitive
capabilities
in
order
to,
if
you,
if
I,
give
him
a
high
level
objective
targets
or
an
intent,
the
way
to
derive
it,
it
requires
I
mean
quite
powerful
mechanism
or
techniques.
Currently,
we
we
have
ideas
for
some
of
them,
but
they
are
not
really
well-defined
and
well
integrated
in
a
system
that
can
really
deliver
that.
But
it's
really
important
that
this.
A
What
the
human
reasoning
capability
we
are,
which
is
very
it
kind
of
simple
or
intuitive
for
us-
is
completely
not
intuitive
and
simple
for
machine,
but
the
transfer
of
that
is
the
key
to
make
an
intern
base
system
feasible
and
also
I'm
working
on
with
intern
basal
networking.
We
also
aim
for
providing
I/o
degree
of
flexibility
and
adaptation
to
the
system,
because
previously
we
were
talking
about
policy
based
approach.
If
you
have
a
enough
imperative
type
of
policy
where
you
have
condition
action
or
even
condition
actions.
A
Essentially,
you
are
pre
thinking
about
all
the
situation
in
which
your
network
can
end
up
in.
So
you
make
these
condition
actions
to
say
if
something's
up
and
I
need
to
know
it
and
I
need
to
plan
what
the
direction
of
my
system,
so
you
code
beforehand
all
the
possibility
of
our
system,
and
then
I
will
say
it's
just
a
matter
of
instantiating
the
policy.
But
then
the
system
has
quite
zero
flexibility,
it's
just
reacting
to
what's
happening
and
it.
A
If
there
is
no
rules
to
cover
the
situation,
then
it
can
raise
abnormal
situation,
and
you
will
inject
a
new
rule
to
cover
this
case
within
turn-based
networking.
This
is
we
want
really
to
take
a
radically
different
approach,
which
is
not
necessarily
new
when
you
have
the
clarity
of
type
of
policy,
but
the
thing
is
that
you
tell
the
system
an
objective,
but
you
don't
prescribe
you
don't
code
beforehand
how
the
system
will
react
to
that.
So
you
give
flexibility.
A
You
give
adaptation
capability
to
the
system
to
address
the
current
situation
and
the
objective
in
the
best
capable
way
and
also
is
in
turn
based
networking.
Something
that
is
not
always
in
the
discussion
is
to
involve
the
user
in
a
virtuous
circle,
I
mean
if
the
user
is
providing
is
intense,
I
mean
in
different
formats
in
it
also
to
be
Oh,
I
mean
aware
of
what's
happening,
so
we
should
find
incentives
way
that
the
user
can
be
in
this
in
this
controller,
a
second
field
or
second
goal
in
this
phase.
A
A
But
then
there
is
the
challenges
so
because
today
there
are
really
some
mainstream
use
of
machine
learning
techniques
because
they
are
successful
in
many
other
fields.
And
what
I
observed
today
is
that
in
our
networking
field
we
try
to
reuse
whatever
machine
learning
techniques
as
a
hammer
to
solve
any
kind
of
network
nail
programs.
So
it
works
in
our
set
some
cases,
maybe
most
of
the
cases.
A
But
we
have
very
few
insight
about
how
useful
it
is
I
mean
if,
if
you
need
to
spend,
if
you
have
a
lot
of
requirements
on
the
collection
of
the
data
centralization
of
the
data,
the
amount
of
energy
computing
energy,
also
power
I
mean
electrical
energy.
You
need
to
consume
in
order
to
get
to
run
this
machine
learning
systems,
maybe
some
time
it's
clearly
an
overkill
and
it's
not
a
sustainable
approach
for
networking
environment.
A
A
We
should
think
on
what
could
be
a
best
way
to
integrate
the
AI
and
machine
learning
elements
in
or
for
the
network,
because
also
all
these
machine
learning
Armour,
they
are
not
Network
specific,
but
network
are
very
specific
environment
to
be
covered
in
the
next
presentation,
and
this
is
the
last
pot
network,
specific
IMO,
so
just
an
out,
but
I,
don't
sure
we
can
have
questions
now,
but
maybe
also
at
the
end
of
this
session.
Future
of
network
management,
so
networks
and
network
management
have
changed
a
lot
in
20
years.
A
I
mean
we
can
have
some
ins,
we
can
have
some
projections
or
some
stuff
may
remain,
so
we
could
predict
to
some
extent,
but
it's
not
an
exact
science,
but
what
is
necessary
is
to
think
about
how
the
network
out
today
is.
We
can
something
we
can
more
of
information
the
near
future.
How
they
are
designed,
deployed
and
operated.
I
was
referring
to
a
sorry
from
it's
completely
way
of
understanding,
networking
or
computer
networking
environment.
A
Are
we
still
talking
about
management
because,
or
is
it
still
the
right
approach?
We
are
still
having
a
network
as
a
common
denominator,
but
is
management
still
something
a
term
that
is
speaking
to
two
people
and
still
the
right
way
to
understand
the
relationship
is
with
this
network.
This
is
just
for
for
as
an
open
question.
A
E
Okay,
yeah,
thank
you
so
I'm
going
to
talk
either
with
about
artificial
intelligence,
machine
learning,
applied
to
network
management
and
I.
Think
that
the
the
wrong
correct
me
if
I'm
wrong,
is
to
try
to
understand
if
the
topics
that
I'm
going
to
present,
they
make
sense
to
to
be
adopted
by
my
desk
research
group
and
the
way
I'm
going
to
do.
It
is
I'm
going
to
do
it
through
a
use
case.
I'm,
not
claiming
that
this
is
the
use
case,
or
that
this
is
the
most
relevant
new
skills.
I'm.
E
Just
saying
that
this
is
one
of
the
potential
use
case,
because
otherwise
it's
very
difficult
to
frame
the
research
topics
without
a
specific
example
in
mind.
So
in
this
case
it's
a
let's
say
exam.
I
management
plane,
which
is
operated
with
artificial
intelligence.
I
will
say
that
this
use
case
it's
quite
common
in
the
the
research
and
we
will
have
two
pieces.
The
first
one
is
a
piece
which
is
a
network
model,
so
this
is
a
box
software
box
that
can
be
understood
as
a
digital
twin.
E
So
it's
a
digital
representation
of
the
real
metro
infrastructure
and
it
is
able
to
answer
questions
very
simple
question
such
as.
Ok,
if
I
have
this
traffic
and
I
have
this
routing
configuration
what
will
be
the
performance
of
my
network,
so
it's
kind
of
a
simulator
right,
a
digital
simulator
of
my
network,
and
then
we
have
a
management
module
which
is
taking
advantage
of
this
network
network
model
to
operate
the
network,
ok
and
to
decide
which
is
the
best
configuration.
E
Let's
say
that
the
user
decides
based
on
some
kind
of
a
natural
language.
Okay,
I
want
the
lowest
latency
possible.
Ok,
then
this
module
will
use
the
network
model
try
to
find,
which
is
the
configuration
that
results
in
the
lowest
source
destination
delay
or
whatever
performance
metric.
You
are
willing
to
find
ok.
So
now,
let's
talk
a
little
bit
about
these
two
boxes,
so
the
first
box
is
a
network
model,
as
I
was
saying,
is
the
digital
representation
of
the
network,
and
it
is
a
box
that
you
can
ask
questions.
E
Ok,
I,
have
this
plastic
I?
Have
this
router
configuration
and
I?
Have
this
topology?
What
will
be
the
losses?
I
will
see
in
my
network
so
these.
Actually
there
are
many
ways
to
build.
This
works.
One
of
the
ways
is
to
use
a
neural
network.
Okay,
so
you
have
a
neural
network
which
is
actually
inside
this
box,
providing
this
answer.
So
this
model
is
already
trained.
Okay
and
I
was
saying
you
can
answer
all
kind
of
questions
regarding
your
network.
E
Then
you
have
the
management
module
which
will
use
the
network
model
and
its
operating
the
network
trying
to
find,
which
is
the
best
configuration
to
fill
the
the
goals
of
the
of
the
user
of
the
administrator.
There
are
many
ways
to
implement
this
box.
One
way
is
to
use
classical
algorithms,
for
instance,
some
sort
of
traffic
engineer,
inaudible
or
shortest
path.
Routing.
There
are
many
ways
to
implement
this
box.
Another
way
to
do
it
is
by
also
using
artificial
intelligence.
E
In
this
case,
informal
learning,
that's
a
new
approach,
and
the
last
item
in
my
use
case
is
the
administrator
setting
up,
which
is
the
goal
of
the
network.
As
well
saying,
you
can
say
the
intent
case
of
this,
some
sort
of
language
that
the
network
module
that
the
management
module
understands
and
that
it
will
be,
it
will
try
to
fulfill
the
administrator
request
by
managing
the
data
in
such
a
way.
E
Just
as
something
which
can
be
interesting
is
that
if
this
module
is
implemented
using
different
from
your
learning,
then
this
language
is
just
reward
function,
which
is
very
simple
to
express
and,
and
that
will
convert
this
revert
into
Network
primitives,
but
I
think
that
this
is
a
topic
for
another
special
okay.
So
this
is
my
use
case.
As
I
was
saying,
let
me
insist
is
not.
The
use
case
is
just
one
use
case
to
show.
E
So
let's
talk
about
what
we
need
in
order
to
build
these
kind
of
use
cases,
but
what
we
need
to
do
in
order
to
build
this
kind
of
boxes.
So
the
first
thing
to
build
the
network
model
is
you
need
to
represent
the
traffic,
the
configuration
and
the
topology
and
represent
the
performance
and
the
cost
in
a
neural
network.
Whenever
you
put
an
input
parameter
or
you
take
an
output
parameter,
you
need
a
mathematical
representation
for
this
parameter.
E
This
is
called
to
engineering
and
unless
you
represent
this
parameter
in
a
meaningful
way,
the
neural
network
won't
be
able
to
understand
it.
So
the
first
thing
you
need
to
do
in
this
kind
of
research
is
to
say:
okay,
what
is
my
input
parameter
traffic
load?
How
I
will
represent
the
traffic
Mathematica
and
that's
the
very
first
thing
you
need
to
do.
E
This
is
called
feature
engineering,
for
instance,
if
you
want
to
represent
an
IP
address,
and
you
want
to
put
an
IP
address
as
an
input
in
a
neural
network,
you
cannot
say
that
an
IP
address
is
just
a
number,
because
actually
the
IP
address
has
a
structure
inside
and
if
you
see
that
the
just
a
number,
the
neural
network
will
not
understand
that
there
is
a
structure.
Okay,
so
that's
the
one
of
the
first
challenges
you
have
in
machine
learning
when
you
apply
to
this
kind
of
a
scenario.
E
Actually
it's
quite
interesting
because
there
are
many
parameters
in
networking
which
are
graphs
right.
The
topology
is
a
graph
routing.
Can
be
understood
as
a
graph
and
many
other
important
parameters
of
a
networks,
a
graph.
Actually,
the
network
is
a
graph
and
actually
representing
graphs
inside
neural
networks
is
extremely
complicated,
extremely
extremely
complicated.
E
So
the
answer
to
this
is
that
there
are
many
kind
of
neural
networks:
okay,
for
instance.
If
you
are
working
with
images,
you
have
one
specific
kind
of
neural
network
attached
or
at
least
need
to
use,
which
is
called
convolutional
neural
network.
If
you
want
to
work
with
information
that
it
is
a
story
in
sequence,
for
instance
like
in
a
sentence,
the
information
is
entered
in
this
story.
E
In
the
sequence
of
events,
then
you
need
another
kind
of
neural
network
which
are
called
recurrent
neural
network,
and
these
two
neural
networks
are
very
it's
a
very
popular.
There
is
lot
of
research,
lots
of
money
being
put
into
that
because
pretty
much
all
the
applications
in
Peter
vision,
they
work
with
computer
convolution,
neural
networks
like
self-driving
cars
and
so
on
so
huge
topic.
Recurrent
neural
networks
are
used
for
text,
processing,
translation
and
so
on.
Huge
topic.
Many
money,
also
many
research
done
into
that.
E
But
until
very
recently
we
didn't
have
a
neural
network
which
has
which
was
specifically
used
to
understand
information
which
was
stored
in
a
graph.
They
are
called
referral
networks
and
you
can
see
here
the
there
is
a
very
recent
publication
of
the
vertices
in
a
team
from
deep
mind
and
other
all
source
proposing
this
new
technology
and
well
we
tried
it
and
we
try
to
apply
it
to
a
computer
network.
You
have
here
the
citation
or
so
it's
an
open
paper
you
can
see
and
it
actually
works.
E
So
we
are
able
to
learn
to
train
a
network
model
with
one
network
and
then
ask
questions
to
this
network
module
with
a
completely
from
a
different
network
with
a
different
topology
which
has
never
seen
it
is
able
to
to
provide
a
Caritas.
But
the
question.
The
question
I
think
that
it's
relevant
for
research
will
appease.
Should
this
research
group
work
on
provided
guidelines
or
mechanisms
to
represent
common
network
parameters?
E
B
F
E
So
this
is
a
very
disruptive
and
profound
change
for
all
of
us
that
we
have
to
decide
if
you
want
to
accept
or
not.
But
that's
that's
a
big
deal
in
in
in
machine
learning.
So
since
we
are
not
unique,
this
is
a
program
for
many
other
applications.
There
are
two
pretty
much
two
approaches
to
solve.
This,
the
first
one,
is
called
explain
ability.
This
is
a
technique
to
try
to
look
inside
your
neural
network
and
try
to
understand
what's
happening
inside.
E
So
look
at
an
understanding
of
why
the
neural
network
has
taking
a
particular
decision.
I
will
say
that
this
is
quite
a
state-of-the-art
and
quite
a
big
research
topic,
but
there
is
another
approach
which
is
accountability,
which
is
you
use
external
boxes
that
will
pre
process
the
action
from
your
neural
network
and
because
you
have
a
better
knowledge
at
you're,
an
expert
on
your
computer
network.
You
can
decide
if
this
action
is
correct
or
not.
E
So
you
have
some
sort
of
accountability,
so
this
box
can
work
in
helping
you
setting
some
operational
bounds
or
limit
the
the
actions
taken
by
that
efficient
interaction.
So
this
is
pretty
much
your
safeguard
that
you
can
use
to
to
when
you're
operating
on
a
computer
network
with
a
neural
net.
Okay.
So
then
the
question
I'm
suggesting
to
the
research
group
is
is:
are
you
interested
in
to
working
on
guidelines
mechanism
and
architectures
to
support
this
kind
of
accountability?
E
I
seen
Xfinity
explained,
aerated
is
really
a
research
topic
and
it's
really
into
machine
learning
in
perturbation
interest,
the
third
one
databases
and
benchmarks.
So
there
is
no
machine.
There
is
no
machine
learning
a
neural
network
without
data.
Okay,
so
that's
you
can
not
do
anything.
Actually,
the
algorithms
are
cheaper
than
the
data
and,
if
you
take
a
look
to
other
well
establishes
well-established
should
fill
such
as
computer
vision.
What
they
do
is
they
have
opened.
E
The
data
set,
for
instance,
imagenet,
is
a
very
well
known
data
set
it's
a
repository
of
images,
okay,
so
anyone
they
see
that
you
want
to
learn
something
about
neural
networks,
vision.
So
when
you
go
to
visit
a
set
and
you
consult,
we
are
playing
around.
So
this
is
a
location.
Let's
say
that
you
want
to
do
research
on
on
this
field
on
computer
vision,
great,
you
have
an
open
repository
of
images,
you
go
there
and
you
download
all
the
repository.
E
You
train
your
neural
network
and
you
try
to
see
how
good
your
neural
networks
and
even
more
importantly,
you
want
to
benchmark
how
good
your
new
neural
network
is
compared
to
the
state
of
the
art.
Well,
there
are
also
well-established
neural
network.
For
instance,
electro
net
is
one
which
then
you
can,
you
can
say,
look
I
did
my
neural
network
in
computer
vision
and
I'm,
comparing
its
performance
to
Alex
net
and
I'm
20
percent
better.
So
now
pretty
much.
Everyone
knows
how
good
you
are
and
having
a
reference.
E
Now
we
don't
have
this
for
computer
networks,
so
students
and
I'm
having
many
emails
from
students,
young
people
that
they
are
happy
to
to
working
in
computer
networks.
They
are
interested
in
to
machine
learning
and
they're
willing
to
do
something.
There
is
no
data
set
to
start
learning
literally,
there
is
nothing
so
education
is
one
factor
a
but
I'm,
not
focusing
on
elevation.
Saying
that
even
for
an
equation
that
that's
an
issue
same
for
research
same
for
same
for
benchmark,
so
should
we
do
something
out
of
it?
E
We
have
a
very
early
attempt
in
this
reference
where
you
have
a
dataset
with
some
data
for
the
students
to
start
data
comes
from
a
simulator
which
is
not
ideal,
but
with
this
is
what
we
can
do
with
the
resources
we
have,
but
should
we
try
to
push
for
a
reference
dataset?
There
are
many
many
network
dimensions,
which
is
the
use
case
they
are
covering
because
with
images.
Well,
you
can
say:
ok,
those
are
images
of
things
then,
and
then
that's
something
right,
but
for
us,
what
is
the
use
case
we
are
willing
to
do?
E
This
is
my
very
personal
suggestion
if
we
decide
to
go
this
path,
I
think
that
we
should
start
by
agreeing
on
one
use
case
and
then
go
for
it
have
one
dataset
and
then
start
seeing
how
things
about
these
also
involved.
Some
of
course,
research,
group
working
group
activity,
because
we
can
go
to
other
working
groups,
research
groups
and
try
to
refine
measurements.
How
to
create
this
dataset.
Let's
say
that
we
decide
one
your
skis,
we
can
go
there
and
ask
them
okay.
E
So
then
the
question
is:
should
research
loop,
work
on
providing
guidelines
and
help
producing
reference,
data
sets
and
benchmarks
and
I
think
that
unless
we
agree
on
a
use
case,
it
doesn't
make
sense
sure
we
should
agree
on
acoustics
and
then
go
for
this,
and
then
this
is
a
more
general
fourth
question
which
is
well.
There
is
no
reference
machine
learning
architecture.
We
there
are
many
many
use
cases
and
there
are
several
documents
describing
distances.
Should
we
try
to
try
to
define
a
reference
architecture?
E
E
G
So
the
the
fifth
I
think
topic,
a
research
question
we
were
discussing
in
in
the
group
was
related
to
what
we
call
this
really
the
artificial
intelligence.
So,
of
course,
we
were
just
produced
in
the
last
night
that
there
is
also
some
market
or
architectural
related
topic
issue
pool
and
some
to
address.
G
And,
of
course,
if
you
we
also
thinking
what
are
the
differences
when
you
want
to
apply
AI
to
our
domain
thousand,
as
you
say,
to
two
images
and
of
course
we
have
to
deal
with
the
networks
that
we
think
are,
of
course,
very
different
and
images.
A
destructive
system
in
the
sense
also
geographically
distributed
system
very
critical,
because
we
want
to
apply,
in
many
cases,
artificial
intelligence
on
real
time
to
change
configuration
and
so
on
and
very
in
a
very
short
time.
G
H
G
Whereas
of
this
treated
AI.
We
can
also
try
to
pursue,
of
course,
have
multiple
objective,
because
if
you
think
of
course,
your
duty
system,
we
have
different,
let's
say
entities
already
systems,
they
want
to
do
the
different
things
configure
different
thing
for
certain,
maybe
application
and
so
on.
They
have
different
objectives,
so
not
necessary
push
to
the
same
objective
and
all
you
accommodate
all
this
objective
to
having
some
things.
That
is
good
for
all.
There's
a
majority.
G
Let's
say
this
is
also
a
problem
and
even
for
a
single
entities
that
want
to
let's
say
fulfill
different
type
of
a
few
is
very,
very
odd:
okay,
something
in
general.
You
says
that
you
just
know:
GT
functions
that
you
put
some
weights,
but
it's
not
so
easy,
sometimes
and
of
course,
come.
If
you
go
for
this
rated
aiu,
you
find
the
same
question
as
we
had
before.
What
are
the
requirements?
Maybe
the
requirements
you
skate
for
the
architectures
that
will
ship
out
this
answer.
So,
yes,
the
tricky
the
I
may
be
alone.
A
So
that's
why
sometimes
we
switched
presenters
because
we
are
more
comfortable
with
the
different
slides.
So
this
is
really
working
progress.
Some
are
maybe
some
not
mature
enough
idea,
but
we
would
like
to
bring
is
as
input
to
the
discussion.
So
this
is
this
two
slides
here
is
a
bit
my
view
about
what
I
was
mentioning
before
already
that,
for
me,
I
am
machine.
Learning,
as
techniques
and
component
will
is
part
of
the
landscape
we
have
to
do
to
do
with.
So
we
are
a
I
elements
and
one
so
I.
A
Don't
think
this
is
the
one
of
the
research
group
to
really
work
on
AI
techniques
themselves.
I
mean
we
can
have
a
good
knowledge,
good
understanding
and
even
maybe
try
to
reflect
on
the
fact
that
post
technique
needs
to
be
adapted
or
evolved.
Like
alga
was
mentioned
in
graph.
Neural
networks
is
something
that
is
that
has
been
specifically
designed
to
address
graph
and
so
could
be
well
applicated
for
for
network
environments,
but
I
don't
think
this
is
where
we
can
have
a
real
differentiator,
but
we
are
network
management.
A
So
I
was
thinking
about
also
presented
with
the
different
other
points
that
we
can
somehow
have
an
external
framework
surrounding
the
AI
elements
and
provide
the
right
so
interfaces,
but
also
the
mechanism
how
operator
can
use
them,
integrate
them
in
his
operational
environment
and
giving
what
I
call
feature
is
not
like
ml
features,
but
networking
management
features
in
the
sense
of
being
able
to
get
some
assessment
on
the
performance,
so
we're
talking
about
testing
or
benchmarking.
So
what
is
the
way?
I
can
get
performance
information,
testing
information.
A
On
my
on
my
element,
the
aspect
of
accountability,
explain
ability,
I
mean
I'm.
An
operator
I
deploy
AI
elements
in
my
system
and
there
is
an
SLA
violation.
My
customer
investigates
and
says
I
want
to
know
why,
and
it's
my
algorithm
that
that
made
it
okay,
but
please
give
me
more
more
insight.
So
this
is
the
kind
of
well.
If
even
if
it's
a
very
simple
example,
but
accountability
I
mean
this
is
as
Jerome
was
mentioned,
in
critical
environments,
we
need
to
be
able
to
make
them.
I
mean
to
have
confidence
in
how
they
run.
A
We
will
not
be
able
to
always
go
inside
those
AI
elements,
because
they
will
come
as
black
boxes
from
many
different
vendors
for
different
reasons.
We
don't
want
to
go
into
that,
but
if
we
have
this
kind
of
external
firm
work,
providing
capabilities
to
to
operate
them
and
to
have
knowledge
and
control
on
them,
it
seems
a
bit
the
picture.
A
All
the
data
are
collected
and
clean,
and
so
on
to
to
to
to
understand
what
will
be
the
behavior
of
the
AI
elements
in
an
operational
network
sort,
which
also
provides
some
level
of
confidence
from
the
operator
point
of
view
on
the
robustness
of
such
elements
to
different
variation
in
the
network,
and
then
again
we
are
a
network
so
that
there
will
be
of
those
elements
spread
out
in
the
network.
Again,
as
Jean
was
mentioning,
you
can
call
them
agents,
but
they
can
have
very
different
goals.
A
I
mean
they
will
do
a
different
task
in
the
network.
They
will
be
designed
for
doing
something,
but
they
can
be
selfish.
They
can
be
non
comparative.
They
can
be
completely
closed
blackbox,
so
we
may
not
always
have
the
way
to
to
to
control
them,
as
we
would
like
to
be,
I
mean
the
full
flexibility
but,
and
they
will
be
deployed
to
address
many
many
functions
across
the
network
and
again
that
is
I
will
not
give
we
come
from
a
single
vendor,
so
they
will
not
be
uniform
in
operation.
A
So
this
is
the
same
template
for
the
topic.
So
what
could
be?
What
should
the
energy
work
on
defining
distributed
AI
for
network?
So
again,
we
will
list
the
same
points
with
the
five
topics.
At
the
end
of
this
talk
to
raise
the
questions
for
the
group
I'll
be
after
me,
okay,
so
this
is
a
this
okay.
This
is.
A
D
Rick
Taylor
Airbus.
Thank
you
guys.
That's
a
really
great
summary.
I
thought
it's
a
really
nice
show
to
presentations
a
couple
of
quick
responses
from
myself.
I
think
it
was
point
number
two
which
is
not
quite
the
same
as
slide
number
two,
which
was
about
the
accountability,
yeah
accountability
and
explain
ability.
I
can't
remember
the
name
of
the
speaker,
but
the
question
was:
does
the
research
group
need
to
look
in?
Are
there
special
use
cases
for
accountability
and
explain
ability
that
apply
to
a
IML
in
networking
network
management?
D
My
question
in
response,
were
it
would
be:
is
there
a
difference
in
accountability
and
explain
ability
when
a
IML
is
applied
to
network
management
as
compared
to
when
AI
is
applied
to
computer
vision
in
cars
or
wherever
else,
a
IML
of
different
flavors?
If
the
answer
is
it's
all
about
accountability
and
explained
ability,
then
going
back
to
one
of
the
somebody
else's
point.
Is
we're
not
about
research?
We're
not
the
AI
research
group
with
a
network
management
research
group.
So
unless
there's
a
particular
accountability,
that's
required
for
network
management
I
would
suggest.
No
sorry.
D
D
So
what
we
can
bring
to
network
management
and
artificial
intelligence
is
our
expertise
in
networking
and
how
the
internet
has
been
built
and
how
it
functions.
Whether
an
agent
is
a
human
being
or
an
AI
or
a
machine
learning
system
or
whatever
doesn't
really
make
any
difference.
We
know
how
to
make
different
ISPs
talk
to
each
other
across
a
gateway
protocol.
We
know
how
to
divide
the
complexity
of
managing
the
Internet
into
small
enough
chunks
that
somehow
it
all
seems
to
run.
D
Despite
the
fact
governments
may
disagree
over
policies
and
corporations
may
disagree
over
who's
going
to
make
the
most
money
out
of
it.
So
I
think
some
of
the
policies
when
it
comes
to
distributed
AI,
so
I'm
jumping
to
sort
of
point
five
I
think
we
know
roughly
how
to
do
that.
I
think
we
have
a
pretty
good
idea
of
best
practice.
D
There
may
need
to
be
some
update
or
some
review
of
current
best
practice
using
humans
to
see
how
that
maps
to
using
a
is,
but
I
would
suggest
that
an
AI
is
really
just
another
example
of
an
agent
who
is
making
independent
decisions,
but
somehow
that
decision
has
to
work
as
part
of
the
greater
good
or
that
to
maximize
benefit
for
the
entire
system.
If
you
want
to
be
AI
NL
about
it,
yeah
oh
shut
up,
but
those
are
my
key
points.
A
E
To
the
first
question
you
made
I:
don't
think
that
there
is
any
difference
between
accountability
and
explain
ability
for
AI
network
compared
to
other
areas,
and
the
reason
is
that
you
may
have
an
agent
operating
an
infrastructure
which
can
be
a
car
or
a
network
and
the
actions
that
it
will
take.
Well,
we
want
them
to
be
accountable.
Unexplainable
I.
D
A
Commenting
sabine,
you
want
to
comment
also
on
those
points
or
not,
because
I
would
just
would
like
to
respond
to
this
question
of
accountability
as
a
participant
and
I
agree
on
the
not
to
go
there.
For
a
fourth
reasons,
and
in
fact
that
so
the
existing
I
might
give
it
leave
it
open
for
one
dimension.
A
What
metrics
or
data
to
collect
that
are
specific
and
how
the
operator
network
operator
can
use
that
this
is
something
we
can
provide
like
an
architecture
or
maybe
the
right
interfaces
which
will
be
different
from
finance
from
elf
from
whatever,
because
this
is
not
the
same
technology,
but
this
is
this
would
be
for
me,
the
only
area
where,
potentially,
we
can
try
to
investigate
something
but
I'm
not
prescribing
anything
in
today's.
In.
D
I
This
is
Sabine
from
Nokia
I,
rather
have
a
question
related
to
the
slides
about
distributed
ai,
so
so
that
raised
a
question
so
actually
distributed.
Ai
is
the
very
point
that
of
potential
difficulty,
because
that
means
a
network
is
not
one
single
topologies,
not
one
single
layer.
It's
it's
it's
a
jungle,
it's
an
ontology
or
whatever
you
call
it
so
I
think
the
big
challenge
would
be
to
if
there
is
to
be
to
define
guidelines.
I
The
first
thing
we
want
to
have-
supposing
we
are
not
scared
about
the
idea
of
a
system
that
takes
probabilistic
decisions,
is
to
make
sure
all
these
all.
These
AIS,
for
example,
that
are
related
working
on
top
of
one
part
of
the
network,
are
one
part
of
the
aspect
we
should
make
sure
they
take
harmonized
decision
so
how
to
how
to
build
a
I.
I
Don't
know
mechanisms
of
framework
that
ensure
that
the
the
the
the
the
the
at
least
the
decisions
that
this
systems
take
are
understandable
among
each
other,
assuming
they
are
each
the
correct
decision,
which
would
mean
how
do
we
ensure
that
they
also
have
a
harmonized
when
I
say
harmonize
I
mean
a
view
of
the
network
of
the
state
of
the
problem
of
also
a
view
of
the
decision
making
that
they
can
share
among
each
other
and
that
that
is
understandable
among
each
other.
I
think
this
would
be
a
crucial
work
to
do.
D
D
Let's,
let's
reinvent
the
way
we
make
the
Internet
hang
together
in
a
different
way,
because
we
want
to
use
AI
rather
than
say:
hey
we've
got
this
tool
called
AI
which
will
allow
us
to
make
probabilistic
decisions
within
partial
portions
of
the
network
or
perhaps
to
bind
portions
of
the
network
together
in
a
distributed
or
a
top-down
or,
however,
we
want
to
do
it.
I
think
we
should
start
with
where
we
are
and
look
at
what
we've
got
before.
We
tear
it
up
and
start
again
because
it'll
be
fun,
but
we
won't
actually
achieve
anything.
A
Just
I
will
leave
you.
The
Floris
I
mean
again
as
a
participant
to
two
elements
of
clarification.
First,
on
the
question,
I
mean
the
comment
from
from
Sabine
in
this
slide
with
destroy
today,
I.
In
fact,
we
are
covering
product.
Two
different
elements
because
distributed
AI
is
something
that
is
existing
today.
Federated
learning
or
different
approaches
to
to
perform
distribute
a
I,
but
one
of
the
aspect
described
in
this
slide.
Is
that
and
it
may
be,
not
explicit,
is
you
may
have
specific?
A
I
mean
it's
that
the
question
between
the
fact
that
network
environment
is
by
definition
distributed
and
the
fight
that
is,
it
relevant
to
consider
a
way
to
distribute
AI,
because
sometimes
data
are
only
local.
They
are
ephemeral,
they
can
be
relevant
only
for
certain
period
of
time,
and
so
certain
approach
of
AI
may
not
really
fit
well
with
thing
that
can
only
be
by
definition
distributed.
A
So
if
this
is
confirmed
to
some
extent,
then
we
can
go
to
the
path
of
distributed
AI,
which
raise
other
questions
in
terms
of
coordination
or
cooperation
of
the
learning
that
can
be
made
made
local
and
it's
I
mean.
Maybe
you
can
transfer
the
model
or
transfer
the
learning
or
all
this
aspect
of
different
approaches
to
approach
the
shoe
today
right.
A
So
this
is
one
one
one
element
I
would
like
to
clarify
and
in
this
element
of
clarification,
what
I,
I
lighted
here
is
a
bit
different
from
this
aspect
of
I
mean
kind
of
techniques
of
distributed
AI.
This
is
more
the
fact
that
my
way
of
representing
that
AI
elements
will
be
spread
everywhere
in
the
network
for
different
goals
and
different
algorithm
insight
cetera,
but
we
will
have
to
to
integrate
them
in
the
operational
processes
of
operator
or
Internet.
A
The
internet,
which
you
see
it's
it's
distributed,
but
it's
not
really
looking
at
how
we
do
this
repeatedly
I.
This
is
these
elements
are
not.
It
doesn't
need
to
really
be
coordinated
in
order
to
share
their
learning,
which
is
more
like
from
integration
from
operational
processes.
Pointers
this
just
to
make
a
clarification,
because
I
think
we
and
animate
your
mentioned
rough
consensus
for.
A
D
A
One
just
quick
comment
and
seven
I
gave
you
the
florist
I
had
a
discussion.
Also.
Yesterday
we
see
new
RDF
chair,
so
we,
this
topic
was
not
a
dust
stage
yesterday,
but
one
element
we
we
came
in
our
discussion
with
whis
Colleen
was
we
can
identify
very
nice
topics
and
really
maybe
try
to
go
into.
This
is
really
where
we
want
to
go,
but
there
is
another
key
point
is:
do
we
have
the
right
competency
in
the
group
to
address
it?
A
I'm,
not
questioning,
if
I
think
a
lot
of
very
smart
people
in
ITF,
but
if
we
go
into
some
directions
and
we
may
not,
we
might
need
to
go
into
other
communities
and
try
to
get
the
right,
also
expertise.
Coming
from
that.
You
know
that,
but
a
bit
like,
similarly
to
the
your
initial
comment
on
DTN
that
you
don't
necessarily
have
the
expertise
of
the
knowledge
of
network
management,
so
you
don't
want
to
reinvent
obvious
things.
This
is
the
same
for
us.
A
I
mean
we
may
say
this
is
the
right
program
from
an
emoji,
but
if
we
don't
have
the
right
skills
from
maybe
specialized
type
of
AI,
then
we
may
not
be
able
to
address
the
topic
properly.
That's
why
some
I
might
comment
from
a
discussion
risk
on
so
please
I
mean
sorry.
It
was
a
bit
longer
than
Brendan.
I
If
you
we
already
have
it,
because
there's
lots
of
stuff
around
but
I,
if
we
can,
if
the
answer
is
yes,
that's
great,
but
I
still
miss
I,
don't
know
some
tools
that
is
able
to
say.
Okay,
I
have
my
big
networks
and
I.
There
is
a
problem
there
by
the
way
when
I
talk
about
distributed,
it's
not
like
the
OSPF
scheme,
it's
also
a
different
level.
So
it's
not
the
information
or
circulation
is
not
among
entity
that
do
the
same
thing
at
the
same
level.
I
It's
also
some
of
the
entity
are
input
to
one
another.
So
if
there
is
a
tool
that
says
that
is
able
to
I,
don't
know
if
it's
translate,
if
I
don't
know
how
to
do
it,
I
mean
it's
I
think
well.
That
may
be
one
of
the
requirements.
If
we
make
a
requirement
list
to
achieve
this,
that
would
be
a
tool
that
enables
us
to
to
represent
the
the
network
views
with
problems
and
decision
in
some
translatable
way.
So
it's
not
about
okay,.
E
I
didn't
want
to
make
this
point
what
I
was
presenting
since
I
was
trying
to
be
as
objective
as
possible
and
trying
to
capture
the
discussion
that
we
had
now.
I
can
be
a
little
bit
more
subjective,
I
guess:
I
have
some
experience
on
IRT,
F,
machine
learning
and
I
know
it
is.
There
is
always
a
certain
extra
girl
on
ok
yeah.
This
is
great,
but
how
these
will
affect
the
protocols
interfaces.
I,
don't
know,
I
have
no
idea.
E
What
I
think
is
that
AI
will
transform
the
way
we
manage
and
operate
networks
I'm
fully
convinced
about
that
I
might
be
wrong,
but
if
I'm
right
this
is
a
big
thing
and
that
we
cannot
just
because
we
don't
see
a
clear
road
map
for
IDF
activity.
In
this
just
say:
no,
that's
a
mistake:
it
is
going
to
transform
our
world,
we
have
to
to
be
ready
and
we
have
to
understand
it
as
much
as
possible.
E
This
brings
to
my
second
point:
I,
think
that
here
we
have
outstanding
protocol
and
networking
experts
at
the
best
in
the
world,
but
maybe
we
don't
have
all
the
AI
experts
in
the
world-
and
this
goes
to
my
specific
proposal,
so
I
think
that
what
we
should
be
working
on
is
try
to
understand
how
this
will
impact
networking.
Try
to
invite
the
artificial
intelligence
experts,
try
to
see
what
people
are
doing,
listen
to
them
and
then
try
to
see
how
this
will
have
an
impact
on
the
network,
but
not
the
other
way
around.
E
F
My
name
is
jung:
un
homme
on
critical
to
last
pacing,
yeah
I,
remember
that
from
ITF
non-being
NN
on
the
dock,
machine
learning,
our
G
and
I
dinero
ambition
of
fist
or
crew,
but
they
happen
and
disappears
so
nowadays,
an
emeritus
of
the
only
place
to
hinder
AI,
poor
Network
AI
aporia
to
measurement.
So
we
also
present
the
sever
time
the
network
much
almost
measurement
for
our
error
mm
mechanism.
F
So,
although
it
is
typical
to
make
quantum
John
to
other
place,
but
it
is
very
happy
to
receive
feedback
what
we
saved,
the
other
person
opinion
what
year
you're
missing
point
and
what
is
your
other
approach
to
solve
your
problem?
It
is
good
to
ours,
so
I
strongly
support.
This
issue
should
be
included
in
the
etymology
at
the
day,
but
I
would
like
to
add
a
couple
of
comment
pro
regarding
the
question
providing
guidelines
or
mechanism.
F
Yes,
if
we
have
this
guideline
mechanism
very
good,
but
you
know
the
docs
network
is
a
kind
of
different
from
the
previous
two
motion
machine
learning
problem.
That
is
not
a
problem
or
regulation.
That
is
not
a
problem
of
the
classification.
It
is
more
complex
and
if
we
talk
about
Delta
machine
learning
to
here
with
some
consensus,
what
is
the
network,
the
network
scope
and
what
is
the
purpose
of
the
net
to
measurement
to
reduce
to
reduce
cost?
What
you
reduces
something
so
I
guess
that
we
need
more
time
to
get
some
common
consensus.
F
J
So
it's
not
clear
to
me
that
they're
not
already
out
in
front
of
us
and
probably
it
will
be
an
interesting
I,
think
science,
experiment
too,
just
like
existing
machine
learning
stuff
will
loot
loose
on
some
of
the
networks
and
they
probably
come
back
and
tell
us
where
we
got
problems,
and
you
know
stuff
like
that.
It's
not
obvious
to
me
that
we
need
to
spend
any
time
thinking
about
stuff
for
them.
I'd
like
to
see
what
they,
the
the
machines,
could
already
tell
us
about
our
existing
networks
and
infrastructure.
J
So,
I
to
say
that
we
should
get
other
people
who
are
doing
research
in
to
come
and
tell
us
what
they'd
already
know,
because
I'm
guessing
that
we're
more
near
fights
here
than
most
people
and
I
might
sense
as
if
they
can
find
you
know
unknown
carcinomas
in
in
you
know,
medical
images.
They
could
find
stuff
that
you
know
we'd
be
surprised
about
quite
trivially.
So
it's
not
clear
that
that
it's,
not
our
ignorance
and
and
their
education.
If
that
makes
any
sense.
J
A
C
A
K
Ideas,
okay,
yeah
Bertolli,
if
I
want
to
follow
up
in
that
point,
because
I
think
it's
a
really
good
point
that
we
have
been
working
expertise
here
and
we
know
what
we
want
to
do
with
the
network
right.
So
maybe,
instead
we
have
espresso
machine
learning
as
well
in
the
room
probably,
but
even
if
we
are
showing
that
si
ATF
will
not
have
the
machine,
their
name
is
parties
which
I
think
we
we
may
have
well.
K
We
said
that
you
can
do
is
expose
our
requirements
for
the
machine
learning
community
to
address
and
to
explore
and
whatnot,
and
even
if
we,
as
a
research
group,
don't
define
any
new
interface
or
language
or
model,
we
said
that
we
can
define
guidelines.
We
can
define
informational
drafts
that
can
help
a
lot
to
to
to
marry
these
two
words
together.
Right,
so
I
think,
there's
there's
another
potential
of
what
we
can
have
tell
the
world
about
with
fact.
Formation
right.
D
Rick
Tyler
again
following
up
on
that
I'm
agreeing
with
everyone
in
the
line
at
the
moment,
I
see
the
ability
to
give
key
I'm
repeating
previous
comments.
But
yes,
if
we
can
reach
out
to
the
air
community
and
say
here,
are
our
problems
and
going
back.
My
appointments
about
data
sets
I
believe
within
the
IETF,
an
IRT.
If
we
have
these
data
sets,
we
just
haven't
told
anyone
about
it.
I
imagine
the
BGP
guys
can
give
you
three
hundred
and
thirty
thousand
node
graphs,
just
on/off
a
command
line.
D
You
know
they
can
pull
this
out
of
out
of
their
data.
I
seem
to
write
about
fifteen
years
ago.
I
was
doing
a
lot
of
work
on
PTP
systems
and
I.
Remember
the
computer
science
department
in
Washington,
Washington
State
University,
had
this
thing
called
the
King
data
set,
which
was
a
massive
data
set
to
arrive
from
DNS.
So
it's
kind
of
a
question
to
the
room:
go
back
to
your
universities
in
your
institutions.
L
Hi
I'm
Sean
from
next
door
I'm
here
mostly
to
hear
all
but
I,
have
some
background
in
network
management,
currently
work
in
vehicle
to
vehicle
network
and
machine
vision,
so
I
have
a
proposal
for
our
birth
and
the
lady
hero.
So
when
we
go
look
for
something
like
a
pothole
or
a
jaywalker
or
slow
down
on
an
obstacle,
a
blockage,
we
first
define
what
it
is
we're
looking
for.
So
something
that
you
can
do
is
just
write
a
meeb
for
the
network.
L
L
A
M
Can't
Perkins
is
gesture,
I
tend
to
think
a
lot
of
the
value
we
have
in
there.
I
RTF
is
bringing
together
the
various
different
communities,
bringing
together
that
the
researchers
and
the
the
network
operators.
The
engineers
I,
can
certainly
see
value
in
bringing
together
people
with
a
who
have
a
a
I
expertise
and
people
with
network
management,
expertise
and
I
can
imagine
there'll
be
interesting
things
that
could
come
out
of
that
I
think
we
would
have
to
see.
M
You
know
both
a
set
of
interesting
topics
and
interesting
research
directions
and
a
community
forming
to
show
that
we
have
something
which
can
be
a
successful
set
of
research
in
this
direction
and
I'm.
You
know
I
think
that
could
that
could
well
be
interesting
things
here,
but
we
need
the
the
community
as
well
as
the
topic.
A
Yeah
you
only
if
you
give
me
a
second
just
also
because
to
react
to
Colleen's
late
statement
and
the
one
off
point
of
Albert
something
we
have.
It
was
quite
difficult
and
still
difficult
that
we
have
tried
to
do
with
listen
always
when,
when
you
want
to
attract
people
or
go
to
people
and
tell
them
look,
we
have
some
some
stuff.
A
Do
you
have
any
proposal
or
ideas
how
to
make
it
a
bit
efficient
I
mean
because
here
in
the
room,
we
have
a
bit
of
a
mix
of
people
that
are
networking
so
may
I
some
other
fields,
and
so
we
have
a
bit
of
a
mix.
But
if
you
want
to
have
a
bit
more
collection
of
the
network,
requirements
on
network
needs
and
present
that
to
the
other
AI
communities
to
say
loudly,
how
can
we
make
this
happen?
A
I
mean
it's
just
not
only
by
going
to
a
ml
conference
that
we
achieve
that
I'm
inviting
them
to
come
here.
I
mean
it's
quite
challenging
to
to
attract
in
the
first
time,
people
that
have
no
no
clear
I
mean
incentives
to
be
here.
It's
costly,
it's
not
academic
conference.
So
it's
really
an
open
I'm,
not
only
targeting
to
you
Albert,
but
if
you
have
an
idea,
it
would
be
welcome.
C
So
you're
sure
well
I'm,
reading
a
message
from
Java
from
Pedro
Martinez
jr..
He
is
that
n
is
et
and
he
wrote
I'm
wondering
if
we
shall
talk
about
AI
or
just
machine
learning
from
what
I
see
here,
it
is
reduced
to
some
sort
of
prediction
of
behavior
from
processing
data.
It
is
not
actually
exploiting
all
AI
methods,
such
as
reasoning
and
planning
in
case
network
management
wants
to
plot
exploit
AI
I
encourage
to
also
consider
these
other
methods,
at
least
reasoning
and
planning
and
research
how
they
really
affect
network
management.
E
Yeah,
this
is
Albert,
so
I
think
we
should
do
the
opposite.
We
should
not
go
there
and
ask
them
about
our
requirement.
I'd
rather
invite
them
here
and
I
will
go
back
to
that,
how
to
invite
them
and
see
what
they
are
doing
and
how
these
kind
of
technologies
will
transform
our
world.
Let
me
put
a
specific
example:
there
is
a
very
recent
paper
about
some
important
researchers.
I,
don't
remember
exactly
a
name
of
the
researcher,
but
it
is
DRL
for
packet
classification.
E
So
it's
pretty
much
a
complete
new
way
to
the
packet
classification
for
forwarding.
Do
we
know
if
this
is
going
to
change
the
way
we
need
to
design
protocols?
This
is
very
related
to
the
talk
we
had
yesterday
on
the
on
the
deep
dive
on
the
technical,
deep
dive.
So
I,
don't
know
really
I,
don't
know,
I
think
it
would
be
a
great
idea
to
invite
them
and
they
show
us
what
they
did
and
then
we
understands
how
this
will
change
protocol
design-
and
this
is
a
very
specific
example.
E
If
you
inviting
them
it's
complex,
then
let's
go
where
they
meet.
I
think
that
we
can
collocate.
There
are
plenty
of
AI
worships
AI
for
machine
learning,
workshops
and
conferences.
Right
now
see
Comcast
like
two
or
three
concurrent,
so
maybe
co-locating
our
research
group
meeting
there
and
then
inviting
those
people
to
present,
and
then
let's
discuss
how
this,
how
what
they
are,
whatever
they
are
doing,
changes,
protocol
design
that
were
design
and
so
on.
A
Believe
the
the
point
raised
by
peddle
is
fully
relevant,
addressing
machine
reasoning
and
other
parts
of
AI
that
are
not
much
learning.
We
had
this
point
in
our
discussion
earlier
this
week,
which
the
participants
providing
these
slices.
This
was
say
a
this
was
mentioned.
I
think
it
was
also
believed
that,
as
a
first
introduction,
we
did
not
want
it
to
open
too
much
other
things
and
also
from
the
people
in
sitting
in
around
the
table.
A
A
So
in
this
last
part
that
will
be
true
to
sir
part
one.
This
will
be
I
give
you
some
high
level
research
group
information
which
I
think
it
will
be
useful,
but
at
the
end
there
is
a
bit
more
that
the
process
for
for
the
evolution,
and
it
will
be
also
open
for
you
to
comment
most
specifically
on
that.
A
So
the
meetings
planned
for
for
this
year
already
been
announced
on
the
mailing
list.
I
will
provide
updates
as
they
come,
but
the
next
very
next
meetings
will
be
in
April,
the
I
Triple
E
I
am
Conference
in
Washington,
so
there
will
be
two
sessions
there.
It's
quite
different
from
the
type
of
session
we
have
been
in
IOT
freaks
because
we
try
to
insert
ourselves
in
the
program,
so
we
have
very
short
to
one-hour
session.
A
One
will
be
on
for
the
kind
of
new
formal
to
be
try
to
educate
and
share,
raise
awareness
of
what
is
IOT,
F,
IETF
and
network
management
in
the
ifip
and
actually
network
management
communities
also
to
try,
try
to
act
a
bit.
The
more
younger
researchers.
True
to
this,
the
second
session
will
be
a
technical
session
and
the
topic
is
not
yet
defined.
In
June
we
will
have
a
this
is
under
preparation.
We
potentially
an
interim
meeting
Santa
Clara
Mountain
View
in
the
u.s..
A
Some
elements
are
still
being
discussed
about
the
hosting
and
the
exact
date.
So
I
will
keep
you
updated
on
that.
But
again
this
should
be
a
kind
of
face
to
face
interim
I
know
it
could
be
challenging
for
people
to
attend,
but
we
have
a
set
of
Parsifal
that
are
a
lotta
base
there,
and
we
would
like
to
take
this
opportunity.
If
you
are
managing
to
humility
to
come,
it
will
be
great
I.
Don't
have
any
clue
about
potential
support
for
remote
participation
to
the
best
extent.
A
But
we
would
like
to
start
trying
to
implement
first
first
aspects
and
try.
Ok,
go
a
bit
more
beyond
the
paper
or
theoretical
aspect
and
try
to
play
with
some
things
there,
and
this
could
be
also
a
kind
of
warm-up
for
some
project,
a
bit
longer
term
to
participate
in
a
cotton
in
ITF.
On
this
idea
and
topic
once
we
have
a
bit
more
experiences
and
results
to
work
with
potentially
also
November
ITF
106
in
Singapore
bits
father.
But
this
is
a
bit
our
usual
plan,
something
we
are
putting
also
in
place.
Monthly.
A
A
Announcement
so
there
will
be
applying
it
working
research
workshop
2019,
which
will
be
in
July
during
the
IETF
week.
If
it's
like
last
year,
it
will
be
on
the
Monday.
There
will
be
no
research
group
scheduled
on
Monday,
so
to
come
for
the
participation
of
this
workshop.
The
colorful
paper
is
out.
One
of
the
topic
is
on
new
approaches
to
network
management,
operation
and
control,
but
some
other
topics
also
related,
so
major
ideas,
the
deadline
is
feasible,
as
you
have
different
forms
of
papers,
short
papers
or
a
bit
longer
paper
I
think
it's.
A
It
would
be
good
to
have
some
participation
from
the
track
management
so
again
submit
your
ideas.
If
you
want
to
discuss
more
at
the
group
level-
and
we
can
also
put
me
in
the
loop
and
I
try
to
see
if
there
are
common
opportunities,
but
it
will
be
nice
to
to
participate
to
this
workshop
research
of
information,
so
Kalina
had
to
leave,
but
I
discussed
with
him
yesterday
about
the
new
culture.
We
found
a
way
and
cleared
the
situation,
so
the
new
culture
will
be
announced
in
a
few
weeks
depending
a
bit
away.
A
We
managed
to
get
interviews
with
the
the
candidates,
but
we
are.
This
will
be
resolved
very
soon.
Also,
anticipating
a
bit
on
the
official
email.
I
will
send
about
that,
but
I
reached
out
to
the
group
to
have
secretaries
appointed
because,
as
we
are
going
into
more
recharter,
a
more
regular
meetings
and
wanting
to
ramp
up
on
the
support
of
the
research
group
to
the
activities,
we
believed
that
it
was
necessary
travels
or
some
some
people
in
support.
A
We
receive
several
candidacy
and
I
selected
Jefferson,
Campo,
snobbery
and
Pedro
Martinez
Julia
in
the
roles
why
to
research
retirees
essentially
because
we
are
operating
one
wide
and
some
since
we
will
have
also
virtual
meetings
in
different
time
zones,
it
will
be
essential
to
provide
supports
for
photos,
meetings
and
also
allow
for
potential
physical
participation
to
the
meetings
of
those
secretaries
which
are
already
active
participant
of
the
research
group.
So
thank
you
for
thank
you
to
them
for
for
standing
up
and
good
luck
for
further
mission
and
to
cope
with
the
research
with
chair.
A
A
So
this
is
to
be
refined,
but
currently
this
is
situation
number
of
documents
on
documents
for
the
research
group.
We
have
some
individual
documents,
no
research
group
document
so
currently
so
number
of
them
that
are
related
or
not
to
the
research
agenda.
This
is
absolutely
not
an
issue
somehow
new
proposal.
Some
are
renewed
proposal
and
this
is
part
of
energy
evolution
discussion,
because
at
some
point
you
will
have
to
clarify
very
specific
I
mean
product.
A
We
would
like
to
have
on
a
research
group
and
so
documents
and
milestone
and
ought
to
adopt
them
as
research
group
documents,
especially
for
the
topic
on
intent-based
networking,
which
is
a
bit
more
mature
on
this
topic
of
intern
base.
This
was
presented
in
the
last
session
that
once
estimate
Ezard
for
the
people
that
were
not
there.
A
The
time
intent
a
design
on
one
time
above
impact
of
the
approach
is
using
C
ICD
interoperability,
which
we
need.
Turn-Based
systems
based
on
the
presentation
from
from
Jeff
the
aspect
of
continuous
validation
of
intent
and
embezzled
analytics,
but
also
something
I
think
is
very
important
for
us
to
go
a
bit
more
into
the
space
of
kind
of
evaluation
of
the
after
concept.
Validating
I
mean
having
having
quantifying
elements
to
validate.
A
A
Deliverable
milestone
and
creatura
for
research
adoption
on
the
IBM
topic
to
clarify
the
set
of
documents
we
would
like
to
produce
in
which
form
and
to
set
some
milestone
in
terms
of
adoption.
Next
meeting
to
meetings
to
now
beat
the
pace
and
our
goals
in
realizing
those
those
objectives,
so
I
will
not
cover
the
last
part
just
support,
but
with
different
forms
for
the
for
the
research
group
for
the
future
of
the
group.
This
was
presented
also
last
time.
A
Some
important
questions
and
I
and
principle
I
would
like
to
remind
where
our
research
group
and
it's
important.
We
are
research
and
we
are
a
group
not
just
a
collection
of
individuals,
so
important
in
the
evolution
to
clearly
understand
what
our
shared
goal.
Watts
collective
outcome
we
want
to
deliver
and
the
discussion
on
the
AI
for
network
management
was
really
illustrative
of
that.
Who
is
the
public
which
are
communities
I
mean?
Who
do
we
address?
Is
it
just
ourself?
Is
it
the
IETF?
Is
it
all
the
research
communities?
Is
it
the
operator
community?
A
So
it's
important
to
understand
who
is
involved
in
order
to
define
the
right
scope,
the
right,
beneficial
reason,
what
we
need
to
deliver
our
approach
and
way
of
working.
A
finish
is
going
quite
well
and
but
always
can
ask
a
question
about
exploring
other
approaches
and
or
changing
a
bit
the
way
we
meet
the
way
we
discuss
together
at
some
point
and
for
some
topic
it
may
be
also
more
obvious.
But
what
is
the
link
with
the
IETF?
At
some
point?
A
It
could
be
also
a
valid
point,
depending
on
the
maturity
of
some
of
our
research,
to
try
to
again
provide
guidelines
or
may
be
a
bit
more
clear,
but
potential
standardization
path
that
could
that
could
come
out
from
from
the
research
of
the
group
element.
Clearly
disciplinarity
and
cross
fertilization.
I
really
believe
that,
and
the
topic
was
of
views
that
we
should
try
to
benefit
from
really
competency.
That
comes
from
different
background.
Different
fields.
A
A
Existing
Charter
is
to
capture
the
current
focus
of
what
has
been
agreed
as
topics
for
the
research
group,
so
the
current
main
topic
is
intent-based
networking
and
we
have
set
of
contributors
proposal
for
I
mean
specific
items
of
investigation,
so
this
will
be
documented
as
a
one
of
the
current
focus
of
the
research
group
and
also
with
milestones
to
deliver
on
that.
This
doesn't
close
the
door
to
other
topics,
but
as
long
as
we
don't
have
something
that
are
bit
more
concrete,
a
bit
more
tangible,
it
will
not
be
part
of
the
Charter.
A
A
A
So
I
will
think
about
some
questions
and,
of
course,
what
I
expect
from
you
is
to
try
to
answer
the
survey
of
course,
but
to
come
with
some
proposals
and
also
why
you
think
those
your
answers
should
be
should
what's
the
motivation
behind
the
answer
and
that's
all
for
this
status
of
the
research
group,
so
I
don't
know
how
much
time
we
have
left
so
much
I
think
10
minutes.
We
still
have
some
time
for
comments,
opinions,
proposal
for
this
evolution
of
the
research
group.
A
Ok,
we
had
a
pretty
good
interaction
for
the
AI
network
management
sessions.
I
think
people
are
a
bit
tired
now.
Thank
you
very
much
for
your
time
for
being
here
and
for
all
the
very
good
inputs
and
feedback
we
close
the
session.
Now
we
will
provide
the
minutes
as
soon
as
possible,
and
so
you
ready
for
the
next
virtual
meeting,
which
would
be
made
of
fibro
something
like
this.
Thank
you
very
much
good
bye,
see
you
next
time.