►
From YouTube: NMRG Interim Meeting, 2020-05-29
Description
NMRG Interim Meeting, 2020-05-29
A
A
A
A
A
A
A
D
D
Basically,
if
we
have,
if
we
want
to
commercialize
a
product,
we
need
to
develop
it
in
the
independence
lab
or
in
networking
lab,
and
basically
we
can
have
a
controlled
test
bed
and
then
artificial
intelligence
model
that
can
be
a
neural
network.
So
in
this
this
scenario,
what
we
can
do
is
to
train
this
model,
and
then
this
can
result
in
a
in
a
final
product.
But
then
this
final
product
should
be
deployed
in
in
the
customers
network
and
and
operate
well.
D
So
the
main
barrier
that
we
find
with
traditionally
a
solution
for
networks
is
that
they
are
not
able
to
generalize
water
networks
that
even
that
they
didn't
see
during
the
during
the
training
phase
and
basically
when
they
operate
in
this.
If
you
apply
this
final
product
in
the
in
the
customer
network,
the
I
product
will
fail
to
operate
well.
D
So,
basically
it
is.
It
has
issues
to
be
commercially
by
our
commercially
viable
product.
Another
alternative
would
be
to
directly
the
AI
tools
in
the
customer
network.
However,
this
is
unfeasible,
since
it
would
require
a
network
instrumentation
and
also
in
optimization
solutions.
It
may
cause
service
disruption
due
to
possible
configurations
applied
by
their
by
the
aie
solution
and
also
it's
not
feasible,
to
replicate
for
every
customer.
The
network
in
a
networking
lab
to
train
the
day.
D
A
problem
I
would
like
to
claim
here
that
it's
to
have
commercializable
aie
solutions,
it's
necessary
to
to
have
a
models
that
are
able
to
analyze
to
other
networks
that
were
not
seen
during
the
training
phase.
In
this
context,
the
recently
proposed
graph
neural
networks,
where
the
unique
a
a
based
models
that
were
able
to
generalize
to
other
networks
that
were
in
seen
during
this
training
phase.
D
With
this
kind
of
models,
we
can
perform
an
offline
training
on
control
test,
but
in
the
independence
lab
and
then
create
a
final
product
that
can
operate
successfully
on
any
customer
network.
So
this,
in
this
context,
graph
neural
networks
has
the
necessary
characteristics,
will
be
a
commercially
viable
AI
solution
for
format.
D
So
each
use
case
in
the
end
requires
to
make
a
mathematical
formulation
to
represent
all
the
different
network
elements
that
are
involved
for
this
particular
use
case,
and
these
elements
should
be
represented
in
the
form
of
glass.
So
at
the
at
the
right,
you
can
see,
for
example,
the
the
mathematical
formulation
of
Ramnath,
which
is
not
a
simple
formulation.
So
it
includes
some
elements
like
the
topology.
The
rotting
configuration
the
traffic
and
it
it
can
be
extended
to
other
elements
such
as
security
policy.
Another
depended
an
important
use
case.
D
What
we
see
is
that
we
need
a
much
learning
experts
nowadays
that
have
high
skills
on
neural
network
programming,
including
or
considering
some
languages
such
as
tensor
flow.
So
the
main
motivation
of
of
this
presentation
is
that
to
boost
the
adoption
of
global
networks
applied
to
networking,
it
is
essential
to
simplify
the
implementation
of
fast
DNN
prototypes.
D
D
Basically,
what
is
this
framework?
This
is
a
generic
framework
where
you
have,
you
can
have
network
use
case.
It
assists
users
to
create
a
custom
graphic
network
detector,
particularly
adapted
to
this
use
case.
So
this
is
an
easy-to-use
DNN
tool
boots
for
networking
researchers
and
practitioners
that
do
not
have
a
specific
background
on
designing
a
neural
network
solutions.
D
So
the
the
main
motivation
to
present
this
framework
is
that
nowadays,
if
you
want
to
design
a
CNN
prototype,
you
have
to
deal
with
with
programming
in
in
some
tensor
based
languages
like
the
one
known
tens
of
long
titles
or
the
more
noble
approach
of
Huawei,
a
minus
4,
but
all
of
them
are
based
in
complex
potential
separations.
So
if
you
want
to
create
your
own
DNN
design,
you
will
need
to
deal
with
this
kind
of
complex
and
survives
operations,
and
also
you
will
have
a
lot
of
problems
because
of
the
complicity
to
debug.
D
So,
in
the
end,
we
face
a
steep
learning
curve
when
we
want
to
implement
effects
dnl
prototype
so
how
it
works.
How
works
this
the
ignition
framework?
Basically,
first,
you
need
to
define
your
network
in
this
case
and
the
main
elements
involved
like,
for
example,
I,
want
to
optimize
the
routing
configuration
security
policy
and
I
have
some
target
metrics,
such
as
improving
the
network
performance
or
detecting
or
making
animal
infection.
D
The
ignition
framework
will
assist
you
by
offering
an
abra
network
abstraction
to
define
the
DNA
in
architecture,
adaptiveness,
network
ministries
and,
finally,
this
framework
automatically
generates
a
tensor
flow
code
with
the
implementation
of
your
last
neural
network
model.
So
you
are
completely
oblivious
of
this
complex
programming
in
tensor
flow,
so
to
show
the
how
is
the
workflow
for
four
users?
Basically,
a
user
can
develop
its
own
and
in
solution,
increasing
poly
steps,
but
are
the
ones
that
you
can
see
here?
First
of
all,
is
to
create
a
DNN
model
description
then
migrate.
D
With
which
is
the
most
complex
part
of
this
process-
and
we
offer
a
very
easy
Dejan
interface
to
create
to
create
this,
so
you
can
abstract
from
all
the
mathematical
elements
of
of
the
graph
neural
network
and
in
in
this
case,
we
offer
an
apple
abstraction,
where
you
can
describe
the
network
elements
that
are
involved
in
the
networking
news
case
and
the
relations
between
them,
for
example,
to
define
the
routing
configuration
in
a
network,
you
can
define
a
link.
The
link
element
on
the
path
element
and
define
relationships
between
n20
paths
are
links.
D
In
this
context,
we
provide
a
template
and
detailed
documentation
to
fill
this
this
template
and
create
the
ideal
model,
so
the
second
step
is
to
migrate.
The
data
set
rotation.
Basically
in
this
case
in
networking
these
cases,
data
sets
have
very
different
structures
depending
on
how
do
you
generate
and
they
can
gather
the
data.
So
what
we
do
is
to
provide
a
nest
under
its
own
interface.
That
is
easy
to
to
be
me,
creative
or
to
migrate
your
data
to
this
this
format
and,
finally,
once
it
is.
E
D
So
we
provide
an
advanced
debugging
tool
where
you
can
visualize
the
architecture
of
your
graphing
and
network
model
in
an
interactive
block.
So
here
you
can
see
some
examples
where
we
saw
some
famous
episodes
of
this
interactive
graph
with
different
parts
of
graph
neural
network
model
called
Rockland
and,
as
you
can
see,
it's
come
to
know
how
its
deadening
internally.
So
in
the
end,
this
framework
also
includes
some
some
properties
and
characteristics
to
identify
potential
levels
and
also
assists
users
to
current
to
correct
them.
D
The
main
advantage
is
that
you
can
make
a
fast
DNN
prototyping,
particularly
targeted
for
for
networking,
practitioners
and
researchers
that
do
not
have
any
background
on
tensorflow
and
related
languages.
So
from
our
experience
we
can
see
that
or
we
could
see
that
when
you
want
to
start
with
tensor
plural
and
these
languages,
it
don't
take
and
some
months
to
create
a
first
DNA
prototype
and
also
it's
very
difficult
to
to
debug
it.
Well
with
our
framework
with
just
few
hours,
you
can
create
a
first
gen
and
prototype
and
run
the
training
of
the
model.
D
D
Additionally,
we
provide
a
public
repository
that,
where
we
are
adding
some
state
of
the
art,
a
graphical
network
models
that
were
already
applied
to
networks
like
the
ones
that
you
can
see
here
as
an
example
I
think-
or
we
claim
that
that
progress
in
education
in
artificial
intelligence
cannot
occur
without
this
kind
of
public
repositories,
including
implementations
and
datasets.
So
this
is
public
repositories
should
serve
as
a
good
material
for
for
educational
advancement
in
engineering.
D
The
model
so
should
predict
the
min
per
packet
delay
on
its
source
destination
path
of
of
the
network.
To
this
end,
we
provide
a
data
set
that
is
simulated
in
innominate
plus
plus,
that
includes
several
topologies,
hundreds
of
configurations
of
routing
scheduling
and
traffic
matrices
and
basically
will,
in
the
valuation
part
of
this
challenge.
What
we
want
is
to
test
the
decision
capabilities
of
the
neural
network
solutions
proposed
by
the
other
participants.
D
So
what
we
do
is
that,
in
the
training
data
set,
we
include
samples
simulated
in
two
particular
Network
topologies,
and
then
we
test
these
solutions
on
sampled
simulated
enough
properly.
That
should
be
unseen
during
during
the
training
phase,
so
the
objective
is
to
test
the
capability
of
these
solutions
may
delay
delay
predictions
in
this
unseen
Network.
D
D
This
challenge
was
organized
as
part
of
broader
challenge
on
the
ideal:
I
a
amount,
in
fact
a
challenge.
So
you
can
check
more
details
at
this
link
that
you
can
see
here
where
there
are
some
other
challenges
from
from
other
companies,
universities,
so
I
would
I
would
like
to
take
the
opportunity
to
encourage
participation
and
dissemination.
D
A
D
They
they
will
provide
neural
network
models,
so
the
idea
is
that
they
can
implement
any
any
model
based
on
neural
networks
that
have
a
simple
the
topology,
the
traffic
and
the
under
configuration
of
the
network
and
the
output
should
be
this.
The
min
delay
estimated
on
each
source
destination
pair.
So
the
idea
is
that
we
will
give
at
the
end
of
the
challenge
a
test
dataset
where
that
is
unlabeled
and
they
have
to
label
this
data
set
and
send
us
the
labels.
D
A
A
D
A
More
because
first
I
mean
not
in
the
challenge
in
the
ignition
framework,
the
present
I
think
in
step
one.
Yes,
when
you
put
the
GNN
with
the
description
and
the
new
idea,
you
have
an
example,
I
think
next
slide,
because
I
think
when
it's
very
interesting,
because
you
say
that
basically,
you
can
I
mean
one
challenge.
Actually
is
that
it's
out
for
expert
people
to
to
design
I
mean
to
define
the
genome,
and
it
has
to
be
edited
which
kind
of
use
case
or
objective.
A
D
Idea
is
that
you
can
have
this
model,
so
the
the
objective
will
depend
on
the
output
values
in
your
in
your
data
set.
So
here
you
define
the
inputs,
the
outputs
on
how
how
how
is
the
GNN
implemented
inside.
So
how
is
the
internal
architecture?
So
then,
once
you
train
your
your
model,
you
need
to
include
a
data
set
and
this
model
will
we'll
learn
based
on
the
on
the
sample,
some
of
your
data
set.
D
How
can
I
know
if
the
neural
network
learned
quickly
about
an
optimized
route?
So
basically
when,
when
you
are
training
or
your
model,
this
is
this
is
part
of
tensorflow
on
other
languages.
You
will
have,
you
will
see
the
evolution
of
your
loss
function
and
the
loss
function
is
basically
a
metric
that
tells
you
how
accurate
are
the
predictions
based
on
the
in
your
in
your
data
set,
so
you
can
consider
some
thresholds
or
for
the
business
function
and
once
the
the
error
of
this
model
is,
is
quite
low
or
is
sufficiently
low.
Different.
D
One
student
that
we
disabled
to
model
or
on
the
other
samples
and
usually
one
methodology
to
achieve
this,
is
to
separate
or
to
split
in
your
data
in
training,
validation
and
test
data
sets
so
to
come.
The
accuracy
of
your
model
usually
do
train
the
data
set
in
Soma
cific
samples
and
then
evaluate
the
accuracy
in
other
samples
that
were
lancing
doing
during
the
training
phase.
D
A
A
A
A
Be
for
and
admonition
single
one
because
the
recommendation
I
think
that's
the
maximum.
You
can
have
actually
it's
a
good
time,
because
at
that
particular
time
and
in
July
and
July
July,
we
have
a
set
of
work
item.
I
mean
my
stones
particular
regarding
AI
challenge,
but
added
I
think
I
be
an
architecture,
new
document
for
the
architecture.
A
Of
course,
in
research
group,
my
stone
are
not
as
strict
as
you
know,
I'm
working
good,
but
cities
are
good,
so
I
got
to
be
there
just
to
recall
also
so
and
to
to
help
us
to
a
bit,
stimulate
and
and
the
activity
of
the
group
so
I
think
it's
it's
good
to
have
this
milestone,
July,
it's
very
short
now
regarding
the
lily,
but
it's
good
for
it
and
that
we
will
do
today.
Discussion
regarding
hackathon
and
I.
A
A
A
So
we'll
continue
with
monthly
virtual
meetings,
meaning
that
we'll
have
another
one
in
June
before
see
the
actual
IETF
meeting,
and
of
course
we
have
always
meaning
list,
and
we
have
set
number
platform
in
number
of
icons,
of
different
platform
on
github,
on
Google
Docs,
at
use,
for
example,
for
Z
a
a
challenge
document,
and
so
we'll
continue
with
this
modification
as
written
there.
And,
of
course
just
you
can
use
any
kind
of
different
tools
that
we
have
to
communicate
with
you.
A
A
We
started
with
some
challenges
in
the
document
Michel
table
of
contents,
and
we
share
this
and
of
course,
then
there
will
be
a
kind
of
call
for
contribution
to
this
document.
I
would
like
to
thank
all
the
contributors
which
you
continue
to
because
a
lot
of
people
reply
positively
but
did
not
have
time
to
to
contribute
to
the
document,
but
at
least
it
looks
like
we
are
doing
in
good
shape.
Regarding
that
we
have
a
small
team
that
can
support
the
document
and
with
his
input.
A
A
Please
do
feel
free
to
interrupt
me
and
tell
me
you
know
feedback
so
looking
at
the
date
on
the
document,
so
as
I
said
so,
there
was
some
discussion
and
some
proposition
regarding
the
let's
say,
introduction
of
the
document
because
of
course,
before
going
directly
into
describing
challenges,
we
wanted
to
have
some
kind
of
introduction.
So
initially
it
was
introduction,
section
I
would
say,
was
for
Hughes
and
objective
of
artificial
intelligence
for
network
management
and.
A
A
So
it's
it's
I.
Think
I
personally,
think
it's
it's
very
relatable
it
is
it
just
the
way
you
present
so
rather
than
seeing
that
different
objective
to
solve
this,
and
this
and
this
problem
it
might
be
more
victims,
even
if
we
start
describing
the
very
out
problem
now
we
can
in
order
to
me,
so
there
are
some
problems
actually
that
have
been
the
Comanche
Iranian
at
least
listed
president.
All
everything
relating
to
resource
education
with
NPR
problems,
realistic
that
are
not
optimal
and
so
on.
A
Yes,
can
he
I
can
do
waiter
and
so
on,
and
although
it
was
I
liked,
it
I
think
and
the
artists
point
of
scalability
accuracy
of
or
techniques,
but
also
the
problem
of
the
bottleneck,
human
bottleneck.
We
hadn't
made
many
operation
network
management
operations.
We
need
human,
we
might
human
operator
to
steal
a
lot
of
human
operation.
To
do
so.
A
This
is
e
means
is
was
so
this
new
section
was
proposed
before
us.
Then
we
go
into
the
goals,
so
they
said
before.
We
need
to
find
the
right,
let's
say
articulation
between
the
two
section,
alright
link.
Of
course,
there
is
clearly
a
link.
It's
not
yet
fully
defined,
we
need
to
or
one
section,
but
anyway
we
need
to
to
motivate
maybe
the
first
two
and
then
go
into
the
second
one
with
the
goals.
A
Here
we
have
a
lot
of
fun,
that's
a
common
description,
or
we
should
organize
because
of
course,
we
should
not
put
for
each
of
the
chain
one
one,
one
in
one
paragraph
and
just
each
after
each,
but
t-shirts
are
organized.
As
a
section.
We
classify
the
different
abilities
kind
of
yeah
structure
on
the
objective
on
the
beginning.
We
just
we
can
maybe
go
through
by
by
later,
but
of
course,
it's
clear
that
ourselves
as
a
layer,
so
maybe
in
something
pralaya
specification
of
the
objective,
but
still
both
objective
trans
are
two
different
later.
A
A
A
B
It's
nice
to
see
that
there
are
I,
mean
people
willing
to
contribute
and
already
several
comment,
that
exchange
on
the
document
relating
to
to
what
you
asked
about
this
presentation
or
organization.
I
think
it
will
be
okay
for
the
moment
to
to
work
with
based
on
them.
What
we'll
bring
us
inputs
and
I
really
think
about
the
best
way
to
present
that
maybe
I
see
it
too
entails
being
a
draft.
B
We
have
ideas
about
sections,
but
we
should
remain
a
bit
flexible
and
even
if
it's
not
completely
stable
now,
if
the
goal
or
if
it's
a
motivation
or
problems
I,
think
it's
a
bit
fluid,
and
we
should
we
keep
it
like
this
as
we
mature
our
thoughts
about
about
the
document,
I
would
not
see
as
as
a
big
issue
concerning
that
we
are
still
at
the
beginning
of
this
document.
Maybe
when
it's
a
bit
more
stable,
we
can
rethink
a
bit
model
if
the
initial
structure
of
the
first
section
needs
to
be
reworked.
A
Okay,
thank
you.
Oh
yeah,
I,
agree,
I,
agree
and
I
think
maybe
it
was
kind
of
a
small
e
word,
because
I
think
in
the
initial
table
of
content
we
actually
put
kind
of
parallel
your
objective.
But
of
course,
at
this
stage
we
are
in
the
early
draft
and
we
can
easily
then
consolidate
the
different
things
into
something
burst
with
more
structure.
A
Rep,
for
which
we
have
contributions
or
comments,
so
one
is
a
regarding
like
weight
at
st.
ingots.
If
you
want
to
come
back
into
small
devices
like,
of
course,
speeches
and
so
on,
this
is
something
that
actually
has
already
been
proposed
and
I
mean
I
were
inscribed
in
I
think,
one
month
ago
that
we
have
this
not
so
much
update
regarding
this.
A
Okay,
we
have
some
some
comments.
We
having
another
challenge
was
also
and
I
think
it's
it's
a
bit
lengthy
technical
presentation,
if
I
think,
when,
when
big
issue
or
challenges
that
we
are
people
that
are
not
even
necessary
expert
in
artificial
intelligence
and
we
need
about
to
be
guided
to
I
mean
to
guide
them
to
use
at
shine,
teach
and
see
if
they
want
to
use
parent
misfortune
operation
right.
A
So
we
all
totally
identified
this
problem
and
it
was
more
on
if
you
want
to
use
after
sharing
teacher,
what
should
be
the
algorithm
to
use
for
your
problem
and,
of
course,
my
vision
here
is
the
two-sentence
description
is
change
little
bit
very
too
strong
and
I
think
the
command
we
receive.
Is
that
probably
finding
the
right
algorithm
is?
Are
we
too
strict?
A
It
really
depends
on
the
number
of
parameters,
because
you
can
say
that
if
you
want
to
I,
don't
know
to
detect
I'd
like
to
predict
the
next,
the
next
traffic
falutin
on
your
link,
see
it's
not
so
easy
that
to
say
that
you
should
use
this
or
this
algorithm
right.
So
maybe
we
should
come
up
here
is
a
formulation
of
the
problem.
A
Oh
to
select
are
all
to
find
the
right
trade-off
between
the
accuracy,
one
plus
the
constraint,
and
so
on
so
I
mean
it
kind
of
methodology
for
helping
somebody
in
not
very
expert
in
advantage
and
set
right
the
the
the
problem
and
help
them
to
be
fine.
Is
you
know
what
what
should
be
then?
Maybe
is
the
best
fit
for
its
own
problem.
E
A
What
a
leveling
so
somebody
that
has
been
it
has
been
written
and
that
has
been
described
here-
is
that
for
many
cases
I
think
we
know
we
always
we
already
had
the
problem
of
details
that
he
be
put
in
the
challenges,
but
more
than
just
eating
the
data
most
of
the
time.
It's
also
a
drive
data
with
letters
that
can
be
used
for
supervised
learning,
so
we're
here.
The
challenge
is
more
about.
Okay,
so
probably
today,
I
think
it's
it's
not
so
so
so
difficult.
We
didn't
think.
A
If
you
are,
if
you
appear
to
own
it
works,
you
can
easily
extract
data
from
the
different
probes
and
so
on,
but
of
presen.
If
you
want
to
do
supervised
learning,
it
could
mean
you
take
them
today
on
mobile
and
more
focusing
on
supervised
learning,
then
you
need
to
have
some
levels
of
unity.
Big
on
the
data
is
not
enough
and
nice,
nothing
in
description,
a
lot.
Of
course.
A
You
can
say
that
when
data
from
one
side
and
then
you
have
problem
not
from
the
other
side
in
the
same
date,
outpost
can
be
used
for
different
woman,
but
should
not
be
at
the
same
level
or
same
level
can
be
used
for
different
problems.
Also,
there
is
not
one
to
one
mapping
between
the
labels,
including
data
and
the
problem,
so
we
can,
it
should
be
adapted.
So
it's
more
exactly
what
I
could
feel
just
fill
the
gap.
A
You
assume
that
you
have
telemetry
or
monkey
20
needs
that
allows
you
to
add
data,
ok,
that
we
can
assume
that
you
learned,
and
then
we
have
a
problem
is
that
we
may
need
to
feed
the
data
to
this
problem
in
particular.
Regarding,
of
course,
our
other
think
was
there
issue
really
the
data
more
related
to
the
I
would
say
a
I
think
it
said,
like
you
know,
pre-processing
normalization
and
so
on,
but
I
think
from
a
cognitive
perspective,
I
mean
you
need
the
label.
If
you
want
I
mean
regarding
subproblem,
you
monitor
Edwards.
A
Why?
The
question
is
that
usually
algorithms
that
can
work
nicely
with
some
offline
data
set
or
on
testbed,
but
it's
very
hard
to
guarantee
that
they
will
continue
to
be
enough.
Accurate
and,
let's
say
under
long-term
perspective,
like
it's
important
to
know.
If
you
put
an
algorithm
in
place
to
variable
to
automate
some
some
operation,
your
network,
we
should
know
how
long
you
need
your
solution
or
can
be
applied
without,
for
example,
we
are
playing
learning,
and
so
it's
important
to
know
at
certain
points
where.
C
So
I
listening
to
your
presentation
and
also
I,
read
the
document
and
I
I'm
one
of
I'm
one
of
the
ones.
That
said
that
was
going
to
contribute
and
I
didn't,
did
it
yet,
but
I
will
and
I
was
thinking
that
also
regarding
Jose's
presentation,
the
document
has
to
some
extent
an
unwritten
underline
assumption
that
AI
will
be
applied
to
a
network
and
training,
for
that
I
will
happen
on
that
network.
C
A
C
C
A
B
To
follow
up
on
this
comment,
I
mean
this
is
the
perfectly
fine
situation.
This
is
something
we
we
are
doing.
I
mean
I'm
speaking
on
behalf
of
ok.
Here
we
have
developed
of
model
on
our
I
will
say:
internal
infrastructures,
and
when
we
deliver
models
to
to
customers,
I
mean
this
is
kind
of
based
models
and
they
need
to
be
I
mean
fine.
You
know
train
to
the
specific
operational
environment
where
they
will
be
used.
So
this
is
a
fully
valid
model.
What
you
were
mentioning,
but.
B
B
We
are
relationship
with
customers
to
get
some
some
train.
That
does
it,
but
once
the
operator
is
deploying
the
specific
solution,
there
is
a
face
of
specializing.
So
we
see
when
you
instantiate
you,
you
will
have
some
level
of
adaptation
of
the
model
to
the
actual
deployment,
and
this
could
be
I
mean
you
can
have.
B
Let's
take
whether
you
have
some
algorithm
that
are
running
for
four
base
stations,
you
can
have
a
kind
of
master
model,
that
is,
that
reside
in
some
some
places
and
when
you
deploy
it
over
100
of
base
stations,
each
of
the
instances
will
be
specific.
They
will
be
closed
because
they
rely
on
the
same
master
model,
but
each
of
the
model
will
be
specific.
B
C
Cool
right,
because
what
you're
telling
me
is
that
it's
not
a
binary
agent
between
on
the
mentors
up
or
you
train
on
the
operators
Network,
but
rather
the
Rize
is
a
exam.
There
are
many
degrees
right.
You
can
fully
trained
on
the
vendor
lab
and
then
don't
do
anything
on
the
operator
just
inference
or
there
are
degrees
in
between.
So
maybe
we
could
reflect
that
I!
Never
thought
about
these
degrees
in
between.
So
you
could
also
reflect
that
on
the
log.
F
C
Just
because
we
are
getting
enjoyed,
everything
is
the
academic
discussion
with
GNN.
You
don't
need
to
do.
Transfer
the
model
should
work
on
on
an
unseen
s.
Energy
without
transferred
luck.
Then
I
agree
with
you.
Then
the
real
world
is
super
complex
and
so
on
and
I
understand
that
in
many
instances
you
will
need
to
adapt,
modify
or
your
own
stuff
to
adapt
it
to
the
real
work
scenario.
But
a
Giants
are
not
transferred.
Not
it's
just
a
model
that
work
but
I
read
with
thruster
learning
you
can
achieve
similar
move.
Oh.
A
C
B
As
you
say,
if
the
problem
fits
with
a
graph
representation,
because
I
don't
want
to
take
too
much
time
for
this
discussion,
but
yeah
I
think
it's
an
interesting
topic
we
be
brought
in
in
the
discussion
for
the
document.
I
just
reflect
that.
We
have
also
investigated
to
combine,
in
fact
a
deep
reinforcement,
learning,
in
fact,
even
yeah,
based
on
neural
network
plus
reinforcement,
training,
plus
transfer
learning
for
different
usage.
B
It
can
be
first
two,
depending
if
you
have
available
data
or
not,
so
you
may
use
in
your
training
brain
model
and
apply
to
the
target
to
the
target
domain,
and
you
can
use
it
also
to
as
an
assistance
to
the
operator
if
you
want
to
be
able
to
plan
or
to
deploy
similar
models
into
the
and
regions
or
different
part
of
your
network.
But
they
have
a
good
level
of
similarity
you
may
have
also
in
on
how
much
gain
you
can
expect
from
from
the
transfer.
B
Learning
of
reusing
existing
model
versus
I
mean
training,
a
model
from
scratch
and
the
cost
associated
with
training
plan,
training,
training,
resources
etc.
So
I
was
I
was
asking
if
it's
very
specific,
to
GNN
the
convener.
In
our
investigation
we
will
show
that,
for
the
professor
different
transport
learning,
plus
by
transfer
learning,
you
can
also
get
very
good
games
on
such
application.
C
Missing
something
I'm
not
saying
anything,
but
our
transformer
I
think
this
ee
technique
I'm
just
saying
that
with
a
GNN
you
don't
need
transfer
learning,
because
it
is
already
understanding
what
a
computer
network
is.
So
if
it
is
a
different
computer
network,
even
if
it
has
not
been
trained
for
other
specific
network,
it
will
be
able
to
understand
it.
I
think
that
those
are
different
techniques
that
apply
to
different
things,
which
is
absolutely
true,
is
that
general
do
not
make
any
sense.
C
A
But
also,
maybe
we
have
to
see
if
it's
this
item
might
be
IBN
specific
man,
it's
more
my
personal
opinion
because-
and
I
read
this
document-
sorry
IBN
specific,
but
so
the
first
one
is
more
related
to
an
interpreting
eye
level
or
on
that
you
language,
intense.
So
it's
it's
more
likely
to
use
NLP
techniques
or
named
on
two
occasions,
I
think
from
NLP
to
twelve
a
better
interaction
with
human.
So
it's
pretty
tall,
so
you
might
in
on
the
loop
and
challenge
so
I.
A
Think
one
question
already
is:
of
course
you
can
interpret
intent,
but
can
you
also
generate
intent
from
current
operation?
Or
can
you
use
these
twelve
diagnosis
or,
in
the
other
ways
that
others
that
the
user
is
providing
in
Tennessee
I
mean
the
inter
top
kind
of
intent?
I
will
say
our
human
and
say
high
level
of
human,
readable
information
operated
to
the
user
based
on
diagnosis
of
the
networks.
A
It
can
be
I,
think
it
wasn't,
the
command
can
be
seen
the
two
two
directions
and
the
other
one
I
think
it's
it's
quite
it's
it's
very
nice,
but
it
can
be
more
generalized,
Omni
and
IBM.
Basically,
so
always
all
you
define
your
action
plan
in
turn,
so
much
Ruby's
action
to
produce
with
it.
Isn't
that
like
design
outcome,
that
is
expressed
in
the
intent
so
as
well
here
for
sure
I
think
it's
very
kind
of
living.
A
Yes,
looking
for
something
hundred
percent,
automated
is
not
possible,
but
it
is
kind
of
the
goal
that
you
would
like
that.
Well,
some
things
that
is
completely
totally
continuous
or
whatever,
but
it
should
be.
It's
a
of
course.
It's.
It's
really
part
of
the
challenging
challenge
of
I
for
network
management.
We
can
mostly
you
just
provide
some
inputs.
What
you
want
and
then
could
use
all
the
action
needed
to
twenty
shipped,
but
yeah.
A
Don't
know
we
can
point
you
to
progress,
so
I
think
it's
good
that
they
said
before
that
you
know
ever
it's
a
list
of
identified
interested
computers,
so
I
think
we
continue
with.
This.
Wolf
is
more
immediately
discussion
with
specific
emails
meetings
to
continue
to
conciliate
the
document.
Of
course
everyone
is
welcome
to
participate.
Maybe
I
will
put
a
bit
less.
The
mailing
list
are
asking
for
contribution,
so
I
will
continue
to.
A
We
will
continue
to
present
the
status
of
the
document
during
meetings
and
everybody
can
contribute,
but,
as
I
said,
maybe
we'll
have
more
dedicated
meetings
and
email
mailing
list
between
the
contributors,
but
yeah
you
want
to
be,
of
course,
so
basically
continue
on
the
same
with
the
kind
of
same
process.
So
with
the
document
we
have
sub
command.
So
we
have
some
discussions
in
the
document
itself,
so
it's
kind
of
its
kind
of
a
document.
A
Let's
come
and
play
discussion,
so
it's
very
nice
because
it's
very
interactive
and
then
there
will
be
the
quick
under
format.
I
think
to
know
we
can
continue
in
shell
document
may
be
some
things
that
we
say
last
time
is
that
we
may
think,
and
what
should
be
there
is
a
good
would
come
at
the
of
the
document,
and
can
you
white
paper
or
can
be
developed,
and
it's
not
probably
I,
think
we
discussed
with
some
some
of
you
and
what?
A
A
A
B
B
So
maybe
there
should
be
some
announcement
about
okay.
The
next
meeting
will
be
all
the
meeting
will
be
every
two
weeks
on
this
slot
and
then
maybe
provide
a
quick
summary
of
the
main
technical
point
that
were
discussed
so
that
people,
if
they
feel
an
interest
to
to
get
involved,
they
can
at
least
get
the
eye
level
summary
to
to
know
what's
happening.
Yeah.
A
A
B
So,
as
discussed
already
a
couple
of
times,
I
think
you're
all
well
aware:
we
have
a
work
plan,
defining
enemy
G
concerning
activities
for
internet
networking,
a
couple
of
those
work
items
already
pretty
well
covered
and
there
will
be
a
point
on
the
internet
classification
activity
at
the
end
of
this
meeting.
But
I
would
like
to
highlight
three
topics
that
also
have
been
shared
on
a
mailing
list
before
this
meeting,
and
we
would
like
to
get
your
opinions
and
see
what
we
could
do
to
move
it
forward.
B
B
So
we
shared
on
the
mailing
list
a
set
of
points
to
try
to
steer
a
bit
the
discussion
on
this.
We
had
already
in
the
past
several
documents
and
that
are
now
expired,
but
also
some
technical
presentations
from
research
group,
participants
and
I
think
we
would
like
to
be
able
to
trigger
activity
on
that
and
and
really
make
progress.
B
So
on
the
consideration,
the
idea
is
that
we
should
concentrate
on
a
clearly
identifying
of
architecture,
documents
or
documents,
the
problem
statement
and
design
goal,
challenges
and
principle
and
requirements,
so
essentially
the
last
part
of
the
sentence
or
the
principle
and
requirements.
What
what
important
to
consider
for
the
scope
of
this
architectural
document.
E
B
About
the
integra
ability
of
our
proposal
with
order,
automation,
framework,
automation,
aspects
that
are
being
developed
in
other
groups
in
ITF,
but
also
other
other
sanitization
activities,
I,
was
in
a
cage
to
try
to
investigate
if
using
principles
from
service
based
designs
of
his
based
architecture
could
be
beneficial
in
making
this
aspect
of
being
flexible
and
extensible.
But
this
is
more
a
specific
consideration.
B
I
think
a
question
that
was
raised
several
times
that
very
important
is
clearly
defined,
agree
on
the
scope
and
system
boundaries
of
what
we
consider
to
be
the
scope
of
this,
where
this
architecture
should
apply.
What
is
on
top
of
this
architecture?
What
do
we
plug
underneath
and
what
is
below
this
architecture?
B
Actually,
what?
What
are
we
controlling
or
interfacing
with
with
this
in
turn
based
system?
So
I
think
this
is
essentially
questions.
We
need
to
clarify
at
the
level
of
the
research
group
in
order
to
make
sure
that
we
output
the
right
types
of
document
for
this
work
item.
Also,
what
are
the
typical
user
types
that
should
be
considered?
There
is
a
draft
on
intern
classification
that
address
those
questions
of
types
of
intents
and
types
of
users
also
technologies
involved.
B
D
B
B
The
the
next
point
is
a
Ibn
coexistence
with
legacy
management
system
and
emerging
management,
automation,
technologies
and
frameworks.
So
this
is
related
what
I
mentioned
before
with
this
integrability,
so
either
greenfield
completely
infinite
deployment
and
integration
with
new
new
approaches
formation.
Also,
if
we
have
existing
deployed
system
or
legacy
systems,
all
the
architecture
and
the
functionality
we
have
to
provide
can
also
interface
and
operate,
which
those
those
existing
systems
so.
B
B
They
may
not
be
fully
imaginary,
so
the
question
of
interoperability,
interoperation,
first
and
interoperability
between
those
different
deployments
may
be
a
questions
to
gate
and
document.
As
part
of
this
initiative,
the
work
item,
then
the
energy
proposition
and
what
is
our
I
will
say
clear
positioning.
They
have
been
already
other
attempts
in
other
groups
and
your
in
open
source,
also
about
proposition
for
reference
models,
architectures
and
different
forms
of
frameworks
for
intent-based
networking.
B
So
what
is
our
clear,
clear,
differentiator
here,
clear
value
that
we
want
and,
of
course,
I
see
connection
with
related
graph
and
presentation?
This
is
to
say
that
this
is
not
beginning
from
scratch.
We
we
have
Indian
energy,
at
least
a
couple
of
different
documents
and
presentation
that
we
can
use
as
a
basis.
The
point
here
is
to
get
a
bit
your
opinion,
what
which,
how
we
can
try
to
bring
this
work
forward
in
the
research
group.
G
Laurent,
this
is
risk
a
maybe
there's
a
possibility
of
driving
this
through
the
use
cases,
maybe
that
that
could
be
because
many
consider
it
an
architectural
in
abstract
terms,
probably
a
little
more
complex
to
to
understand
what
the
specific
necessities
to
be
addressed.
Probably
going
through
different
use
cases
could
help
to
understand
what
will
be
there.
It
means
the
kind
of
information
that
is
required
the
kind
of
procedures
and
they
can
even
of
elements
to
check
and
consider
as
well.
C
B
E
Of
course
you
have
to
start
from
the
problem
statement,
requirements
and
use
cases,
but
the
question
is
kind
of:
do
we
go
with
some
kind
of
generic
use
cases
or
specific
use
cases
in
terms
of
that,
then
you
say
here:
should
we
consider
use
case
specific
Ibn
architecture
or
try
to
have
more
generic
common
architecture,
and
you
say
use
case:
do
you
mean,
like
you
know,
in
terms
of
provisioning
closed-loop
things
like
that
to
do
mean
different
solutions
like,
for
example,
5g
or
or
or
optical,
or
something?
But
what?
E
B
I,
don't
specific
preference
here,
especially
because
I
was
thinking
about
some
use
cases,
and
this
is
that
the
next
point
in
the
discussion
I
mean
to
suggest
potential
use
cases
that
are
different
from
each
other,
so
that
we
can
learn
different
things
and
I.
Don't
if
I,
potentially
commonality
I'm,
not
sure
that
we've
qualify
as
a
very
specific
location
but
more
in
Eric
ones.
B
But,
as
I
said,
I
don't
have
a
clear
preference.
What
should
be
the
process
to
try
to
derive
this
architecture,
but
I
would
like
to
see
happen.
Is
that
manage
to
deliver
a
document
or
a
set
of
documents?
That
is
valuable
in
the
sense
that
it's
reused
either
in
the
research
community
as
being
or
seen
as
a
reference.
That
means
that
we
are
quite
complete
and
and
useful
in
what
we
relevant
in
what
we
have
identified
as
being
the
key
component
and
interaction
of
an
architecture
and
also
could
be
reuse.
B
I
will
say
wider
in
the
industry
as
some
guidelines
to
understand
what
what
is,
because
I,
don't
expect
that
this
architecture
will
be
only
a
one-shot
that
they
will
be
I
would
say
derivation
of
those
of
those
of
those
potential
architecture,
so
that
people
really
the
blob
solution,
did
make
implementation
out
of
that
and
can
use
it
as
guidelines
to
understand.
Okay,
we
have
those
key
functionalities
to
develop
and
they
have
been
some
guiding
principles.
B
Oh
man,
a
key
key
aspect:
I
saw
from
data
structure
or
from
the
type
of
interactions
and
the
type
of
mechanism
that
should
be
thought
about
when
we
design
the
system-
and
this
should
be
a
kind
of
help
to
those
implementation
and
and
also
with
the
goal
that,
ultimately,
even
if
we
are
done
doing
standards
in
the
research
group
and
that
this
is
a
stepping
stone
to
to
add
towards
interoperable
in
base
system.
If
you
have
multiple
vendors,
multiple
solution
providers,
the
bloating
there
are
solutions.
G
B
Can
still
also
aim
to
have
some
level
of
interoperability
between
different
functionality
or
between
different
systems,
if
we
can
contribute
to
that.
For
me,
this
will
be
a
very
good
goal
of
this
architecture
dispersion.
This
is
why
I
was
also
considering
very
important
question
about
the
scope
and
what
is
our?
What
is
the
our
value
proposition?
B
Because
if
we,
if
we
draft
a
complete
architecture,
then
we
will
just
compared
to
what
the
orders
are
done
and
then
it's
just
a
matter
of
maybe
they
will
add
more
momentum,
adoption
we
don't
know.
But
if
we
manage
to
say
this
is
something
that
was
missing.
This
is
very
addressing
a
gap
or
need
I
think
that
we
contribute
to
something
of
value,
because.
E
Compared
to
thanks
for
that,
but
compared
to
the
you
know
what
we
did
before
in
the
standards
in
terms
of
the
use
cases
and
different
architectures,
you
know.
The
main
difference
here
is
that
we
are
bringing
all
of
those
two
cases
together
to
support
the
intent
based
approach
so,
for
example,
in
other
standards
like
TMF
m
EF
or
whatever
they
could
kind
of
pick
the
use
case
and
say
and
prioritize
it
and
say:
ok,
I'm
looking
at
provisioning
now,
but
I
will
have
it.
Monitoring
would
be
late
or
assurance
would
be
late
for
us.
E
What
we
are
bringing
it
in
10,
driven
or
intern
based
approach
is
putting
that
all
this
together
and
supporting
in
10
driven
life
cycle.
So
from
that
perspective,
Pacific
use
cases.
If
you
are
saying
oh
we'll
support
this
so
this,
so
this
may
not
be
possible.
We
have
to
kind
of
bring
them
all
together
to
support
intent,
driven
approach
on
top
of
them.
So.
F
H
So
another
Mario
speaking
another
point
to
take
into
consideration
would
be
that
if
we
follow
specific
use
cases
right
and
in
architecture
that
this
has
a
specific
scope
right,
we
may
get
into
the
trap
of
doing
something
quite
similar
to
Polish,
based
management,
that
the
solutions
were
quite
customized
to
specific
sort
of
use
cases
all
right.
So
I'm
also
Pro
a
generic
sort
of
architecture.
H
E
B
E
B
So
for
me,
when
we
say
general
or
generic
I,
think
it
will
be
beneficial
for
us
in
the
energy
to
say.
Is
this
in
scope
of
what
we
consider
to
be
applicability
of
the
framework?
Even
if
we
meet
will
it
will
be
applicable
afterward,
but
at
least
at
the
beginning
of
our
design
process,
to
say
we
consider
this
to
be
in
scope,
because
this
is
where
we
will
exercise
the
design,
and
this
is
out
of
scope.
B
The
same
can
be
I
mean
horizontally
in
terms
of
our
type
of
technology
or
a
type
of
use
cases.
It
can
be
also
vertical.
I
will
say:
where
do
we
start
and
when
do
we
stop
I?
Think
in
the
plantation
from
marinos
the
southbound
boundary
was
on
as
soon
as
the
management
intent
based
management
interfacing
towards,
for
instance,
Sdn
controllers
manual
cetera.
So
those
were
considered
to
be
out
of
scope
of
the
incan
based
system
and
we
were
interfacing
with
those.
B
They
are
in
turn
based
system
that
considered
purely
Sdn,
so
I
think
we
should
really
clarify
the
positioning
of
this
architectural
work
in
terms
of
scope,
because
otherwise
we
may
not
share
the
same
understanding
of
what
the
scope
is.
So
it
will
be
difficult
to
agree
on
the
on
the
design
and
then
for
the
applicability
and,
as
alec
was
mentioning
trying
to
exercise
the
the
deployment
scenario.
E
C
B
B
Another
thing
is
to
use
news
cases
a
bit
like
Alex
was
mentioning
that
let's
say
we
have
an
architecture
and
you
want
to
and
try
to
exercise
tech
chure
is
speaking
to
some
particular
use
cases,
and
then
we
see
the
adaptation
process.
Can
you
recap
picture
to
a
deployed
architecture?
It
can
be
the
same
set
of
use
cases
or
the
two
part
of
the
process.
They
could
be
a
kind
of
fully
bottom-up
approach
to
say
we
just
take
a
ground
set
of
abuse
cases
and
we
identify
commonalities
to
derive
the
common
functional
blocks
and
interactions.
B
Also,
for
the
sake
of
time,
what
you
got
I
think
we
have
also
our
minute
and
recording
it
that
will
help
us,
but
I
heard
several
inputs.
Thank
you
for
that.
We
try
to
summarize
that,
after
after
the
meeting
and
come
back
to
the
research
group
is
some
sort
of
proposal
to
move
this
forward.
Of
course
everything
will
be
contribution
driven,
so
this
will
just
be
guidelines
at
the
level
of
the
research
group.
Try
to
see
how
we
want
to
address
this.
B
This
work
item
on
IBM,
but
I
think
we
already
have
some
indication
of
what
people
will
be
willing
to
do
and
how
to
do
that
progress.
This
work,
let's
move
on
to
the
next
slide
and
I
think
we
we
partially
touch
that,
even
if
here
the
discussion
will
be
slightly
different
too.
So
the
second
point
is
on
the
IBN
use
cases.
We
have
also
a
specific
port
item
dedicated
to
that
essentially
towards
validation
scenario
and
use
cases
for
intent,
expression
and
function
then,
to
assess
the
quality
and
completeness
of
beyond
of
the
specification.
B
So
in
the
considerations
that
we
have
shared
partially
in
the
mailing
list-
and
we
some
of
the
people
involved
in
pickle
use
cases-
and
this
is
a
proposition
that
that
we
would
like
to
make
and
potential
use
cases
of
interest.
This
is
you
can
open
list
and
we
have
to
see.
We
will
be
really
at
least
the
way
I
see
that
is,
you
can
have.
B
You
know
typical
use
cases
that
will
be
in
different
setting,
but
actually
a
5g
or
slash
network
slicing
type
of
use
case,
because
this
will
represent
certain
set
of
actors
and
the
vertical
industry,
vertical
actors,
the
telecommunication
operator
or
multiple
operators
and
typical
types
of
infrastructure.
Ism.
Support
of
that
the
second
category
could
be
around
enterprise
or
data
center
networking,
because
this
is
could
involve
different
type
of
actors.
B
The
type
of
infrastructures
and
problems
that
are
faced
from
the
operation
perspective
could
be
quite
different
of
commentary,
at
least,
but
could
be
in
the
first
category
of
use
cases
and
finally,
the
last
family
or
last
people
type
of
use.
Cases
would
be
more
to
what
smart
eggs,
so
this
could
be
for
quick
early
digit,
but
on
smart
city
that
could
involve
again
another
type
of
factors.
B
If
you
take,
for
instance,
a
type
of
environment,
may
we
have
the
user
of
the
system
that
are
nope,
no
qualification,
no
technical
qualification
on
the
operation
of
the
network,
so
the
type
of
intent
and
the
type
of
operation
is
expecting
on
the
network
and
the
automation
level
to
be
quite
different
from
if
you
have,
for
instance,
specialists
that
are
from
in
the
connection
operator
that
are
here
to
operate,
and
the
type
of
intense
could
be
quite
different.
B
This
will
be
interesting
to
see
if
we
come
back
to
a
previous
discussion,
how
well
this
generic
architecture
may
adapt
and
fit
the
applicable
to
this
different
Emilia
of
use
cases,
but,
of
course,
I
think
wrong.
Truth
about
the
dis
use
cases
that
we
need
to
have
participants
that
are
willing
and
capable
of
raising
those
two
sketching
excuse
cases
as
part
of
the
work
of
the
researcher.
B
That
said,
I
think
we
also
have
more
recently
a
concrete
proposal
coming
from
Luis
and
finalities
on
the
wrong
spot
slice
intact,
which
is
the
targeted
to
be
a
specific
application.
I
think
this
could
be
also
a
basis
of
of
the
discussion
on
the
use
cases
in
the
group
and
see
if
there
are
other
participants
willing
to
join
them,
or
there
are
other
teams
of
defense
that
are
willing
to
develop
additional
use
cases
and
so
to
have
more
than
one
use
case
we
can
play
with
and
learn
from
in
the
research
group.
B
The
second
bullet
point
in
the
consideration
is
again
a
proposition
from
the
chairs
on
the
for
the
research
group
to
potentially
define
guidelines
and
common
structure
of
an
expectation
and
even
eventually
a
use
case
template
in
order
to
help
support
the
development
of
this
activity.
We
have
seen,
for
instance,
experiencing
or
the
group
that,
if
they
are
several
use
cases
that
are
proposed
it
a
share,
at
least
a
common
structure.
This
could
be
useful
for
the
comparison,
but
also
to
to
stimulate
and
help
in
the
reduction
of
in
the
writing
of
will
use
cases.
B
So
this
is
something
we
could
provide
for
the
research
group.
Babies
are
interesting
to
that
and
I
think
there
is
an
important
connection
to
be
made
with
the
document
the
work
on
intern
classification,
because
in
this
document
there
is
underneath
the
classification,
but
on
AB
approach
to
have
those
different
technologies,
user
types
in
ten
types
that
is
think:
okay,
obvious
relationship
with
the
work
on
you
escape.
B
B
B
B
I
B
B
B
B
B
B
Overdue,
so
the
last
point
of
this
Ibn
discussion
is
on
implementation
and
the
VT
run
a
control.
So
we
have
another
work
item
in
the
idæan
work
plan
which
dedicated
to
implementation
and
proof
of
concept
we
had
last
October,
they
keep
meeting
and
first
very
nice
interactions,
concept,
demonstrations
and
some
outcome
we
wanted
to
to
deliver
further,
and
so
the
idea
is
is
to
continue
on
those
on
the
spread
of
activity,
and
we
have
again
shared
on
the
mailing
list
and
asking
now
for
you
about
a
set
of
considerations.
B
What
are
the
expectation,
the
objectives
from
the
activity
around
implementation
on,
in
ten
days,
networking
and
especially
kind
of
akhaten
activity,
whether
we
should
also
combine
with
first
work
on
documents,
so
typically,
for
instance,
the
youth
just
talked
about,
and
the
architecture
and
at
which
stage
we
need
to
really
consider
I
will
say,
put
faces
on
the
on
the
implementation
and
development
of
prototypes
of
Ibn.
And
how
can
this
be
organized
and
develop
Putra
like
a
times
virtual
teams,
with
rickety
platforms,
to
support
these
using
the
gig
Ericsson?
B
Also,
whatever
we
can
make
available
from
available
research
group,
we
will
try
to
route
support
and
also
who
is
willing
to
take
part
in
this
activity,
and
there
could
be
different
types
of
participation,
actually
develop.
The
tools
and
the
demos,
but
also
you
may
also
want
to
provide
some
development,
environment
or
infrastructure
in
support
of
the
simulation
at
the
trial.
So
your
participation
is
not
only
to
to
develop
code.
There
could
be
also
some
other
form
of
support,
also
to
help
us.
You
know
running
a
bit
how
we
see
this
activity
moving
forward.
B
We
think
we
should
at
least
try
to
share
this
common
understanding
of
what
will
actually
with
this
plantation
aspect
of
IBM
and
see
if
this,
if
we
can
move
forward
with
this
one
and
also
typically,
what
kind
of
outcomes
you
expect
from
the
short
term
to
long
term
from
very
individual
I
will
say:
simple
type
of
put
patent
demos
who
more
integrated
and
also
more
our
second
Prince.
If
type
of
reference
tools
and
frameworks,
oh
defining.
B
G
So
the
run
just
to
be
so,
they
implemented
the
use
cases
that
we
could
use
cases,
architecture
that
we
could
work
out
the
previous
work
items,
or
do
you
foresee
something
that
could
be
running
in
parallel,
not
necessarily
linked
with
the
with
the
other
work
items
just
to
understand?
Well,
what
is
your
birth
date
on
this?
So.
B
Here,
it's
just
my
fault,
I
mean
necessarily
as
a
research
chair
is
just
that
yeah
I
can
see
a
lot
of
flexibility
here.
What
we
had
in
October
was
really
we
had.
We
made
a
call
for
demo
as
part
of
the
conference
dedicated
to
mmog
ibn
topic,
and
we
got
four
four
four
demos
not
correlated
to
it
solid
and
we
already
had.
B
It
was
already
a
very
good
basis
of
discussion
and
further
as
a
lot
of
contention
of
attack
in
that
sense,
for
me
will
be
already
quite
nice,
but
we
also
identify
potentially
that
there
could
be
connections
between
part
of
the
tools.
For
instance,
one
was
providing
a
part
of
the
functionality
to
define
intent,
and
maybe
some
do
some
verification
on
intent.
Rob
is
an
interactive,
intuitive
approach,
some
other
demos
were
really
more
on
the
I
will
say,
and
they
will
intent
into
specific
innately
by
an
MVP
by
systems
or
Sdn
based
systems.
B
Some
other
were
related
to
oh
we're,
having
specific
functionality
that
were
not
seen
in
order
like
a
calendaring
or
scheduling
tab
functionality,
so
that
could
be
depending
this
is
really
dependent.
Also
on
what
the
teams
are
we
willing
to
do
and
how
much
effort
they
would
and
Internet
they
will
see
in
collaborating
together,
but
I,
think
having
individual
demos,
not
necessarily
to
the
architecture
or
to
the
use.
Cases
is
possible
if
they
are
linked
to
the
use
cases
and
the
architecture.
B
I
think
this
is
more
consistent
forward,
but
I
don't
want
to
create
too
much
dependency
or
burden
to
make
this
happen,
having
also
connection
between
the
different
demos
or
imperfect,
that
could
also
be
a
plus,
because
then
we
we
have
more
pieces
of
the
puzzle
together
and
we
can
work
also
on
this
Intel
operation
and
adaptation
function
between
the
different
pieces.
So
we
can.
We
may
also
be
able
to
demonstrate
a
more
complete
in
turn-based
system
that
just
with
individual
pieces
and
maybe
also
some
actor
reusing
components
between
them.
B
B
Also
in
this
globally
for
for
all
these
different
disc
discussion
item
on
ibn
sure
whole
myself,
we
will
make
some
summary
of
the
output
of
this
discussion
and
make
a
feedback
to
do
to
the
research
group,
and
maybe
some
recommendation
of
what
you
expect
will
happen.
And
we
will
use
that,
as
at
the
next
step
to
proceed.
B
B
A
Disgusting,
the
last
time
we
actual
meeting-
and
so
we
requested
do
two
last
round
of
command
before
requesting
the
call
for
adaptions
I
think
we
are
now.
We
know
in
the
process
to
to
send
us
the
mainly
the
core
production
for
this
document
so
as
they
actually
ask
for
commands
and
address
some
of
them
in
the
expensive,
ill-advised
user
in
the
next
iteration
and
we're
waiting
for
the
of
course,
as
we
go
for
all
the
others,
the
document.
So
we
got
it
yesterday,
so
in
the
next
it's
today
or
tomorrow.
A
E
E
I
are
their
drafts.
We
asked
for
the
adoption.
We
got
some
comments.
Thank
you
additional
comments.
So
thank
you
very
much
yeah
here
Magda
for
new
comments,
and
our
initial
proposal
is
that
we,
we
replied
in
those
comments,
but
we
also
have
a
list
that
we
created
for
the
comments
that
we
will
address
in
the
next
version
of
the
document
after
the
adoption.
So
we
created
that
list
that
we
would
address.
I,
don't
know,
I,
wouldn't
go
back
in
what
we
changed
and
what
he
added
I
think
he
presented
it
several
times
and
discussed
around.
A
K
A
E
B
B
So
for
me,
and
maybe
a
message
that
I
would
like
to
convey,
this
is
really
chair,
but
this
is
applicable
to
any
research
group
document,
not
only
specifically
to
this
one
is
that,
for
me,
I,
wouldn't
change
that
once
the
document
change
from
being
an
individual
proposal
to
a
research
group
proposal,
it
becomes
the
the
baby
of
the
research
group.
Somehow
and
then
we
see
collective
responsibility
and
especially
from
the
offers
that
now
become
I
will
say
initially
there
were.