►
From YouTube: SimPEG Meeting February 28, 2017
Description
SimPEG meeting on Feb 28. Thibaut Astic presents on Designing objective functions: a probability approach
A
C
Okay,
everybody,
so
not
that
much
call
for
me
today,
like
I'm
working
on
it,
but
I
still
have
some
numerical
issue
to
figure
out
so
today,
I
mean
mostly
focused
on
showing
you
like,
yellow
chair
and
shoes.
So
far
for
a
project.
I
gave
me
so
with
so
basically
I
really
I
I
try
to
like
unfollow
the
assumption,
we're
making
about
simply
a
few
key
function
and
like
the
zipper,
the
the
main
ability
with
that
was
basically
to
have
to
design
a
framework
where
I
can
work.
C
C
Under
and
chargeable,
overburden
too
so
like
when
I
do
no
more
to
the
low
morning
barrel.
Aha,
it's
singing
nothing
like,
as
we
can
see
in
the
video,
the
middle
slices.
So
that's
my
model
and
that's
what
I
will
try
from
inversion
enzymatically,
what
I
see
with
my
real
my
realtor-
and
this
is
I'm
using
this
new
data-
cause
this
one
for
this
story,
but
instead.
C
C
The
main
goal
at
hand
will
be
also
to
be
to
say
that
if
we
pump
rented,
for
example,
the
Metro
PD
of
the
structure
saline,
we
don't
have
to
put
an
artificial
distance
of
X
waiting
on
a
row
I
for
in
our
physics.
So
because
that's
what
I
will
drive
so
I
don't
see
anything.
But
we
see
that
for
a
number
I
just
stood
face,
see
like
my
love.
C
Collectivity
like
a
high
quality
can
be
much
higher
than
what
I
have
in
my
photo
and
if
you
had
I
could
say
like
a
detective
et
Ossa
of
ice
field
data.
So
we
have
the
overburden
connectivity
here,
the
swivel
utility
and
I'm
somewhere
around
me.
For
my
father's
suicide,
like
I,
was
our
change
of
at
least
ten
to
fifteen
percent
of
the
data
due
to
the
sphere
so
I
detected.
C
The
thing
is
that
can
I,
we
drive
after
that
since
the
inversion,
and
so
that's
kind
of
like
the
motivation
and
the
beginning
of
what
I'm
going
to
show
you
today.
So
let's
game
moving,
so
I
try
to
unfold
it
for
provari
probability.
Part
of
you
like
there
is
nothing
too
crazy,
but
it's
from
it
was
really
like
your
high
opener
on
what
all
the
Assumption
been
making
with
the
with
your
objective
function.
C
So
the
first
thing
is
that
we
had
a
will
daddy
and
we
basically
gonna
basically
consider
it
as
a
realization
of
multivariate
probability
distribution.
So
this
observe
that
is
like.
So
it's
a
realization
of
poor
distribution,
P
of
the
low
in
the
Tigers,
my
physics,
my
model
and
and
some
co
violence.
On
my.
B
C
C
C
B
C
B
B
C
D
C
Be
that
I
want.
The
other
axis
would
be
that
that
tooth,
okay
and
then
on
the
under
said.
We
have
the
probability
density.
Ok,
so
here
file
and
use-
and
you
see
it
like
I
put
it
shifted
here
and
it's
not,
and
it's
actually
a
bit
skewed
to
so
my
maximum
probability
will
be
right
will
be
here
so
like
around
two
for
this
guy
in
the
room.
They
were
for
this
for
this
one.
C
B
C
C
So
let's
start
to
look
familiar
at
some
point.
Yes,
so
that's
so,
as
I
said
before,
we
want
to
find
out
the
model
M
that
maximize
the
probability
to
get
this
data,
so
I'm
looking
for
Z
for
the
ad
max
of
em
of
this
probability
and
a
way
to
do
that
is
that
it's
actually
equivalent
is
that
it's
if
I
want
to
maximize
the
potion
soma
likelihood
it's
it's
like
minimizing
the
negative
of
the
rug.
Also
liked
you,
because
negative
log
will
be
a
convert.
C
A
C
B
B
C
So
if
I
want
to
minimize
that
now
the
negative
log
log
likelihood
so
I
just
drop
the
ad
I
just
drops
the
exponential
here
and
minus
10
minus
here
got
plus
so
I'm
looking
for
the
minimum
of
the
of
the
l2,
none
of
n
minus
D
lowering
the
covariance
matrix
here.
So
we
get
back.
We
got
back
basically,
what's
what
we
usually
use,
but
we
just
derive
it
from
from
the
form
a
coherent
account
of
you,
saying
that
okay
I'm
your
soon
that
my
nose,
my
like
made
my
basically
and
never
tell
that.
C
C
C
So
what
we
look,
instead
of
looking
for
the
polarity
of
dealing
a.m.
and
similar
we're
looking
for
the
probability
of
our
model,
moving
our
physical
data
and
our
covariance,
which
can
be,
but
that
thing
is
much
harder
to
choose,
because
we
don't
know
anything
about
the
model.
At
first,
and
so
this
approach
of
like
maximizing
the
poverty
of
your
mat
perimeter,
that's
what
people
usually
call
maximum
a
posteriori
estimation.
C
C
So
this
guy
is-
and
I
have
a
normalizing
term
here-
that
we
don't
need
to
to
carry
just
like
the
probability
of
the
data
over
like
the
wall
space
of
like
or
possible
other
things,
so
just
a
normalization
term
that
we
usually
just
compute
as
the
end
we
compute
the
probability
of
things
and
after
at
the
end,
we
just
make
sure
that
the
sum
of
everything
is
equal
to
one.
So
it's.
C
We
care
about,
and
it's
not
dependent
of
our
model
too.
So
if
I
want
to
maximize
this,
it's
equivalent
to
maximizing
just
a
guy
are
on
top
so
now
I
have
this
probability
of
a
man's
ass.
This
one
is
open.
Let's
I
may
need
to
make
some
assumption
for
this.
One
from
infant
assumption
we
can
make
is
that
my
M
is
now
also
a
realization
of
a
group
of
a
Gaussian
distribution
of
mean
my
fun
and
covariance
one
of
a
better.
This
looks
familiar
again
and
we
do
with
this
one.
C
C
So
it's
going
to
it's
gonna
drop,
so
I
didn't
do
the
intermediate
thing
but
accident
when
I
minimize,
the
minus
log
of
my
probability
of
em
I
get
back
so
by
that
a
fitting
tell
me
so
before
and
my
regularization
term
that
is
so
assuming
I
like
so
basically
so
we
got
so
from
the
probability
we
got
back.
The
chicken
of
regularization,
so
why
I
do
or
for
me,
was
actually
really
nice
to
develop
to
develop
that
likes.
C
C
C
It's
drawn
from
a
Gaussian
distribution
and
we
have
made
actually
a
tons
of
assumption
to
get
to
get
to
here.
So
l2
and
ocean
are
very
like
interweb.
It's
it's.
The
same
thing
just
want
two
different
paltalk
you
and
one
thing
was
nice
to
look
at
too-
is
that
when
we
do
actually
when
we
do
with
better
cooling,
narrow
inversion,
that's
what
we
do
actually
like
that
at
first
like
when
beta
is
high
with
users.
C
C
Okay,
so
we
got
back
what
we
all
already
all
known,
but,
as
I
said
for
me,
I
found
it
really
nice
to
to
derive
it
from
this
point
of
view
and
I
just
pretty
much
I'm,
just
under
the
spot
or
the
search
or
the
accenture
behind
it.
So
what
what
can
we
do
with
it?
First
thing:
I,
you
can
think
of
it
like
you,
just
use
a
different
probability
distribution.
C
So
if
you
use
a
lot
last
priority
distribution,
you
get
back
your
zl1
normal,
so
that
she'll,
don't
you
use
successfully,
for
example,
if,
if
you
have,
if
you
want
to
recover
a
sparse
model
in
some
basis,
such
can
be
in
some
at
cancer
that
can
be
in
like
in
just
in
the
space
in
the
space.
But
if
you
want
to
have
like
a
your
model
that
is
fast
in
full
year
in
web,
blades,
etc,
that
would
work
the
same.
C
Another
distribution
that
is
really
cool
is
a
student's
t-distribution.
It's
like
a
Goshen
but
with
if
your
tail,
so
it's
very
nice
for
outliers
detectors
or
if
you
know
that
most
of
your
data
are
good,
but
you
have
some.
You
have
also
a
lot
of
outliers
that
are
hard
to
detect,
but,
and
you
have
enough
out
layers
that
is
basically
skewed.
Your
your
distribution,
like
am
I,
should
have
never
made
this
distribution
for
that.
C
But
imagine
that
you
have
you
even
like
leave
you
in
over
the
dome
in
case
you
have
tons
of
the
point
that
for
almost
a
linear
trend,
but
then
you
have
a
lot
of
very
high
value
on
top,
like
you
like
in
l2
norm,
it
will
just
put
you
in
the
middle
and
if
you
then
outliers
are
the
ones
that
I
don't
fit.
You
can
actually
threw
away
all
the
good
data
and
I
can
enter
yes.
C
So
yeah
I
have
like
tons
of
bone
disease
likely
that
pepper
tree
behind
at
CTU
vs.
line,
but
have
also
talked
about
Tigers,
do
too
to
some
it.
So
this
is
just
looking
at
it.
I
would
like
to
do
in
your
exes
right,
but
if
I
do
a
to
know
that
is
that
supposed
to
get
to
lacy
square,
give
me
right
and
then
I
said.
Oh,
so
all
the
other
two
years
are
our
sides
at
winco.
C
B
B
C
So
that's
so
find
out
what,
if
you
have
a
if
you're
busy,
because
basically
an
end
to
in
the
end
to
know
if
you
are
like
I
outliers
means
we
push
you
out
because
you
use
square
00.
That
really
is
get
even
video
ends,
but
I
ready
for
your
distribution.
Anyone
just
look
at
the
absolute
value
of
your
you
just
forget.
The
absolute
value
goes
everything
so
many
you'd.
Like
sir
this
one
looks
at
the
log
of
the
square
of
the
week,
so
I.
C
C
B
C
A
B
D
Yeah
yeah
no
I
totally
are
I
mean
that
I
think
it
they
can.
I
want
to
do
this
when,
when
they
re
waiting
after
a
few
songs
right,
the
crew
started
implementing
it
with
the
with
the
UPC
codes
with
yeah
I,
don't
know
it's
just
another,
it's
another
layer
on
top
of
it
and
I.
It
needs
to
mean
something:
it's
give
you
something
out
of
it.
I
guess,
work
yeah,.
C
This
one
will
be,
as
I
said
like
if
you,
if
your
data,
if
you,
if
you
have
good
data,
there's
also
some
very
noisy
data,
and
it's
can
be
in
it
scanned
and
l2
is
not
able
to
sort
to
sort
it
properly.
I
could
send
others
it's
a
it's
one,
another
tool
you
can
use.
If
you
have
this
type
of
this
type
of
thing,
ya,.
C
C
This
one
is
like
small
slide
for
very
complex
things
that
I
try
to
look
at
it.
It's
like,
like
people
in
machine
learning,
call
that
a
structural
sparsity
and
it's
basically
like
the
basic
idea,
is
that
you
apply
different
known
to
different
part
of
your
model
and
you
can
even
compose
the
different
and
you
can
even
put
the
same
parameter
in
different
nodes.
So
that
is
so.
C
Then
you
can
deal
with
dependency
between
your
model
like,
for
example,
if
you
have
you,
oh,
if
you
push
then
make
em
20
and
if,
if
both
are
nonzero,
then
we
utilize
them
in
the
l2
norm
and
stern
stuff
like
and
stuff
like
that,
so
you
can
actually
structures
bus
like
put
actually
impose
structure
in
our
model,
with
that
I
haven't
seen
it
in
done
in
geophysics,
yet
but
I'm
I'm.
I
Explo
that,
but
I
try
to
read
like
some.
It's
something
very
sad
someone
we
knew
like
I
read
the
disease
from
like
2015
like.
C
Is
this?
Yes,
that's
kind
of
what
it
is,
but
it's
been.
It's
been
used
for
like
machine
learning
application
before
so.
Actually
I
just
made
like
a
very
random
fuse.
We're
like
11
parameter
II,
like
is
drawn
from
the
Gaussian
and
the
other
perimeter
is
drawn
from
the
Laplace.
So
that's
something
that
is
the
thing
to
do
about
this.
This
one
is
religious
like
okay,
every
some
things
that
people
do
this,
but
I
haven't
dive
into
it
that
much,
but
it
what
I've
been
working
more.
C
These
days
is
working
on
the
prior
that
we
talked
about
like
so
far.
We
we
think
that
if
we
use
it,
if
you
use
like
a
single
motion
or
single
Laplace,
we
can
I
get
l2
l1.
No,
but
in
my
case,
for
my
data,
for
example,
are
tons
of
petrophysical
data
and
my
goal
would
be
to
make
it
falls
into
like
that.
C
My
inversion,
we
drive
model
by
metals
that
falls
into
these
bins
that
are
defined
by
my
petrophysical
data,
so
then
so
that,
for
me,
the
opposed
our
approach
I
choose
right
now
is
to
define
my
prior
as
a
sum
of
probability.
Distribution
channel
laplace
or
whatever
you
like,
but
let
goshen
is
good
enough.
It's
I
can
derive
it
if
I
can
divide
it
as
many
time
as
I
want
and
it
can,
if
I
have
and
if
I
put
in
our
ocean
I
can
fit
basically
any
type
of
distribution.
So
that's
quite
nice.
C
C
Because
the
Sun
is
inside
the
log
now,
which
make
it
more
painful,
that's
where
are
some
issue
is
that
I
will
show
it
later
was,
as
that,
what
that
looks
like
so
I
should
put
the
keys
of
the
prod.
Is
that
with
that
I
can
fit
any
parameter
distribution
and
it's
easily
scalable.
That's
not
like
something
in
1d,
but
that's
something
to
the
that
you
can.
C
You
can
fit
and
you
can
go
in
many
dimension
up
to
us
up
as
long
as
your
data
keep
up
with
the
dimension
and
actually
passed
four
dimensions,
probably
hopeless
in
a
way
that,
for
example,
in
in
one
day,
if
you
want
to,
if
you
it's
basically
an
app
instagram,
if
you
want
to
think
of
it
and
in
1d,
if
you
want
to
have
10
mins,
for
example,
you
need
dancer.
Thanks
up,
you
can
start
with
10
sample,
but
2d.
You
want
n
beans
for
each
of
the
two
parameters
unit.
C
You
need
hundred
three
dimensions
as
a
three
dimensional
parameter
like
a
cross
street
physical
property,
you
will
need
more
than
a
thousand
and
it
just
scales
and
scaling
scales.
So
imagine
that
DC
IP
data
I
have
resistivity
and
22
times
IP
time
windows.
If
I
want
to
do
a
fitting
on
all
this
parameter,
I
would
to
be
able
to
start
to
do
a
good
job.
C
B
C
God
that
by
tacky
bug
is
actually
important,
scalable,
but
mathematically
and
with
this
formulation,
it's
actually
englobe
k-means
and
thousand
window
I'll.
Just
special
case
where
you
came
in,
for
example,
would
be
put
your
covariance
matrix
all
the
same
and
that
are
just
proportional
to
identity
and
buzzing
windows
is
like
you
have
a
50.
You
have
a
500
data
pound
I'm
going
to
feed
500
Goshen.
So
it's
going
to
fit
perfectly.
C
There
are
some
twink
behind
that,
but
I
don't
need
to
take
two
to
see
it
right
now,
and
so
that
was
an
early
resort,
but
it's
not
quite
there
yet
like
this
one.
My
I
admit
that
some
of
my
gradient
was
wrong
or
anything
but
was
like
kind
of
still
beginning
to
give
something
something
nice.
You
know
in
a
way
that
you
see
here
that
I
didn't.
C
I
didn't
change
the
scale,
but
just
I
just
get
the
same
scale
for
my
perimeter
back
and
with
the
same
with
the
same
survey
and
in
putting
this
non-local
need
is
non-local
information
into
my
inversion
and
before
the
sphere
was
invisible
and
it
was
putting
me
very
conduct
it
stuck
on
top.
Now
I
got
some
high
at
the
surface,
but
I
start
to
distinguish
that
so
that's
kind
of
the
idea.
C
This
was
a
early
resolved
and
now
my
code
is
all
broken,
but
that
wasn't
doing
that
was
encouraging
at
that
point
and
I'm
still
working
on
that
and
so
like
now
to
show.
What's
my
function,
objective
looks
like
that's
what
it
looks
like
and
that's
when
it's
become
become
nicely,
so
my
data
fitting
term
is
0
still
the
same,
but
now
my
instead
of
a
Norma,
how
the
negative
log
of
a
song
of
waiting
exponential.
C
C
My
put
my
the
probability
here:
it's
it's
always
positive,
but
it
can
get
as
close
to
zero
are
like
as
we
want.
So
that
means
that
I
take
the
log
of
something
that
is
always
positive,
but
can
can
delete.
Get
close
to
zero
may
be
very,
very
fast
too.
So,
let's
can
impose
that
caused
me
or
so.
Some
numerical
issue
because
I
took
that
as
a
log.
C
Very
close
to
zero,
so
sometimes
I
can
glue
it
can
blow
up
and
that's
why,
even
if
it's
infinitely
divisible
the
grenye,
the
nation
can
get
very
big
very
fast.
That's
that's
one
of
my
in
the
program
a
happy
face,
or
so
right
now
and
the
last
point
I,
don't
know
if
it's
a
problem
or
not
it
just
is
that
in
most
very
most
cases
like
this,
this
thing,
like
the
probability
density,
gonna,
be
less
than
less
than
1.
C
So
my
negative
log
going
to
be
positive,
but
this
this
thing
ear
is
like
we
talked
about
probability,
but
it's
we
should
not.
We
should
not
be
talking
about
probability.
We
should
be
talking
about
probability,
density
distribution
and
if
I'm
ABT
stay
between
0
and
1
probability
density
does
not
so
this
term
can
sometime
be
more
than
one.
So
I
get
something
in
my
objective
function
that
is
negative
as
you
see
on
them,
which
should
not
be
a
problem.
C
If
you
will,
if
you
really
go
mom
that
you
can
just
shift
by
your
constant
to
make
sure
that
everything
stay
positive,
it
just
batted
you.
It's
just
like
an
annoying
detail
to
look
at
when
you
look
at
your
blog
result
and
like
you
see
that
your
phi
m
is
to
get
you
and
and
this
one
I
haven't
like
I,
haven't
included
in
the
presentation.
But
we
can
you
like.
Like
all
the
same
way,
we
include
those
small
nexus
cetera.
C
We
can
include
the
gradient
into
the
admin
to
that
thing,
and
one
thing
why
I
mentioned
to
you
that
I
was
really
interested
to
the
composite
board.
Objective
function
is
that
for
sister,
like
better
cooling
comes
very
naturally
when
we
do
the
when
we
have
just
a
Gaussian
drawn
from
a
Goshen
distribution,
but
this
one
I
don't
have
anything
with
it
doesn't
mean
that
I
don't
want
to
put
anything.
C
I
would
probably
put
some
ways
like
we
find
that
I'll
find
the
getty
blog
like
he
would
or
something
like
another
alpha,
but
I
will
definitely
use
another
bit
like
a
different
bit.
A
strategy
for
like
I,
don't
know
like
what
say
what
judge
did
in
CSM
is
that
evident,
I,
don't
think
use
this
framework
to
derives
came
in.
He
basically
took
them
the
objective
function
and
add
the
and
add
the
objective
function
you
usually
use
for
chem
in
which
is
like
a
special
case
of
that.
C
Once
we
got
some
simplifications,
we
just
adding
term
so
after
your
objective
function,
so
we
still
have
a
small,
nice
and
derivative,
etc
and
I
think
maybe
it
helps.
Maybe
it
else
also
like
stabilizes
the
problem
when
you,
when
this
guy
become
that
become
nasty
but
but
definitely
I
will
like
a
compostable.
C
C
I
will
do
a
better
cooling
on
the
small
exit
strategy,
meaning,
but
this
guy
would,
I
will
try
and
make
it
back
stays
the
same,
because
it's
there
is
nothing
appearing
on
that
side
that
they
will
eat
like
mathematically
speaking
in
that
framework,
I
don't
see
any
reason
to
do
a
code-
cooling-
maybe
maybe
numerically,
but
we
see
that
letter
so
just
to
that's
just
an
example.
So
that's
my
model
here
and
I
just
have
5
Goshen
to
fit
that.
That's
and
that's
quite
nice,
like
I,
mean
it's
all
the
web
or
etc
in
blue.
C
B
C
C
C
D
C
Right
now
how
I
do
that
is
that
I
created
a
new
regularization
class
that
I
could
fit
for
utilization
and
that,
like
her,
so
it's
and
so
now
in
the
inversion
like
in
the
eval
function,
like
figure
ization,
like
it
evals
and
negative
log
line
likelihood
in
the
gradient
and
asian
of
the
negative
log
likelihood.
Okay.
C
A
E
D
A
A
A
A
Awesome
and
then
the
other
pull
request.
I
wanted
to
go
over
and
maybe
I'll,
let
Rowan
jump
in
briefly.
Here
too
is
pulling
out
the
mesh
class,
and
so
now
it's
in
the
discretized
package.
So
now
there's
no
machine
inside
of
sim
peg.
This
is
on
dev
right
now,
but
I
would
like
to
get
it
down
to
master
soon,
ish
so
yeah.
So
all
the
meshing
changes
can
be
done
strictly
and
discretized
and
then
Rowan
updated
the
Richards
stuff.
E
Yeah,
which
I
mean
there's
it
it
works
better
now,
so
you
can
invert
for
any
of
the
different
sort
of
empirical
parameters.
I
might
do
that,
like
your
talk
about
that
at
a
later
time,
because
I
think
it
has
some
application
to
IP
and
any
time
we
want
to
do
sort
of
none
any-any
mapping
that
depends
on
the
model
or
on
the
fields
and
I'm
sure
we'll
be
getting
into
more
of
those
sorts
of
things,
especially
as
we
get
into
joint
in
versions
as
well.
E
E
Yeah
I
could
I
could
do
do
that
sort
of
prepare
something
for
the
next
next
couple
weeks,
yeah
and
then
there's
that
one
other
thing
about
the
number
hindsight
I
are
no
longer
your
site.
I
sparse,
are
no
longer
being
exported
by
this
impact
import
start.
So
that
means
you'll,
probably
if
you're,
if
you're
relying
on
those
you'll,
have
to
add
to
other
imports
at
the
top
of
various
functions.
A
E
Yeah,
I
think
this
it's
it's
pretty
simple.
I
don't
think
they're
ibly
any
breaking
changes
other
than
the
numpy
sci-fi
thing.
The
discretized
package
is
exactly
the
same
ins.
It
it
sounds
like
Joe
is
working
on
some
improvements
to
speed,
especially
on
the
tree
mesh
side
of
things.
I
don't
have
a
an
idea
of
when
that
sort
of
stuff
will
land,
though.
A
D
D
E
E
And
this
sort
of
allows
people
to
sort
of
jump
in
at
different
levels,
because
I
mean
so,
for
example,
Joe
is
jumping
in
on
the
discretized
side
of
things,
which
is
like
finite
volume,
with
derivatives
in
mind,
and
that's
like
what
that
packages
for
it
doesn't
have
anything
to
do
with
inverse
theory
or
inverse
frameworks.
Just
you
want
to
take
a
derivative
with
respect
to
your
properties
on
the
mesh
yeah.
D
E
D
The
same
yeah,
it's
the
same
old,
it's
just
a
man
is
bug
bug
hunting
yeah.
One
thing,
though
they
would
be
nice
to
changes,
is
the
print
statement
when,
when
we're
solving,
can
we
print
like
at
the
end
of
the
end
of
everything
after
the
directives
or
applied
basic
one,
I
should
say
before
the
directives
are
applied.
E
E
E
I,
don't
have
any
strong
preferences
without
without
sort
of
seeing
one
way
or
the
other
don't
like
it
held
it
up,
sweet
and
then
also
with
your
ear
at
the
printing
I've
been
looking
at
some
other
table
stuff,
there's
some
good
stuff
out
of
a
trope
I
that
we
could
possibly
rely
on
or
ask
them
to
break
it
out
or
something
like
that.
I
don't
know
for
like
print
to
file
your
man
yeah
print
to
file
and
like
console
login,
which
is
similar
to
what
kootenay
talked
about
last
week
or
two
weeks
ago.
E
Yeah.
So
there's
there's
some
maintenance
there,
because
our
iteration
printers
are
terrible.
I
got
a
trick,
is
just
a
little
bit
yeah.
They
do
the
trick,
but
especially
as
we're
going
to
this
compensated
object,
objective
function,
stuff,
it's!
It
is
sort
of
difficult
to
see,
see
all
of
the
various
pieces
all
at
once.
Oh.
C
D
E
A
And
from
talking
with
Doug
and
soggy,
it
sounds
like
mt
is
one
of
the
things
that
is
like
been
quite
important,
both
in
Indonesia
and
in
India
and
so
like
getting
some
of
those
examples
built
up
and
have
you
saw
on
the
cob,
we'll
just
call
him
out
would
be
that
would
have
a
big
impact.
So
that's
exciting.