Peter it would be.

Really cool or
brown best using brown color or

Jason is the first right.

Or if we.

Really are in a good.

Workout.

We get back next year so we're fortunate
he's with us Jay says the talk about

architecture and
I don't know it's very easy for

them in that you can back into it and

record agreement with it he.

Or he didn't really care
who actually cared.

And that's not very.

Clear.

Here.

And with no.

Thanks thanks Well yeah I do care about
architecture but I don't have any three

hundred D.P.I. renderings you know what
really fine lines but I'll try and

make it you know kind of relevant and
the tack I'm going to take on this.

Comes from some of the research
that some students and

I are doing that's geared towards
design and making design decisions and

I just realized I neglected
to put a name on here but

a lot of the work that you'll see later
on was done by my Ph D. student Roy.

And at the very end we'll talk a little
bit about what he's probably going to work

her Ph D. thesis on so
I'm going to talk about making decisions

in the presence of not having perfect
information in the kind of decisions

are applied to architecture but
are geared towards performance and

by performance I mean energy energy
performance kinds of decisions

and before I get all into that I kind
of want to lay a little groundwork so

I've got some slides here to just kind of.

Give everybody.

A picture of where I'm
going to be coming from and

first I'm going to talk
about models right so

water models what's a model.

A representation of.

Something right.

Yeah actually this was
the very first thing I said in

in the very first class
that we had together.

Yeah a model is
a representation of something.

In this case it's a model of reality or
in the case of making designs it's a model

of a imagined future reality that
we want to bring about right so

that's what a model is and we can use
models for different purposes right and

the kinds of models that you
are right here trying to produce.

A final set of models that
represent a design right and so

those models are a specification
their use is intended as a as

a specification and
Rekha might I hope I don't.

You know give you spasms right now no
no this won't do it this is last year

here's one representation
this is a design of a high

performance net zero
energy housing project.

From one of our M.S.H.A. P B student
teams last year entered in a competition

it's a specification on how to build
a building in this case a small house but

you can give this to a builder and
they'll go out and build this and

a lot of other documents as well but
they'll go out and build it but

how did we get to this we started
out with a blank sheet of paper and

some notion of where what we wanted to do
but in order to get from the blank sheet

of paper to this we had to make
a whole lot of decisions along the way

thousands and thousands of decisions
get made along the way and

what do we do to help us
make those decisions.

And what do we use to help
us make those decisions.

In.
Pretty much in all case models right so

a model is not just a specification as
a representation of reality it's also

a representation of something else
that we used to think with so

a model is also a thing to think
with right as a thing that helps

us support making our decisions and
there's lots of different kinds of models

in fact you could probably argue that
everything that we ever do in our head is

some kind of modeling we can't do anything
without modeling right crossing the street

requires us making a model of what's going
to happen if I step into that street with

a car that far away going some speed
towards us right everything is a model.

So it's a thing to think with.

Here's some other kinds of models that
are architecturally designed based

that were used to work through some
problems and make some design decisions.

This is from two thousand and
seven when we did the sort of kept on.

And you know one of the things that we
wanted to make decisions about is how to

build the P.V. array and this went through
a lot of different kinds of models from

physical mock ups at a small scale
to a larger scale to help work out

you know different problems and figure out
how this thing needs to be put together

really quick models not made with the same
materials but relevant nevertheless

representative of what we want to do and
that gets keeps getting refined

as you go along and so where you can
alternately the models that you're making.

That you're using to think with become
models that are the end product

specification that you can go out and
build the thing for real right there and

if anybody wants to see this it's
up to tell us museum and Carter.

Right now so
it's still there so use models.

To work up through problems but what's
the thing course the deal with models.

Every model is wrong they're all wrong but
some are useful some are useful and

wasn't actually what does it mean
to be useful for a model to be use.

Steers you in the right
direction yeah I like that

notice he never said anything
about how close it is to reality

he said just the model helps
you make the right decision.

Said nothing about how close
it is to reality OK so

it's kind of a nice point but.

Yeah.

Well.

It's something something that
helps you ask questions.

Right yes right the right.

OK yeah that's right don't do that it
doesn't say do this or that well yeah I've

actually got maybe some examples towards
the end that might say what not to do

in addition to saying what
to do on the same time.

OK We're still working on that so
it's a it's a little bit speculative OK so

yeah I can tell you what not to
do as well as what to do OK.

So some are useful and I'm going to just
kind of look at the kind of models that I

work with from day to day and
if you've had me in classes

before you kind of have some
idea what what models I mean so

this is my bedroom circa
summer two thousand and nine.

Putting on an extension and
let's take a model.

Of just a little section this
way of this wall right so

what does that look like it's
a regular old stud frame wall and

we can represent it maybe like this in
from a person performance point of view or

one aspect of performance point
of view we might look at this and

ask the question what's the heat
transferred across the wall

right from the some outside temperature to
some inside temperature and there's lotsa.

Different ways to construct a model
to help you answer questions related

to that type of performance and one of
them if you've had me in class before and

systems one looks like that so
that's one out

of many different kinds of thermal
performance models we can construct for

the physical reality that already
exists or will exist in the future and

this is a graphical representation of it
each one of these little squiggly lines if

you're familiar with electric circuits is
a resistance that resists the flow of heat

between the inside and outside and
they'd come from different things.

Inside temperature and outside temperature
that's just the weather inside and

outside this are six and that are one
our resistances that have to do

with the film of air that's
close to either the inside and

outside surfaces of the wall
these other resistances are.

In coatings a particular kind
of view of the geometry and

the materials that go into that wall so
these resistors are things that we've

actually made choices about or made
decisions on right there parameters of

the model that we specify we
have control over these we

really don't have a whole lot of control
over we can't control the weather and

we can't you know we don't we can't say we
can't dictate how the air flows on either

side of this wall to very
much much of a degree.

Now.

Here you can really represent this in
another form as an equation I told

Tristan I tried not to write any
equations down I tried but I violated it.

We'll see later but we can take all
of these different variables and

put them together in
an equation to compute

the he transferred through here right and
so if we have numbers for each one of

these variables we just stick it in there
turn the crank and we get a number out.

Right.

All done.

Of course not the problem is this is
about unknown is that we don't really

know what is going on inside or outside
we really don't know what the air is

doing on either side of the wall it
could be blowing it may not be it

depends on the other buildings around us
it depends on where we are it depends on

the time of year all sorts of things that
we really don't know about right these

things these resistors that we've actually
made a choice on and have decided on you

might think that we know what they are but
we really don't because we can go and

buy some insulation and then look up
what it's thermal properties are but

those may not be right right there's
always some error and how these things

are actually measured there's also Aaron
how these things are constructed and

if you've ever been in a building under
construction you know it's not exactly.

High craft going into
everything especially houses

if we look back at my house.

You know we see this insulation
here it's not really filling

the hole cavity right there or there you
know it might be smooshed in one place and

not another and sag over time and
all kinds of things like that so

we don't necessarily know what these
guys are at least not with complete and

total confidence right now
as a result Well let's.

Scroll down a little bit we
can take these these these.

The fact that we don't know about these
that we don't know everything perfectly

about these different variables and
we can kind of categorize.

These types of variables in this
case into two kinds here in the blue

going to call these scenario variables
that are uncertain and I'm going to

call that uncertainty associated with the
scenario scenario uncertainties if you're

mathematically inclined these are boundary
conditions we don't always know what's

going on outside or even inside right so
these are these are uncertain.

In the scenario with or the context
within which this model finds itself we

don't know everything about that we also
don't know what these resisters are with

complete confidence and
because these are variables or

parameters that we make a decision about
them and call them parameter uncertainties

this distinction will come in and a little
bit as a result of all of these different

variables that go into this function F.
being uncertain that means this is

uncertain does well we don't really
know what it is with confidence.

So what do we do we want to be able
to quantify how much we know and

how much we don't know both on what
goes into this model is the input and

what comes out on it as well so

the first part of the exercise of dealing
with uncertainty in this kind of.

View of a building is the first thing
is to quantify the uncertainties that

go into our input into our model right and
so that's a whole effort it's not

the point of this talk let's just say that
we can do that and here's just an example

of one way to represent that if it's been
a while since you've had statistics I'll

just remind you this is a problem these
are two probability distributions you kind

of think of it as a histogram if
you remember those that have been.

Shrunk down and
done some limits on and made tenuous.

And their plotted with the value of
the variable here on the X. axis and

the Y. axis given as I'm going to
call in quotes likelihood it's not

using the term likely in
a kind of aloose way but

you can just as a conceptual we can think
of this is how likely something is so

here this room this resistor is that
outside air film it turns out that

that's pretty uncertain we don't have a
whole lot of informational around that and

it can take a broad range of values
take a very broad range of values so

here these are to.

You know a cartoon.

Version of these distributions but
here for

this are one it looks like
the most likely value for

this resistance is four point
five right is four point five But

there's a chance that it could be three
there's a chance it could be five there's

a chance it could be eight in the chance
that it could be eight is smaller than

the chance that it's four and
a half right so broad spread and

so when you see something like this that
means this is quite uncertain and we don't

have a whole lot of information about it
this other resistance is the resistance of

the outer sheathing this is something that
we've made a design decision about and

it turns out that we do know more about
that then we know about the outside but

it's still not perfectly certain here in
this case the most likely value is five

but it could be four it could be six
it could be something in between and

the probability that it's between you
know four and a half and five and

a half in this case is greater than
the probability that this is in between.

Say three and a half and five and.

OK so this there we have more information
associated with this one more little bit

more certain about the value of this but

it's still uncertain because there's
a spread to this value of completely

certain this would just be a straight line
going way up and down OK All right so

you do that for all these different
variables that are in our model and

you get a whole collection
of these things.

Might look something like this
I'm kind of making this up.

So what do we do how do we find out
the corresponding construct for cube.

Here's where I use quotes and
do some hand-waving we take all

the uncertainties that we've quantified
in the scenario in the parameters and

we run it through our model we would
call it propagating the uncertainty.

Through and
we end up with a probability distribution

of the heat transferred through here so
this.

Quantified is the uncertainty
in the ultimate.

Reason that we've constructed this model
in the first place the question is what do

we do with that now that this is uncertain
how do we make decisions based on certain

information or get to that in a little
bit OK So here we've talked about this

one model that those of you that have had
me in systems one are familiar with but

a dirty secret I didn't tell you in that
class is that's not the only model for

that that's just the thing that we used
for that class in some cases that model is

really wrong they're all wrong in some
cases that model is not true wrong and

is useful in other cases
it's not really wrong and

not useful so
you can model this not just this way but

also that way it's two different
views of the same idea and

one is maybe appropriate for some
situation one might be appropriate for

other situations but there's not a clean
division between that kind of overlap

there are there realms of usefulness
might overlap right so a choice is

which one of these do we use now this is
a completely different view of this and

it's a completely different model it's
a different function that goes along with

this if you take this and translate it
into equation it's going to be a different

function G.
than what we started out with F. right.

Now the choice between this and
that actually constitutes another kind

of uncertainty modeling
uncertainty a model for months or

just the fact that we make
an abstraction out of reality means and

it's not perfect means that
there is an uncertainty.

In that modeling itself just in the choice
of the model and we can compare that

between models although it's hard
to compare that to to reality OK.

So to kind of wrap up just
some of the ground work.

Groups it's going the wrong way yeah OK.

Go through the same exercise run
it through both of the models and

you're going to get different results for
each one right we're going to get

different results for each one one model
I give you something like that your model

might give you something like that now the
question is What do you do which one is

right which one is more useful to
get to that in a minute as well OK.

OK So just to kind of sum up a little
bit we've looked at modeling uncertainty

modeling on through your model for months
certainty has to do with the choice of F.

or G. scenario uncertainty and
parameter uncertainty have to do with what

goes in these three dots here so
the different kinds of uncertainty.

And for us I'm going to break up this
parameter uncertainty into two different

other kinds that are specific to the sort
of work I do all the rest of this stuff is

fairly standard stuff what's kind of new
is what's coming up I'm going to break

these this kind of parameter uncertainty
up into two different kinds of parameters

uncertainty one is decided parameter
uncertainty and that's the.

Parameter uncertainty that I've been
talking about up to till now this is this

is uncertainty dealing with parameters
that we've made a decision about we've

said I want that are three to be
this value even though it's not

that value but
I've made a decision about it.

If I've called it decided
parameter uncertainty

what does that mean
the other one has to be.

Undecided What is that one.

OK this is the uncertainty
that doesn't exist.

With the value of a variable that you've
decided upon it's the uncertainty

that exists in the evolution of the design
you don't know what decision you're going

to make that in itself is an uncertainty
so this is an uncertainty that exists in

a dimension that's orthogonal to the ones
that I've been talking about right here

right so this is uncertainty
that's associated with the.

That I don't know what
I'm going to do next.

Write that in itself is an answer
is something that's uncertain OK.

And to kind of illustrate
that let's take a look at

a design process this is just a cartoon of
a design process where we have the same

kind of graph that I was looking for
before frequency here or likelihood and

the value here of let's say Q.
the performance indicator and

then I've what I've done is I've just
taken this that graph that we were looking

at before from this direction and
just added the dimension of time to it or

the dimension of sequential decision
making I know it's not necessarily

sequential but the design process
going through this way right so

when we first start out we don't have
a whole lot of information and we haven't

made a whole lot of decisions so there's
a lot of things that are on no right so

we may not know very much and so
these distributions could be very spread

out now let's take a case where we're
looking at two different options

we're trying to make a choice of what to
do next and let's just say we're we're

acting like we're at the ophthalmologist
and we're getting glasses and what do

they say they don't say do you
need minus to die off there or

minus one and a half diopter they
say is this better than that.

So let's say if the let's just look at
this in a comparative analysis in this

case if this is energy demand Q Let's
say in this case it looks like

a is probably preferred to be

right probably because we can't say for
sure.

Because there is a chance that they could
be worse than B. there's actually a chance

to get whenever there's an overlap between
these distributions that means that what

looks like might be better at right here
if we're only looking at a single value

assuming that we had perfect information
and no one certainty may actually flip.

Right there's a chance that that could
happen in this case it's a small chance

that it could happen and that's what we
see as we go along here we make a couple

of other a few other decisions
the distributions should tighten up

because as we've made decisions we've
generated information about what

the building is going to be so
these distributions ought to tighten up

they'll never stop you know they'll
never become straight lines but

they should tighten up because
we have more information and

less uncertainty because we've
made decisions and here now B.

is preferred today in this case there
was a chance that that could happen and

by God here it happened and
as we go along now in this case once we

get to this final design right here it's
very clear that B. is preferred to A and

we can say that with pretty complete
confidence because these two distributions

don't overlap at all this says there's no
chance that A could be better than B..

So there's an uncertainty not
only here but here as well

that's the undecided parameter
uncertainty that I'm talking about and

that's the one that I'm going to
present some stuff about Turns out it's

really hard to quantify
that it's really hard.

And we'll see that in a little bit so the
question is can later decisions undo early

decisions right that's kind
of the crux of this of this.

This problem and

what you're seeing here is why it's really
hard to make early design decisions.

And this is also something people come up
with energy models to support early design

decisions without ever talking are dealing
with this uncertainty that exists in this

dimension right here right lots of people
talk about it in this direction to my mind

to my knowledge nobody's gone that way
right it's also really hard to do.

So the questions that we might ask for

is this this is something that's most
relevant in early design so some question.

My ask is Can architectural
decision be made in an early design

phase with confidence that the performance
rated outcome of that decision

will be a preferred outcome by
the time we get the final design.

Right and then the other question is
what models can support those decisions

what's a good model for right so these are
two questions that that my student Roy and

I kind of set out to try and
try and answer.

And if these are the questions that we're
looking at then the kinds of uncertainty

that we're really looking at is some kind
of some view of this modeling uncertainty

for trying to compare two models and
then clearly this undecided parameter

uncertainty why am I ignoring the US I'm
ignoring those because we're going to

be doing the ophthalmologists question
as we go on is a better than B. and

in that case you have a lot of
commonality all things being equal

the assumption and we have a little
bit of anecdotal evidence.

To support it that all things being
equal these kinds of uncertainties don't

have any impact in the decisions that you
make as long as you're doing a comparative

analysis a is a better than B.
right is this better than that as long

as that's what we're trying to do we don't
necessarily have to say this notice that

we're not trying to predict the future
we're trying to support the decision.

If you ever wonder why you lead buildings
are not necessarily energy saving

buildings it's a fifty fifty shot whether
or not they are this is kind of built into

that OK All right so
let's take a look at some of our process

it just kind of recap a little bit
here we take two design options and

run them through an energy model a very
big version of Q equals F dot dot dot

write a lot of that string them all
together you get an energy model for

doing this in a deterministic fashion
under the assumption that we have perfect

information we get single numbers for
the outcomes and here is.

We prefer to be if less is better.

The Way We Work is we take our options and
then add these probability distributions

this is that quantification
that we go through.

Call it that design option X.

the gray box just means it's just
a parameter with you know uncertainty

quantified and run it through something
similar we have to have some extra

infrastructure to run through the synergy
model we propagate the distributions and

here through this box and we get
the outcomes just like we saw before here

a seems to be preferred to be by going
by the most likelihood thing but

there is a pretty decent chance
that he could outperform at.

That possibility exists
OK A in we would say that

A does not dominate be OK.

So again the thing that we're asking
is option a better than Option B. S.

really not what we're asking what we're
really asking are is are we confident that

option A will be better than B.
Once we get to the end right so

that's really the kind of question
that we're that we're looking at OK.

So what do we need to do I
mean kind of the missing link

in all of this is coming up with these
probability distributions that encode

undecided parameter on certainty.

If we have lots and
lots of data about lots and

lots of design process as in the decisions
that people made we could in sit down and

spend a lot of time and lots of cups
of coffee and figure this out but

unfortunately that data is not there OK So
we resorted to plan B..

And you'll probably see Plan B.
has a couple of assumptions built into it.

And some limitations that we'll
talk about in a little bit but

before we can do any of this we have
to quantify these distributions we have

to figure this out so
how did we do it well the way we did it.

It is the only really data that is
really available out there are these

surveys of energy consumption data of
different kinds of buildings in the U.S.

There's something called the commercial
energy consumption survey in

the residential energy
consumption survey we do this for

red is residential buildings and
we looked at this.

Energy consumption for

different kinds of buildings in
different climates in the United States.

And we were able to
construct this distribution

are at least a distribution like
this say let's just say for

office buildings or hospitals or
different types of buildings for

different climate zones actually different
sense of zones in the United States so

we can get this total energy outcome
consumption data out there but

the problem is this is what we know for
actual built buildings right that exists

what doesn't exist is this right so
what do we do.

Well the kind of a nice thing about
a model is I talk about them in

terms of inputs and outputs or what's
an input can switch places with an output

under certain conditions so if I run
a model in a foot's called a forward

mode from what we think of as inputs to
outputs it turns out that you can actually

run it backwards as well and
that's called an inverse model and

has lots of applications in different
fields so what we did is we

took this outcome and then tried to find
out what these distributions are right in

here by reversing the directions
of these arrows Basically yeah.

An example of inverse modeling OK.

Talk about the chalk
Let's come back up you.

Know it's OK.

In this problem in this problem when I
talked about this I was talking about

a forward model these are all.

Presumed to be inputs and
they're all known and that's an output.

An inverse model and

well kind of different ways to think of
an inverse model one way you could think

of an inverse model is if I solve for this
if I know that I can just algebraically

switch a few things right now it's kind
of one way of running it backwards and

something you can do that with a simple
function like this if you have lots of

different equations in a you
know simultaneous model.

It's harder to do but you can invert
things like that the kind of inverse model

that I'm talking about would be an ill
posed problem we know that in this case

we know all of these but we don't know
that it's easy to turn the crank and

solve that in the inverse problem we know
that but we don't know any of these or

we may have a little bit of
information the forward mode is is

well this inverse mode is kind
of like a one to many problem

right it turns out there's ways that you
can take this a distribution on this and

having a little bit not much but just
a little bit of information on the bounds.

You know the part the range of possible
values of these there's ways that

you can estimate the distributions of each
one of these parameters are variables

that lead to the same distribution
here it may not necessarily be unique.

Or very unique but
you can at least kind of quantify that.

Right.

There OK.

I'll actually give you one of the end
OK I'll give you one of the end but

basically the idea is that you can if
you know this our distribution of this

you can estimate distributions of these
without a whole lot of information and

if you're interested that's you know
that tends to be posed as some kind

of optimization problem OK so what we did

is we just reverse the directions on
these arrows where it turns doing.

This inverse problem is very
confrontational intensive so what we did

is we made a baby version of this we did
a sensitivity analysis on the energy model

to find the parameters that are the most
important to have the biggest impact and

once you do that turns out you can take
a lots of parameters and energy model and

shrink them down to
the most important say.

Eighty percent and we built the baby
model that we built here is a very

simple statistical model it actually does
reproduce the outcomes if you go back and

and check it reasonably well but
with a lot less computational cost and

then we just ran this backwards and
we found probability distributions.

For these most important parameters and
this is I think four think this is for

office buildings in Atlanta
these are some different.

Parameters that in normal
forward mode would go into

an energy model the the probably the best
interpretation of each one of these is

that given the energy consumption let's
just say of office buildings in Atlanta

these are the likely distributions of
these different parameters so this is.

How many appliances you have.

The window to wall ratio we can't really
nail that down very well it turns out

this is the cooling set point
kind of broad building height

there actually the correct interpretation
of these distributions is that they

are representative of the actual
building stock out there right now.

That's really what that's really
the correct interpretation

where it starts to get a little dicey is
when we assume that these distributions

also represent undecided
parameter on certainty.

That's kind of hard because you have to
in order to really make that leap you

have to tease out all the old buildings

before you do that which is
kind of hard and another thing.

Well I'll talk about in a bit I get to
that all right so kind of the leap that

we're making the assumption that we're
making is that these distributions that

we get which represent the distribution
of these parameters in the building stock

are also at least roughly representative
of undecided parameter uncertainty.

OK.

OK Good yeah.

OK.

Great you.

Know it's it's it's not it's it's

looking at the stock of building models
that's the data that goes into it so

what comes out of it is really
in reference to the stock.

Yes Right right OK.

With that.

Well there are predators that we've
left out from the sensitivity analysis.

Yeah so there's there's actually a lot
more but it turns out that these

are the top twelve most that have
say I don't remember what the number

is off the top of my head but contribute
to eighty percent of the impact right

because this is just a way
to make it tractable.

Yeah right we do it all the time here
to you know open up any journal and

the every other word sensitivity analysis.

OK so that's dealing with the parameters
dealing with the parameter the underside

of parameter uncertainty right the other
thing that we're looking at is the model

uncertainty and so our take on
this is what do we use a model for

when it comes to making a design and
making design decisions at the early phase

it's not to protect it's really not to
predict the future because there's just

too many uncertainties in this
all we really want to know or

we really want to do with
this model is to support.

What do we do next what's the best
choice of action under uncertainty what

has the least risk for what we want to do
right so that's really what we're after so

we're not we're going to look at
what we call model the little not

in its comparison with necessarily
reality but comparison with correct

decisions which is the purpose of
the model in the first place right.

And this is a quote that I love to you
know to throw out I love George Box quotes

and I love George Hazel read quotes
this is a program manager at N.S.F.

who wrote a pretty cohesive book about
engineering design under uncertainty but

it could be any kind of design really.

And you know he had this
nice statement about model

right a model is valid if when used in
a specific decision making situation

with a given set of
available alternatives and

this is something I haven't talked about
yet the decision makers beliefs and

preferences these are subjective
an entirely up to you right decision

makers do so there's freedom in this this
is an optimization press a button and

you get a design and you must do that
right so it's up to you the decision maker

is certain that is preferred choice is the
choice that indeed yields the outcome that

is most preferred from among the outcomes
that could have been obtained

from the set of available alternatives
a lot of words but I don't know if it

could be made any more concise without
leaving out any information right so

notice that model validity in this case
when it comes to decision making really

has nothing to do in general
with prediction of the future.

It might but it doesn't have to write
it leaves open that possibility but

you're not you know restricted that OK
there is a point where you want to try and

predict reality as certainly I don't want
to diminish that but it's not necessarily

the end all be all when it comes
to making design decisions OK.

So how do we know if we're
confident about this this is

you know why I broke my rule and

said I wasn't going to have any equations
but here's how we quantified quote con.

And so I'm going to use those quotes.

But basically what this says is that
we can be this percent confident

using a given model that one option A or
B is

that percent better than the other.

B.R.T. here is probability of
a relative difference right so

these two things are your preference
it's how confident you want to be and

how much better you want one
option to be than the other.

What's the likelihood of
that difference being

some percentage that's
all the way up to you OK.

I'll have some graphs
of this later that are.

Up to you in this case what we did
better is lower thermal cooling and

heating loads.

We left out actual energy consumption
from the power plant of the gas coming

in just because that's a whole other
uncertainty that we didn't really want to

deal with and plus you're not making those
decisions in early design anyway right

you're mostly worried about Lowe's from
a thermal performance point of view OK OK.

So let's take a look at a case study the
building that's right out there that just

got fixed let's see if
they made a good decision.

So boy it's different now.

So let's take a look at this
we looked at this building

with two different design options and
two different models one

of the options that we chose are the ones
that they actually did OK So

Option A What I said is exactly the way
it's built it's been renovated right now

the other one is a variation on that do
you have the transparent East facade but

you also have some transparency
on the north and south and

instead of using that big
overhang use fans for shading.

Which one do we do right when we're
just making trying to come up with what

the scheme you know the conceptual scheme
of the building is going to be OK.

The two models that we used.

If you've had me in a class
we've done this if you've had

the either of the systems courses for me
we've used this model F. I'm calling than

the normative model which I'm calling
model F. here this is just that

spreadsheet that you got in those classes
is the same one that I've given you.

It's a normative model meant that it's
not really meant to be accurate but

it's meant to judge against something else

it's a very simple model it's built off
an international standard and it's so

simple that you can just
program a spreadsheet for.

The other model is a dynamic energy
simulation model built Fortran with

thousands and thousands of lines of code
called energy plus it's been developed

over you know a couple of decades
if you go back to it's its roots by

the Department of Energy and it is
a dynamic sub hourly simulation model.

Dynamic meaning.

Not a lot of calculus under
the hood this one doesn't OK

this is a very detailed model and
it's actually very appropriate for

the end of stage when you actually
are trying to predict reality and

it has a better representation of
certain things than this model so

let's test it to these are the two
models we're going to look at.

And if we take all those distributions
that we had before those two options and

those two models and we run it all
through our statistical tools and

our energy models this is the output
that we get you can care but

compare between models tween here and
here between cooling and heating here and

here and between options a here and
options be there so

lots of different comparisons if we
look between models are really if we

look between models between here and
here this is a normative model for

cooling energy plus for
cooling well we can see that the two do

not line up option A here is different
than Option A they're going.

Model seems to well I'm not going to say
under predict because I don't know what

the actual reality is but comes in with
cooling that tends to be a little bit less

than the cooling that energy plus does but
if we look at the comparison between

options within a model they're not that
different they're not that different so

for not trying to predict reality exactly
where on this line the distribution is but

only say do we do A or B.
It looks just by eyeballing it

that both models support.

Which is what they did.

Right.

At least when it comes to cooling.

When it comes to heating it's a pretty
similar story option A and Option B.

don't really seem to vary that much
between one or another although

the normative model seems to have a lot
more on certainty coming out of it for

some reason and then the heating model
probably has to do with the fact that

there's probably a lot more
dynamics happening during the day

with heating which would not
get picked up by this model so

there's a model uncertainty that
you can see right there but

they still seem at least by eyeballing
it to support A over B. in the heating.

If you just kind of eyeball.

Right well we don't really
always eyeball things.

We calculate things and

if I tried to calculate this
confidence that I showed before with

those two equations and plotted it's
going to look something like this.

Here's cooling in the blue heating in the
red normative basically what this says is

the X. axis here is how much better you
want one option to be than another right

ten percent better twenty percent
better thirty percent better.

And then this is the kind of
the confidence level that you want to be

somewhere from zero to one hundred
percent knows we cannot be because these

distributions overlap one another none
of these curves can ever go up to one.

Right none of these crew.

Can ever go up to one.

If we compare the two model wise it's
roughly similar seems like cooling for

energy plus gives you a little bit more
confidence about making decision a for

cooling than does the norm of model but

they still both lead you to the same about
the same decision again depending on on

what your confidence level is that
you want to be and this is this is.

Looking at Condell if this current
program is this educational building

officer class right if we change this and
we just change our

program to make it a gallery let's
just see what happens when we do that

great thing about models you can just
change something impressive button and

see what happens if it's a gallery
things actually change we still make

choice a Looks like we would make
choice and in both cases for

both models although Well actually
the heating question here is

kind of kind of interesting but they still
show it what's interesting is that these

distributions spread for
both models right.

The normative model is still showing
cooling loads a little bit lower than

energy plus but it's still showing
a being more better than B.

than it was before when this is an
educational building we go by a calculator

confidence that way it stands out
because we have a larger spread in these

distributions we have more quote
confidence from both models but

both models and this program tend to you
know show pretty decent correspondence

in decision making between the two
both models provide fairly similar

levels of decision support but we haven't
said anything about which one actually

predicts reality but that's not important
right now we're just doing it A or

do we want to do big Right turns
out that this normative model is

something that we push a lot because
it's a lot easier to use and the number

of parameters that it needs to that
you need to have in order to make him.

To make the model run is a lot less It
runs a lot faster it takes a lot less time

to to collect the information for it it's
just you know really nice from a kind

of an operational point of view let
me energy plus takes a long time.

Takes longer to run but it even takes
longer to just build the input model for

it takes an awful lot of time which is why
nobody does it in practice until the very

end when they need to check code
compliance they don't use it for design.

And the other thing that we did if we well
it doesn't help you unless it's something

that's accessible so we've also built a
read it plug and that lets you connect to

all this other infrastructure so you can
build something in rabbit right here and

have a little button right here to
quantify this design uncertainty it opens

up a little box like this we set our
distributions on the parameters we also

tell this thing whether we've
decided on that parameter or not.

It takes all this information and
ships it out to that model and

it comes back with curves like this
comes back with curves like that so.

You know about your choices OK.

The questions about that are going to move
on to one more thing who we are out of

time one more thing really quickly
talked about inverse modeling right and

we use the inverse modeling as a tool to
ultimately be able to work with models in

a forward mode the next question we asked
is What if we get rid of the forward mode

and just do everything in inverse In other
words usually when we're dealing with

forward mode with design we're
looking at we cook up a design and

we say does this thing meet my preferences
or does this thing satisfy my preferences

and even more interesting question is
what designs satisfy my preferences.

That's the inverse design problem and

you can pose it as an inverse modeling
problem so what happens if we do that and.

What we did instead of.

You know doing all of this to get
that what if we could make this

a product use this is
our design support So

what we did is just went
backwards whereas before this.

What was once output now being input was
imperial data about the energy consumption

of buildings what if we construct a curve
like this that expresses our preference

on the ultimate energy performance of
the building then run it backwards and

it'll tell us what designs need this right

the other way of saying it is what are
the distribution of the parameters that I

have control over that
lead to what I want.

And then within that you have freedom to
choose another way you can look at it is

what scenarios in my limited you can
almost think of that as if I want to do

this with these design choices what
climate zones can I put it in Right OK so

I can go through all of these and just.

This is Roy is trying
to work this up as her.

Thesis topic and one of the things that's
kind of hard about this is how do you

test this it's actually really hard to do.

But to just give you some idea of
where it goes is let's say we take

just as a as an example.

A Chicago office building.

And let's say that we're aggressive
about its energy consumption

let's say that we want to
have an eighty percent chance

that it's going to say cooling I
don't know what this is that that

it's cooling load is less than this
number it's kind of hard to read this but

that's actually a pretty aggressive target
you can express that preference with

a distribution that looks like this is
called a triangle distribution it's not

a bad thing to use you know if you don't
have a whole lot of information but

this distribution does express the idea
that we want an eighty percent.

It's that the consumption of this
building is at this value or less so

twenty percent chance it could be over but

we as long as we've you know eighty
percent chance low the Silex press that OK

we take this we run it backwards
through again a baby model and

we end up with distributions like this
let's say in this case study that we have

these different parameters different views
of of the design of the building and

let's say that we've made a decision or
for or for

because of the program all we know is that
the gross floor area needs to be a certain

value that's something we can actually be
pretty certain about is going to actually

come about there's not going to be a whole
lot of spread around that distribution so

we use this we fix this
parameter to be that value and

then we let these other ones be calculated
from the inverse model given this right so

what this is saying is if we want this in
Chicago for office building given this

this is our choices these are the choices
of these parameters that we have and

it tells us which parameters are most
likely by the height of these histograms

to give us what we want right now one of
the big limitations of the last thing that

I told showed you were I was using those
distributions of the building stock as

as distributions of undecided
parameter uncertainty

one thing that that doesn't
include is the conditionality of

those probabilities right if you
choose one parameter then that might

influence the probability distribution
of other parameters right if you have

lots of parameters that's going to be a
real common an authorial headache you know

even if you can't figure out how all
those things link together right

this gets around that problem this well
it was we don't really worry about that

work as a designer we usually do we'll
just make decisions about what we want and

then update.

It's let's say we make choices about
some of these other parameters

if we do that and then run the model again
these distributions chain's again so

if we make choices about the number of
floors the building heights aspect ratio

these distributions change to represent
to reflect those choices we don't have to

worry about that conditionality anymore so
if we flip from here to here it looks like

the building has to become tighter lips
has to become a little bit tighter.

Our set points have to be we have a little
less freedom in what we choose for

our set points and we have just cow
less freedom over on this case if

we look at a different set of options or
excuse me as I switch between this and

yeah if we choose other values for
these Instead we have different

distributions as well and we can keep
marching on as we go and it tells us

what you know how much constrained we are
in the other decisions that we have yet

to make as we go along.

So that's.

Just starting right now.

So I mean with that
that's it I'll take any

questions.

Or.

Yes this.

Is.

All.

A do.

Yeah.

They're going to.

Look if.

You look at what they followed this.

Model.

But look right right.

Right or
I mean if all that new research and

knowledge was right you might suspect you.

Are right or that.

You will get work like
this out of you he said.

Or did not make me write that
it is going to read it with you

is like her over the last.

Year so
maybe that helps you well you know.

So well that is all.

I do enjoy myself probably only.

Right but some are.

All we have last year we graduated
to students of R M S H P B

programme Sandeep and Patrick and
their trainers architects

their architects from the beginning and
then they came and went through R.M.S.

H.P.V. and now they work at Perkins and
will and they're they're

the evangelists at Perkins and will for
maybe I should say evangelists but

well actually yes I should say Vangelis
they their job is getting Perkins and

will to work with this in A is
sufficient a way is that they

can make better decisions and so
that's that's their path they're not so

much designing buildings they're helping
other people design buildings with these

things in mind right and also Yeah.

We romanticize Yeah.

Well all.

Right right really to the end
of it all are there are no.

Laws.

There must be involved here with.

You know.

Yeah.

You know in your other point about
maybe this just helps you think through

a problem to organize information.

Really the only way to
make decisions under

uncertainty if you really want to be.

Rigorous about it there's only one theory
that we know of that lets you do that.

But it's really hard to do right and

if you look through it there's actually
I mean there's a mathematical proof that

says this is what you need to do
this is the only thing that work

under these certain assumptions that
encode being rational The problem is that

it's hard to do right so nobody really
does it it's really time consuming but

when I went through it I became convinced
by it I'm pretty much in the tank for

the method not that I do it but it helps
me think through the problem in a much

clearer way right.

Where is.

Right.

Right.

OK.

OK Well in that case you don't necessarily
when you make a decision about this using

this method you're not restricted to
making that parameter be a straight line

you can shape that
distribution if you want.

Some way yeah there's some leeway in it

what we would want to do with this is
what's called a basic inverse method and

what the and I mean this is kind of
a double edge sword a base in method just

basically says that it
incorporates your prior knowledge.

Incorporates your prior knowledge and
now it's going to try and

modify that prior knowledge
in the face of evidence.

And it may do that correctly or
it may not you know so

it's a double edged sword but
that you know that's.

Basine ism is kind of
a buzz word out there.

You know for for some good reasons but.

We yell a bit like Anyway like I said
you're not restricted to make not

a straight line.

Yes.

Yes.

He.

Is yes.

Thanks thanks or was he saying.

He.

Has his.

Or.

Her Yeah OK.

Right the.

Way I see it was was that.

He.

Was.

More.

Like they.

Were No.

More.

Normal.

Us You're more no.

Less if you're so.

You know.

He was.

More right I say.

While I'll defer to you on
the on the on the deeper.

Issues of that from a mathematical
sparked of it you can hear it's operable

whether or not it's actually.

Capital C. correct yeah that's
that's actually a good point.

It's one step.

Well.

Yeah sure so good.

So smart what not blah blah
blah it's good it's got.

It's good.

Money.

More on.

More.

Be more go.

There all of.

Us.

With.

More go away or.

This.

Is me.

Yeah yeah yeah yeah.

This is.

Right.

So so trade on US trade offs is here
yeah there's lots of different ways to

look at trade offs the you know the
theoretically correct way of dealing with.

A.

Decision that is limited to only
a single person making a decision

about things that are quantifiable right.

It's only it's really only
valid of that circumstance and

there is a way to deal with trade
offs if you're a single person

making a decision about
something that's quantifiable.

It counts and can be counted.

And you're the only person making
a decision of that on the presence of

uncertainty with trade offs if if you're
trying to work with things that either

are not quantifiable along with things
that are quantifiable I'm not sure how

to handle that that's almost kind of
getting into a game theory which is how

people interact making their own decisions
And there's an inverse to that something.

They call it mechanism design if you want
to have a team decision making process

that has a good chance of leading to
a good outcome you design a game for

that to happen but the mathematics
of that are intractable pretty much

I don't even know about that you
might be able to incorporate

things that are not quantifiable
in that but I'm not sure.

It was a well yeah yeah.

Yeah.

OK.

Well.

We were OK.

OK.

So you know you're right.

Well that that would be encoded what
I haven't shown is when this inverse

modeling you actually have to prime it a
little bit you start out with a theme and

find a graphic here if I
understand your question right.

So it sounds like the way I I
realize this now it sounds to what

I said would what I said was that you
start with this and nothing else here and

then cook up something over here that's
not entirely true you start with this but

then you do put some distributions
in these based on prior knowledge

those distributions get worse.

Depending on this one problem
with that though is so

you can put those distributions in
there saying that a wall has to be for

either code reasons or physical reasons
has to have a parameter value within this

range right and then that gets
shaped according to this and

the model right the problem is
that that reshaping may be.

Common physical especially if it's
a Basie technique that you're using

This is one of the downfalls of Basine
techniques the what's called the Post your

probability may not
necessarily be true right so

you do have to have some to start this
you do have to tell it something about

where what you think these things ought to
be at least in the range does that answer

is that the question or answer OK.

All right you know.

You're.

All working.

With.

What's going on.

There so are you are you asking.

Is this only a sequential process or

you know only looking at
two decisions at a time.

Or you have an alternative in my.

Area or
you have a set of alternative right yeah.

You know it is a matter how many that
you have we just used to hear but

you can have more you could have A B.
C. and then you can rank order to us.

Right now we just did two you know
just for Illustrated purposes but

it's not limited to it all.

Right.

With it OK.

The.

Well.

I would say well let's see some
of these parameters though

are you know maybe a little
bit more interesting.

I would say like window to
wall ratio in aspect ratio.

But you kind of bring up something that
that we have kind of heard from people

before when we presented this to different
audiences is one on one kind of question.

Haven't you just made a.

Made a big machine out of design
guidelines was right as I.

Write write.

Do.

You get there.

These cars or if.

There was no one real well yeah.

Well that actually that question is I
think it's actually a pretty good research

question and when we get done here
mental Roy about it maybe the thing that

tells you whether this is a good method or
not to test it is to test it

against guidelines.

Yes.

They're.

Right there.

Here's.

Where.

Yeah.

Yeah.

Yeah.

They're.

Here.

They're here.

I think you're right.

Yeah.

Sure.

Yeah they're.

Really.

Really.

One thing you know what you're.

Right yeah.

It's a Big Ten problem in fact that I mean
that's why I came to architecture with.

The.

Right yeah I mean it's it's
a there's a lot more going on in

putting together a building
than mine narrow take on.

This or yeah it's but it's just one piece
yeah if I think of all of that there

is worth all right but it was an.

Hour.

Earlier than me that.

You're right yeah.

You don't have a whole lot of right your
freedom gets gets constrained right that.

You know.

You're hearing.

A lot of the last thirty.

Years with no.

Good you know you're right leave
it there we all thought of this.

As easy I still.

Think that you know in the longer run

it is circular very very
Yeah you know you're.

You're living in the world we have
with you know we you know we.

Yeah I do it's.

Basic method.

We go out and wish.

We did our issues with the whole lot or
you know.

Yeah.

Yeah yeah it's with all those right sure.

They're right hard to get.

You know early on especially.

If insert a river out.

There.

OK.

Earth is the.

Country where the rock I thought was
it like you know the information

level and ability to change the design
it and there's this crossing Yeah I

think I know what you're talking
about yeah yeah right here.

Yes there.

You.

Are Yeah right.

Yeah and it's too late to go back yeah.

We get to the Great Lakes.