Choose Language Hide Translation Bar

Design of Experiment’s Crucial Step 0: Choosing the Right DOE Option - (2023-US-30MP-1432)

JMP has a wealth of design of experiments (DOE) options from which to choose. While this array is incredibly powerful, it also has the potential to be a bit intimidating to those who are new to this area. What category of design should I choose from the many possibilities? How do I know what the best one is for my experimental objectives? This talk provides some ideas for how to strategically tackle Step 0 of the process of constructing the right design by considering the following questions:

  1. What are the goals of the experiment?
  2. What do we already know about the factors, responses, and their relationship?
  3. What are the constraints under which we need to operate?

 

Once these questions are answered, we can match our priorities with one of the many excellent choices available in the JMP DOE platform.

 

 

Hi.  I'm  here  to  talk  today  about the  crucial  Step  0

of  Design  of  Experiments.

Really  the  idea  of  this  is  to  take full  advantage

of  the  wealth  of  different  tools that  are  under  the  DOE  platform  in  JMP.

I'll  walk  through   what  we should  be  thinking  about

in  those  early  stages  of  an  experiment.

If  you  look  at  the  DOE  platform  listing,

what  you'll  see  is  that  there's a  lot  of  different  choices.

Within  each  choice, there's  many  more  choices.

Within  some  of  those, there's  nested  possibilities.

If  you're   an  expert in  design  of  experiments,

this  wealth  of  possibilities  really  feels like  such  a  wonderful  set  of  tools.

I  love  all  of  the  options that  are  available  in  JMP  that  allow  me

to  create  the  design  that  I  really want  for  a  particular  experiment.

But  if  you're  just  getting  started, then  I  think  this  set  of  possibilities

can  feel  a  little  bit  intimidating and  sometimes  a  bit  overwhelming.

It  may  be  a  little  bit  like  going

to   a  new  kind  of  restaurant that  you've  never  been  to  before.

Someone  who's  a  seasoned  visitor to  those  kinds  of  restaurants,

loves  all  the  possibilities and  the  wealth  of  options  on  a  big  menu.

But  if  you're  there  for  the  first  time,

it  would  be   nice  if  someone  guided you  to  the  right  set  choices

so  that  you  could  make  a  good  decision

for  that  first  visit and  have  it  be  successful.

Here's  what  I'm  planning on  talking  about  today.

First,  I  think  the  key  to  a  good experimental  outcome  is  to  really  have

a  clear  sense  of  what  the  goal of  the  experiment  is.

I'll  talk  through  some  different possibilities  of  common  goals

for  experiments  that  really  help  us  hone in  on  what  we're  trying  to  accomplish

and  what  will  indicate a  success  for  that  experiment.

Then  I'll  do  a  quick  walk- through  of  some

of  the  more  common  choices  of  design of  experiment  choices  in  JMP,

and  then  I'll  return  to  how  do  we  interact with  those  dialog  boxes  that  we  get

when  we've  chosen   a  design for  what  factors  to  choose,  the  responses,

and  the  relationship  between the  inputs  and  the  outputs.

That's   where  we're  headed through  all  of  this,

and  I  will  say  that  the  first and  the  third  steps  really  need

a  tremendous  amount of  subject  matter  expertise.

If  you're  going  to  be  successful  designing an  experiment,

you  really  need  to  know as  much  as  possible

about  the  framework  under which  you're  doing  that  design.

We  want  to in  fact  incorporate subject  matter  expertise  wherever  possible

to  make  sure  that  we're  in  fact  setting  up the  experiment  to  the  best  of  our  ability.

What  are  we  trying  to  do?

I've  listed  here  six  common experimental  objectives.

I  think  that  sort  of  gives  you,

a  checklist  if  you  like  of  different options  of  things

that  you  might  be  thinking of  accomplishing  with  your  experiment.

We  might  start  with  a  pilot  study,

so  we're  just  interested  in  making  sure

that  we're  going  to  get  data of  sufficient  quality

for  the  experiments  and  answering the  questions  that  we  want  to  have.

We  might  be  interested in   exploration  or  screening.

We  have  a  long  list  of  factors,

and  we  want to  figure  out  which  one  seem

to  make  a  difference for  our  responses  of  interest,

and  which  ones don't  seem  particularly  important.

We  also  might  want  to  do  some  modeling.

Actually  formalizing that  relationship  that  we're  seeing

between  inputs  and  responses, and  capturing  it  in  a  functional  form.

Sometimes  we  don't  get  the  level

of  precision  that  we  need, and  so  we  need  to  do  model  refinement,

and  so  that  might  be  a  second  experiment.

Then  once  we  have  a  model, we  want  to  use  that  to  actually  optimize.

How  do  we  get  our  system  to  perform  to the  best  of  its  capability  for  our  needs.

Then  lastly,   there's  a  confirmation  experiment

where  we  make  that  transition

from  the  controlled  design of  experiments  environment

that  we're  often  doing our  preliminary  data  collection  in

to  production, and  making  sure  we  can  translate

what  we've  seen  in  that first  experiment into  a  production  setting.

You  can  see  from  this  progression that  I've  outlined  here,

that  we  may  actually  have  a  series  of small  experiments  that  we  want to  connect.

We  may  start  off  with  a  pilot  study  to  get the  data  quality  right,

then  we'll  figure  out  which  factors  are important,  then  we'll  want to  model  those,

then  we'll  want to  use  that  model to  optimize,

and  then  lastly  translate  those  results

into  the  final  implementation in  production.

We  can  think  of  this  sequentially

or  for  an  individual  experiment  just tackling  one  of  these  objectives.

Now  we  have  some  framework  for  what

the  goals  of  the  experiment  are and  how  to  think  about  that,

we'll  now  transition  to  looking

at  what  some  of  the  common  choices are  in  JMP

and  how  they  connect  with  different  goals.

I'll  open  up  the  DOE  tab  in JMP, and  you  can  see  that  we've  got

the  list  of  possibilities  here where  we've  got  the  nested  options

tucked  underneath  some  of  the  main menu  items  that  we  have  here.

The  talk  is  only  half  hour,  and  so  I won't  be  able  to  cover  all  of  the  tabs.

I've  given  a  brief  description  of  some

of  the  tabs  that  I  won't have  time  to  talk  about.

Design  Diagnostics  is  all  about  having a  design  or  maybe  several  designs

and  comparing  and  understanding the  performance.

Sample  Size  Explorer  is  all  about  how  big

should  the  experiment  be and  some  tools  to  evaluate  that.

Consumer  Studies  and  Reliability  Designs are  really  kind  of  specialized  ones.

I'm  setting  those  aside

for  you  to  do  a  little  research on  your  own  about  that.

In  Consumer  Studies, we're  usually  asking  questions

of  consumers,  about  what  their priorities  are,  what  features  they  like.

That  tends  to  be a  comparison  between  two  options

and  how  they  value  those  choices.

Reliability  is  all  about   how  long our  product  will  last.

That's  a  little  bit  different

than  things  that  I'll  talk  about in  the  rest  of  the  talk.

I'll  start  off  with  some of  the C lassical  Designs

or   the  general  designs  that  we  have that  have  been  developed.

Then  I'll  finish  with  some of  the  JMP  specific  tools

that  are  much  more  flexible  and  adaptable to  a  broader  range  of  situations.

I'll  start  with   that  bottom  portion of  the  tab.

Here  we  are  in  JMP  in  the  DOE  tab, and  I'm  going to  start  with  Classical.

You'll  see  that  I'm  tackling  this

in  a  little  bit  different  order than  the  list  is  presented  by  JMP.

I  think  those  ones  are  presented  by  JMP

in   their  order  of  popularity, and  I'm  choosing  to   tackle  them

more  from   principles about  how  they  were  developed.

In  Classical  Designs, a  Full  Factorial  design

is  looking  at  all  combinations of  all  factors  at  all  levels.

That  works  nicely

if  we  have  a  small-ish  number  of  factors,

but  it  can  in  fact  get  a  little  bit out  of  control

if  we  have  a  large  number  of  factors, but  it's  exploring

the  entire  set  of  possibilities very  extensively.

The  next  one  that  I'll  talk  about is  a  Two-Level Screening  design,

and  essentially,  what  that's  doing is  it's  choosing

a  subset  of  the  two factorial  possibilities,

and  it's  a  strategic  subset

that  allows  us  to  explore  the  space, but  keep  the  design  size  more  manageable.

You'll  notice  that  those first  two  possibilities

I've  shown  at  two  levels, and  that's  typical  for  screening  designs.

Usually,  we  just  want to  get a  simple  picture

of  what's  happening  between the  input  and  the  responses.

When  we  want to  start  modeling,

then  a  Response  Surface  Design  typically allows  for  exploring  curvature.

When  we're  modeling,  three levels or  sometimes  more  than  three  levels

can  be  a  good  way  to  understand  curvature and  also  understand  interactions  between

the  factors  and  how they  impact  the  response.

Alright. T hat's  three  of  the  items under  the  Classic  tab.

The  other  ones  are   Mixture  Design.

Typically  in  all  the  other  possibilities,

what  we  have  is  that  we  can  vary

the  individual  factors  separately from  each  other.

But  in  a  Mixture  Design  where  we're talking  about  the  composition

or  the  proportion  of  the  ingredients, they're  interdependent.

If  I  increase  the  amount of  one  ingredient,

it  probably  reduces  the  proportion

of  the  other  ingredients that  are  in  that  overall  mixture.

A bit  of  a  specialized  one

when  we're  looking at  putting  together  ingredients

into   an  overall  mixture.

Taguchi Arrays,  I've  listed  here as  a  kind  of  optimization,

and  the  optimization that  they're  interested  in

is  making  our  process  robust.

Typically  when  we're  in  a  production environment,  we  might  have  noise  factors.

These  are  in  fact, factors  that  we  can  control

in  our  experiment

but  when  we  get  to  production, we're  not  able  to  control  them.

Then  we  have  a  set  of  factors

that  we  can  control  both  in  the  experiment and  in  production.

The  goal  of  Taguchi Arrays  is to  look  for  a  combination

of  the  controllable  factors that  gets  us  nice  stable  predictable

performance  across  the  range of  the  noise  factors.

You  can  see  C1  here has  a  pretty  horizontal  line

which  means  it  doesn't  matter which  level  we  are  at

for  the  noise  factor, we'll  get  a  pretty  consistent  response.

Those  are  the  classical  options.

The  next  of  the  items on  this  JMP  design  tab

that  I'll  talk  about are  Definitive  Screening  Designs.

These  are  specialized  designs that  were  developed  at  JMP,

and  they   are  a  blend of  an  exploration  or  screening  design,

so  a  focus  on  a  lot of  two- level  factor  levels,

and  modeling.

You  can  see  with  the  blue  dots,

we  have  some  third  levels, so  a  middle  value  for  the  factors

that  allows  us  to  get  some curvature  estimated  as  well.

It's  a  nice  compact  design

that's  primarily  about  exploration and  screening,

but  it  does  give  us  the  option

for  an  all  in  one  chance  to  do some  modeling  as  well.

That's  very  popular  in  a  lot of  different  design  scenarios.

The  next  tab  is  Special  Purpose, and  you  can  see  there's  quite  a  long  list

of  possibilities  there,

and  I'll  hit  some  of  the  more  popular  ones

that  I  think  show  up in  a  lot  of  specialized  situations.

A Covering  Array  is  often  used  when we're  trying  to  do  testing  of  software.

A  lot  of  times  what  causes  problems

in  software  is  when  we  have the  combinations  of  factors.

This  is  a  pretty  small  design

that's  typical  for  Covering  Arrays, so  13  runs,

and  we're  trying  to  understand things  about  10  different  factors.

What's  nice  about  these  Covering Arrays  is  that  it  gives  us  a  way

to  see  all  possibilities of,  in  this  case,  three  different  factors.

If  I  take  two  levels  of  each  factor, a  zero  and  a  one,

there's  eight  different  combinations for  how  I  can  combine  those  three  factors.

All  zeros,  all  ones, and  then  a  mixture  of  zeros  and  ones.

I've  highlighted  those with eight  different  underlined,

what's  really  nice  about these  Covering  Arrays

is  whichever  three  factors  I  choose,

I  will  be  able  to  find  all  eight of  those  combinations.

There's  10  choose  three  different combinations  of  those  3  factors

that  I  might  be  interested  in,

and  all  of  them  have  all  of  those possibilities  represented.

That's   a  very  small  design

that  allows  us  not  so  much  estimation, but  to  check  possibilities  for  problems

that  we  might  encounter particularly  in  software.

Next,  a  very  important  category of  Space  Filling  Designs.

Compared  to  the  other options that  I've  talked  about,

which  are  model- based,

this  one  just  says,  I  maybe  don't  know what  to  expect  in  my  input  space.

Let  me  give  even  coverage  throughout

the  space  that  I've  declared and  just  see  what  happens.

You  can  see  that  I  have  many more  levels  of  each  of  the  factors.

There's  a  lot  of  specialized  choices in  here,  but  they  all  have  this  same  feel

of  nice,  even  coverage throughout  the  inputs  face.

I  think  these  are  often  used

in  computer  experiments or  in  physical  experiments

where  we're  just  not  sure  what the  response  will  look  like.

I'll  talk  a  little  bit  more  about  that

when  we  get  to  the  decision  making  portion in  Step  3  of  the  talk.

Next  is   MSA  Design or  a  Measurement S ystem  Analysis,

and  this  typically is  associated  with  the  Pilot  Study.

Before  I  dive  in  and  really  start

to  model  things  or  do  some  screening, it's  helpful  to  understand  some  basics

about  the  process  and  the  quality of  the  data  that  I'm  getting.

Here,  I  can  sort  of  divide the  variability  that  I'm  seeing

in  the  responses and  attribute  it  to  the  operator,

the  measurement  device,  or  the  gage, and  the  parts  themselves.

Sort  of  understand  the  breakdown of  what's  contributing  to  what  I'm  seeing.

That's  very  helpful  before launch  into  a  more  detailed  study.

Finally, G roup  Orthogonal  Super saturated  Designs

are  in  fact,  really  compact  designs.

In  this  example,  where  you  have  six  runs,

and  we're  trying  to  understand  what's happening  with  seven  different  factors.

That  may  seem  a  little  bit  magical,

but  it's  a  very  aggressive  screening  tool that  allows  us  to  understand

what's  happening  with  a  lot  of  factors in  a  very  small  experiment.

It's  important  with  these  designs to  not  have  a  lot  of  factors.

If  all  seven  factors  are  doing  something,

and  I  only  have  six  runs, I'll  end  up  quite  confused  at  the  end.

But  if  I  think  two  or  three  of  them may  be  active,

this  may  be  a  very  efficient  way

to  explore  what's  going  on  without spending  too  many  resources.

Those  are   the  start  here  ones that  I've  talked  through  a  little  bit.

Now  I'm  going to  finish with   these  wonderful  tools  in  JMP

that  are  more  general  and  more  flexible for  different  scenarios.

Custom  Design,  I  think  is  just an  amazing  tool  for  its  flexibility.

What's  really  nice  in  Custom  Design

is  that  I  have  this  wealth of  different  possibilities

for  the  kinds  of  factors that  I  can  include.

Continuous  factors,  maybe  I'll  add  in,

Discrete  Numeric  ones, and  then  also  Categorical  Factors.

I  have  a  lot  of  different  choices so  I  can  put  together  the  pieces,

and  if  I'm  not  sure  what  the  design should  look  like

in  that  bottom  portion  of  the  list,

this   gives  JMP  some  control to  help  guide  me  to  a  good choice.

On  the  next  page,  I  have  the  option  about

whether  I'm  just  interested in  Main E ffects,

whether  I  want to  add some  two  factor  interactions,

and  whether  I  want  to  build a  Response S urface  Model,

so  more  the  modeling goal  of  the  experiment.

This  is  sort  of  an  easy  way to  build  a  design,

and  I  have  flexibility  here  to  specify whatever  design  size

I feel  would  be  helpful  and  is  within my  budget  to  make  a  design

and  the  expertise  of  the  JMP  design  team

are  going to  guide  me to  a  sensible  choice.

This  is  a  great  way  if  you're  not  sure about  how  to  proceed,

but  you're  still  making  some  key  decisions

about  what  the  goal  of  the  experiment should  look  like.

Next,  the  Augment  tab.

If  you  think  back  to  what  I've  talked about  for  the  Experimental  Objectives,

you  see  that  there's  this  connection between  the  stages.

Maybe  I've  done  some  exploring or  screening,

and  then  I'd  like to  transition  to  modeling.

Well,  this  allows  me  to  take  an  experiment

that  I've  already  run and  collected  data  for,

and  then  connect  it  to  the  Augment D esign, assign  the  roles  of  what's  a  response

and  what's  the  factor, and  then  add  in  some  additional  runs.

There's  some  specialized  ones  here,

but  if  I  choose  the  Augment  portion,

that  allows  me  to  specify a  new  set  of  factors,

perhaps  a  subset  of  what  I  have or  an  additional  factor

and  then  also  what  model I  would  now  like  to  design  for.

This  is  a  flexible  tool  for connecting  several  sets  of  data  together.

Lastly,  Easy  DOE  is  a  great  way  to  get started  for  your  very  first  experiment.

It  allows  you  to  build  sequentially

and  it  guides  you  through the  seven  different  steps

of   the  entire  experiment.

It'll  allow  us  to  design  and  define,

and  so  that's  figuring  out what  the  factors  are,

what  the  levels  are,

their  general  nature, then  we  can  select  what  kind  of  model

makes  the  most  sense  for  what  we're  trying to  accomplish,

then  progress  all  the  way  to  actually running  the  experiment,  entering  the  data,

doing  the  analysis and  then  generating  results.

This  is  a  wonderful  progression that  walks  you  all  the  way  through

what  am  I  trying  to  do?

To  having  some  final  results to  be  able  to  look  at.

What  I  will  say  is  that  this  is designed  for  a  model- based  approach.

What  you'll  see  is  that  all  of  these

look  like  they're  going  to  choose a  polynomial  form  of  the  model.

That  needs  to  make  sense as  a  starting  point.

But  if  that  does  make  sense and  it  does  in  a  lot  of  situations,

then  this  is  a  wonderful  option.

Just  to  finish  things  up  here, what  are  some  of  the  other  key  questions

now  that  I  have  a  goal  I  know  a  particular choice  that  I  want  to  use  in  JMP,

what  are  some  of  the  other  key  questions before  I  actually  generate  that  design?

A whole  category  is  about  the  factors.

We  need  to  use our  subject  matter  expertise

to  figure  out  which  factors we  should  be  looking  at.

If  we  have  too  long  of  a  laundry  list of  factors,

then  the  experiment  necessarily  needs

to  be  quite  large  in  order to  understand  all  of  them.

That's  going to  have  an  impact  on  how expensive  our  experiment  will  be.

If  we  have  too  few  factors,

then  we  run  the  possibility of  missing  something  important.

What  type  are  they  going  to  be?

We  need  to  think  about getting  the  right  subset.

As  I  showed  you  in  C ustom  Design, we  have  quite  a  wide  variety  of  different

types  of  roles  for  the  different factors  that  we're  looking  at.

That's  another  set  of  choices.

How  much  can  we  manipulate  the  factors?

Are  they  naturally  categorical, or  are  they  continuous?

Then  we  need  to  think  about  the  ranges or  the  values  for  each  of  those.

Let's  go  to  DOE  and  Custom  Design.

Then I'll  just  start  off

and  I'll  have  three different  continuous  factors.

What  you  can  see  is  I  can  give  a  name

to  each  of  the  factors, but  I  also  get  to  declare  the  range

that  I  want  to  experiment  in for  each  of  those  factors.

A s  you  can  imagine,

this  has  a  critical  role  in  the  space that  I'm  actually  going  to  explore.

I  need  to  hone  in  on  what's  possible

and  what  I'm  interested  in to  get  those  ranges  right.

If  I  make  the  range  too  big,

then  I  may  actually  have  a  lot  going  on across  the  range  of  the  input

and  I   may  not  be  able  to  fully capture  what's  going  on.

If  I  make  the  range  too  small, then  I  may  miss  the  target  location

and  I  may  get  a  distorted  view of  the  importance  of  that  factor.

Here,  this  input  actually  has a  lot  going  on  for  that  response,

but  if  I  sample  in  a  very  narrow  range, it  looks  like  it's  not  doing  anything.

Lastly,  if  I'm  in  the  wrong  location, I  may  miss  some  features

and  not  be  able  to  optimize  the  process for  what  I'm  doing.

Again,  the  choice  of  which  factors and  the  ranges,

relies  a  lot  on  having some  fundamental  understanding

about  what  we're  trying to  do  and  where  we  need  to  explore.

The  next  piece  to  talk  about  is

the  relationship  between inputs  and  responses.

I  will  say  that  one  of  the  common  mistakes

that  I  often  see  is  that  we  run an  experiment,

and  then  after  the  fact,  people  realize, oh,  we  should  have  collected  this.

In  textbooks,  a  lot  of  times,

it  looks  like  there's  a  single  response that  we're  interested  in

and  we  run  the  experiment to  just  collect  for  that  response.

In  practice,  I  think  most  experiments have  multiple  responses

and  so  this  is  a  key  decision, is  to  make  sure

before  we  collect  that  first  data  point

that  we  actually  include the  right  set  of  responses

so  that  we  can  answer  all of  the  questions  from  that  one  experiment.

Then  we  need  to  think  about  what we  know  about  the  relationship.

Is  it  likely  to  be  smooth?

Is  it  going to  be  continuous in  the  range  that  we've  selected?

How  complicated  are we  expecting  it  to  be?

A ll  of  these  have  an  impact on  the  design  that  we're  going  to  have.

A  couple  of  common  mistakes about  the  relationship  is,

one,  being  a  little  too  confident,

so  we  assume  that  we  know  too much  about  what's  going to  happen,

and  then  don't  build  in  some protection  against   surprises.

Then  also  if  we  have  multiple  responses,

not  designing  for  the  most complicated  relationship.

If  one  of  them  were  interested in  Main E ffects

and  the  other  one  we  think there  might  be  curvature,

we  need  to  build  the  design so  that  it  can  estimate  the  curvature

because  that's the  more  complicated  relationship.

A  first  key  decision  that  I  think

is  a  little  bit  hidden  in  JMP  is  that  we have  to  decide  between  model- based,

and  that's  usually  sensible if  we're  confident

that  our  responses  will  be  smooth and  continuous,

and  that  we're  not  investigating too  big  of  a  region,

or  should  we  do  space  filling?

Space  filling  can  be  a  good  safety  net if  we're  not  sure  what  to  expect,

if  we're  exploring  a  large  region,

or  if  we  want to  protect against  surprises.

I'm  pointing  here  on  the  last  slide.

I  have  more  details  about  that  to  a  paper

that  I  wrote  with  a  colleague,  Dr. Lu Lu

at  the  University  of  South  Florida,

where  we  talk  about   the  implications

of  that  first  fork  in  the  road, how  do  we  choose  between  model- based

and  space  filling, and  what  are  the  repercussions?

Then  lastly,  we  need  to  think a  little  bit  about  constraints.

Our  input  region,

if  we've  declared  some  ranges for  the  different  inputs,

that   naturally  seems  like it's  a  square  or  a  rectangle.

But  in  that  region,

there  may  be  some  portions  where  we  can't

get  a  response   or  we  just  don't  care about  what  the  responses  look  like.

Imagine  if  I  am  doing  an  experiment  about baking  and  I'm  varying  the  time

that  the  cookies  are  in  the  oven and  the  temperature  of  the  oven.

I  might  know  that  the  coolest  temperature

for  the  shortest  amount  of  time won't  produce  a  baked  cookie.

It'll  still  be  raw, or  it  might  be  the  hottest  temperature

for  the  longest  time  will overcook  the  cookies.

I  want  to  sort  of  chop  off  regions

of  that  space  that  aren't  of  interest or  won't  give  me  a  reasonable  thing.

In  JMP,  there's  easy  ways to  specify  constraints

to   make  the  shape of  that  region  match  what  you  want.

The  last  thing  is  all  about  budget,

how  big  should  my  experiment  be, and  that's  a  function  of  the  time

that  I  have  available and  the  cost  of  the  experiment.

In  JMP,  we  jump  to  here.

Maybe  I  specify  a  response  surface  model,

you'll  see  that  there's  a  new  feature called  Design  Explorer,

which  when  I  activate  that,

it  allows  me  with  a  single  click of  a  button  to  generate  multiple  designs.

I  can  optimize  for  good  estimation, so  D  or  A-O ptimality,

or  good  prediction of  the  responses  with  I-O ptimality.

I  can  vary  the  size  of  the  experiment and  center  points  and  replicates.

If  I  click  Generate  All  Designs, it  will  generate  a  dozen  or  so  designs,

which  then  I  can  compare  and  consider

and  figure  out  which  one makes  the  most  sense.

I  think  understanding  the  budget, thinking  of  that  as  a  constraint,

is  an  important  consideration that  we  need  to  have.

To  wrap  things  up, just  a  few  helpful  resources.

The  first  one  is   a  JMP  web  page

that  talks  in  a  little  more  detail about  the  different  kinds  of  designs.

It  fills  in  a  lot  of  the  details that  I  wasn't  able  to  talk  about  today

about  those  individual choices  on  the  DOE  tab.

The   Model-B ased  versus  Space-F illing , that's  the  paper  I  referenced  earlier,

where  we  need  to  understand

the  implications  of  choosing a  model- based  design

or  doing  space- filling, which  is  a  little  more  general

and  a  little  more protective if  we  are  expecting  some  surprises.

Then  the  last  two  things  are, two White  Papers  that  I  wrote,

the  first  one  talks  about  how you  can  use  Design  Explorer

to   consider  different  design  sizes

and  different  optimality  criteria and  then  choose  between

the  different  choices  by  looking at  the  compare  design  option  in  JMP.

Then  lastly, everything  I've  talked  about  here

is  dependent  on  subject  matter  expertise.

The  why  and  how  of  asking  good  questions,

give  some  strategies  for  how  to  interact with  our  subject  matter  experts

to  be  able  to  target  those  conversations and  make  them  as  productive  as  possible.

I  hope  this  has  been  helpful,

and will  help  you  have  a  successful  first experiment  using  JMP  software.

Thanks.