Title: Experimental Design, Response Surface Analysis, and Optimization
1Experimental Design, Response Surface Analysis,
and Optimization
2Outline
- Motivation and Terminology
- Difficulties in Solving the Basic Problem
- Examples of Factors and Responses
- Types/Examples of Experimental Design
- Full Factorial Designs
- Randomness of Effects
- Example Full Factorial Design
- Situations with Many Factors
- Response Surfaces and Metamodels
- Regression Analysis
- Response Surface Methodology
3Motivation and Terminology
- Useful when there are many alternatives to
consider (e.g., numerous capacity levels of
various types, numerous parameters for a proposed
inventory system) - Two basic types of variables factors and
responses - Factors input parameters
- -controllable or uncontrollable
- -quantitative or qualitative
- Responses outputs from the simulation model
- -uncertain in nature
- Basic problem find the best levels (or values of
the parameters) in terms of the responses - Experimental Design can tell you which
alternatives to simulate so that you obtain the
desired information with the least amount of
simulation
4Difficulties in Solving the Basic Problem
- Multiple Responses
- Uncertain Responses
5Examples of Factors
- 1. Mean interarrival time (uncontrollable,
quantitative) - 2. Mean service time (controllable or
uncontrollable, quantitative) - 3. Number of servers (controllable,
quantitative) - 4. Queuing discipline (controllable,
qualitative) - 5. Reorder point (controllable, quantitative)
- 6. Mean interdemand time (uncontrollable,
quantitative) - 7. Distribution of interdemand time
(uncontrollable, qualitative)
6Examples of Responses
- 1. Mean daily production rate
- 2. Mean time in the system for patients
- 3. Mean inventory level
- 4. Number of customers who wait for more than 5
minutes
7Types/Examples of Experimental Designs
- Completely Randomized Designs
- Randomized Complete Block Designs
- Nested Factorial Designs
- Split Plot Type Designs
- Latin Square Type Designs
- Full Factorial Designs
- Fractional Factorial Designs
82k Factorial Designs
- Suppose that we have k (k gt 2) factors. A 2k
factorial design would require that two levels be
chosen for each factor, and that n simulation
runs (replications) be made at each of the 2k
possible factor-level combinations (design
points). For 3 factors, this yields a Design
Matrix -
92k Factorial Designs
- Design Matrix for 3 Factors
-
- Factor 1 Factor 2 Factor 3
- Design Point Level Level
Level Response - 1 - - - O1
- 2 - - O2
- 3 - - O3
- 4 - O4
- 5 - - O5
- 6 - O6
- 7 - O7
- 8 O8
- refers to one level of a factor and -
refers to the other level. Normally, for
quantitative factors, the smallest and largest
levels for each factor are chosen.
102k Factorial Designs Estimating Main Effects
- The main effect of factor 1 is the change in the
response variable as a result of the change in
the level of the factor, averaged over all levels
of all of the other factors - If the effect of some factor depends on the level
of another factor, these factors are said to
interact. - The degree of interaction (two-factor interaction
effect) between two factors i and j is defined as
half the difference between the average effect of
factor i when factor j is at its level ( and
all factors other than i and j are held constant)
and the average effect of i when j is at its -
level for example,
11Example Full Factorial (2k) Design
- Consider a simulation model of reorder point,
reorder quantity inventory system. The two
decision variables, or factors, to consider are
the reorder point (P) and the reorder quantity
(Q) for the inventory system. The maximum and
minimum allowable values for each are given
below - Suppose that the response variable output by
the model is the long-run average monthly cost
(composed of three components holding cost,
shortage costs, and ordering costs) in thousands
of dollars.
12Example Full Factorial (2k) Design
- A 22 factorial design matrix with simulation
results (for 10 replications at each design
point) might be given by - where a factor level of - indicates the
minimum possible value for that factor, and a
factor level of indicates the maximum
possible level for example, design point 2 has
P40 and Q15.
13Example Full Factorial (2k) Design
- The response given is the average cost over the
10 replications. Now, the main effects are given
by - The interaction effect (ePQ) is given by
- Therefore, the average effect of increasing P
from 20 to 40 is to increase monthly cost by 1.2,
and the average effect of increasing Q from 15 to
50 is to decrease monthly cost by 5.3. Hence, it
would be advisable to set P as low as possible
and set Q as high as possible.
14Example Full Factorial (2k) Design
- Also since the interaction effect is positive, it
would seem advisable to set P and Q at opposite
levels. (Of course, all of the above could be
inferred from a cursory analysis of the responses
for the various design points.) Note also that
the literal interpretation of main effects
assumes no interaction effects (pages 669 and 670
of Law and Kelton, 1991).
15Randomness of Effects
- Note that the main and interaction effects
computed in the previous examples are just random
variables. To determine if the effects are
significant or real, and not due to random
fluctuations, one could compute values for the
main and interaction effects 10 times (once for
each replication) and form confidence intervals
for each of the main effects, and the
interactions effect. If the confidence interval
contains 0, then the effect is not statistically
significant. (Note that statistical significance
does not necessarily imply practical
significance).
16Situations with Many Factors
- When there are many factors to consider, full
factorial, or even fractional factorial designs
may not be feasible from a computational
standpoint. - Other types of design (e.g., Plackett-Burman
designs or supersaturated designs) may be
appropriate (Mauro, 1986). - Another approach is to reduce the number of
factors to consider via factor screening
techniques, involving, for example, group
screening in which a whole group is treated as a
factor.
17Response Surfaces and Metamodels
- A response surface is a graph of a response
variable as a function of the various factors. - A metamodel (literally, model of the simulation
model), is an algebraic representation of the
simulation model, with the factors as independent
variables and a response as the dependent
variable. It represents an approximation of the
response surface. - The typical metamodel used in a simulation
application is a regression model. -
18Response Surfaces and Metamodels
- A metamodel through the use of response surface
methodology can be used to find optimal values
for a set of factors. It can also be used to
answer what if questions. (Experimentation
with a metamodel is typically much less expensive
than using a simulation model directly). - An experimental design process assumes a
particular metamodel, e.g.,
19Basic Concepts of Regression Analysis
- Regression is used to determine the best
functional relation among variables. - Suppose that the functional relationship is
represented as - E(Y) f (X1, ..., Xp / B1, ..., BE)
- where E(Y) is the expected value of the response
variable Y the X1, ..., Xp are factors
and the B1, ..., BE are function
parameters e.g., - E(Y) B1 B2 X1 B3 X2 B4 X1 X2
20Basic Concepts of Regression Analysis
- The observed value for Y, for a given set of X
s, is assumed to be a random variable, given by - Y f (X1, ..., Xp/B1, ..., BE)
- Where is a random variable with mean equal
to 0 and variance . The values for
B1,...,BE are obtained by minimizing the sum of
squares of the deviations.
21Response Surface Methodology
- Source (Fu, 1994)
- Response surface methodology (RSM) involves a
combination of metamodeling (i.e., regression)
and sequential procedures (iterative
optimization). -
22Response Surface Methodology
- RSM involves two phases
- Fit a linear regression model to some initial
data points in the search space (through
replications of the simulation model). Estimate
a steepest descent direction from the linear
regressions model, and a step size to find a new
(and better) solution in the search space.
Repeat this process until the linear regression
model becomes inadequate (indicated by when the
slope of the linear response surface is
approximately 0 i.e., when the interaction
effects become larger than than the main
effects). - Fit a nonlinear quadratic regression equation to
this new area of the search space. Then find the
optimum of this equation.
23Terminology in Experimental Design
- Source(Ostle, 1963)
- Replication - the repetition of the basic
experiment - Treatment - a specific combination of several
factor levels - Experimental Unit - the unit to which a single
treatment is applied to one replication of the
basic experiment - Experimental Error - the failure of two
identically treated experimental units to yield
identical results - Confounding - the mixing up of two or more
factors so that its impossible to separate the
effects
24Terminology in Experimental Design
- Randomization - randomly assigning treatments to
experimental units (assures independent
distribution of errors) - Main Effect (of a factor) - a measure of the
change in a response variable to changes in the
level of the factor averaged over all levels of
all the other factors - Interaction is an additional effect (on the
response) due to the combined influence of two or
more factors