| Title: | Wrapper Package for Design of Experiments Functionality |
|---|---|
| Description: | Various kinds of designs for (industrial) experiments can be created. The package uses, and sometimes enhances, design generation routines from other packages. So far, response surface designs from package 'rsm', Latin hypercube samples from packages 'lhs' and 'DiceDesign', and D-optimal designs from package 'AlgDesign' have been implemented. |
| Authors: | Ulrike Groemping [aut, cre], Lenth Russ [ctb] |
| Maintainer: | Ulrike Groemping <[email protected]> |
| License: | GPL (>= 2) |
| Version: | 0.12 |
| Built: | 2024-11-13 06:36:40 UTC |
| Source: | CRAN |
This package creates various kinds of designs for (industrial) experiments. It uses, and sometimes enhances, design generation routines from other packages. So far, response surface designs from package rsm, designs for computer experiments (latin hypercube samples etc.) from packages lhs and DiceDesign and D-optimal designs from package AlgDesign have been implemented.
Currently, the package provides classical response surface designs from package
rsm: Box-Behnken designs by Box and Behnken (function bbd.design
and central composite designs by Box and Wilson (ccd.design) are implemented.
For the latter, there is also a function for augmenting 2-level fractional factorials
into central composite designs (ccd.design).
Furthermore, latin hypercube samples and other designs for computer experiments from packages lhs and DiceDesign are provided.
Furthermore, D-optimal designs have been implemented, using package AlgDesign. This implementation is currently in beta shape.
All designs created by this package are class design objects,
which are data frames with attributes.
Apart from providing designs, the package also provides functions for comparing
the quality of several designs with quantitative variables (function compare).
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Box, G.E.P. and Behnken, D.W. (1960). Some new three-level designs for the study of quantitative variables. Technometrics 2, 455-475.
Box, G.E.P., Hunter, J.S. and Hunter, W.G. (2005, 2nd ed.). Statistics for Experimenters. Wiley, New York.
Box, G.E.P. and Wilson, K.B. (1951). On the Experimental Attainment of Optimum Conditions. J. Royal Statistical Society, B13, 1-45.
NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/pri/section3/pri3361.htm, accessed August 20th, 2009.
Myers, R.H., Montgomery, D.C. and Anderson-Cook, C.M. (2009). Response Surface Methodology. Process and Product Optimization Using Designed Experiments. Wiley, New York.
See also ccd.design, ccd.augment, bbd.design, lhs.design,
Dopt.design, Dopt.augment,
lhs-package, rsm, optFederov,
optBlock, compare, FrF2
Function for generating Box-Behnken designs, making use of package rsm
bbd.design(nfactors, ncenter=4, factor.names = NULL, default.levels=c(-1,1), block.name=NULL, randomize=TRUE, seed=NULL, ...)bbd.design(nfactors, ncenter=4, factor.names = NULL, default.levels=c(-1,1), block.name=NULL, randomize=TRUE, seed=NULL, ...)
nfactors |
number of factors |
ncenter |
integer number of center points for each block |
factor.names |
list of scale end values for each factor
(the middle value is calculated); |
default.levels |
default levels (vector of length 2) for all factors for which no specific levels are given; must consist of two numeric values for the scale ends, the default middle level is calculated |
block.name |
name of block factor that distinguishes between the blocks;
blocks are usable for nfactors=4 and nfactors=5 only, |
randomize |
logical that indicates whether or not randomization should occur |
seed |
optional seed for random number generation in randomization |
... |
reserved for future usage |
Function bbd.design creates a Box-Behnken design, which is a design for quantitative
factors with all factors on three levels. Box-Behnken designs should not be used if the combinations
of the extreme levels of the factors are of interest (cf. also Myers, Montgomery and Anderson-Cook 2009).
There are designs for 3 to 7 factors, and the unreplicated versions of these have
14 (3 factors), 24 (4 factors), 40 (5 factors), 48 (6 factors), and 56 (7 factors) runs
plus the specified number of center points ncenter.
Function bbd.design is an interface to function bbd from package rsm
that makes this design accessible using similar syntax as used in packages DoE.base and FrF2
and creates an output object that is also treatable by the convenience functions available in package
DoE.base.
Currently, creation of replications and repeated measurements - as would be usual for other design functions - is not implemented. This is planned for the future, but does not have high priority.
The function returns a data frame of S3 class design
with attributes attached. The data frame itself is in the original data scale.
The matrix desnum attached as attribute desnum contains the coded data,
the attribute run.order contains the standard order and the actual run order
of the design (contrary to package rsm, the row names of the design refer to
the actual rather than the standard run order).
The attribute design.info is a list of design properties.
The element type of that list is the character string bbd.
Besides the elements present in all class design objects,
there are the elements quantitative (vector with nfactor TRUE entries),
and a codings element usable in the coding functions available in the rsm
package, e.g. coded.data.
Since R version 3.6.0, the behavior of function sample has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a randomized design that was produced with an earlier R version,
please follow the steps described with the argument seed.
This package is still under (slow) development. Reports about bugs and inconveniences are welcome. bbd.design is based on version 1 of package rsm.
Ulrike Groemping
Box, G.E.P. and Behnken, D.W. (1960). Some new three-level designs for the study of quantitative variables. Technometrics 2, 455-475.
Box, G.E.P., Hunter, J.S. and Hunter, W.G. (2005, 2nd ed.). Statistics for Experimenters. Wiley, New York.
Box, G.E.P. and Wilson, K.B. (1951). On the Experimental Attainment of Optimum Conditions. J. Royal Statistical Society, B13, 1-45.
NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/pri/section3/pri3361.htm, accessed August 20th, 2009.
Myers, R.H., Montgomery, D.C. and Anderson-Cook, C.M. (2009). Response Surface Methodology. Process and Product Optimization Using Designed Experiments. Wiley, New York.
See also FrF2, ccd.design, lhs-package, rsm
plan1 <- bbd.design(5) ## default for 5 factors is unblocked design, contrary to package rsm plan1 ## blocked design for 4 factors, using default levels plan2 <- bbd.design(4,block.name="block",default.levels=c(10,30)) plan2 desnum(plan2) ## design with factor.names and modified ncenter bbd.design(3,ncenter=6, factor.names=list("one"=c(25,35),"two"=c(-5,20), "three"=c(20,60))) ## design with character factor.names and default levels bbd.design(3,factor.names=c("one","two", "three"), default.levels=c(10,20))plan1 <- bbd.design(5) ## default for 5 factors is unblocked design, contrary to package rsm plan1 ## blocked design for 4 factors, using default levels plan2 <- bbd.design(4,block.name="block",default.levels=c(10,30)) plan2 desnum(plan2) ## design with factor.names and modified ncenter bbd.design(3,ncenter=6, factor.names=list("one"=c(25,35),"two"=c(-5,20), "three"=c(20,60))) ## design with character factor.names and default levels bbd.design(3,factor.names=c("one","two", "three"), default.levels=c(10,20))
Function for augmenting an existing fractional factorial with a star portion in case of a late decision for a sequential procedure.
ccd.augment(cube, ncenter = 4, columns="all", block.name="Block.ccd", alpha = "orthogonal", randomize=TRUE, seed=NULL, ...)ccd.augment(cube, ncenter = 4, columns="all", block.name="Block.ccd", alpha = "orthogonal", randomize=TRUE, seed=NULL, ...)
cube |
design generated by function |
ncenter |
integer number of center points,
or vector with two numbers, the first for the cube and the second for
the star portion of the design. |
block.name |
name of block factor that distinguishes (at least) between blocks; even for unblocked cubes, the ccd design has a cube and a star point block |
alpha |
“orthogonal”, “rotatable”, or a number that indicates the position of the star points; the number 1 would create a face-centered design. |
randomize |
logical that indicates whether or not randomization should occur |
seed |
NULL or a vector of two integer seeds for random number generation in randomization |
... |
reserved for future usage |
columns |
not yet implemented; it is intended to later allow to add star points for only some factors of a design (after eliminating the others as unimportant in a sequential process), and columns will be used to indicate those |
The statistical background of central composite designs is briefly described
under CentralCompositeDesigns.
Function ccd.augment augments an existing 2-level fractional factorial
that should already have been run with center points and should have resolution V.
In exceptional situations, it may be useful to base a ccd on a resolution IV design
that allows estimation of all 2-factor interactions of interest. Thus, it can be
interesting to apply function ccd.augment to a cube
based on the estimable functionality of function FrF2
in cases where a resolution V cube is not feasible.
Of course, this does not allow to estimate the aliased 2-factor interactions
and therefore generates a warning.
The function returns a data frame of S3 class design
with attributes attached. The data frame itself is in the original data scale.
The data frame desnum attached as attribute desnum is the original data frame
returned by package rsm. The attribute design.info is a list of various design properties.
The element type of that list is the character string ccd.
Besides the elements present in all class design objects,
there are the elements quantitative (vector with nfactor TRUE entries),
and a codings element usable in the coding functions available in the rsm
package, e.g. coded.data.
Note that the row names and the standard order column in the run.order
attribute of ccd designs based on
estimability requirements (cf. also the details section) are not in conventional order
and should not be used as the basis for any calculations. The same is true for
blocked designs, if the blocking routine blockpick.big was used.
Since R version 3.6.0, the behavior of function sample has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a randomized design that was produced with an earlier R version,
please follow the steps described with the argument seed.
This package is still under (slow) development. Reports about bugs and inconveniences are welcome. ccd.augment is based on version 1 of package rsm.
Ulrike Groemping
Box, G.E.P., Hunter, J.S. and Hunter, W.G. (2005, 2nd ed.). Statistics for Experimenters. Wiley, New York.
Box, G.E.P. and Wilson, K.B. (1951). On the Experimental Attainment of Optimum Conditions. J. Royal Statistical Society, B13, 1-45.
NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/pri/section3/pri3361.htm, accessed August 20th, 2009.
Myers, R.H., Montgomery, D.C. and Anderson-Cook, C.M. (2009). Response Surface Methodology. Process and Product Optimization Using Designed Experiments. Wiley, New York.
See also ccd.design, FrF2,
lhs-package, rsm
## purely technical examples for the sequential design creation process ## start with a fractional factorial with center points plan <- FrF2(16,5,default.levels=c(10,30),ncenter=6) ## collect data and add them to the design y <- rexp(22) plan <- add.response(plan,y) ## assuming that an analysis has created the suspicion that a second order ## model should be fitted (not to be expected for the above random numbers): plan.augmented <- ccd.augment(plan, ncenter=4) ## add new responses to the design y <- c(y, rexp(14)) ## append responses for the 14=5*2 + 4 star points r.plan.augmented <- add.response(plan.augmented, y, replace=TRUE) ## for info: how to analyse results from such a desgin lm.result <- lm(y~Block.ccd+(.-Block.ccd)^2+I(A^2)+I(B^2)+I(C^2)+I(D^2)+I(E^2), r.plan.augmented) summary(lm.result) ## analysis with function rsm rsm.result <- rsm(y~Block.ccd+SO(A,B,C,D,E), r.plan.augmented) summary(rsm.result) ## provides more information than lm.result loftest(rsm.result) ## separate lack of fit test ## graphical analysis ## (NOTE: purely for demo purposes, the model is meaningless here) ## individual contour plot contour(rsm.result,B~A) ## several contour plots par(mfrow=c(1,2)) contour(rsm.result,list(B~A, C~E)) ## many contourplots, all pairs of some factors par(mfrow=c(2,3)) contour(rsm.result,~A+B+C+D)## purely technical examples for the sequential design creation process ## start with a fractional factorial with center points plan <- FrF2(16,5,default.levels=c(10,30),ncenter=6) ## collect data and add them to the design y <- rexp(22) plan <- add.response(plan,y) ## assuming that an analysis has created the suspicion that a second order ## model should be fitted (not to be expected for the above random numbers): plan.augmented <- ccd.augment(plan, ncenter=4) ## add new responses to the design y <- c(y, rexp(14)) ## append responses for the 14=5*2 + 4 star points r.plan.augmented <- add.response(plan.augmented, y, replace=TRUE) ## for info: how to analyse results from such a desgin lm.result <- lm(y~Block.ccd+(.-Block.ccd)^2+I(A^2)+I(B^2)+I(C^2)+I(D^2)+I(E^2), r.plan.augmented) summary(lm.result) ## analysis with function rsm rsm.result <- rsm(y~Block.ccd+SO(A,B,C,D,E), r.plan.augmented) summary(rsm.result) ## provides more information than lm.result loftest(rsm.result) ## separate lack of fit test ## graphical analysis ## (NOTE: purely for demo purposes, the model is meaningless here) ## individual contour plot contour(rsm.result,B~A) ## several contour plots par(mfrow=c(1,2)) contour(rsm.result,list(B~A, C~E)) ## many contourplots, all pairs of some factors par(mfrow=c(2,3)) contour(rsm.result,~A+B+C+D)
Function for accessing central composite designs from package rsm, with automatic creation of an appropriate cube portion
ccd.design(nfactors=NULL, factor.names=NULL, default.levels=c(-1,1), ncube=NULL, resolution=if (identical(blocks,1) & is.null(ncube)) 5 else NULL, generators=NULL, ncenter = 4, alpha = "orthogonal", replications=1, block.name="Block.ccd", blocks=1, randomize=TRUE, seed=NULL, ...)ccd.design(nfactors=NULL, factor.names=NULL, default.levels=c(-1,1), ncube=NULL, resolution=if (identical(blocks,1) & is.null(ncube)) 5 else NULL, generators=NULL, ncenter = 4, alpha = "orthogonal", replications=1, block.name="Block.ccd", blocks=1, randomize=TRUE, seed=NULL, ...)
nfactors |
number of factors |
factor.names |
list of cube corner values for each factor;
names are used as variable names;
the names must not be x1, x2, ..., as these are used for the variables
in coded units; |
default.levels |
default levels (vector of length 2) for all factors for which no specific levels are given |
ncube |
integer number of cube points (without center points for the cube) |
resolution |
arabic numeral for the requested resolution of the cube portion
of the design; cubes for ccd designs should usually be at least of resolution V.
the default value for resolution is therefore 5, unless |
generators |
generators in the form allowed in function |
ncenter |
integer number of center points for each cube or star point block, or vector with two numbers, the first for the cube and the second for the star portion of the design |
alpha |
“orthogonal”, “rotatable”, or a number that indicates the position of the star points; the number 1 would create a face-centered design. |
replications |
the number of replications of the design; currently, only proper replications can be generated; these are randomized in blocks within the center point and star blocks. The same number of replications is used for both the cube and the star blocks. |
block.name |
name of block factor that distinguishes between blocks; even for unblocked cubes, the ccd design has at least one cube and one star point block |
blocks |
the same as in function
If If the experiment is randomized, randomization happens within blocks. For the statistical and algorithmic background of blocked designs, see |
randomize |
logical that indicates whether or not randomization should occur |
seed |
NULL or a vector of two integer seeds for random number generation in randomization |
... |
reserved for future usage |
The statistical background of central composite designs is briefly described
under CentralCompositeDesigns.
Function ccd.design creates a central composite design from scratch.
It proceeds by generating a cube design with function FrF2 and then
augmenting this cube design using functions add.center from package
FrF2 for adding center points to the cube and subsequently function
ccd from package rsm for generating the star portion of
the design.
There are two main purposes for this function: one is to provide
central composite designs within the same syntax philosophy
used in packages DoE.base-package and FrF2.
The other is to automatically identify good (=resolution V) cube portions,
which can be achieved by using the resolution parameter.
In comparison to direct usage of package ccd, the functions make the syntax closer to that of the other packages in the DoE.wrapper suite and allow automatic selection of fractional factorials as cubes.
Function ccd.design does not allow direct use of the estimable functionality
that is available in function FrF2. Nevertheless, ccd designs with a cube
based on the estimable functionality can be generated
by first using function FrF2 and subsequently applying
function ccd.augment. It may for example be interesting to use designs based on
estimability requirements for 2-factor interactions in cases where a resolution V cube
for the ccd is not feasible - of course, this does not allow to estimate the full second order model
and therefore generates a warning.
The function returns a data frame of S3 class design
with attributes attached. The data frame itself is in the original data scale.
The data frame desnum attached as attribute desnum is the coded design.
The attribute design.info is a list of various design properties.
The element type of that list is the character string ccd.
Besides the elements present in all class design objects,
there are the elements quantitative (vector with nfactor TRUE entries),
and a codings element usable in the coding functions available in the rsm
package, e.g. coded.data.
Note that the row names and the standard order column in the
run.order attribute of ccd designs
are not in conventional order,
if the blocking routine blockpick.big was used.
In such situations, these should not be used as the basis for any calculations.
Since R version 3.6.0, the behavior of function sample has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a randomized design that was produced with an earlier R version,
please follow the steps described with the argument seed.
This package is still under (slow) development. Reports about bugs and inconveniences are welcome. ccd.design is based on version 1 of package rsm.
Ulrike Groemping
Box, G.E.P., Hunter, J.S. and Hunter, W.G. (2005, 2nd ed.). Statistics for Experimenters. Wiley, New York.
Box, G.E.P. and Wilson, K.B. (1951). On the Experimental Attainment of Optimum Conditions. J. Royal Statistical Society, B13, 1-45.
NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/pri/section3/pri3361.htm, accessed August 20th, 2009.
Myers, R.H., Montgomery, D.C. and Anderson-Cook, C.M. (2009). Response Surface Methodology. Process and Product Optimization Using Designed Experiments. Wiley, New York.
See also ccd.augment, add.center,
FrF2, lhs-package, rsm
ccd.design(5) ## per default uses the resolution V design in 16 runs for the cube ccd.design(5, ncube=32) ## uses the full factorial for the cube ccd.design(5, ncenter=6, default.levels=c(-10,10)) ## blocked design (requires ncube to be specified) ccd.design(5, ncube=32, blocks=4) ## there is only one star point block ## for usage of other options, look at the FrF2 documentationccd.design(5) ## per default uses the resolution V design in 16 runs for the cube ccd.design(5, ncube=32) ## uses the full factorial for the cube ccd.design(5, ncenter=6, default.levels=c(-10,10)) ## blocked design (requires ncube to be specified) ccd.design(5, ncube=32, blocks=4) ## there is only one star point block ## for usage of other options, look at the FrF2 documentation
Brief description of the statistical background of central composite designs
Central composite designs (ccd's) were invented by Box and Wilson (1951) for response surface experimentation with quantitative factors. They are used for estimation of second order response surface models, i.e. models that allow to estimate linear, quadratic and interaction effects for all factors.
Central composite designs consist of a cube and star points (also called axial points). Both the cube and the star portion of the design should have some center points. The cube is a (fractional) factorial design and should be at least of resolution V. The line between the center points and the star points intersects the faces of the cube in their middle (see the link to the NIST/Sematech e-Handbook for a visualization). There are two star points per factor, i.e. the number of runs for (each block of) the star portion of the design is twice the number of factors plus the number of center points in the star portion.
The tuning parameter alpha determines whether the star points lie on the
faces of the cube (alpha=1, face-centered), inside the cube (alpha<1,
inscribed) or outside the cube (alpha>1, circumscribed).
The latter case is the usual one. The value of
alpha can be chosen such that the design is rotatable (may be useful if the scales
of the factors are comparable) or such that the design is orthogonally blocked
(i.e. the block effects do not affect the effect estimates of interest). The default
is to generate orthogonally blocked designs.
Central composite designs are particularly useful in sequential experimentation, where a (fractional) factorial with center points is followed up by a star portion of the design. While the cube can already estimate the linear and interaction effects, the center points can only estimate the sum of all quadratic effects. If this indicates that quadratic effects are important, a star portion can be added in order to investigate the model more deeply.
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Box, G.E.P., Hunter, J.S. and Hunter, W.G. (2005, 2nd ed.). Statistics for Experimenters. Wiley, New York.
Box, G.E.P. and Wilson, K.B. (1951). On the Experimental Attainment of Optimum Conditions. J. Royal Statistical Society, B13, 1-45.
NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/pri/section3/pri3361.htm, accessed August 20th, 2009.
Myers, R.H., Montgomery, D.C. and Anderson-Cook, C.M. (2009). Response Surface Methodology. Process and Product Optimization Using Designed Experiments. Wiley, New York.
See Also ccd, ccd.design, ccd.augment
Function for comfortably augmenting a design with D-optimal additional points; this functionality is still somewhat experimental.
Dopt.augment(design, m=1, formula=NULL, candidates=NULL, constraint=NULL, center=FALSE, nRepeats=5, seed=NULL, randomize=TRUE, ...)Dopt.augment(design, m=1, formula=NULL, candidates=NULL, constraint=NULL, center=FALSE, nRepeats=5, seed=NULL, randomize=TRUE, ...)
design |
an experimental design of class |
m |
integer number of additional points to add to design |
formula |
a model formula (starting with a tilde), for the estimation of which a D-optimal design is sought; For quantitative factors, functions |
candidates |
data frame of candidate points; if not specified, candidates
are constructed as a full factorial from the
|
constraint |
a condition (character string!) used for reducing the candidate
set to admissible points only. |
center |
requests that optimization is run for the centered model; the design is nevertheless output in non-centered coordinates |
nRepeats |
number of independent repeats of the design optimization process; increasing this number may improve the chance of finding a global optimum, but will also increase search time |
seed |
seed for generation and randomization of the design (integer number); |
randomize |
logical deciding whether or not the design should be randomized;
if it is |
... |
additional arguments to function |
Function Dopt.augment augments an existing design by m D-optimal
additional points (unblocked designs, no split-plot, no parameter or crossed design,
no repeat.only replications), i.e. by points that make the design particularly efficient
for the intended model.
Option center, which is available for both blocked and unblocked designs as part of the ... argument,
requests optimization for the centered model; the design that is created is nevertheless an uncentered design.
NULL entries in the arguments are filled with automatic values that are determined
from design.
The function returns a data frame of S3 class design
with attributes attached.
The data frame contains the experimental settings.
The matrix desnum attached as attribute desnum contains the
model matrix of the design, using the formula as specified in the call.
Function Dopt.augment preserves additional variables (e.g. responses) that
have been added to the design design before augmenting. Note, however, that
the response data are NOT used in deciding about which points to augment the design with.
The attribute run.order provides the run number in standard order (as returned from
function optFederov in package AlgDesign) as well
as the randomized actual run order. The third column is always identical to the first.
Note that the first n runs (the ones that are already present before augmentation)
have run numbers in standard order from 1 to n (i.e. their original run numbers
in standard order, if they were also generated by Dopt.design are lost).
The attribute design.info is a list of various design properties, with type resolving to “Dopt.augment”.
In addition to the standard list elements (cf. design), the element
quantitative is a vector of nfactor logical values or NAs,
and the optional digits elements indicates the number of digits to
which the data were rounded.
The list contains further entries regarding the optimality that has been achieved
(D, Dea and A).
Note that the original design is contained in the first rows of the new data set.
The original design also contains columns that are not directly part of the
design, e.g. comment columns.
Note that replications is always set to 1, even if the original design was
replicated, and repeat.only is always FALSE.
These elements are only present to fulfill the formal requirements for class design.)
Since R version 3.6.0, the behavior of function sample has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a design that was produced with an earlier R version,
please follow the steps described with the argument seed.
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Atkinson, A.C. and Donev, A.N. (1992). Optimum experimental designs. Clarendon Press, Oxford.
Federov, V.V. (1972). Theory of optimal experiments. Academic Press, New York.
Wheeler, R.E. (2004). Comments on algorithmic design. Vignette accompanying package AlgDesign. ../../AlgDesign/doc/AlgDesign.pdf.
See also optFederov, fac.design,
quad, cubic,
Dopt.design. Furthermore, unrelated to function Dopt.augment,
see also function gen_design from package skpr
for a new general R package for creating D-optimal or other letter optimal designs.
## a full quadratic model with constraint in three quantitative factors plan <- Dopt.design(36,factor.names=list(eins=c(100,250),zwei=c(10,30),drei=c(-25,25)), nlevels=c(4,3,6), formula=~quad(.), constraint="!(eins>=200 & zwei==30 & drei==25)") summary(plan) y <- rnorm(36) r.plan <- add.response(plan, y) plan2 <- Dopt.augment(r.plan, m=10) summary(plan2) ## add the new response values after conducting additional experiments y <- c(y, rnorm(10)) r.plan2 <- add.response(plan2,y, replace=TRUE) summary(r.plan2, brief=FALSE)## a full quadratic model with constraint in three quantitative factors plan <- Dopt.design(36,factor.names=list(eins=c(100,250),zwei=c(10,30),drei=c(-25,25)), nlevels=c(4,3,6), formula=~quad(.), constraint="!(eins>=200 & zwei==30 & drei==25)") summary(plan) y <- rnorm(36) r.plan <- add.response(plan, y) plan2 <- Dopt.augment(r.plan, m=10) summary(plan2) ## add the new response values after conducting additional experiments y <- c(y, rnorm(10)) r.plan2 <- add.response(plan2,y, replace=TRUE) summary(r.plan2, brief=FALSE)
Function for comfortably creating a D-optimal design with or without blocking based on functions optFederov or optBlock from package AlgDesign; this functionality is still somewhat experimental.
Dopt.design(nruns, data=NULL, formula=~., factor.names=NULL, nlevels=NULL, digits=NULL, constraint=NULL, center=FALSE, nRepeats=5, seed=NULL, randomize=TRUE, blocks=1, block.name="Blocks", wholeBlockData=NULL, qual=NULL, ...)Dopt.design(nruns, data=NULL, formula=~., factor.names=NULL, nlevels=NULL, digits=NULL, constraint=NULL, center=FALSE, nRepeats=5, seed=NULL, randomize=TRUE, blocks=1, block.name="Blocks", wholeBlockData=NULL, qual=NULL, ...)
nruns |
number of runs in the requested design |
data |
data frame or matrix of candidate design points; |
formula |
a model formula (starting with a tilde),
for the estimation of which a D-optimal design is sought; |
factor.names |
is used for creating a candidate set (for the within Block factors)
with the help of function
|
nlevels |
can be omitted if the list |
digits |
is used for creating a candidate set if |
constraint |
a condition (character string!) used for reducing the candidate
set to admissible points only.
|
center |
requests that optimization is run for the centered model; the design is nevertheless output in non-centered coordinates |
nRepeats |
number of independent repeats of the design optimization process; increasing this number may improve the chance of finding a global optimum, but will also increase search time |
seed |
seed for generation and randomization of the design (integer number); |
randomize |
logical deciding whether or not the design should be randomized;
if it is |
blocks |
a single integer giving the number of blocks (default 1, if no blocking is needed) |
block.name |
character string: name of the blocking variable (used only if blocks are requested) |
wholeBlockData |
optional matrix or data frame that specifies the whole block characteristics; Since wholeBlockData must be completely specified by the user, optimization is for the split-plot portion of the design only. The rationale is (assumably) that the characteristics of the available blocks are known. If this is not the case, users may want to try out various possible whole block setups, or to proceed sequentially by first optimizing a whole block design for a model with the whole block factors only and subsequently using this model for adding split-plot factors. |
qual |
optional logical (length 1 or same as number of factors); ignored, if |
... |
additional arguments to functions |
Function Dopt.design creates a D-optimal design, optionally with blocking,
and even as a split-plot design. If no blocks are required, calculations are carried
out through function optFederov from package AlgDesign.
In case of blocked designs, function optBlock from package AlgDesign
is behind the calculations. By specifying wholeBlockData, a blocked design becomes
a split-plot design. The model formula can refer to both the within block data (only those
are referred to by the “.” notation) and the whole block data and interactions between both.
In comparison to direct usage of package AlgDesign, the function adds the possibility
of automatically creating the candidate points on the fly, with or without constraints.
Furthermore, it embeds the D-optimal designs into the class design.
On the other hand, it sacrifices some of AlgDesigns flexibility; of course, users
can still use AlgDesign directly.
The D-optimal designs are particularly useful, if the classical regular designs are too demanding in run size requirements, or if constraints preclude automatic generation of orthogonal designs. Note, however, that the best design in few runs can still be very bad in absolute terms!
When specifying the design without the data option, a full factorial in the
requested factors is the default candidate set of design points. For some situations - especially
with many factors - it may be better to start from a restricted candidate set. Such a candidate set
can be produced with another R function, e.g. oa.design or FrF2,
or can be manually created.
If there are doubts, whether the process has delivered a design close to the absolute optimum,
nRepeats can be increased.
For unblocked designs, it is additionally possible to increase maxIteration.
Also, improving the starting
value by nullify=1 or nullify=2 may lead to an improved design.
These options are handed through to function optFederov
from package AlgDesign and are documented there.
The function returns a data frame of S3 class design
with attributes attached.
The data frame contains the experimental settings.
The matrix desnum attached as attribute desnum contains the
model matrix of the design, using the formula as specified in the call.
Function Dopt.augment preserves additional variables (e.g. responses) that
have been added to the design design before augmenting. Note, however, that
the response data are NOT used in deciding about which points to augment the design with.
The attribute run.order provides the run number in standard order (as returned from
function optFederov in package AlgDesign) as well
as the randomized actual run order. The third column is always identical to the first.
The attribute design.info is a list of various design properties, with type resolving to “Dopt”,
“Dopt.blocked”, “Dopt.splitplot”.
In addition to the standard list elements (cf. design), the element
quantitative is a vector of nfactor logical values or NAs,
and the optional digits elements indicates the number of digits to
which the data were rounded.
For blocked and splitplot designs, the list contains additional information on numbers and sizes of blocks or plots,
as well as the number of whole plot factors (which are always the first few factors) and split-plot factors.
The list contains a list of optimality criteria as calculated by function optFederov,
see documentation there)
with elements D, Dea, A and G.
(Note that replications is always 1 and repeat.only is always FALSE;
these elements are only present to fulfill the formal requirements for class design.
Note however, that blocked designs do in fact repeat experimental runs if nruns and blocks
imply this.)
Since R version 3.6.0, the behavior of function sample has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For reproducing a design that was produced with an earlier R version,
please follow the steps described with the argument seed.
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Atkinson, A.C. and Donev, A.N. (1992). Optimum experimental designs. Clarendon Press, Oxford.
Federov, V.V. (1972). Theory of optimal experiments. Academic Press, New York.
Wheeler, R.E. (2004). Comments on algorithmic design. Vignette accompanying package AlgDesign. ../../AlgDesign/doc/AlgDesign.pdf.
See also optFederov, fac.design,
quad, cubic,
Dopt.augment. Furthermore, unrelated to function Dopt.design,
see also function gen_design from package skpr
for a new general R package for creating D-optimal or other letter optimal designs.
## a full quadratic model with constraint in three quantitative factors plan <- Dopt.design(36,factor.names=list(eins=c(100,250),zwei=c(10,30),drei=c(-25,25)), nlevels=c(4,3,6), formula=~quad(.), constraint="!(eins>=200 & zwei==30 & drei==25)") plan cor(plan) y <- rnorm(36) r.plan <- add.response(plan, y) plan2 <- Dopt.augment(r.plan, m=10) plot(plan2) cor(plan2) ## designs with qualitative factors and blocks for ## an experiment on assessing stories of social situations ## where each subject is a block and receives a deck of 5 stories plan.v <- Dopt.design(480, factor.names=list(cause=c("sick","bad luck","fault"), consequences=c("alone","children","sick spouse"), gender=c("Female","Male"), Age=c("young","medium","old")), blocks=96, constraint="!(Age==\"young\" & consequences==\"children\")", formula=~.+cause:consequences+gender:consequences+Age:cause) ## an experiment on assessing stories of social situations ## with the whole block (=whole plot) factor gender of the assessor ## not run for saving test time on CRAN ## Not run: plan.v.splitplot <- Dopt.design(480, factor.names=list(cause=c("sick","bad luck","fault"), consequences=c("alone","children","sick spouse"), gender.story=c("Female","Male"), Age=c("young","medium","old")), blocks=96, wholeBlockData=cbind(gender=rep(c("Female","Male"),each=48)), constraint="!(Age==\"young\" & consequences==\"children\")", formula=~.+gender+cause:consequences+gender.story:consequences+ gender:consequences+Age:cause+gender:gender.story) ## End(Not run)## a full quadratic model with constraint in three quantitative factors plan <- Dopt.design(36,factor.names=list(eins=c(100,250),zwei=c(10,30),drei=c(-25,25)), nlevels=c(4,3,6), formula=~quad(.), constraint="!(eins>=200 & zwei==30 & drei==25)") plan cor(plan) y <- rnorm(36) r.plan <- add.response(plan, y) plan2 <- Dopt.augment(r.plan, m=10) plot(plan2) cor(plan2) ## designs with qualitative factors and blocks for ## an experiment on assessing stories of social situations ## where each subject is a block and receives a deck of 5 stories plan.v <- Dopt.design(480, factor.names=list(cause=c("sick","bad luck","fault"), consequences=c("alone","children","sick spouse"), gender=c("Female","Male"), Age=c("young","medium","old")), blocks=96, constraint="!(Age==\"young\" & consequences==\"children\")", formula=~.+cause:consequences+gender:consequences+Age:cause) ## an experiment on assessing stories of social situations ## with the whole block (=whole plot) factor gender of the assessor ## not run for saving test time on CRAN ## Not run: plan.v.splitplot <- Dopt.design(480, factor.names=list(cause=c("sick","bad luck","fault"), consequences=c("alone","children","sick spouse"), gender.story=c("Female","Male"), Age=c("young","medium","old")), blocks=96, wholeBlockData=cbind(gender=rep(c("Female","Male"),each=48)), constraint="!(Age==\"young\" & consequences==\"children\")", formula=~.+gender+cause:consequences+gender.story:consequences+ gender:consequences+Age:cause+gender:gender.story) ## End(Not run)
Functions for comfortably accessing latin hypercube sampling designs from package lhs or space-filling designs from package DiceDesign, which are useful for quantitative factors with many possible levels. In particular, they can be used in computer experiments. Most of the designs are random samples.
lhs.design(nruns, nfactors, type="optimum", factor.names=NULL, seed=NULL, digits=NULL, nlevels = nruns, default.levels = c(0, 1), randomize = FALSE, ...) lhs.augment(lhs, m=1, type="optAugment", seed=NULL, ...)lhs.design(nruns, nfactors, type="optimum", factor.names=NULL, seed=NULL, digits=NULL, nlevels = nruns, default.levels = c(0, 1), randomize = FALSE, ...) lhs.augment(lhs, m=1, type="optAugment", seed=NULL, ...)
nruns |
number of runs in the latin hypercube sample; |
nfactors |
number of factors in the latin hypercube sample |
type |
character string indicating the type of design or augmentation method;
defaults are “optimum” for |
seed |
seed for random number generation; latin hypercube samples from package lhs are random samples. Specifying a seed used to make the result reproducible for early versions of package lhs - lately, results are reproducible within a package version, but reproducibility between package versions cannot be guaranteed. |
factor.names |
list of scale end values for each factor;
names are used as variable names; |
digits |
digits to which the design columns are rounded; one single value
(the same for all factors) or a vector of length |
nlevels |
used for type |
default.levels |
scale ends for all factors; convenient, if all factors have the same scaling that deviates from the default 0/1 scale ends. |
randomize |
logical that prevents randomization per default. The option has an effect
for types |
lhs |
design generated by function |
m |
integer number of additional points to add to design |
... |
additional arguments to the functions from packages lhs
or DiceDesign.
Refer to their documentation. |
Function lhs.design creates a latin hypercube sample,
function lhs.augment augments an existing latin hypercube sample (or in case of
type optSeeded takes the existing sample as the starting point but potentially modifies it).
In comparison to direct usage of package lhs, the functions add the possibility
of recoding lhs samples to a desired range, and they embed the lhs designs into
class design.
Range coding is based on the recoding facility from package rsm and the
factor.names parameter used analogously to packages DoE.base and FrF2.
The lhs designs are useful for quantitative factors, if it is considered desirable to uniformly distribute design points over a hyperrectangular space. This is e.g. considered interesting for computer experiments, where replications of the same settings are often useless.
Supported design types are described in the documentation for
packages lhs
and DiceDesign.
Both functions return a data frame of S3 class design
with attributes attached.
The data frame contains the experimental settings as recoded to the scale ends defined in factor.names (if given),
rounded to the number of digits given in digits (if given).
The experimental factors in the matrix desnum attached as attribute desnum contain the
design in the unit cube (all experimental factors ranging from 0 to 1) as
returned by packages lhs or DiceDesign.
Function lhs.augment preserves additional variables (e.g. responses) that
have been added to the design lhs before augmenting. Note, however, that
the response data are NOT used in deciding about which points to augment the design with.
The attribute run.order is not very useful for most of these designs, as there is no standard order.
It therefore is present for formal reasons only and contains three identical columns of 1,2,...,nruns.
For designs created with type=fact or type=faure, the standard order is the order
in which package DiceDesign creates the design, and the actual run order may be different in case of
randomization.
In case of lhs.augment, if the design to be augmented had been reordered before,
the augmented design preserves this reorder and also the respective numbering of the design.
The attribute design.info is a list of various design properties, with type resolving to “lhs”.
In addition to the standard list elements (cf. design), the subtype
element indicates the type of latin hypercube designs and possibly additional augmentations, the element
quantitative is a vector of nfactor logical TRUEs,
and the digits elements indicates the digits to which the data were rounded.
For designs created with package DiceDesign, special list elements from this package
are also added to design.info.randomize is always TRUE for designs generated by random sampling, but may be FALSE for
designs created with type=fact or type=faure. coding provides formulae for making the designs comfortably usable with
standard second order methodology implemented in package rsm.
replications is always 1 and repeat.only is always FALSE;
these elements are only present to fulfill the formal requirements for class design.
Since R version 3.6.0, the behavior of function sample has changed
(correction of a biased previous behavior that should not be relevant for the randomization of designs).
For using code that randomizes a design interchangeably between a new R version (3.6.0 or later)
and an older one, please follow the steps described with the argument randomize.
Note also: Package lhs does not promise to keep designs reproducible between package versions. Thus, please make sure to store important designs for the future, if needed (of course, this is always wise anyway!).
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Beachkofski, B., Grandhi, R. (2002) Improved Distributed Hypercube Sampling. American Institute of Aeronautics and Astronautics Paper 1274.
Currin C., Mitchell T., Morris M. and Ylvisaker D. (1991) Bayesian Prediction of Deterministic Functions With Applications to the Design and Analysis of Computer Experiments, Journal of the American Statistical Association 86, 953–963.
Santner T.J., Williams B.J. and Notz W.I. (2003) The Design and Analysis of Computer Experiments, Springer, 121–161.
Shewry, M. C. and Wynn and H. P. (1987) Maximum entropy sampling. Journal of Applied Statistics 14, 165–170.
Fang K.-T., Li R. and Sudjianto A. (2006) Design and Modeling for Computer Experiments, Chapman & Hall.
Stein, M. (1987) Large Sample Properties of Simulations Using Latin Hypercube Sampling. Technometrics 29, 143–151.
Stocki, R. (2005) A method to improve design reliability using optimal Latin hypercube sampling. Computer Assisted Mechanics and Engineering Sciences 12, 87–105.
See Also compare for comparing optimality criteria for various
designs, lhs-package and
DiceDesign-package for the packages that do the calculations,FrF2, oa.design,
fac.design, pb for other possibilites of generating designs
## maximin design from package lhs plan <- lhs.design(20,7,"maximin",digits=2) plan plot(plan) cor(plan) y <- rnorm(20) r.plan <- add.response(plan, y) ## augmenting the design with 10 additional points, default method plan2 <- lhs.augment(plan, m=10) plot(plan2) cor(plan2) ## purely random design (usually not ideal) plan3 <- lhs.design(20,4,"random", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) plot(plan3) cor(plan3) ## optimum design from package lhs (default) plan4 <- lhs.design(20,4,"optimum", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan4) cor(plan4) ## dmax design from package DiceDesign ## arguments range and niter_max are required ## ?dmaxDesign for more info plan5 <- lhs.design(20,4,"dmax", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, range=0.2, niter_max=500) plot(plan5) cor(plan5) ## Strauss design from package DiceDesign ## argument RND is required ## ?straussDesign for more info plan6 <- lhs.design(20,4,"strauss", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, RND = 0.2) plot(plan6) cor(plan6) ## full factorial design from package DiceDesign ## mini try-out version plan7 <- lhs.design(3,4,"fact", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan7) cor(plan7) ## Not run: ## full factorial design from package DiceDesign ## not as many different levels as runs, but only a fixed set of levels ## caution: too many levels can easily bring down the computer ## above design with 7 distinct levels for each factor, ## implying 2401 runs plan7 <- lhs.design(7,4,"fact", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan7) cor(plan7) ## equivalent call plan7 <- lhs.design(,4,"fact",nlevels=7, factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) ## different number of levels for each factor plan8 <- lhs.design(,4,"fact",nlevels=c(5,6,5,7), factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan8) cor(plan8) ## equivalent call (specifying nruns, not necessary but a good check) plan8 <- lhs.design(1050,4,"fact",nlevels=c(5,6,5,7), factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) ## End(Not run)## maximin design from package lhs plan <- lhs.design(20,7,"maximin",digits=2) plan plot(plan) cor(plan) y <- rnorm(20) r.plan <- add.response(plan, y) ## augmenting the design with 10 additional points, default method plan2 <- lhs.augment(plan, m=10) plot(plan2) cor(plan2) ## purely random design (usually not ideal) plan3 <- lhs.design(20,4,"random", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) plot(plan3) cor(plan3) ## optimum design from package lhs (default) plan4 <- lhs.design(20,4,"optimum", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan4) cor(plan4) ## dmax design from package DiceDesign ## arguments range and niter_max are required ## ?dmaxDesign for more info plan5 <- lhs.design(20,4,"dmax", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, range=0.2, niter_max=500) plot(plan5) cor(plan5) ## Strauss design from package DiceDesign ## argument RND is required ## ?straussDesign for more info plan6 <- lhs.design(20,4,"strauss", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, RND = 0.2) plot(plan6) cor(plan6) ## full factorial design from package DiceDesign ## mini try-out version plan7 <- lhs.design(3,4,"fact", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan7) cor(plan7) ## Not run: ## full factorial design from package DiceDesign ## not as many different levels as runs, but only a fixed set of levels ## caution: too many levels can easily bring down the computer ## above design with 7 distinct levels for each factor, ## implying 2401 runs plan7 <- lhs.design(7,4,"fact", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan7) cor(plan7) ## equivalent call plan7 <- lhs.design(,4,"fact",nlevels=7, factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) ## different number of levels for each factor plan8 <- lhs.design(,4,"fact",nlevels=c(5,6,5,7), factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) plot(plan8) cor(plan8) ## equivalent call (specifying nruns, not necessary but a good check) plan8 <- lhs.design(1050,4,"fact",nlevels=c(5,6,5,7), factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2) ## End(Not run)
One function for calculating the S-optimality criterion is provided here. This help file documents this function and also describes optimality criteria from other related packages, referring to space filling optimality like the S criterion or to model-based optimality like the well-known D-criterion.
Scalc(design) compare(...)Scalc(design) compare(...)
design |
matrix, often normalized to unit cube, |
... |
two or more designs, either all of type |
Function Scalc calculates the S criterion for optimality, which is
employed in package lhs for most optimization purposes (exception:
maximin designs). The criterion is the harmonic mean of all pairwise
interpoint distances, and space-filling optimization tries to maximize it.
Function eval.design from package AlgDesign
calculates various model-based optimality criteria:
a confounding matrix of effects, if requested
the k-th root of the determinant of Z'Z/N, where Z is the model matrix of the model under investigation, k is the number of columns of Z and N the number of rows; this is the quantity optimized for D-optimal designs
the arithmetic mean of coeffient variance multipliers, i.e. the average diagonal element of the inverse of Z'Z/N, intercept included
the average prediction variance over a space X; calculated only, if X is specified
the minimax normalized variance over X; calculated only, if X is specified
A lower bound on D efficiency for approximate theory designs. It is equal to exp(1-1/Ge), i.e. is also calculated only, if X is specified.
the k-th root of the ratio of the determinant of M1 divided by the product of diagonal elements of M1, where M1 is Z'Z with the column and row referring to the intercept removed, and k the number of columns of M1; if this is 1, the coefficient estimates are uncorrelated.
the geometric mean of normalized coeffient variance multipliers (intercept excluded), i.e. the geometric mean of the diagonal elements of the inverse of Z'Z/N, without the first element, if an intercept is in the model.
Several functions from package DiceDesign calculate optimality criteria
regarding the space filling qualities of a design. These functions
normalize the design to lie in the unit cube,
if it does not yet do so. Application of these functions
to designs with qualitative factors does not make sense and
yields errors. The following functions are available:
mindist calculates the minimum distance between
any pair of design points. This is the criterion which is maximized by
maximin designs, i.e. should be large for a space-filling design.
For the next two distance metrics, it is helpful to define g_i as the minimal
distance of design point i to any other design point.
meshRatio calculates the ratio of the
maximum g_i to the minimum g_i
(a small mesh ratio indicates a similar minimal
distance for all design points).
coverage calculates the coefficient of variation
of the g_i, however using the denominator n instead of n-1
for the standard deviation (a large coverage indicates that the average minimal
distance for of design points is large relative to their standard deviation;
large values are desirable).
Function discrepancyCriteria calculates
several versions of L2 discrepancy.
Function link[skpr]{eval_design} calculates effect and parameter powers
for a design for a specified model and a specified significance level alpha.
Function Scalc returns a single number: the harmonic mean of all pairwise interpoint distances
is calculated, based on the matrix design or the desnum attribute of the design design.
(This value should be as large as possible for a space-filling design.)
Note that the resulting S value differs from the printed optimum value by function lhs.design for
type optimum in two respects: the printed optimum value is the sum of inverse distances,
i.e. the denominator of the harmonic mean. choose(nruns, 2) divided by the printed final optimal value
is approximately equal to the calculated S; perfect equality cannot be achieved
because the underlying the printed final optimum refers to an interim latin hypercube of integers
that is subsequently rescaled to the unit cube and scrambled by random numbers.
Function compare returns a matrix, with rows representing the criteria
and columns the different designs. Apart from many of the criteria mentioned above, the
determinant of the correlation matrix is shown, which should ideally be close to one for
a near-orthogonal design (at least in terms of linear effects).
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Beachkofski, B., Grandhi, R. (2002) Improved Distributed Hypercube Sampling. American Institute of Aeronautics and Astronautics Paper 1274.
Currin C., Mitchell T., Morris M. and Ylvisaker D. (1991) Bayesian Prediction of Deterministic Functions With Applications to the Design and Analysis of Computer Experiments, Journal of the American Statistical Association 86, 953–963.
Santner T.J., Williams B.J. and Notz W.I. (2003) The Design and Analysis of Computer Experiments, Springer, 121–161.
Shewry, M. C. and Wynn and H. P. (1987) Maximum entropy sampling. Journal of Applied Statistics 14, 165–170.
Fang K.-T., Li R. and Sudjianto A. (2006) Design and Modeling for Computer Experiments, Chapman & Hall.
Stein, M. (1987) Large Sample Properties of Simulations Using Latin Hypercube Sampling. Technometrics 29, 143–151.
Stocki, R. (2005) A method to improve design reliability using optimal Latin hypercube sampling. Computer Assisted Mechanics and Engineering Sciences 12, 87–105.
See Also lhs-package,
DiceDesign-package,
eval.design,
eval_design
## optimum design from package lhs (default) plan <- lhs.design(20,4,"optimum", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## maximin design plan2 <- lhs.design(20,4,"maximin", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## purely random design (usually not ideal) plan3 <- lhs.design(20,4,"random", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## genetic design plan4 <- lhs.design(20,4,"genetic", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## dmax design from package DiceDesign ## arguments range and niter_max are required ## ?dmaxDesign for more info plan5 <- lhs.design(20,4,"dmax", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, range=0.2, niter_max=500) ## Strauss design from package DiceDesign ## argument RND is required ## ?straussDesign for more info plan6 <- lhs.design(20,4,"strauss", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, RND = 0.2) ## compare all these designs compare(plan, plan2, plan3, plan4, plan5, plan6)## optimum design from package lhs (default) plan <- lhs.design(20,4,"optimum", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## maximin design plan2 <- lhs.design(20,4,"maximin", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## purely random design (usually not ideal) plan3 <- lhs.design(20,4,"random", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## genetic design plan4 <- lhs.design(20,4,"genetic", factor.names=list(c(15,25), c(10,90), c(0,120), c(12,24)), digits=2) ## dmax design from package DiceDesign ## arguments range and niter_max are required ## ?dmaxDesign for more info plan5 <- lhs.design(20,4,"dmax", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, range=0.2, niter_max=500) ## Strauss design from package DiceDesign ## argument RND is required ## ?straussDesign for more info plan6 <- lhs.design(20,4,"strauss", factor.names=list(torque=c(10,14),friction=c(25,35), temperature=c(-5,35),pressure=c(20,50)),digits=2, RND = 0.2) ## compare all these designs compare(plan, plan2, plan3, plan4, plan5, plan6)
These functions support response surface analysis with package rsm. Function rsmformula creates a model formula for use with function rsm, using the FO, TWI and PQ notation. Function code.design prepares a fractional factorial 2-level design with center points from package FrF2 or a ccd, bbd or lhs design from this package for convenient use with package rsm functionality, function decode.design reverses the coding action.
code.design(design) decode.design(design) rsmformula(design, response=NULL, factor.names=NULL, use.blockvar = TRUE, degree=2, coded=TRUE, ...)code.design(design) decode.design(design) rsmformula(design, response=NULL, factor.names=NULL, use.blockvar = TRUE, degree=2, coded=TRUE, ...)
design |
a response surface design of class |
response |
character string specifying the response variable to be analysed (default: the first response of the design) |
factor.names |
character vector specifying the factors to be included (default: all experimental factors) |
use.blockvar |
logical indicating whether or not the block effect (if available) is to be included into the model |
degree |
default is 2 for a full second order model.
For a first order only model, specify 1;
for a model with main effects and 2-factor interactions, specify 1.5.
|
coded |
logical indicating whether the formula is to be provided in
coded names (x1, x2 etc., |
... |
reserved for future usage |
Function code.design rescales the variables of a design with quantitative
variables according to the information stored in the coding element of
the design.info attribute of design,
function decode.design rescales a coded design to original units.
Function rsmformula creates a formula for use with function rsm.
If this function is created in coded variable names (which is the default),
it can be used in function rsm
together with the coded data object created by function code.design
for creating a response surface model, which can be post-processed by the utilities provided
in package rsm, especially the [rsm:rsm]{methods} for class rsm objects
and functions steepest or canonical.path.
IMPORTANT: coded vs. original units
The text below assumes that the design has been entered using the default.levels or the
factor.names option to specify the factor levels in original units.
The usual steepest ascent analysis is done
in coded units, i.e. if e.g. factor X1 has original units 10 (code -1 = (10-20)/10)
and 30 (code +1 = (30-20)/10)
and factor X2 has original units 0.1 (code -1 = (0.1 - 0.2)/0.1)
and 0.3 (code +1 = (0.3 - 0.2)/0.1),
an increase of 10 for a change in factor X1 from 10 to 30 is considered
steeper (slope 10/2) than an increase of 9 for a change in factor X2
from 0.1 to 0.3 (slope 9/2). If this behavior is
desired, usage of rsmformula with option coded=TRUE and a design generated
by code.design is needed.
Otherwise, i.e. when assessment is desired
in original units, the ascent for factor X2 (9/0.2) would of course be much steeper
than for factor X1 (10/20) in the above example. For obtaining an assessment based
on the original units, one can simply use rsmformula with option coded=FALSE
and the design itself in original units in the
rsm model. This only makes sense for first order models:
function steepest always assesses the slope at the origin;
for first order models, it does not matter where the slope is assessed. For models
with order (=degree) 1.5 or 2, the steepest analysis in original units
is adequate only for the exceptional case that 0 is the point of interest.
Function code.design returns a coded.data object
for usage with function rsm; this object can be returned to its
original state by applying function decode.design.
Function rsmformula returns a formula with an FO (=first order) portion,
for degree=1.5 additionally a TWI (=two factor interactions, 2fis) portion,
and for degree=2 also a PQ (=pure quadratic) portion.
This representation of the model formula is needed for response surface analyses
with package rsm. Per default, the formula comes
in coded variable names (x1, x2 etc.).
This package is still under (slow) development. Reports about bugs and inconveniences are welcome.
Ulrike Groemping
Lenth, R.V. (2009). Response-Surface Methods in R, using rsm. Journal of Statistical Software 32(7), 1-17. URLhttp://www.jstatsoft.org/v32/i07/.
Myers, R.H., Montgomery, D.C. and Anderson-Cook, C.M. (2009). Response Surface Methodology. Process and Product Optimization Using Designed Experiments. Wiley, New York.
See also rsm,
steepest, canonical.path,
contour.lm,
.
The formula method for class design objects
creates equivalent model formulae in standard model notation.
## an artificial example with random response ## purely for demonstrating how the functions work together with rsm plan <- ccd.design(5, ncenter=6, factor.names = list(one=c(10,30),two=c(1,5),three=c(0.1,0.9), four=c(2,4),five=c(-1,1))) set.seed(298) plan <- add.response(plan, rnorm(38)) ## coding plan.c <- code.design(plan) plan.c decode.design(plan.c) ## first order analysis ## formulae needed for first order models: rsmformula(plan, degree=1) ## coded rsmformula(plan, degree=1, coded=FALSE) ## original units ## steepest ascent: steepness assessed in coded units, ## results also presented in original units linmod1 <- rsm(rsmformula(plan, degree=1), data=plan.c) summary(linmod1) steepest(linmod1) ## steepest ascent: steepness assessed in original units!!! ## this is different from the usual approach!!! ## cf. explanation in Details section linmod1.original <- rsm(rsmformula(plan, degree=1, coded=FALSE), data=plan) summary(linmod1.original) steepest(linmod1.original) ## second order analysis (including quadratic, degree=1.5 would omit quadratic ## formulae needed for second order models: rsmformula(plan, degree=2) ## coded rsmformula(plan, degree=2, coded=FALSE) ## original units ## the formulae can also be constructed analogously to the FO formulae ## by using SO instead of FO ## rsmformula returns the more detailed function because ## it can be more easily modified to omit one of the effects ## the stationary point is not affected by using coded or original units ## neither is the decision about the nature of the stationary point ## a subsequent canonical path analysis will however be affected, ## analogously to the steepest ascent (cf. Details section) ## analysis in coded units linmod2 <- rsm(rsmformula(plan, degree=2), data=plan.c) summary(linmod2) ## analysis in original units linmod2.original <- rsm(rsmformula(plan, degree=2, coded=FALSE), data=plan) summary(linmod2.original) ## the contour plot may be nicer when using original units contour(linmod2, form=~x1*x2) contour(linmod2.original, form=~one*two) ## the canonical path is usually more reasonable in coded units canonical.path(linmod2) ## coded units canonical.path(linmod2.original) ## original units ## analogous analysis without the special formula notation of function rsm linmod <- rsm(rnorm.38. ~ Block.ccd + (one + two + three + four + five)^2 + I(one^2) + I(two^2) + I(three^2) + I(four^2) + I(five^2), data=plan) summary(linmod) contour(linmod, form=~one*two) ## contour plot is possible ## steepest or canonical.path cannot be used, ## because the model is a conventional lm ## contour will not work on the convenience model ## lm(plan), which is otherwise identical to linmod ## (it will neither work on lm(formula(plan), plan)) ## or lm(rsmformula(plan), plan)## an artificial example with random response ## purely for demonstrating how the functions work together with rsm plan <- ccd.design(5, ncenter=6, factor.names = list(one=c(10,30),two=c(1,5),three=c(0.1,0.9), four=c(2,4),five=c(-1,1))) set.seed(298) plan <- add.response(plan, rnorm(38)) ## coding plan.c <- code.design(plan) plan.c decode.design(plan.c) ## first order analysis ## formulae needed for first order models: rsmformula(plan, degree=1) ## coded rsmformula(plan, degree=1, coded=FALSE) ## original units ## steepest ascent: steepness assessed in coded units, ## results also presented in original units linmod1 <- rsm(rsmformula(plan, degree=1), data=plan.c) summary(linmod1) steepest(linmod1) ## steepest ascent: steepness assessed in original units!!! ## this is different from the usual approach!!! ## cf. explanation in Details section linmod1.original <- rsm(rsmformula(plan, degree=1, coded=FALSE), data=plan) summary(linmod1.original) steepest(linmod1.original) ## second order analysis (including quadratic, degree=1.5 would omit quadratic ## formulae needed for second order models: rsmformula(plan, degree=2) ## coded rsmformula(plan, degree=2, coded=FALSE) ## original units ## the formulae can also be constructed analogously to the FO formulae ## by using SO instead of FO ## rsmformula returns the more detailed function because ## it can be more easily modified to omit one of the effects ## the stationary point is not affected by using coded or original units ## neither is the decision about the nature of the stationary point ## a subsequent canonical path analysis will however be affected, ## analogously to the steepest ascent (cf. Details section) ## analysis in coded units linmod2 <- rsm(rsmformula(plan, degree=2), data=plan.c) summary(linmod2) ## analysis in original units linmod2.original <- rsm(rsmformula(plan, degree=2, coded=FALSE), data=plan) summary(linmod2.original) ## the contour plot may be nicer when using original units contour(linmod2, form=~x1*x2) contour(linmod2.original, form=~one*two) ## the canonical path is usually more reasonable in coded units canonical.path(linmod2) ## coded units canonical.path(linmod2.original) ## original units ## analogous analysis without the special formula notation of function rsm linmod <- rsm(rnorm.38. ~ Block.ccd + (one + two + three + four + five)^2 + I(one^2) + I(two^2) + I(three^2) + I(four^2) + I(five^2), data=plan) summary(linmod) contour(linmod, form=~one*two) ## contour plot is possible ## steepest or canonical.path cannot be used, ## because the model is a conventional lm ## contour will not work on the convenience model ## lm(plan), which is otherwise identical to linmod ## (it will neither work on lm(formula(plan), plan)) ## or lm(rsmformula(plan), plan)