Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Cap 8

Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

8

EXCHANGE ALGORITHMS
Methods for Non-Standard Designs

I. Introduction to exchange algorithms


Recognising the exchange algorithm situation
II. Fundamental properties
Choice of mathematical model and model discrimination
Candidate points
Quality criteria for an experimental design
Principle of the exchange algorithm method
III. Examples of exchange algorithms
Process study for a wet granulation
Another wet granulation process study
A screening design
D-optimal designs at 2 levels for factor influence studies

I. INTRODUCTION TO EXCHANGE ALGORITHMS

It is only during the last 10 years that the techniques described in this chapter have
begun to be widely used by experimenters. Even now they are often considered
difficult to use and of marginal utility. We intend to show that this prejudice is
unjustified and we will demonstrate, with the help of several examples, that,
because of their flexibility and general applicability, they have an important place
in the formulator's and developer's tool kit along with the more traditional and
better known methods. The long and complex calculations needed are taken care
of by specialised computer programs, of which a large number are now available.

A. Recognising the Exchange Algorithm Situation

The so-called "classical designs" exist independently of the problem being treated.
They consist of a set of points in factor space, their arrangement being optimal for

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


the determination of a given type of model. So in order to use these designs we
need to carry out a projection between the numerical values corresponding to the
coded variables (these are the formal values of the design) and the real levels of the
true factors, corresponding to the experimental conditions. Thus this projection
allows a temperature of 40°C to be associated with a value -1 of a factorial design,
for example, or a mixer rotation speed of 1200 rpm to correspond to a level 0.866
in a Doehlert design, etc.
The main problem with these designs is that they do not take into account
specific experimental constraints which may render them inefficient or impossible
to use. We will now outline various situations where this might occur. Exchange
algorithms are used in all of these.

1. Constraints on the experimental domain

The domains so far studied have been symmetrical - cubes, spheres and regular
simplexes. Classical designs are used, provided that there are no limitations on the
experimental conditions that prevent one or more of the experiments being carried
out. It is by no means unusual that certain combinations of values of the
experimental conditions are excluded for reasons of cost, stability, feasibility, or
simply interest to the researcher. In a tableting experiment there might be whole
regions where tablets cannot be obtained because of incompatibility with the
machine, lack of cohesion, etc. Thus the experimental domain is transformed from
a regular polyhedron to a more or less irregular convex one, in which the
experiments may no longer be placed according to a standard design. Figure 8.1
shows two examples of this, where the initially regular domains [a cube in case (a),
a simplex for a mixture in case (b)] have been reduced by the constraints to
irregular shapes. It is of course possible to construct standard designs within these

(a) (b)

Figure 8.1 Experimental domains with constraints.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


domains, but then considerable portions of the design space would remain outside
the domain of the new design. The problem is well illustrated by Snee (1).
The use of exchange algorithms to construct D-optimal designs for mixtures
with constraints is described in some detail in chapter 10.

2. Discontinuous experimental domain

Quantitative factors may sometimes be allowed only to take certain levels, not
always evenly spaced. Examples are sieve sizes, volumes of mixer-granulators,
speed controls, etc. These factors may be treated as qualitative, but this would cause
potentially interesting information to be lost to the analysis. If these levels do not
correspond to those of the standard designs, the methods of chapter 6 (RSM) would
no longer be applicable.

3. Limitations on combinations of levels

Only certain combinations of levels of some of the variables may be possible. Two
examples of this are given later in this chapter, where there were 5 factors in a
process study, but only 6 combinations of the variables X{, X2 were allowed. This
precluded all standard RSM designs.

4. Limited number of experiments

For reasons of cost, time, or amount of material available, the number of possible
experiments might be limited, and the standard design might require too many. We
show examples of this for a process study and for a screening study.

5. Incorporation of previous experiments

When preliminary experiments have already been done (under reliable conditions),
and when the cost of an experiment is very important, it is often a more economical
and better strategy to build up a design by adding experiments to those already
carried out, rather than to start completely afresh with a standard design. All the
same, it must be noted that the results of the complementary experiments are
assumed to be homogenous with the preliminary ones. Otherwise, a block effect
variable must be introduced into the model.

6. Use of an "incomplete" mathematical model

So far we have generally assumed complete mathematical models: first-order,


synergistic, or second-order polynomials. In chapter 5 we supposed either that the
model contained only first-degree terms, or it included all the second-order terms
(square and rectangular), and the designs then proposed corresponded to those
models. The experimenter may be led to postulate the presence or absence of
certain terms in the model. For example, during a preliminary study he may have
tested a factor at several levels, keeping other conditions constant, and seen a
curvature in the response. He would then postulate a second-degree term in that
variable. On the other hand, by analogy with related studies, he might have good

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


reason to expect that another factor has a purely linear effect. He would then
include only this first-order term in the model. The number of terms in the model
is therefore less than in the second-order model and the use of classical
experimental designs thus becomes expensive, with respect to the limited number
of coefficients in the model to be determined.
We saw in chapter 7 that models for minimizing variability include
interaction terms between noise variables and control variables, but omit
interactions between two noise variables. They also are incomplete models. The
experimenter may occasionally wish to use a "mechanistic", non-polynomial model,
one resulting from, or inspired by a thermodynamic, or kinetic equation.

7. Complement to a design

In spite of careful reflection and discussion before carrying out the experiments, it
may well happen that the phenomenon appears more complex than was initially
believed. Interaction or curvature effects that were not considered significant a
priori, may appear during or after the experimentation. A more complex
mathematical model must then be postulated and it is then almost invariably
necessary to add the most pertinent (informative) experiment(s) to those carried out
already. As we have already mentioned, one would not normally consider starting
afresh.
An example is the addition of third-order terms to a second-order model
(chapter 5, section VII.B, and chapter 7, section III.A.2).

8. Repairing a design

The most careful attention must be paid to the choice of experimental domain, to
the levels of each factor, and to their ranges of variation. The whole experimental
plan (expressed in terms of the real factor levels and not the coded variables) must
be examined closely before starting the experiments, in case certain experimental
combinations are not feasible. In fact it is often wise to carry out any "doubtful"
experiment first.
In spite of this, one or more experiments may be found to be impossible, or
the results unreliable. In general, an experimental design can only be analysed if the
whole of it has been carried out. Lacking a single result it could be totally unusable.
There are statistical methods for replacing missing values under certain
circumstances, but there is no treatment which can really replace an experimental
datum. A design is repaired by adding one or several experiments that provide the
same quantity of information as the failed experiment would have done (1).

II. FUNDAMENTAL PROPERTIES

The objective of using an exchange algorithm is to construct the "best" design


within the proposed experimental domain for determining the proposed

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


mathematical model. We will first clarify and quantify these intuitive ideas.

A. Choice of Mathematical Model and Model Discrimination

The experimenter does not usually know which is the most adequate mathematical
model — that which best represents the phenomenon being studied in the
experimental domain. From time to time, his theoretical knowledge might lead him
to postulate certain special models (kinetic, mechanical, or thermodynamic models,
for example). He must also take into account any partial knowledge that he has on
the system, to postulate a curvature or a particular interaction between two factors.
And by analogy with other systems already tested, he may put forward models
which "work" in similar cases. Otherwise he may continue to postulate one of the
empirical models we have used before, the first-order, second-order, or synergistic
model, according to the circumstances. The models used for mixtures and
formulation problems will be introduced in the following chapters.

There is no universal experimental design allowing us to carry out the


experiments and then to select the best mathematical model.

The only exception to this rule would be to cover the domain with a
sufficiently dense network of experiments, but this leads very rapidly to an
excessive number. If s is the number of levels to be tested on each axis, and k the
number of factors, it requires a complete factorial sk design.
It is an illusion to imagine that one might choose the mathematical model
on the basis of experiments carried out without any mathematical structure. The
question is not, in general, one of finding a mathematical model which represents
the experimental results already obtained, but rather to choose a model to treat the
experiment that one intends to carry out.
We must therefore select experiments that are the most representative of all
the possible experiments within the domain and which provide the maximum of
information about the coefficients of the model. Now, it is well known that not all
experiments provide the same amount of information. For example, the centre
points provide very little information for a first-degree model, points midway
between two axes are rich in information on the first-order interactions of two
factors, etc.

To be optimal, a design must be constructed after choosing the mathematical


model.

And, as we have already seen a number of times, the experimenter does not
know what will be the best model. He therefore has the following dilemma: to set
up the design he must postulate a model, but if he makes the wrong choice of
model, the design loses much of its interest.
Techniques exist allowing identification of the experiments which provide
the most information distinguishing the various models. These would therefore

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


allow us to postulate the best model. They do not allow that model to be
determined; for this the design must be completed by carrying out complementary
experiments. These model discrimination methods (2, 3) are as yet little used and
we will not discuss them here. The computer programs they use are not yet of very
high performance, nor are they available generally.
There remains the traditional solution to the problem, adopted here: to
postulate several models, the most likely ones, and either:
• determine the best design for each and combine them or try to find a
compromise between them, or
• construct the optimum design for the most probable model and complete it
so that the other models may be determined satisfactorily.

B. Candidate Points

Once the mathematical model has been postulated, there are two possible
approaches.

• Use an algorithm to place the experiments directly in that position of the


factor space where they are expected to provide the most information.
Programs for these quite recent methods - genetic algorithms (4), simulated
annealing (5, 6, 7) - are again not widely available and we will not
describe their use here.
• Cover the domain with an assembly of possible experiments, called
candidate points or candidate experiments, from which the algorithm
chooses the "best" ones. The simplest solution is to set up a network or
lattice, more or less dense, according to the required precision and the
means of calculation available. Fortunately, it is possible to set up this
network intelligently, allowing us to impose the properties we seek on the
final solution and at the same time limiting the number of candidate points
and the amount of computation required (8).

C. Quality Criteria for an Experimental Design

The main quality criterion for choosing a design is that of D-optimality. This is
combined with the R-efficiency and normalized to give the D-efficiency. However
there are other properties that are useful for assessing the design quality, though
they are not used directly in determining the design.

1. D-optimality (8, 9, 10, 11)

In chapter 4, section III.C.3, we introduced the calculation of the confidence


interval for calculating the coefficient. We recollect that:

b,-tvMxsbi < (3, < fc; + rVia/2 x sw (8.1)

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


This widely used formula is only true for (3, if all the other estimations bj are equal
to P;. Figure 8.2a shows the confidence interval calculated for 2 coefficients p, and
P2 according to the formula 8.1, for a given value of a. More generally, a joint
confidence region must be considered. For a model with 2 coefficients this is an
ellipse (figure 8.2b) and for more coefficients it is an ellipsoid (the value of a being
assumed to be the same for all coefficients).

"least-squares
solution
(b !( b 2 )
>1 t I

individual 1
confidence
interval
for p.,
(a given)

.* , _._,. __ w ft
joint
confidence confidence
interval interval
for p, for (pv p2)
(a given) (a given)

Figure 8.2 (a) Individual confidence limits, (b) Joint


confidence limits or confidence ellipse (ellipsoid).

The smaller the volume of the confidence ellipsoid, the more precise will
be the estimates of the coefficients. It may be shown that, for a significance level
a and a given experimental variance, this volume is proportional to the determinant
I (X'X)"1 I of the dispersion matrix, defined in chapter 4, section II.C.4.

An experimental design of N experiments is said to be D optimal within a given


domain if the determinant of the dispersion matrix I (X'X)"' | is minimal, and thus
the coefficients are estimated with maximum overall precision.

The determinant of a matrix is equal to the reciprocal of the determinant of


its inverse, and so:

|X'X|=.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


This property is a particularly useful one because it enables the quality of two
designs to be compared, by comparing the determinants of the information matrices
X'X, without inverting them. This leads to a considerable saving in computing time.

2. R-efficiency

We recollect that the R-efficiency is the ratio of the number of coefficients of the
model to be determined, p, to the number of experiments in the design, N:
RR - P
«« - 77
This function is not particularly useful on its own. The objective in experimental
design is not to do the minimum number of experiments, but to use the available
resources in carrying out each experiment as effectively and as efficiently as
possible, and thus to arrive at an optimum use of the allocated budget. There is no
point in doing a minimum number of experiments, if the results obtained are
unreliable, imprecise, and do not properly answer the question.
Let us consider a design of N experiments, to which we add one experiment,
giving a design of N + 1 experiments. The determinant of the information matrix
of the second design is greater than that of the first design.

'JV+

Simply adding another experiment therefore leads to more precise overall estimates
of the coefficients, because the volume of the confidence interval ellipsoid is
diminished. So one method of improving a design is to add experiments to it.
However these designs are not strictly comparable. Because the one design has one
more experiments its cost is greater, and we may in effect ask ourselves whether
it is worth carrying out that extra experiment. This depends on whether the
information provided by that one experiment is great or small. When we wish to
compare designs with different numbers of experiments (in the same domain) we
"combine" the criteria of D-optimality and R-efficiency by the D-efficiency.

3. D-efficiency

The moments matrix M is defined by:

M = W x X'X

and its determinant is therefore:

|M| = ix'xl/w
A given design A has a greater D-efficiency than a design B if I MA I > I MB \ .
This property may be considered as a description of the design's "value for
money".

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


To the above we may add several other criteria (10). They are not used
directly for the construction of experimental designs in the way that the D-optimal
and the D-efficiency criteria are used, but they are often useful for comparison once
several D-optimal solutions, of the same or different numbers of experiments, have
been found.

4. A-optimality

A-optimality refers to the Average variance of parameters.


The confidence ellipsoid is characterised by its volume I (X'X)" 1 1, but also
by its shape - whether it is flattened or not and its direction with respect to the
axes bt. The trace of the dispersion matrix (X'X)"1 is the sum of its diagonal
elements £c". For a given volume of ellipsoid (given by the determinant), the trace
takes its minimum value when all the terms c" are equal, corresponding to a
spherical confidence ellipsoid, as in figure 8.3a, where the estimations of the
coefficients are independent (uncorrelated).

Figure 8.3 Confidence ellipsoids

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


Figure 8.3b shows a case where the estimations are uncorrelated, but of very
different precisions. |32 is estimated with a much greater precision than the
coefficient p:.
In figure 8.3c the estimations of the coefficients are correlated, as are the
confidence intervals. This correlation will also become clear on examining the
dispersion matrix, (X'X)"1 by the presence of a non-zero, non-diagonal term, in this
case c12 > 0. See chapter 4 for an example.
For equal I (X'X)"1 I or I X'X I, we may select the design with minimum
rr(X'X)-1.

5. G-optimality

This refers to the General variance.


If the mathematical model has been determined, it is possible to calculate
the theoretical value of the response and its confidence interval for every point in
the domain. We saw in chapter 5 that dA, the variance function at any point A,
along with the experimental variance a2, describes the precision of prediction at that
point. dA depends only on the model and the experimental design and no measured
response data are required. Within the domain there is at least one point where the
precision of prediction of the response is the worst of all, so the variance function
is maximum at this point: dmiK. A design is called G-optimal when Amax is as small
as possible.
In practice, it may be difficult to find the exact point where the variance
function is maximal and it is more usual to investigate only the candidate points,
and find which of these has the highest variance function. Another solution that is
sometimes adopted is to calculate the variance function at a large number of points
selected randomly within the domain.
The G-efficiency is given by:
r, = P
Nd

D. Principle of the Exchange Algorithm Method

After the mathematical model has been chosen and the candidate experimental
design (that is to say, all the experiments possible within the domain) has been
generated we must be able to answer two questions:

• How many experiments are needed to obtain estimates of the coefficients


as precisely as possible and at the least cost?
• Given we know the answer to the first question, which experiments do we
choose?

It is in fact easier to reply to the second question than to the first! Therefore, we
will first assume the number of experiments in the design to be fixed, equal to N.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


Once we have described how the N experiments are chosen, we will be able to
answer the first question.

1. Initial phase

The first stage is the choice of a sub-set of jV experiments from the set of nc
candidate experiments. Knowing the mathematical model, the information matrix
X'0X0 for this initial TV-experiment design may be calculated and its quality can be
quantified as I X' 0 X 0 1. This initial choice may be random, or using an algorithm to
find an homogeneous distribution of points in factor space (9).

2. Iteration

The objective is to progressively replace experiments within the sub-set so as to


increase the determinant of the information matrix.
Fedorov has shown how it is possible to choose a pair of points (A, B), A
being within the design and B a candidate point, so that their exchange gives rise
to a maximum increase in the determinant of the information matrix (9, 11). The
details of the calculation are outside the scope of this book. Other methods for
optimizing IX'jXj I have been developed by Mitchell (12) and Wynn (13).
If I X'jXj I > I X' 0 X 0 1 the exchange of points is accepted and the iteration
is begun again, at the end of which a new exchange is carried out.

3. Stopping condition

If at step/, IX'jX, I = I X'UXM I, that is the determinant has stopped increasing,


the iteration is stopped.

4. Validation of the solution found

As is the case for many iterative methods where the starting point is chosen
randomly, the final solution may depend on the initial choice. Although the
algorithm is monotone, with the determinant increasing, it may still converge to a
local maximum. It is then necessary to begin the treatment again, starting from
several initial random starting choices.
If the solutions found are identical, or equivalent (that is, different
experiments but each design having the same determinant) this increases the
probability that the maximum that is found is the highest possible value. The design
of the final solution is therefore the D-optimal design.
If the solutions are very different, this may indicate an instability, which
must be analysed.

5. Protected points

Certain points in the candidate design are protected. This means that they are
selected for each run and they are maintained in the design throughout the iteration,

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


not being replaced. This is normally because:

• the experiments have already been carried out as preliminary exploratory


experiments, (see section I.A.5);
• they have been done, or it is intended to do them, as a preliminary design
(see section I.A.7 and also chapter 10),
• they are part of a (standard) design, also already carried out, but requiring
"repair", either because of experiments which proved impossible, or because
of the need to introduce extra terms in the model (see section I.A.8).

6. Exhaustive and non-exhaustive selection of points

Some computer programs give the possibility of 2 modes for choosing experiments
from the candidate design (8).

Non-exhaustive selection
Once any point has been chosen, it remains within the candidate design and can
therefore be selected a second or a third time. By this means the exchange
algorithm gives the design with maximum I X'X I without any further consideration.
When the number of experiments is greater than the number of coefficients in the
model, N > p, it is highly probable that some experiments will be replicated within
the final D-optimal design. It is quite frequent that when, for example N = 2p, the
final design consists of p duplicated experiments. Use of such a design and analysis
by regression and ANOVA will give an unbiased estimation of the experimental
variance. However the design is likely to be either saturated or nearly saturated,
there being only as many distinct experiments as there are coefficients (or just a
few more). It will not be possible to test for the goodness of fit of the model, as
there can be no estimate of the variance by the regression.
The method is also useful where a number replications of each experiment
is required for reasons of precison, for example, in vivo experiments on
formulations. The "best" experiments are thus replicated more times than those less
well placed in the design space.

Exhaustive selection of points


Here any point chosen is eliminated from the candidate design. Thus all points in
the final design are distinct from one another. If N > p, the determinant of I X'X I
will often be smaller than that obtained by non-exhaustive selection. Analysis of
variance allows estimation of the variance about the regression equation, but the
validity of the model cannot be tested because of the lack of replicated points. The
model validity (lack of fit) can only be tested provided there are more distinct
points than coefficients, and some points are repeated.
This mode of running the exchange algorithm is especially recommended
when the experimenter suspects that not all points belong to the same population,
fearing a possible discontinuity or unstable zone. In such a case there is no point
in giving excessive weight to experimental points which may, theoretically, be rich
in information according to the information matrix | X'X |, but for which that

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


information may be biased.

Combined selection
As elsewhere, it is generally better to find a compromise between the two solutions.
A possible approach, allowing estimation of the experimental error, the significance
of the regression and testing of the model's goodness of fit would be:

• non-exhaustive selection of enough points to allow a certain number of


repetitions. Select for example N large enough (at least p + 5) to have 5
duplicated points in the design, to allow estimation of the pure error
variance with 5 degrees of freedom,
• then protect the duplicated points, eliminating them from the candidate
design, and continue the exchange algorithm in the exhaustive choice mode,
to ensure that more distinct candidate points are chosen within the domain.

III. EXAMPLES OF EXCHANGE ALGORITHMS

A. Process Study for a Wet Granulation

We illustrate the setting up of a D-optimal design with the example of a process


study (14) already discussed at length in chapter 6, where the experimental results
of this example were analysed. We will now show how the design itself was
obtained. The process was a wet granulation, which was followed by drying,
sieving, lubrification, and compression (see table 6.2).

1. Constraints imposed by the problem

There are three reasons why this problem could not be solved using a standard
design, and required the use of an exchange algorithm.

• The factors X^ (residual humidity) and X4 (sieve size) took only 2 levels and
could be treated as qualitative factors. The mathematical model could not
contain square terms in these factors.
• Even apart from this, the mathematical model (equation 6.2) was not a full
second-order one, as only 2 interactions, XtX2 and X^XS, were postulated.

y = Po + Pl*l + 02*2 + P3-*3 + P**4 + P5*5 + Pi A" + (W + P55*52

The possible number of experiments was limited because a fixed quantity


of active substance, a 30 kg batch, was available for the study. Since the
product contained 33% active substance, this allowed 90 kg to be
formulated. The mixer-granulator required about 15 kg of product for each
granulation, so no more than 6 granulations could be carried out. On the

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


other hand, lubrification could be done on a scale of only 5 kg. Each batch
of 15 kg of granules might therefore be divided into 3 parts (sub-batches)
for drying and sieving, followed by lubrification. Thus 18 experiments might
be carried out, but with certain restrictions in the combinations of levels.

2. Candidate points

The construction of the candidate design is very important. An apparently simple


solution is to postulate as many experiments as possible. The humidity X^ and the
sieve size X4 are limited to 2 levels each, but the remaining 3 factors are continuous
and may take any number of levels. If we allow 10 equidistant levels for each, there
will be a total of 4000 combinations of factors. Between 11 and 18 experiments are
to be selected from these 4000 possible ones. Under these circumstances, it is by
no means certain that the algorithm would converge to the optimum solution,
though it would probably be close. It is probable that certain programs will not
accept so many data, and the calculation times may be long, particularly if we wish
to determine several solutions for different numbers of experiments. The solution
may not respect the different constraints. This approach should therefore not be
used.
It is far better to propose only the candidate points which satisfy the
conditions, and that is what we shall do here. Consider the factors Xl and X2
(granulation time and added liquid). The optimal design can only contain 6 different
combinations (since only 6 granulations may be carried out). So the candidate
design itself should contain only 6 combinations. Also, the model is second-order
with respect to those 2 variables, so we must find an experimental design allowing
us to determine a second-order model, with only 6 separate combinations of Xl and
X2. Only two of the designs described in chapter 5 may be used. The first is a
design of 6 experiments, for the cubic domain (BD206 in table 5.28, see appendix
III). The second is an equiradial design, a regular pentagon with one centre point
(table 5.8, column 3, and figure 5.5), and it is this one that was chosen.
The effect of the lubrification time X5 includes a square term, so it must be
set at 3 levels. The humidity X3 and the sieve size X4 are each allowed 2 levels.
Thus each granulation batch is divided into 3 sub-batches of 4 kg, for the remainder
of the processing, each with a different combination of levels of X3, X4, and X5.
There are 12 possible treatments for each of the 6 granulations. The candidate
design is therefore the product of the pentagonal design and a 223' factorial design
and contains 72 experiments.

3. D-optimal design

The number of experiments is fixed at 18, the maximum possible. The D-optimal
design thus obtained by running the exchange algorithm is described in table 6.3,
and in figure 8.4. The experimental results for each run are given in chapter 6 (see
table 6.3), where they were used for determination of the model and optimization
of the process.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


Figure 8.4 D-optimal design for wet granulation process study.

4. Determination of the optimum number of experiments

We have shown the principle of the exchange algorithm in paragraph II.D, for a
fixed number of experiments. The next question to be answered is: how do we
decide on the number of experiments? Subject to possible external constraints, this
depends on the statistical properties of the design. We demonstrate this by
determining D-optimal designs for different values ofN, in the above process study,
and examine trends in their properties.
There are 11 coefficients in the model so the design must contain at least
11 experiments. The maximum number of experiments is 18. D-optimal designs

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


were therefore determined for 11 to 18 experiments. Twenty D-optimal designs
were calculated for each value of N, each starting from a different random selection
from the 72 experiments. The "best" design (that with the highest value of I X'X I)
was selected in each case. The trends in the design criteria are shown in figures
8.5a-d, I X'X I and I M I being normalized as I X'X I '* and I M 1 1/p respectively,
where p is the number of coefficients in the model.

/•.
k | XX | i , IM |

^
„-•• ''
0.38 _i
5
0.36 h~-
--" (
u^^- ———

034

11 12 13 14
J
15 16 17 18
N
0.32

11
/

12 13 14 15 16 17
-*•
18
N

(a) (b )
,
> tr (X tI ^
\
10 '• \

\
'-.,.^ ,_......
N *— ._, ^
1——— ..-H

'~---— v
? 1———————
N
0
1 12 13 14 15 16 7 18 •1 12 13 14 15 16 17

(c) (d

Figure 8.5 Design properties : (a) I X'X 1 1/n ; (b) I M 1 1/n ; (c) tr (X'X)''; and
(d) dma,

The determinant of the information matrix \ X'X I


Figure 8.5a shows a steady increase in I X'X I with increasing N. Addition of any
point to a design increases the determinant of the information matrix. If the decision
criterion is that the absolute precision of the coefficient must be maximized, the
determinant I X'X I must be maximal and 18 experiments must be carried out.

The determinant of the moments matrix \ M I


I M I represents the quantity of information per experiment. Figure 8.5b shows it
also increasing with N, but it is clear that on going from 11 to 12 experiments the
information brought in by each experiment is considerably increased. The value
then continues to increase from N = 12 to 18. We conclude that at N = 18 each
experiment provides the maximum of information for determining the coefficients.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


The trace of the dispersion matrix tr (X'X)'
Figure 8.5c shows the trace of (X'X)"', diminishing by a factor of 2, when going
from 11 to 12 experiments. The one additional experiment greatly improves the
precision. Going from 12 to 18 experiments leads to very little further
improvement. This signifies that the sum of the variance functions c" stays constant,
whereas the volume of the confidence ellipsoid, defined by I X'X I, increases (figure
8.5a). It seems that the ellipsoid becomes more spherical and thus the precision of
estimation of the coefficients becomes more homogenous and the estimations
become more independent (orthogonal). The addition of experiments 12 to 18
improves the symmetry of the distribution of experiments chosen within the
experimental domain.

The variance function dmai


Amax also diminishes very considerably between 11 and 12 experiments and remains
practically constant up to N = 17 (figure 8.5d). With the addition of the 18th
experiment we see a further decrease to a value of less than 1, which means that
the variance of prediction of the model is equal to the experimental variance, since
var(yA) = dX-

5. Problems analysing the "split-batch" design

The part of the model describing the effect of the granulation parameters in this
problem is determined by a saturated design. Thus, there are no degrees of freedom
for determining its coefficients.
The 18 experiments are not independent. The errors in each experiment may
be attributed partly to the granulation (and these are therefore considered the same
in each of the 3 sub-batches of a given granulation batch) and partly to the
succeeding operations and measurements. These latter errors are probably
independent. Analysis of variance, therefore, gives an error that is representative
only of the post-granulation steps in the process. It is for this reason, that although
multi-linear regression is used correctly to estimate best values for the coefficients,
analysis of variance may not be used here to estimate the significance of either the
model or the coefficients for this problem, as already noted in chapter 6.
This is an example of the split-plot design (15), discussed briefly in chapter
7, with reference to correlated errors.

B. Another Wet Granulation Process Study

Another very similar process study was described by Chariot et al. (16). The same
factors were studied and the constraints were very similar. It was possible to carry
out up to 7 granulations and each granulation could be divided into up to 4 sub-
batches. The model was also similar, except that it included one more interaction
term and no square term in the granulation time. The solution chosen by the authors
was to take a 2'3' factorial design for the granulation experiment, and multiply this
by a 223l factorial, as in the previous example, to give a candidate design of 72

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


experiments.
D-optimal designs were determined between 12 and 28 experiments. There
was a maximum in | M I for 22 experiments, so it was this design that was carried
out. For 5 of the 6 combinations of granulating conditions (batches), there were 4
combinations (sub-batches) of the other conditions. The remaining batch was split
into only 2 sub-batches. This is therefore another example of the split-plot design.
If the model had included a square term in X2 it would have been possible
to treat the problem using the pentagonal design, or (since there was material for
7 granulations, although only 6 were in fact done), the hexagonal design, of 7
experiments, or if it was desired to work in a cubic domain, the 6 experiment
design of Box and Draper (BD306), tabulated in appendix III.

C. A Screening Design

We saw in chapter 2 that where factors took different numbers of levels it was not
always possible to obtain a suitable design by collapsing of a symmetrical design
as described in chapter 2, section V.A. This is especially the case where a limited
number of experiments may be carried out. In chapter 2, section V.C we saw how
a 12 experiment design (table 2.25) was proposed for treating a problem in
excipient compatibility screening. There were 4 diluents, 4 lubricants, 3 levels of
binder (including no binder), 2 disintegrants, glidant and no glidant (2 levels),
capsule and no capsule (2 levels) to be studied. A 16-experiment design may be
derived by collapsing A and B. If 16 experiments are too many, a D-optimal design
must be sought.

1. Proposed model

This is a screening model, an additive (first-order) model for 6 factors, where the
s, levels of the i0* qualitative variable are replaced by st - 1 independent (presence-
absence) variables:

y = Po + PlA*lA + PlB*lB + PlC*lC + P2A*2A + P2B*2B + P2C*2C

There are thus 12 coefficients in the model. We therefore construct D-optimal


designs with 12 or more experiments.

2. Candidate design

The full factorial 423'23 design, of 384 experiments ( 4 x 4 x 3 x 2 x 2 x 2 ) ,


consisting of all possible combinations of the levels of the different variables, is a
solution to the problem. This set of experiments is therefore used as candidates,
from which we extract exhaustively a representative sub-set, of 12 to 20
experiments.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


For each number of experiments N = 12 to 20 we determined the final design, to
which the exchange algorithm converged. Figures 8.6a-d show the trends in the four
properties of the design that we have already used. I X'X I and I M | are again
normalised as I X'X I l / p and IM I l l p , where p = 12. All of these show that the most
efficient solution contains 16 experiments:

• The determinant IM I (quantity of information per experiment) is maximal.


• The trace of (X'XV1 remains stable from A' =16.
• The determinant I X'X | increases steadily.
• The value of dmax reaches a value of less than 1 (for each point of the full
factorial design) at N = 16, but increasing the number of experiments
beyond this does not lead to any further decrease in dmax.

1i i M ""
022

..-\
_ A 0.2 1
1
""-I ....-•1

r ,,-- ..-•-• 1
1
/

N N
n is
12 13 14 15 16 17 18 19 20

(a) (b)
• trQC'X)' 1 1 . d™
2.4

^ik%ii_ 20
"^1 *•
1.6
^i — ,
'
1,2

0.8

0.4
N N
nn —fe-
12 13 14 15 16 17 18 19 20

(c) (d)

Figure 8.6 Properties of D-optimal designs for excipient compatibility screening


problem: (a) I X'XP"2; (b) IMP"2; (c) tr (X'X)-'; and (d) dmai.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


In conclusion, a D-optimal screening design of 12 experiments is possible
for treating this problem, but if the resources are available it is better to carry out
16.
Figures 8.7a-b show the distribution of the determinants I X'X I obtained for
the designs resulting from the 30 runs at N = 12 and N = 16. For N = 12, 30 % of
the runs converged to give the same value of I M |. These were probably not
exactly the same designs, but equivalent. The best solution for N = 16 was only
obtained once! Nothing proves that those at W = 12 are the best solutions possible,
but the probability of finding a better solution is small and they are sufficiently
close to one another to be considered acceptable. The best result at N = 16
experiments is identical to the one derived from the 44//42 symmetrical screening
design, by collapsing A followed by B (see chapter 2, table 2.22).

" freq. freq.

1II
2,12 2.16 220 2.24 2.28 2.32 3.18 322 326 3.30 334 3.38
214 2.18 2.22 2,26 2 3 0 2 3 4 320 324 328 332 3.36

(b)

Figure 8.7 Distributions of I X'X 1 1/12 for solutions to the screening problem:
(a) N = 12 and (b) N = 16.

D. D-optimal Designs at 2 Levels for Factor Influence Studies

In chapter 3 we showed that a factorial design of resolution V is needed to


determine the constant term, the main effects and the first order interactions in the
model. We may thus refer to any design as being of resolution V, should it allow
these effects to be estimated, even if it is not a fractional factorial design.

1. Designs for 4 factors

In chapter 3, we gave 3 different saturated or near saturated designs for determining


a second order synergistic model for 4 variables:

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


• a Rechtschaffner design (11 experiments),
• a % design of a 24 full factorial design (12 experiments), and
• a D-optimal design (12 experiments).

We will now show the construction of the D-optimal design, and compare the
different designs.

Postulated model and candidate design


The model is the second order synergistic model of 1 1 coefficients:

y=

The candidate experiments are those of the complete 24 factorial design (16
experiments), which corresponds to the above model. We select the sub-set of the
most representative experiments from these.

D-optimal designs
We have determined the D-optimal designs for 11 to 16 experiments, each
experiment in the candidate design being selected no more than once. There were
10 trials for each value of N, each one starting with a random choice of points.
Each converged to the same solution. Figures 8.8a-b show the trends in I M I and
dmar The trace of (X'X)"1 decreases steadily and I X'X I increases with N.

1
' d™
2.8
1 ..-.._..H
L
2.4
"""-1

2.0

1.6

1.2

0,8

0.4
N

nn
•2 13 M 15 16 12 13 14 15 16

(a) (b)

Figure 8.8 Properties of 2 level D-optimal designs for the 4 factor synergistic
model of degree 2.

All criteria indicate that the full factorial design is the best solution. There is no

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


design more efficient than this. Note that if experiments had been allowed to be
selected more than once from the candidate matrix (non-exhaustive choice), and D-
optimal designs had been constructed for N > 16, the plot of I M I as a function of
TV would have shown a maximum at N = 16.

Comparison of the different designs at 11 and 12 experiments


Table 8.1 shows the characteristics of the different designs constructed. The D-
optimal design for 11 factors is not given, as it is identical to the Rechtschaffner
design (which is thus D-optimal).

Table 8.1 Designs for the 4 Factor Synergistic Model: Statistical Properties and
Optimality Criteria

Design N 1 X'X I "" 1 M 1 1/n rr(X'X)'1 cU VIF


Rechtschaffner 11 9.172 0.834 1.486 2.56 1.52
D-optimal design 12 10.293 0.858 1.313 2.50 1.37 - 1..50
4
% of 2 design 12 10.293 0.858 1.313 2.50 1 - 1.50
4
2 design 16 16 1.0 0.688 0.688 1.0

The two designs for TV = 12 have the exactly the same characteristics in I X'X |,
I M I, fr(X'X)"1, and Amm but they are not equivalent. It is not possible to obtain one
from the other by permutating lines and columns. They also differ in their variance
inflation factors, which are slightly more constant in the D-optimal than the %
design. However, they are very close and there is no reason to choose one rather
than another on the basis of these properties.

2. Two level designs for 5 to 8 factors

The second order synergistic model (with main effects and all possible first-order
interactions) is postulated in each case. As in the previous example, the candidate
design is the full factorial, 2s, 26, 27, 28.

Full and fractional factorial designs


A similar analysis to the above shows that the most efficient designs, according to
the D-efficiency criterion, are the full and fractional factorial designs of resolution
V and better:

• 25 and 25'1 (I = 12345, resolution V: N = 16 runs)


• 26 and 25"1 (I = 123456, resolution VI: N = 32 runs)
• 27 and 27'1 (I = 1234567, resolution VII: N = 64 runs)
• 28 and 28'2 (I = 12347 = 12568, resolution V: N = 64 runs)

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


• 29 and 29'2 (I = 134689 = 235679, resolution VI: N = 128 runs)

However, a number of these designs have low R-efficiencies, so alternative


solutions may be sought if resources are limited and a minimal design is required.

Saturated and near designs


Of the saturated designs, we have already seen (chapter 3) that the Rechtschaffner
design for 5 factors is identical to the resolution V 25"1 fractional factorial. The 6
factor Rechtschaffner design, with 22 experiments to determine 22 coefficients, is
also D-optimal, like the above 4 factor saturated design.
The Rechtschaffner designs from 7 factors onwards are no longer D-optimal,
due to their lack of balance. In terms of precision of estimation of the coefficients
and VIF values, the 7 factor Rechtschaffner design is only slightly inferior to the
best design found using the exchange algorithm. Because it is easily constructed
and all coefficients are estimated equivalently, it is probably to be preferred to the
D-optimal design. In contrast to this, the D-optimal design for 8 factors has clearly
better properties than the corresponding Rechtschaffner design.
For 7 and 8 factors, increasing the number of experiments from 29 or 37 to
64 gives a monotonic increase in the determinant of M, in both cases. There
appears to be no ideal number of experiments, and no particular advantage in
carrying more than the minimum number of experiments.
For 9 factors, the efficiency of the D-optimal design, quantified as det(M),
increases with the number of experiments from the saturated design of 46 runs, to
reach a peak at 50 runs, and then steadily decreases. We would therefore advise use
of the nearly saturated 50 experiment D-optimal design for treating this problem.

3. Number of runs for determining the D-optimal design

It should be noted, just as for the screening designs, that the carrying out of a large
number of different runs with the exchange algorithm is highly recommended, if at
all possible, in order to increase the probability of finding the true D-optimal design
at least once.

References

1. R. N. Snee, Computer-aided design of experiments - some practical


experiences, /. Qual. Tech. ,17, 222-236 (1985).
2. A. C. Atkinson and D. R. Cox, Planning experiments for discriminating between
models, /. Roy. Statist. Soc., Ser. B, 36, 321-348 (1974).
3. A. C. Atkinson and V. V. Fedorov, Optimal design: experiments for
discriminating between several models, Biometrica, 62(2), 289-303 (1975).
4. A. Brodiscou, Contribution of genetic algorithms to the construction of optimal
experimental designs, Thesis, University of Aix-Marseilles, 1994.
5. J. H. Holland, Adaptation in Natural and Artificial Systems, University of
Michigan Press, Ann Arbor, 1975.

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.


6. S. Kirkpatrick, C. D. Gelatt, and M. K. Vecchi, Optimization by simulated
annealing: quantitative studies, Science, 220, 671-680 (1983).
7. P. Franquart, Multicriteria optimization and methodology of experimental
research, Thesis, University of Aix-Marseilles, 1992.
8. D. Mathieu, Contribution of experimental design methodology to the study of
structure-activity relations, Thesis, University of Aix-Marseilles, 1981.
9. P. F. de Aguiar, B. Bourguignon, M. S. Knots, D. L. Massart, and R. Phan-Tan-
Luu, D-optimal designs, Chemom. Intel!. Lab. Syst.,30, 199-210 (1995).
10. M. J. Box and N. R. Draper, Factorial designs, the IX'X I criterion and some
related matters, Technometrics, 13(4), 731-742 (1971).
11. V. V. Fedorov, Theory of Optimal Experiments, Academic Press, N. Y., 1972.
12. T. J. Mitchell, An algorithm for the construction of "D-Optimal" Experimental
Designs, Technometrics, 16, 203-210 (1974).
13. H. P. Wynn, The sequential generation of D-optimum experimental designs,
Ann. Math. Stat., 41(5), 1644-1664 (1970).
14. L. Benkerrour, J. Frances, F. Quinet, and R. Phan-Tan-Luu, Process study for
a granulated tablet formulation, unpublished data.
15. D. C. Montgomery, Design and Analysis of Experiments, 2nd edition, J. Wiley,
N. Y., 1984.
16. M. Chariot, G. A. Lewis, D. Mathieu, R. Phan-Tan-Luu, and H. N. E. Stevens,
Experimental design for pharmaceutical process characterisation and
optimization using an exchange algorithm, Drug Dev. Ind. Pharm., 14, 2535-
2556 (1988).

Further reading

• A. C. Atkinson, The usefulness of optimum experimental designs, J. Roy. Stat.


Soc. Set: B, 58, 58-76 (1996).

TM

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

You might also like