Cap 8
Cap 8
Cap 8
EXCHANGE ALGORITHMS
Methods for Non-Standard Designs
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.
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
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)
TM
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.
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.
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.
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.
TM
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).
TM
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.
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.
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
B. Candidate Points
Once the mathematical model has been postulated, there are two possible
approaches.
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.
TM
"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)
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.
|X'X|=.
TM
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
M = W x X'X
|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
4. A-optimality
TM
5. G-optimality
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:
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
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
3. Stopping condition
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
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.
TM
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:
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.
TM
2. Candidate points
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
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
/•.
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,
TM
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.
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
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:
2. Candidate design
TM
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)
TM
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.
TM
We will now show the construction of the D-optimal design, and compare the
different designs.
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
Table 8.1 Designs for the 4 Factor Synergistic Model: Statistical Properties and
Optimality Criteria
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.
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.
TM
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
TM
Further reading
TM