Professional Documents
Culture Documents
OR - 9522A Mixed Integer Programming Model For Solving A Cutting Stock Problem in The Fashion Clothing Industry
OR - 9522A Mixed Integer Programming Model For Solving A Cutting Stock Problem in The Fashion Clothing Industry
OR - 9522A Mixed Integer Programming Model For Solving A Cutting Stock Problem in The Fashion Clothing Industry
ECONOMISCHE WETENSCHAPPEN
ONDERZOEKSRAPPORT NR 9522
by
Zeger DEGRAEVE
Martina V ANDEBROEK
by
Zeger DEGRAEVE
Martina V ANDEBROEK
D/1995/2376/23
KA THOLIEKE UNIVERSITEIT LEUVEN
by
Zeger Degraeve *
Martina Vandebroek*
July, 30 1995
We would like to thank W. Gochet for constructive comments and encouragements and R. Bemelmans and J.
Abstract
The cutting operation In the high fashion clothing industry· has some particular
features that make it especially interesting and complex. The process essentially involves
putting several layers of cloth on a long cutting table and fixing templates of the parts of
several articles on top of it before the actual cutting can be initiated. This is a very time
consuming task giving raise to high setup costs in addition to waste production resulting from
cutting. Total production costs can then be optimized by minimizing the number of these
cutting operations while at the same time producing little or no waste. In this paper a mixed
integer programming model is proposed that searches for optimal cutting patterns, each
giving the height of the stack of cloth and the corresponding combination of articles to be cut
in one operation.
2
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
1 Introduction
For several years now, the Belgian clothing industry has been suffering from strong
competition of low wage countries. Large series production and stitching work for low cost
substantial increase in the domestic female unemployment. The remaining clothing industry
focuses on high fashion designer garments, e.g. skirts, blouses, suits and dresses, in small
quantities. This clothing is sold in a few specialty stores which carry frequently at most one
item of various designs. Typically, the products make use of exclusive and therefore
expensive fabrics requiring waste, resulting from the production process, to be kept at a
minimum. One of the most costly, waste producing and time consuming operations is the
preparation for and actual cutting of the fabric into the various parts of an article of clothing.
This paper addresses the optimization of this operation by presenting a mixed integer
The cutting problem was presented to us by INSYS N.V., a software consulting firm
with customers in the high fashion garment and textile industry. Several of their clients were
struggling with this problem for quite some years now. Using some straightforward simple
heuristics they had been able to derive feasible solutions, however, neither themselves nor
cutting process and introduce the problem. Section 3 presents the derivation of the mixed
extension of our formulation to allow for multiple colors and corresponding computational
3
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
In this section we give a detailed description of the cutting process and introduce the
Before production is initiated, the demand data is collected partly from fixed orders
and partly from sales forecasts. Demand for a specific article of clothing typically consists of
the number of units of each size. This contains all necessary input for the cutting process. An
example of the demand data for an article is given in Table 1. The size information is for
female adults in European measurements. The example in Table 1 represents a real-life case
as it was found at one of the client companies of INSYS N.V .. In addition, the order of
magnitude of this demand data remains essentially the same for other data sets reflecting the
low quantities and consequently exclusive character of the various clothes. Observe that the
distribution of the demand data is centered around the mean size of 42.
Size 38 40 42 44 46
Demand 54 84 91 60 29
Next, an efficient way of cutting the different parts of an article, e.g. the front, the
back, the sleeves, the pockets, etc., out of the width of cloth supplied on rolls is needed for
each size. Most companies are using commercially available software, called Pattern Design
System (PDS), for this decision. The software visually constructs templates indicating how to
place the parts on the cloth such that the width of the roll is used up as much as possible.
These templates are called stencils and are laid out on the cloth to guide the cutting process.
4
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
As a result, good stencils effectively use the width of the roll of cloth as much as possible.
The length of cloth needed for an article does not vary substantially from size to size and is
generally assumed to be equal for all sizes. As a result, the different stencils for the various
It would be inefficient to cut each stencil separately, therefore the cloth is spread out
in several layers on a long cutting table of a specific length. The cutting operation can start
after a particular combination of the stencils are fixed on top of this stack, see Figure 1 for a
simplified illustration. The spreading of the fabric on the cutting table, the fixil}g of the
stencils and layers and the cutting itself are extremely delicate and very time consuming.
Therefore, the people in charge of cutting want to minimize the number of these operations.
The problem then consists of finding the optimal combination of the number of layers of cloth
on the cutting table and the associated combination of stencils that will result in the minimum
number of cutting setup operations while the demand is satisfied with little or no excess.
Figure 1 : Stencils fixed on top of a stack of layers of cloth ready for cutting.
The number of layers that can be cut in one operation is limited by the length of the
knives and the thickness of the fabric. This upper limit on the number of layers has to be
taken into account. A second restriction on the cutting process is introduced by the length of
the cutting table which limits the number of stencils that can be cut in one operation.
5
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Given an upper bound on the number of layers of 35 and a cutting table's length
allowing for at most 4 stencils to be cut in the same operation, a solution to the problem
introduced in Table 1 is given in Table 2. This solution was termed 'excellent' by the
specialists at INSYS N.V .. The column labeled 'Number of Layers' gives the height of the
stack; the other entries give the number of stencils of the corresponding size that will be cut
in the same operation. Such a combination will be referred to as a cutting pattern in the
remainder of this paper. In the industry, a cutting pattern is also termed a marker or lay plan.
Number Sizes
of 38 40 42 44 46
Layers
Pattern 1 27 2 2 0 0 0
Pattern 2 31 0 1 2 1 0
Pattern 3 29 0 0 1 1 1
Production 54 85 91 60 29
Demand 54 84 91 60 29
Excess 0 1 0 0 0
The row labeled 'Production' gives then the total production of the different sizes
using the proposed cutting patterns with the corresponding number of layers of cloth. In order
to compute the excess production, labeled 'Excess', we have also included the demand data
of Table 1. Given that 3 patterns are to be used, the minimum excess is 1 as size 40 will be
cut one more time than was demanded. In most cases however, this excess is not really
considered as waste because it is generally accepted that the client will buy a few extra pieces
of an article upto 5 % of the demand. In some cases however, the surplus production is really
6
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
useless, e.g. when a client is ordering uniforms for its employees and has already included
The problem we are presented with, is a variant of the well known one dimensional
cutting stock problem discussed throughout the literature. Materials such as paper, cellophane
and metallic foil are manufactured in rolls of a standard width. The widths of the rolls
specified by different customer orders may vary widely. When a large number of orders has to
be filled, finding the most economical way of cutting the standard rolls into the desired final
The objective function is often stated in terms of minimizing waste or minimizing the
number of standard rolls to be used but it has also been stated in terms of minimizing the
number of different cutting patterns or, equivalently, minimizing the setup cost. For our
problem, we will need a combination of these two objectives as we are minimizing the
7
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
H : maximum height of the stack, maximum number of layers of cloth that can be cut at
C : setup cost that is associated with the spreading of the cloth and the fixing of the
subject to
Vj E R (4)
8
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
(v) integrality.
Yj E {a, I}
Zj E {a, 1, 2, 3, ...}
Due to the fact that all sizes have equal length 1, we can scale the length of the table b and use
The model (CSP) is difficult to solve for two reasons. First, due to the fact that both aij
and Zj are general integer variables appearing as a product in the objective function (1) and in
the demand constraints (2), model (CSP) is a general integer nonlinear programming
problem. Second, the presence of the knapsack constraints (3) complicate matters also.
The first problem can be resolved by generating all possible combinations of stencils
in advance such that the aij are no longer variable. In the resulting model only the Zj, the
thickness of the stacks, have to be optimized for each pattern. However, the model will only
be solvable if the length of the cutting table is not too large compared to the length of the
stencils. For the example of Table 1, consisting of 5 different sizes and a cutting table which
can hold at most 4 stencils, the possible number of patterns is 124 (I R I = 124). This can
easily be computed with the following recursion:
eij = number of different patterns with i stencils and size j being the largest size, i = 1, 2,
9
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
1 = 1, 2, ... , b
1 J = 1, 2, ... , IPI
= 2, 3, ... , b; j 2,3, ... , Ipi
We tried in vain to solve this problem for finding 3 different patterns on a 486, 33
MHz IBM PC with LINDO (Schrage, 1994). After more than 24 hours the program had not
found a feasible solution yet while the solution reported in Table 2 was found after about 6
minutes on a 486, 66 MHz IBM PSNP PC with the model to be presented next.
generation was developed by Gilmore and Gomory (1961). This iterative procedure starts
with a master program consisting of the demand constraints and a few prespecified cutting
patterns. The decision variables are then the number of layers that will be cut according to the
currently generated patterns. Using information provided by the master (i.e. the dual prices),
the subproblem consisting of the knapsack constraint (3) implicitly considers all possible
cutting patterns to select the one(s) to be added to the master program likely to improve its LP
objective. This new pattern is then included in the master program etc .. The procedure stops
when an optimal solution is found to the LP relaxation of the master program. Then an integer
program is solved over the patterns on hand. Consequently, the resulting integer solution may
be arbitrarily bad. Empirically however, it has been found that the column generation
procedure produces heuristic solutions close to optimality while at the same time giving a
We have used this procedure for trying to solve our problem but the resulting cutting
patterns appeared far from being optimal. We speculate that this might be due to the fact that
10
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
all stencils have equal lengths and that the demand for the different sizes shows no extreme
high or low values. For such a cutting problem no cutting pattern is especially favorable and
the optimal combination of patterns can hardly be found by extending an existing set of
patterns with an extra pattern at a time. We therefore had to resort to a different solution
approach.
We cope with the nonlinearity in the demand constraints by discretizing the variables
Before writing the discrete expansion of the variables aij, we need an additional set and
variable definition:
\IkE B
To clarify the set definition B, consider a cutting table of length 4 (b = 4), then B is the set
{O, 1, 2, 3, 4}. The discrete expansion of variable aij is then as follows:
\I i E P, \I j E R (6)
L, W ijk =1 \I i E P, \I j E R
kEB (7)
W ijk E {O, I} \I i E P, \I j E R, \I k E B
We use (6) to substitute variables aij out of the demand and knapsack constraints (2) and (3)
11
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Obviously, the discretization did not resolve the nonlinearity problem. However, the product
of a 011 variable with a general integer variable, Wijk Zj, can easily be linearized. Before giving
the linearization constraints, we define the variable which will capture the nonlinearity as
follows:
Vijk :::: Zj if Wijk:::: 1,0, otherwise, effectively Vijk:::: Wijk Zj, ViE P, V j E R, V k E B
and the objective function is changed accordingly. The linearization constraints are then as
follows:
By discretizing the variables aij and the linearization constraints, we have effectively
eliminated the nonlinearity from the model (CSP), a general integer linear program remains.
Alternatively, we could have chosen to discretize the variables Zj. However, as the upper
12
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
bound on the variables aij is always lower than the one on the variables Zj, we have chosen to
discretize the variable which gives the lowest cardinality of the set B. It is now possible to
reduce the number of constraints (8b), knowing that Vijk =Wijk Zj and summing this expression
L
keB
Vijk L Wijk
= keB Zj L Wijk = Zj
= Zj keD ViE P, V j E R (9)
However, we still have to zero out the variables Vijk in case Wijk equals zero. Therefore, we
Eliminating the nonlinearities did not remove the difficulty presented by the knapsack
constraints (3). We are able to resolve this difficulty because of the particular property that
the different stencils all have equal length 1. This property fundamentally reduces the
complexity of the· knapsack problem. For a knapsack problem, a state is defined to be the
length remaining after some sizes have been cut. The state space or the number of states of a
regular knapsack problem grows exponentially with the number of sizes, which actually
happens if the sizes have different lengths. However, with stencils of equal length, the state
space only grows linear, specifically it is (b + 1) Ipl. This observation led us to model the
knapsack constraint using its network or dynamic programming formulation (Dyckhoff,
13
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
1981), resulting in a model which remains linear in size. In order to do so, we need the
Xijrs = 1, if we allocate s copies of size i to pattern j when r positions are still free, 0,
The network formulation for the knapsack problem, one for each possible pattern j, "i j E R,
is then as follows:
This formulation is derived by considering the sizes sequentially, allocating first copies of
size 1, then copies of size 2, and so on until the pattern is fully allocated. Essentially, this
allows us to eliminate alternative solutions at the pattern level from being investigated by the
branch and bound search tree. We believe that this is one of the major strengths of our
formulation. Note also that when pattern j is not being used (Yj = 0), the corresponding Xijrs
variables are all zero. Due to the introduction of the variables Xijrs, we are able to substitute
the variables Wijk out of our model using the following expression:
14
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
In addition, constraints (7) can be eliminated from the model by observing that through
b b b
L W ijk L L X ijrk = Yj \:;j i E P, \:;j j E R (13)
k=O k=O r=k
To derive (13), we have taken into account the network constraints (11). Consequently, we
have shown that constraints (7) are implicitly represented in our model.
Although constraints (11) through the variable definition Xijrs, effectively eliminate
alternative solutions at the pattern level, it is still possible to obtain alternatives across
patterns, i.e. any renumbering of the patterns will create such an alternative solution. It is also
possible to eliminate those alternatives from being considered by the branch and bound
constraints will enforce a solution to look only as illustrated in Table 3. The solution should
satisfy the conditions that patterns are ranked such that the number of copies of the first size
Table 3). Observe also that the solution reported in Table 2 also satisfies those conditions.
15
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Sizes
38 40 42 44 46
Pattern 1 2 2 0 0 0
Pattern 2 1 3 0 0 0
Pattern 3 0 1 1 2 0
Pattern 4 0 1 3 0 0
For the first size (i = 1) the constraints eliminating alternatives across patterns are as follows:
b b
L X 1jbr :::;; L xlCi-llbr Vj E R\{l}, Vk E B (14)
r=k r=k
for all other sizes, the constraints are quite similar except for an additional term between
b b
LX jjbr
r=k
:::;; LXjCi-llbr
r=k
+ (1 - x(i-1l(j-llbO) ViE P \ {I}, V j E R \ {I}, V k E B (15)
Only if X(i-l)(j-l)bO is equal to 1, implying that size i is the first size effectively appearing in
patternj-l, will the constraint be enforced otherwise it will be redundant. The constraints (14)
and (15) tremendously increase the density of the constraint matrix. In order to do a density
transformation define the slack variables Uijk on (14) and (15) resulting in :
16
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
b b
L,X 1jbr + U1jk L, x1G-l)br V j E R \ {I}, V k E B (14a)
r=k r=k
b b
L,X ijbr + Uijk L, xiG-Ilbr + (1- X(i-IlG-IlbO) ViE P \ {I}, V j E R \ {I}, V k E B (15a)
r=k r=k
The density transformation on (14a) and (15a) respectively is performed by subtracting row
k+I from row k for k = 0, 1, ... , b-I, giving the following result:
compute a lower bound on the minimum number of different patterns necessary to satisfy
demand, we have altered the objective function (1) by fixing the number of different patterns
(see below), i.e. by specifying values for the 0/1 variables yj, V j E R. We then use the model
to create the pattern composition by only minimizing the excess, the amount by which the
For the example in Table 1 it is easy to calculate the minimum number of patterns to
have to cut size 38 at least twice, sizes 40 and 42 at least 3 times, size 44 at least twice and
2 + 1) which, taking the length of the cutting table (b = 4) into account, require at least 3
17
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Yl = Y2 = Y3 = 1, Yj = 0, V j E P \ {I, 2, 3}
To summarize, the formulation for our cutting stock problem in the fashion clothing
MIN L L L
iEP jER kEB
k V ijk
-
(CSTFI) (17)
subject to
X Ijb(b-r) L
= s=o x 2jrs (19)
b-r r
L
s=o
xij(r+s)s L
= s=o x(i+l)jrs
ViE P, V j E R, V k E B (20)
'v'j E R (21)
18
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
X ijbk + U ijk - U ij(k+11 = Xi(j-Ilbk 'v'i E P, 'v' j E R \ {I}, 'v' k E B \ {b} (22)
X ijrs E {O, 1}
yj E {O, 1}
Zj E {O, 1, 2, 3, ...}
For the data of Table 1, generating model CSTFI yields a formulation with 326
variables, amongst which 201 are integer variables, 211 constraints and 1035 nonzeroes.
Optimality of the reported solution was proven after about 20 minutes on the 486, 66 MHz
IBM PSNP Pc. Clearly, with the objective function (17), the objective function value of the
LP relaxation is simply the sum of the demands which is 321 for the example. The objective
function value of the MIP is 322 giving the excess production of one copy for size 40.
We also tried to solve the problem with two different models. One model in which we
kept the original knapsack constraints (3) with the discrete expansion of the variables aij.
Another model where we tried the network or dynamic programming formulation for the
knapsack constraints (11) but instead of using the discrete expansion of the variables aij (6)
and (7), the binary expansion was used. However, in those few cases where we were able to
19
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
find solutions, computation times with both shorter models far exceeded those obtained with
4 Model Implementation
We have implemented our model using OMP (Beyers and Partners, 1995) on an IBM
PSNP 486, 66 MHz Pc. OMP is a professional integrated menu driven LP/IP optimization
through its automatic pilot which is close to a high level programming language. The
automatic pilot is used to build a shell around OMP that will guide users through an
application from data input to solution interpretation without them even knowing that linear
programming is being used to solve their problem. Besides the automatic pilot, OMP consists
permits prototyping at a realistic scale. The model generator allows for separation between
model and data. It also contains a state-of-the-art LP/IP optimization engine enhanced with
provides the expert user with great flexibility for choosing alternative branching strategies to
complement its build in procedures for solving integer programs. Moreover, it contains some
advanced integer programming implementations such as tightening through the use cutting
planes. Finally, a report generator gives the user complete freedom for defining tables, lists
and full information reports on variables and constraints selected. People at INSYS N.V.
20
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Being very pleased with a solution for a problem they had been struggling with for
years, people at INSYS N.V. asked us if it would be possible to also include layers of cloth of
different colors in the stack. It was quite straightforward to extend our model to cope with this
additional complication. Vie define an extra set S with index c, modeling the number of
different colors and adding the index to all relevant variables and parameters. The multi-color
subject to
L L
jER kEB
k V icjk ~ d ic ViE P, V C E S (25)
X Ijb(b-r) L
s=o
x 2jrs (26)
b-r r
L
s=o
x ij(r+s)s L
= s=o X (i+I)jrs
L V icjk \i i E P, V j E R, V k E B, \i c E S (27)
kEB
b
V icjk ::::; H IX jjrk
r=k
21
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Vj E R (28)
ViE P, V j E R, V k, r, s E B, s ~ r, V c E S
yj E {O, I}
the minimum number of patterns required so, we solved the problem with a given number of
patterns ; this is also common practice in the industry. Table 4 summarizes computational
results for a three color demand data set using 4 cutting patterns. The upper bound on the
number of layers of cloth was 100 (H = 100) and the cutting table's length permitted at most
8 stencils to be cut at the same time (b = 8). The solution reported was the best found after
having had the PC run overnight, however, the particular solution was obtained after 1 hour
and 21 minutes of computing time on the 486, 66 MHz IBM PSNP Pc. In general, we find
that intermediate solutions with low excess are obtained within minutes, surely much shorter
time than needed by an experienced person constructing cutting patterns with similar excess
production heuristically. To obtain this result, we had to exclude constraints (29). From our
22
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
computational experiments we learned that constraints (22) were useful for proving
optimality for the single color problems, however, for the multiple-color version, constraints
(29) prevented us from finding feasible solutions within reasonable computation times. The
particular cutting pattern proposal of Table 4 was again termed 'excellent' by the consultants
at INSYS N.V ..
23
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Table 4 : Results for a three color data set using 4 cutting patterns.
H= 100 Sizes
b=8 38 40 42 44 46
Demand Blue 54 84 91 60 29
Red 30 61 89 76 45
r"!..-o.on ').c:; '7(\ ;::;'2 <;:11 '2{\
'-Jl. \..1\..111 .<"J IV VJ J'-t J7
Number
of Layers
Pattern 1 Blue 27
Red 15 2 0 1 1 0
Green 13
Pattern 2 Blue 28
Red 21 0 3 0 0 1
Green 24
Pattern 3 Blue 32
Red 37 0 0 2 1 0
Green 25
Pattern 4 Blue 1
Red 24 0 0 0 1 1
Green 16
Production Blue 54 84 91 60 29
Red 30 63 89 76 45
Green 26 72 63 54 40
Excess Blue 0 0 0 0 0
Red 0 2 0 0 0
Green 1 2 0 0 1
24
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Although the consultants at INSYS N.V. were quite impressed with the convenience
with which good patterns, giving low excess production, could be computed with our model,
they soon started to express some secondary, implicit criteria which had previously been
unvoiced. Obviously, we had optimized the problem given the key criteria they had presented
to us, however, it became apparent that they are considering additional criteria to finally
select the cutting patterns to be used from among the list generated with the simple heuristics
they were using before. Moreover, it turned out to be extremely hard for them to explicitly
describe those secondary criteria. After sitting down with the consultants and comparing
several cutting pattern proposals for the same demand data set, we were able to extract the
First, for various cutting pattern proposals, given the same number of cutting patterns
and low excess, they give preference to the proposal which has fewer stencils per pattern and
consequently a higher height of the stack of layers of cloth. As can be observed from the
example problem in Table 4, they felt that our solution satisfies this criterion extremely well.
The cutting table's length allowed for 8 stencils to be used per pattern (b = 8) but the
Second, in those cases where the cloth is really extremely expensive, e.g. some cloths
are· woven from both textile and gold threads, and when the excess production is not too bad,
they prefer to completely use the full length of the roll of cloth. If cloth remains on the roll it
is either stored in inventory for later use or, as happens in most cases, it is just considered as
waste. To avoid such waste, they modify the cutting pattern proposal such that the total
production will use up the complete roll of cloth as much as possible, allocating excess
production appropriately over the sizes. For example, if cloth remains for the problem in
25
A Cutting Stock Problem in the Fashion Clothing Industry Degraeve and Vandebroek
Table 2, they would prefer to add an additional layer to the stack of pattern 2 because
resulting excess production with this pattern is smallest relative to the demand.
The third and most interesting observation is the fact that for some articles, the
stencils generated with the layout software are not very efficient with respect to the usage of
the width of the cloth. mostlv because there iust does not exist a better wav of constmctinlY
' . . 1 oJ - - -J - - - - --- - -- - ----0
the stencil. In those cases, they try to manually reallocate the various parts of the articles over
the cloth across stencils, keeping the generated combinations of stencils fixed, in order to
save on the length of the cloth needed for the pattern. This results in a length of the cutting
pattern which is not an integer multiple of the length of the stencils, ultimately giving less
waste thus higher efficiency. The efficiency is then defined as the ratio of the new length of
the pattern after reallocation of the parts to the original length of the pattern before
reallocation. In reality, this reallocation process is almost always carried out actually. The
problem stems from the fact that they are really facing a two dimensional cutting stock
problem which is solved by tackling two one dimensional problems consecutively. Clearly,
total efficiency can only be optimized by considering the two dimensional problem as a whole
although constructing good patterns using their efficiency measure is already an interesting
problem.
particularly challenging.
26