Applied Intelligence 23, 9–20, 2005

c 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Genetic Algorithm Coding Methods for Leather Nesting

ALAN CRISPIN, PAUL CLAY AND GAYNOR TAYLOR

School of Technology, Leeds Metropolitan University, Calverley Street,
Leeds LS1 3HE, England
[email protected]
[email protected]

TOM BAYES
SATRA Technology Centre, Rockingham Road, Kettering, Northamptonshire, N16 9JH, England
[email protected]

DAVID REEDMAN
R&T Mechatronics Ltd. The Cottage, Main Street, Wartnaby, Melton Mowbray,
Leicestershire, LE14 3HY, England
[email protected]

Abstract. The problem of placing a number of specific shapes in order to minimise waste is commonly encountered
in the sheet metal, clothing and shoe-making industries. The paper presents genetic algorithm coding methodologies
for the leather nesting problem which involves cutting shoe upper components from hides so as to maximise material
utilisation. Algorithmic methods for computer-aided nesting can be either packing or connectivity driven. The paper
discusses approaches to how both types of method can be realised using a local placement strategy whereby one
shape at a time is placed on the surface. In each case the underlying coding method is based on the use of the no-fit
polygon (NFP) that allows the genetic algorithm to evolve non-overlapping configurations. The packing approach
requires that a local space utilisation measure is developed. The connectivity approach is based on an adaptive graph
method. Coding techniques for dealing with some of the more intractable aspects of the leather nesting problem such
as directionality constraints and surface grading quality constraints are also discussed. The benefits and drawbacks
of the two approaches are presented.

Keywords: computer-aided nesting, genetic algorithms, encoding, leather, image processing, packing, connec-
tivity, optimisation

1. Introduction cutting stock (or trim loss) problem generally refers to

the case where the surface and parts are rectangular.
The general two-dimensional cutting stock problem However, in practical problems as found in the textile
can be defined as follows. Given a set of N shapes find and leather shoe making industries the parts are irreg-
the optimal non-overlapping configuration of all shapes ularly shaped and in the case of shoe manufacture the
on a sheet. The N shapes can consist of a smaller set of surface shape is irregular being a hide. Consequently,
shapes that are repeated a number of times to provide the leather nesting problem involves placing a set of
consistent parts for a number of assembled items. The plane irregularly shaped parts on a plane irregularly
10 Crispin et al.

shaped surface which is also characterised by further fied as unsuitable because of the variation in hide qual-
constraint requirements (see Section 2). ity with position e.g. defects such as scar tissue and
Many packing-driven solutions to the cutting stock holes that may occur randomly. In order to make deci-
problem have been reported in the literature [1]. Anand sions about placement the skilled cutter (and any pro-
et al. [2] and Bounsaythip et al. [3] have explored ge- posed automation) requires implicit knowledge of the
netic algorithm approaches. One of the main difficulties subsequent processes. During the lasting or 3D shap-
reported was finding an appropriate encoding strategy ing of the shoe upper certain components undergo large
for mapping shapes evolved by the genetic algorithm elastic and visco-elastic distortions. The presence of
to non-overlapping configurations. Anand [2] used an singular defects in such components may lead to a me-
unconstrained approach whereby the genetic algorithm chanical failure during lasting and will certainly be un-
was allowed to evolve overlapping configurations and acceptable cosmetically in the finished product. There
post processing was subsequently used to check and is also a need to understand the anisotropy of the elas-
correct for shape intersection. Bounsaythip et al. [3] tic properties of leather and their effect on the shap-
have used an order based tree encoding method to pack ing process. This is known as identifying the lines of
components in strips. tightness, which are generally held to follow the skele-
Kahng [4] discusses the preoccupation with packing- tal structure. Conventionally shoe components are cut
driven as opposed to connectivity-driven problem for- “tight to toe” so that, when assembled, the stiffer elas-
mulations in the context of floor planning (i.e. auto- tic axis runs consistently along the shoe. Any proposed
mated cell placement for VLSI circuits) and suggests computer-aided nesting system must take into account
that approaches based on a connectivity centric should the constraint criteria outlined above. In addition any
be used. cosmetic defects present in the leather must either be
This paper explores genetic algorithm based coding discarded or must appear in areas of the shoe which are
methods for both a packing driven and connectivity hidden from casual view.
driven solution to the leather-nesting problem. Section The hide data files used for this research contain
2 describes the leather-nesting problem. Section 3 point co-ordinate information on the edge outline,
describes shape placement using the no-fit polygon. holes and grade regions (quality zones) from real hides
Section 4 develops a packing driven genetic algorithm scanned at a shoe factory. Region grading conforms
solution based on developing a local space utilisation to a grading system commonly used in shoemaking,
measure. In Section 5, a connectivity driven genetic al- which ranges from grade 1 regions where material
gorithm solution is presented based on creating a graph quality is rated between 95–100% down in steps of
at time of placement such that rules are used to prevent 5% to grade 5 regions where material quality is rated
invalid positioning. Section 6 discusses results and the between 75–80%. An image-point dual was used in
benefits and drawbacks of the two approaches. Finally the work allowing images of shapes to be created
conclusions are drawn. from part co-ordinates and vice versa [5]. Figure
1 shows a typical hide image with grade regions
marked.
2. Problem Description To constrain the placement of shapes that require
high quality leather a threshold can be applied to the
The shoe manufacturing process starts with the cutting grade image on a level determined by the lowest grade
of shoe upper components from a hide with parts nested that a shape may be placed within. Any lower grades
together as far as possible to maximise material usage. are removed from the hide. This approach completely
A significant proportion of the manufacturing cost of a restricts a shape from entering a poorer grade area than
pair of shoes is invested in the natural raw material and required.
so the efficient utilisation of this resource is of prime This work has used a tightness vector map, which
importance. Small improvements in yield can have a generically follows the skeletal structure (see Fig. 1).
dramatic effect on profitability. This allows an error difference angle to be calculated
In leather nesting constraints are placed on the se- between the tightness vector and shape direction vector.
lection process by external practicalities. In particu- A shape is placed so its direction is constrained to a
lar certain areas of the hide may be unsuitable for maximum deviation value either side of the tightness
a given class of components. Such areas may be identi- vector.
 Genetic Algorithm Coding Methods for Leather Nesting 11

Figure 1. Hide image with directionality and grade constraints shown. Darker regions represent poorer quality grade material. Tightness vector
arrows indicate the required placement direction.

3. Shape Placement Using NFP edges of the individual shapes [8]. The basic algorithm
only works correctly for convex polygons and a number
Optimal nesting can only be achieved when shapes are of adjustments have been made to accommodate the
touching since any small gaps between two shapes are relationship between both simple non-convex [9] and
usually unusable and so are wasted. The no-fit polygon more complicated non-convex shapes [10].
(NFP) [6, 7] is the path traced by any selected refer- Our approach to calculating the NFP is based on im-
ence point on one shape as it circumnavigates around age processing techniques [5, 11]. The NFP is found
another shape. Selection of the reference point will from the Minkowski sum, which is the convolution
translate the NFP relative to the other shape. In Fig. between two shape images at a set orientation. The
2 the reference point has been taken as the shape cen- calculated Minkowski sum for two leather compo-
troid. The NFP can be used to ensure that each shape nents used in shoe making is shown in Fig. 2. The
placed touches the hide edge or another shape without NFP is the boundary between the area covered by
overlap. the Minkowski sum and those areas not covered by
A number of different methods have been devised the sum. Four-connected shape boundary images are
to calculate the NFP between two shapes involving the used to calculate the Minkowski sum with the NFP
addition of ordered vectors describing the polygonal extracted using morphological operators. As the angle

Figure 2. Minkowski sum found by convolution between two shape images at a set orientation. The NFP is the boundary and is dependent on
the angle between the two shapes as shown in the second diagram.
12 Crispin et al.

Figure 3. An example of a shape-hide no-fit polygon (NFP) and window region centred on point labelled A.

between shapes changes the NFP jumps from one the previously placed shape and the process of find-
state to another (i.e. there is a non-linear discontinuous ing the best local placement solution within the new
change). window region is repeated.
Our window placement strategy imitates the place-
ment strategies of a professional lay-planner and pro-
4. Packing Driven Solution motes clustering and runs of similar shapes. The NFP
calculations are only performed within the local win-
A packing driven solution requires that shapes are dow region, thereby reducing the number of compu-
placed such that space utilisation is maximised. The tational operations required as compared to those that
method developed for this work is based on a local would be required if the whole hide image is used.
placement strategy where shape parts are placed on the Figure 4 shows an example of shape placement
hide one at a time in sequence. Once a shape is placed it against a hide edge within a window area. The window
is not changed becoming part of a dead region in which is centred on a current point labelled A. The average
it is not possible to place further parts. The local place- directionality angle is calculated using the tightness
ment approach requires using a local measure for space vector map in a box positioned at the centre point of
utilisation such that the sum approximates to the final
space utilisation measure. The alternative approach is
to use a global placement method in which whole lays
are repeatedly created on a hide and tested directly
against a global space utilisation measure. Heistermann
and Lengauger [12] discuss the advantages of using a
local placement strategy over that of a global placement
strategy in terms of nesting time requirements.
The strategy of our packing algorithm is based on
placing a shape at a point on the shape/hide NFP (or a
shape/shape NFP) to maximise local space utilisation.
Calculations are performed within a window region.
The window is initially positioned around the point at
the rear of the spine of the hide nearest to the image
corner. In Fig. 3 this point is labelled A and the first
shape is placed in the window region shown such that
the shape which fits closest to the edge is chosen. The Figure 4. An example of shape placement within a window region
window is moved (i.e. stepped) so that it is centred on showing the relationship between shape angle θ and lay angle ϕ.
 Genetic Algorithm Coding Methods for Leather Nesting 13

the line projected from A at lay angle ϕ to the window tionality) lined up along the length with either orienta-
edge. The box dimensions for determining the direc- tion acceptable. The number of shoes required is used
tion angle are made equal to the window step distance. to determine how many of each shape type should be
A shape angle θ measured from the horizontal axis is placed.
chosen such that it is constrained to ±5◦ of the aver-
age tightness angle found. The shape is placed at the
4.1. Genetic Algorithm Packing Optimisation
point on the shape-hide NFP which is closest to the
line projected from A at lay angle ϕ. By generating lay
The key features of a simple genetic algorithm are:
angle sample populations a number of placement solu-
tions are created which can each be assigned a fitness
value so that the best placement can be determined. The – They start with an initial random population of po-
specification requirement that the maximum deviation tential solutions.
for the shape angle should be ±5◦ round the tightness – The population is indexed in terms of fitness using
angle found in the region of placement is also used in an objective function.
the optimisation process (see Section 4.1). – High fitness members are randomly selected and re-
Efficient utilisation of space requires that shapes are combined to produce an offspring.
located close to one another. An indication of the prox- – A small proportion of the offspring mutates to extend
imity of a shape to the hide edge or another shape can the search.
be calculated for optimisation purposes. Our approach – Fitness values for the offspring are calculated.
to proximity measurement is based on filtering the bi- – The offspring and their corresponding fitness values
nary images of the shape and the hide using a mask replace poorly performing members of the original
of unity-values. A threshold is applied to generate an population.
outer border around both the hide edge and shape (see
Fig. 5). It is observed that shapes that fit closest to the Iteration is used to generate more and more offspring
edge have the largest overlap area. Consequently, a fit- solutions to yield better approximations to the final op-
ness value can be calculated as the overlap area between timal solution, see Goldberg [13]. For specific coding
the two borders as this provides a measure of proximity details see [14, 15].
to the edge or another shape. The genetic algorithm generates a real value popu-
The shapes are placed in order of size since it is lation consisting of lay angles and directionality ratios.
easier to find suitable positions for the smaller shapes As previously discussed the shape angle is calculated
once the available area becomes limited. It is noted from the local tightness vector map using a box region
that the largest shape, the vamp (toe piece), is also centred on the lay angle line. The directionality ratio
the shape that requires the highest-grade leather and value is used to allow a constrained variation of shape
is to be placed with the tightness of the material (direc- angle from the local tightness value calculated in the
box region.
To utilise the mechanics of evolution a selection of
the current set of solutions is chosen to become the
parents of the next set of offspring. The tournament
selection process is used to generate random pairs of
indices to the current solutions and their corresponding
fitness values compared, the greater of the two being
chosen as a parent. Pairs of these successful parents are
then combined using linear recombination to create a
new population. The offspring is 60% of the size of
the population. A mutation operator randomly alters
5% of the genes (lay angle, directionality ratio) in the
offspring to extend the search area and so reduce the
chances of becoming stuck at sub-optimal solutions.
An elitist reinsertion strategy is used. A typical lay-
Figure 5. Fitness value calculation. The fitness value is equal to the plan obtained using the packing coding approach is
free space overlap region shown as the darker region. shown in Fig. 6.
14 Crispin et al.

Figure 6. Lay plan of 14 shoes giving 61.6% coverage (with directionality and grading constraints applied).

5. Connectivity Driven Solution

An alternative coding strategy can be realised based

on a connectivity driven approach. Rather than pack
shapes to maximise a local space utilisation measure,
the shapes can be placed by calculating the intersection
points between NFPs.
This image processing technique for NFP calcula-
tion allows direct Boolean manipulation for fast iden-
tification and removal of crossing points of NFPs. The
crossing points of two individual NFPs are required to
find the points at which a shape can touch two other
shapes at the same time. Possible crossing points of the
NFPs may lie within the Minkowski sum between the
shape to be placed and shapes other than the ones it
is required to touch. These Minkowski sums are used
Figure 7. No-fit polygons for three shapes. Shape 1 is placed at the
as masks to remove these crossing points from the list origin and Shape 2 is placed on a chosen point on NFP 1-2, shown
of possible solutions and so prevent overlapping place- as a long dashed line. Shape 3 must be placed on a crossing point of
ments being chosen. NFP 1-3 and NFP 2-3. The crossing points are shown as asterisks.
Figure 7 shows three shapes with fixed angles and
their associated NFPs. In this example, shape1 is placed
at a set position and all other shapes connect to it and 5.1. Graph Topologies
each other. Shape 2 can be moved around shape1 along
NFP 1-2 and one point on this boundary is selected. This approach can be described using graphs that spec-
Valid positions for shape 3 to touch both shapes 1 and 2 ify the connectivity between shapes. A graph [16, 17] is
are at any crossing points between NFP 1-3 and NFP 2- a data structure used to model objects and connections
3. Therefore the three shapes are connected via the NFP. between them and is defined by a set of nodes and a
This connectivity approach can be scaled to any num- set of edges that connect the nodes together. Figure 8
ber of shapes by computing NFP intersection points. If shows a graph of four connected shapes. The proper-
a fourth shape is required to touch shapes 1 and 3, pos- ties of the nodes determine shape type and angle. The
sible solutions may lie such that shape 2 is overlapped edges of the graph determine how shapes touch.
by shape 4. To prevent overlapping, all points within The graph of Fig. 8 dictates that shapes 1 and 2 touch
the area of the Minkowski sum between shapes 2 and all other shapes but shapes 3 and 4 touch only shapes
4 are removed using a mask. 1 and 2 but not each other. Shapes 3 and 4 may be
 Genetic Algorithm Coding Methods for Leather Nesting 15

Figure 8. A graph to shape placement relationship which shows

the way a graph represents the placement of 4 connected shapes.

transposed and the graph is still valid. Consequently,

the positioning of shapes can vary for a single combi- Figure 9. An example of a shape connectivity graph for 16 shapes
nation of graph and node properties. The graph repre- created where a rule is used to check that edges between nodes can re-
alistically exist. Adjusting node properties (shape type, shape angle,
sentation describes how shapes connect with each other
angle for next placement) varies the graph and lay.
but not where.
The key concept of a graph-based approach is that
the edges represent touching shapes. The no-fit poly- lapping shape and those after are placed around this
gon can be used for finding placements for a set of new anchor shape until overlapping occurs again. Us-
shapes that touch one another and allows the graph ing this adaptive graph method to create a valid non-
connectivity concept to be realised. overlapping lay of shapes the properties of the nodes
A number of graph topologies can be used to de- (shape type, shape angle, angle for next placement) can
scribe the connectivity between shapes. Fixed topology be optimised to improve coverage by applying a genetic
graphs set the connectivity, while the shape type and algorithm.
orientations are adjusted to create a valid placement
within the graph constraints. With this type of topol-
5.2. Genetic Algorithm Connectivity Optimisation
ogy the complete graph may not always be realised
with fixed node properties, since the existence of NFP
The genetic algorithm has to evolve a combination of
crossing points are dependent on the properties and the
shape order, shape angle and NFP intersection point
placement of those nodes placed previously.
position in order to nest shapes inside the hide area.
Alternatively, a graph can be created at time of place-
Vectors can be used to store shape order (represented
ment using rules that prevent invalid shape placement.
as a list of N unique numbers) and shape angles (i.e. a
This is the approach used in this work. The graph edges
list of angles associated with each shape). To select an
are varied to fit the lay that develops using the rule
NFP intersection point requires generating a placement
that each shape node is connected so that it touches
angle. The closest valid crossing point on the NFPs to
the previous node and an anchor node. An example of
this angle is selected for placement.
the specific case of this adaptable topology is shown
The genetic algorithm chromosome is separated
in Fig. 9. In this specific case shape 1 is the anchor
into three separate sections:
for shapes 2 to 4, while shape 5 is unable to touch both
shapes 4 and 1 without overlapping another shape. Con- Shape order set of 20 unique
sequently, shape 2 becomes the new anchor for shapes non-repeating integers.
5 to 8 until shape 9 is unable to touch both shapes 2 Angle of shape 20 modulo 2π real values.
and 8. Similarly shape 3 is the anchor for shapes 9 Angle for next placement 19 modulo 2π real values.
to 11 and shape 4 is the anchor for all other shapes
shown. Each chromosome section has its own population,
The shapes build up in a ring around a central an- all with the same number of members. The members
chor shape until no more shapes can touch it and the in all three sections are indexed by a single position
preceding shape without overlapping another shape. in the population so that a single member will consist
The shape placed directly following the anchor node of three separate sections from the chromosome lists.
shape now becomes the new anchor node and the over- On placement the radian value at a chosen locus in
16 Crispin et al.

Figure 10. Typical connectivity based lay-plan of 84 shoe parts giving approximately 58% coverage obtained using the adaptive graph
connectivity strategy.

the shape angle chromosome section rotates the shape ing shapes in the list to recombine. This requires the
indicated in the corresponding locus in the shape order angles to be reordered so that the corresponding shapes
chromosome section. from all chromosomes are in the same order and return-
Each of the angles in the shape angle chromosome ing their offspring to their parent’s individual orders
section is associated with the shape at the correspond- after creation. The result would be that the shape angle
ing locus in the shape order chromosome section. When properties of a shape would only recombine with a like
the shapes are reordered, the associated angles are also shape irrespective of shape order. The angles are then
reordered to the same positions as the corresponding reordered to follow their respective loci in the shape
shapes. Consequently, a shape angle is a property of a order chromosome section as described before.
shape since it is linked to the shape order. The angles for next placement describe a method of
The partially matched crossover (PMX) operator as positioning a shape relative to the last shape placed.
described in Goldberg [13] is used to reorganise the or- They remain independent of the shape properties. Ob-
der in which the shapes are placed. PMX combines the servation and analysis of the effect of these values sug-
sequences contained in the two parent solutions to cre- gest that the angles in the first two loci show the great-
ate two new offspring solutions consisting of elements est importance to the final lay, while the rest have more
of both parents while ensuring that no index within the subtle consequences.
order sequence is repeated. The PMX operator respects The fitness value is calculated as the hide area multi-
the absolute positions of individual indices in the order plied by the square of the ratio of the number of shapes
sequence promoting chromosome structures dependant placed to the number of shapes required. If all the re-
on positional attributes. quired shapes are placed the value is equivalent to the
The real valued line recombination operator is used absolute area of the hide, any less increases the fitness
for recombining shape angle vectors and placement an- value significantly. Figure 10 shows a typical result of
gle values. Real valued line recombination [18] is a nesting components using the connectivity coding ap-
mathematical approach to the combination of real val- proach. The graph spirals in from the edge of hide since
ued solutions. Offspring are created by finding a ran- the coding method attempts to place shapes so that they
domly weighted average of each individual position in connect to the previously placed shape and an anchor
the chromosome. The line recombination operator uses shape that is initially the hide edge.
a single weighting value for all positions in the chro-
mosome so that solutions are developed on a straight 6. Results
line between each pair of parents.
To associate the properties of shape types it was de- An empirical analysis to measure statistically the
cided to only allow the shape angles from correspond- quality of the final solution (coverage, time) for the
 Genetic Algorithm Coding Methods for Leather Nesting 17

Figure 11. (a) Change in average percentage coverage for different window sizes (b) change in average time taken for typical lay of whole
hide verse window size (1.2 GHz Athlon running Windows 2000).

developed approaches has been undertaken for a range

of parameter settings.
For a representative hide, the results for the packing
approach can be summarised as follows.

(i) Changing window size from no window (search

whole image) to a size one third of the image
height results in a small change in coverage (from
59 to 60%) for a fixed population size (15) and
number of generations (25) See Fig. 11(a).
(ii) A window size that is a third of the image height
takes 50% less time to obtain the same coverage
than the case where no window is used for fixed
population size (15) and number of generations
(25). It was concluded that a window size one third
of the image height yielded good coverage in an Figure 12. Average percentage coverage verses population size for
packing algorithm with window size equal to 1/3 image height over
acceptable computational time. See Fig. 11(b).
25 runs.
(iii) For a fixed window size (1/3 of image height) there
tends to be a linear increase in average percent-
age coverage for the range of populations tested 47% with a population size of 5 to 65% with a
(5,15,25,50). See Fig. 12. The algorithm was eval- population size of 50. When the algorithm is eval-
uated with a setting of 25 generations. The in- uated with a setting of 25 generations the average
crease in coverage is coupled with a linear increase directionality percentage increases from 57% with
in time. See Fig. 13. When the algorithm was eval- a population size of 5 to 67% with a population
uated with a setting of 10 generations the average size of 50.
percentage coverage curve shows a similar trend.
(iv) The average percentage of shapes requiring di- The connectivity driven approach requires that a
rectionality that are within the ±5◦ specification shape be placed so that it touches an anchor shape (ini-
requirement increases with population size. When tially the hide edge) and the shape previously placed.
the algorithm is evaluated with a setting of 10 gen- This promotes a spiral lay arrangement. Initial testing
erations the average percentage of shapes meet- was based on the compaction of 20 shapes from 6 dif-
ing the directionality specification increases from ferent shape types [19]. The results showed that
18 Crispin et al.

shapes first as used in the packing algorithm approach

could increase coverage. It should be noted that be-
cause shapes are place at NFP crossing points the GA
is tuning an initial good placement. The drawback of
the approach is the difficulty of developing coding con-
straints and recombination methods that prevent graphs
creating invalid lays when grading and directionality
constraints are applied.

7. Conclusions

In order to develop a practically effective solution for

leather nesting the research has investigated two cod-
ing methodologies both based on no-fit polygon cal-
culations. The first approach is based on a local pack-
Figure 13. Average time taken versus population size for packing ing paradigm and demonstrates how a window area
algorithm with window size equal to 1/3 image height over 25 runs
is used to imitate the actions of a human lay-planner.
(1.2 GHz Athlon running Windows 2000).
The method takes into consideration multiple irregu-
lar shape types and part-placement using directionality
(i) Modest coverage improvements (of order 10%) (lines of tightness) and grading constraints. The GA
could be obtained by the genetic algorithm opti- uses a population consisting of lay angles and direc-
misation of graph node parameters from the initial tionality ratios to optimise part placement within the
lay (see Fig.14). window region.
(ii) Large sample populations result in good initial The second approach is based on a graph connec-
lays. tivity strategy. Graph nodes are considered to have
properties shape order, shape angle and placement an-
gle. Graph edges describe connections between shapes.
Small percentage coverage improvements occur
Graphs are created by checking that edges between
when the graph connectivity approach is applied to
nodes can realistically exist using rules to connect
leather nesting. It was also found that placing largest
shapes at NFP crossing points so that non-overlapping
placements are always made. A GA is used to evolve
node properties (shape order, shape angle and place-
ment angle) for leather nesting.
The results of experiments have shown that both
methods can be used for leather lay planning. How-
ever, it has been found that the packing driven approach
provides a simpler coding solution and produces better
lay-plan results for the test cases used. Also the imple-
mentation of grading and directionality constraints is
more readily realised using this approach. Experiments
show that the packing method is competitive with hu-
man nesters in terms of material useage.

Acknowledgments

The authors gratefully acknowledge the support of this

Figure 14. Compaction area improvement for connectivity algo-
rithm taken over 30 generations as a ratio of area of convex hull research by the U.K. E.P.S.R.C. grant on “A Genetic
to total area of shapes placed. Population sizes of 5 (solid line), 15 Algorithm Approach to Leather Nesting Problems”,
(dotted line) and 25 (dashed line) are shown. reference number: GR/M82110.
 Genetic Algorithm Coding Methods for Leather Nesting 19

References for the Breeder Genetic Algorithm,” Evolutionary Computation,

vol. 1, no. 1, pp. 25–49, 1993.
1. K.A. Dowsland and W.B. Dowsland, “Solution approaches to 19. A.J. Crispin, P. Clay, G.E. Taylor, R. Hackney, T. Bayes, and D.
irregular nesting problems,” European Journal of Operations Reedman, “Genetic algorithm optimisation of part placement
Research Elsevier, vol. 84, pp. 506–521, 1995. using a connection based coding method,” in 15th International
2. T. Anand, S. McCord, C. Sharma, and R. Balachander, “An Conference on Industrial and Engineering Applications of Arti-
integrated machine vision based system for solving the non- ficial Intelligence and Expert Systems, Cairns, Australia, 2002.
convex cutting stock problem using genetic algorithms,” Jour-
nal of Manufacturing Systems, vol. 18, no. 6, pp. 396–415,
1999.
3. C. Bounsaythip, S. Maouche, and M. Neus, “Evolutionary search
techniques: Application to automated lay-planning optimisation
problem,” IEEE international conference on Systems, Manufac-
turing and Cybernetics (SMC), Vancouver, Canada, 1995, vol.
5, pp. 4497–4503.
4. A. Kahng, “Classical Floorplanning harmful?” in International
Symposium on Physical Design, Proceeding ACM/IEEE, 2000,
pp. 207–213.
5. P. Clay and A.J. Crispin, “Automated lay-planning in leather
manufacturing,” in 17th National Conference on Manufacturing
Alan Crispin is a Principal Lecturer in the School of Technology at
Research, Cardiff, 2001, pp. 257–262, ISBN 1-86058-325-3.
Leeds Metropolitan University. His current research interests have fo-
6. M. Adamowicz and A. Albano, “Nesting two-dimensional
cussed on component lay-out optimisation using genetic algorithms
shapes in rectangular modules,” Computer-Aided Design, vol.
and other heuristics. He is a Chartered Engineer and obtained his
8, no. 1, pp. 27–33, 1976.
Ph.D. from the University of East Anglia. He has authored one book
7. E. Burke and G. Kendall, “Evaluation of two-dimensional bin
and has published 40 papers.
packing using the no fit polygon,” in Computers and Industrial
Engineering Conference, Melbourne, Australia, 1999.
8. R. Cunninghame-Green, “Geometry, shoemaking and the milk
tray problem,” New Scientist, vol. 12, 1677, pp. 50–53 1989.
9. P.K. Ghosh, “A unified computational framework for Minkowski
operations,” Computer and Graphics, vol. 17, no. 4, pp. 119–144
1993.
10. J.A. Bennell, K.A. Dowsland, and W.B. Dowsland, “The irreg-
ular cutting-stock problem—a new procedure for deriving the
no-fit polygon,” Computers and Operations Research, vol. 28,
pp. 271–287, 2000.
11. A.J. Crispin, P. Clay, G.E. Taylor, R. Hackney, T. Bayes, and D.
Reedman, “Genetic Algorithms applied to material utilisation
in leather shoe manufacturing,” in 18th National Conference on Paul Clay is a doctoral student currently working as Product De-
Manufacturing Research, Leeds, 2002, pp. 267–272. ISBN 1- velopment Manager at Pulsonic Technologies Ltd. His Ph.D. studies
86058-378-4. involve the application of heuristic optimisation to leather nesting. He
12. J. Heistermann and T. Lengauer, “The nesting problem in the received his B.Eng. (Hons) in Electrical and Electronic Engineering
leather manufacturing industry,” Annals of Operations Research, from Leeds Metropolitan University.
vol. 57, pp. 147–173, 1995.
13. D.E. Goldberg, Genetic Algorithms in Search, Optimisation and
Machine Learning, Addison Wesley, Longman Inc. 1989, ISBN
0-020-15767-5.
14. M. Mitchell, An Introduction to Genetic Algorithms,
Massachusetts Institute of Technology Press, 1996, ISBN 0-262-
13316-4.
15. P. Clay, A.J. Crispin, and S. Crossley, “Comparative analysis of
search methods as applied to shearographic fringe modelling,” in
Proceeding of 13th International Conference on Industrial and
Engineering Applications of Artificial Intelligence and Expert
Systems, New Orleans, June 2000, pp. 99–108.
16. B. Brassard and P. Bratley, Fundamentals of Algorithmics, Gaynor Taylor has recently retired as Deputy Vice-Chancellor at
Prentice-Hall: New Jersey, 1996, ISBN 0-13-335068-1. Leeds Metropolitan University. She has been involved in robotics
17. J. Gross and J. Yellen, Graph Theory and Its Applications, CRC and automation since the early eighties when she was a member
Press: Florida, 1998, ISBN 0-8493-3982-0. of the Robotics Research Group at Hull University. She received a
18. H. Mühlenbein and D. Schlierkamp-Voosen, “Predictive Models B.Sc., M.Sc. and Ph.D. from UMIST.
20 Crispin et al.

Tom Bayes is team leader of the Advanced Concepts section within David Reedman, recently deceased, was formally the research man-
the Shoe and Allied Trade Research Association (SATRA). He has ager of British United Shoe Machinery with particular expertise in
been working on the problem of leather cutting for several years automatic control and system engineering. He had been working in
having developed one of the first computer-aided nesting systems to the shoe industry for over 30 years.
meet industry requirements.