See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228691239

A Genetic Algorithm for 2D Glass Cutting Problem

Article · October 2005

CITATION READS

1 2,900

5 authors, including:

Will Benfold Antonio Manso

3 PUBLICATIONS   6 CITATIONS   
Polytechnic Institute of Tomar
35 PUBLICATIONS   60 CITATIONS   
SEE PROFILE
SEE PROFILE

Salvatore Spinella
Università degli Studi di Torino
27 PUBLICATIONS   147 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Algoritmi View project

VITASENIOR-MT: Senior healthcare assistance in Medio Tejo (Assistência aos cuidados de saúde de idosos no Médio Tejo) View project

All content following this page was uploaded by Salvatore Spinella on 19 May 2014.

The user has requested enhancement of the downloaded file.

 Université Libre de Bruxelles
Institut de Recherches Interdisciplinaires
et de Développements en Intelligence Artificielle

A Genetic Algorithm for 2D Glass Cutting

Problem

William Benfold, Max Manfrin, António Manuel

Rodrigues Manso, Salvatore Spinella

IRIDIA – Technical Report Series

Technical Report No.
TR/IRIDIA/2005-13
September 2005
 IRIDIA – Technical Report Series
ISSN 1781-3794

Published by:
IRIDIA, Institut de Recherches Interdisciplinaires
et de Développements en Intelligence Artificielle
Université Libre de Bruxelles
Av F. D. Roosevelt 50, CP 194/6
1050 Bruxelles, Belgium

Technical report number TR/IRIDIA/2005-13

The information provided is the sole responsibility of the authors and does not necessarily
reflect the opinion of the members of IRIDIA. The authors take full responsability for
any copyright breaches that may result from publication of this paper in the IRIDIA –
Technical Report Series. IRIDIA is not responsible for any use that might be made of
data appearing in this publication.
 A Genetic Algorithm for 2D Glass Cutting Problem
William Benfold [email protected]
University of Southampton, Southampton, Great Britain
Max Manfrin [email protected]
IRIDIA, Université Libre de Bruxelles, Brussels, Belgium
António Manuel Rodrigues Manso [email protected]
Instituto Politécnico de Tomar, Tomar, Portugal
Salvatore Spinella [email protected]
Università della Calabria, Arcavacata di Rende, Italy

August 2003

Abstract
A Genetic algorithm is a heuristic approach which applies analogies from nature to tackle
complex optimization problems. Existing solutions (the ancestors) are randomly modified to
produce offsprings which might replace their parents.
We apply a genetic algorithm to a special case of a 2D bin packing problem, using a
permutation-style representation decoded by a modified first-fit heuristic

1 Introduction
A glass maker usually starts from a set of desired pieces that have been commisioned to him, and
a set of sheets to cut from. Loss is created when a sheet can not be exactly covered by the desired
pieces. Loss is constituted by small glass pieces, called scraps, that can not be utilized in any way.

For economical reasons, glass makers require a given list of pieces to be placed over the smallest
number of sheets, and the total glass loss to be minimized. There are also several technological
constraints, deriving from how the glass sheet is actually cut by automatic or semi-automatic
cutting machines (We don’t treat these).

2 Problem Description
Given sheets with H hight and W witdh and given a set of pieces
m1 with h1 hight and w1 witdh
m2 with h2 hight and w2 witdh
... mn with hn hight and wn witdh

We wish to find a way to cut the pieces from the sheets which minimizes the area of wasted
glass.However, there are additional constraints:

• All cuts must be either horizontal or vertical

• All cuts must be “guillotinable”; cuts must run across the whole length of the glass (2)

1
2 IRIDIA – Technical Report Series: TR/IRIDIA/2005-13

Figure 1: An example of a possible layout

Guillotine cutting pattern non−Guillotine pattern

Figure 2: An example of a guillotinable layout

• There are also other industrial constraints which we shall not consider

A solution to this problem takes the form of a cutting tree. This is a tree for which each
node represents one guillotinable cut, with each branch representing one of the two new subsheets
created; an example is shown in 2). In industry, the glass is cut in stages, with each stage producing
the cuts for one level of the tree. The depth of the tree is therefore limited to the number of stages
used by the glassmaker. For the problems we shall be considering, there are three stages, so our
cutting tree can have no more than three levels.
The 2D glass cutting problem is a particular case of a more general problem, namely the 2D
Bin Packing Problem.
Lodi et al. [7] proposed the following topology for the four possible cases produced by orien-
tation and/or guillotine cutting for 2D Bin Packing Problem:

2BP—O—G the items are oriented (O) and guillotine cutting (G) is required;

2BP—R—G the items may be rotated by 90 degree (R) and guillotine cutting is required;

2BP—O—F the items are oriented and cutting is free (F);

2BP—R—F the items may be rotated by 90 degree and cutting is free

The problem we shall consider is thus 2BP—R—G. The following references are examples of
industrial applications involving the above variants. A problems of trim-loss minimization in a
crepe-rubber mill, studied by Schneider [10], induces subproblems of 2BP—O—G; fuzzy two-
dimensional cutting stock problems arising in the steel industry, discussed by Vasko et al. [11], are
related to 2BP—R—G; the problem of optimally placing articles and advertisemens in newspaper
and yellow pages, studied by Lagus et al. [12] falls into the 2BP—O—F case; finally, several
applications of 2BP—R—F are considered by Benbtsson [13].
IRIDIA – Technical Report Series: TR/IRIDIA/2005-13 3

1. row

2. row

3. row

Figure 3: An example of a layout which can be produced using only two stages

Sheet
4
2 3
1 34
2
1

2 2
3 3
1

Figure 4: Diagram of a cutting tree

4 IRIDIA – Technical Report Series: TR/IRIDIA/2005-13

3 A Simple Heuristic
3.1 The First Fit Heuristic
A very simple way to construct a cutting tree is to use a “First Fit” heuristic. The rectangles are
first sorted into height order, with the orientation being chosen so as to maximise height. The first
rectangle is placed in the upper left corner of the first sheet. Subsequent rectangles are placed to
the right of this until there is insufficient space to add another. The first row is now complete,
with a height equal to that of the first rectangle. This process is repeated until no more rows can
fit on the sheet, and is then repeated for the other sheets until no rectangles remain.
This heuristic is a “greedy algorithm”, which considers only the heights of the rectangles.
Within a row, the columns produced by the vertical cuts will each contain only one rectangle; no
attempt is made to stack several rectangles together in a single column.

3.2 An Improved Heuristic

We present an improvement on the First Fit heuristic, whereby we attempt to fit multiple rectan-
gles into the same column. Note that due to the restriction on the depth of the cutting tree, only
rectangles of the same width can be placed in the same column.
When we have completed a row, we do not immediately proceed to the next. Instead, we
iterate over all remaining unplaced rectangles, looking for any which have the same width as one
of the columns, and are sufficiently small to be placed in the same column without increasing the
height of the row.

4 The GA
4.1 The fitness function
The fitness function is derived from the criterion to be optimized and serves as a measure to
evaluate individuals. Then, a successive improvement of the population’s mean quality can be
achived. However since genetic algoritms still are heuristic strategies, it cannot be guarateed that
an optimal solution will be found.

The simplest bound for 2D Glass Cutting Problem is the Continous Lower Bound
!n
wj hj
L∅ = i=1 (1)
WH
where

• n is the number of the piece cutted

• wj hj is the area of piece j

• W H is the total area of the sheet

In our application a secondary objective is the maximization of th unused area in the last
sheet, so as to produce a possibly large trim to be used later. The same secondary objective was
used for solving 2BP—R—F in [13].
We combine the two objectives by defining our fitness function in terms of the number of sheets
used and the fraction of the last sheet which is wasted. The function we use is:
Nused − L∅
Nused
where Nused is the number of sheets completely used plus the fraction of the available height
used on the last sheet. Hence our fitness function should be minimized to produce optimal results.
IRIDIA – Technical Report Series: TR/IRIDIA/2005-13 5

4.2 Representation
We use a permutation representation, but with the exception that we use duplicated labels to rep-
resent identical elements, rather than assigning every element a unique label. In the chromosome
we also include an additional entry for the rotated version of each rectangle. The cutting tree
is created by decoding this permutation using the Improved First Fit heuristic described above.
Instead of sorting rectangles by height order, we sort them in order according to the chromosome.
Thus the chromosome is a queue of rectangles to be placed by the heuristic. We track the number
of rectangles of each type that we have placed, so that we do not place a rotation of an element
which has already been placed.

4.3 Operators
4.3.1 Mutation operator

Similar, to nature, the task of mutation in a genetic algorithm is to effect random variations of
the genes which are lost in the current population and which cannot be gained if only the existen
material is combined. In general, mutation aims at preveservating the genetic diversity within a
population of individuals.
We shall use two different mutation operators:

• Swap: exchange two randomly chosen elements

• Move: extract a random element and reinsert at a randomly chosen postion

4.3.2 crossover operator

To simulate the sexual propagation of individuals a recombination operator is included into genetic
algorithms. its task is to combine the genetic material of offsprings which inherits certain good
characteristics of its ancestors.
We shall explore two different crossover operators:

• Uniform order based, modified to accept duplicate labels

• Elitist:

– Adopt genes appearing at the same positions in both parents

– Choose the remainder from alternate parents where possible

Where we are unable to choose genes from alternate parents (because no genes of that type
remain to be placed), we arbitrarily choose to insert the lowest-numbered gene which still remains
to be placed.

4.4 Local Search

We define the neighbourhood of a permutation to be all other permutations which can be reached
through a single swap. Because our chromosomes contain duplicated labels, we choose to ignore
swaps which involve identical elements. The neighbourhood is still too large to explore exhaus-
tively, so we a attempt a randomized first-improvement search. A swap is applied, and the fitness
of the new permutation evaluated. If an improvement has been made, we terminate the search,
and overwrite the original permutation with the improved one. Otherwise we undo that swap and
apply a different one. If we try 30 swaps without any improvement being found, we terminate the
search.
6 IRIDIA – Technical Report Series: TR/IRIDIA/2005-13

0.6
"m0c0_avg"
"m0c1_avg"
"m1c0_avg"
"m1c1_avg"
"m0c0_l_avg"
0.5 "m0c1_l_avg"
"m1c0_l_avg"
"m1c1_l_avg"
"orig"
0.4

0.3

0.2

0.1

0
20 25 30 35 40 45 50 55

Figure 5: A plot showing average solution quality. The horizontal axis gives the problem number,
the vertical axis gives fitness. Lower fitnes is better. The “orig” series shows the results obtained
from the simple heuristic.

5 Experimental Setup
The experiments was carried out on with the follow features

Processor Intel(R) Pentium(R) 4 CPU 2.60GHz

Memory 500 MB
Compiler environment gcc version 3.3

6 Results
We observe that in many cases there is no significant difference in quality between the solutions
produced by each variation of the algorithm. In a large number of cases, the solution quality is
identical to that of the First Fit heuristic. Such cases appear to arise when the problem being
considered is sufficiently simple that all algorithms, including the simple heuristic, are able to find
the global optimum. For instance, there are a large number of problems containing rectangles
which occupy more than half of a sheet; it is usually trivial to find a way to place the remaining
rectangles without requiring any additional sheets. In addition, there are a few problems where
the all of the rectangles fit on a single sheet, and there are so few rectangles that problem could
easily be solved by enumerating all possible placements.
The problems where the algorithms differ in performance seem to be the most difficult instances.
Such problems usually have a large number of different shapes of rectangle.
 IRIDIA – Technical Report Series: TR/IRIDIA/2005-13 7

7 Conclusion
It would appear that both crossover and both mutation operators yield similar results, except on
the hardest problems. Similarly, local search does not produce a consistent increase in the quality
of solution. There appears to be, when all problems are considered, no particular advantage to be
gained from using a particular operator, or from using local search. We do note, however that in
almost all cases, the GA produces solutions of quality better than or equal to those of the simple
heuristic.
Acknowledgements

Whe thank the organizers and teachers of the EvoNet Summer School 2003 for all the advices.
In particular we thank Günther Raidl of the Vienna University of Technology, for providing us
with the 2DBP framework and the GA library that we have used as starting point for this work.
The computational experiments have been executed at the Laboratori CEDI- Facoltá di Ingegneria
- Universitá di Parma

References
[1] F. Corno, P. Prinetto, M. Rebaudengo, M. Sonza Reorda, S. Bisotto Optimizing Area Loss in
Flat Glass Cutting.
[2] E. Hopper and B. C. H. Turton, An Empirical Investigation of Meta-Heuristic Algorithms for
a 2D Packing Problem, European Journal of Operation Research 128/1,34-57, 2000.
[3] E. Hopper and B. C. H. Turton, A Review of the Application of Meta-Heuristic Algorithm to
2D Strip Packaging Problems, Artificial Intelligence Review, vol. 16, 257-300, 2001.
[4] Shian-Miin Hwang, Cheng-Yan Kao, Jorng-Tzong Horng, On Solving Rectangle Bin Packing
Problems Using Genetic Algorihms, IEEE 1994.
[5] Berthold Kroger, Guillotineable bin packing: A genetic approach, European Journal of Oper-
ational Research 84 (1995) 645-661.
[6] A. Lodi, S. Martello, D. Vigo, Recent advances on two-dimensional bin packing problems,
Discrete Applied Mathematics 123 (2002) 379-396.
[7] A. Lodi, S. Martello, D. Vigo, Heuristic and metaheuristic approaches for a class of two-
dimensional bin packing problems, INFORMS J. Comput. 11 (1999) 345-357.
[8] D. E. Goldberg, Genetic Algorithm in Search, Optimization and Machine Learning, Addison-
Wesley, Reading, MA 1988.
[9] M.R. Garey, D. S. Johnson, Computers and Intractability: A Guide to the Theory of NP-
Completness, Freeman, San Francisco, CA, 1979.
[10] W. Schneider, Trim-loss minimization in a crepe-rubber mill; optimal solution versus heuristic
in the 2 (3)-dimensional case, Eur. J. Oper. Res. 34 (1988) 273-281.
[11] F.J. Vasko, F. E. Wolf, K. L. Scott, A pratical solution to a fuzzy two-dimensional cutting
stock problem, Fuzzy Sets and System 29 (1989) 259-275.
[12] K. Lagus, I Karanta, J. Ylä-Jääski, Paginating the generalized newspaper: A comparison of
simulated anniling and a heuristic method, Proceedings of the Fifth International Conference
on Parallel Problem Solving from Nature, Berlin, 1996, pp 549-603.
[13] B. E. Bengtsson, Packing recangular Pieces - a heuristic approach, Comput. J. 25 (1982)
353-357

View publication stats