Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

A New Guillotine Placement Heuristic for the

Orthogonal Cutting Problem*

Slimane Abou Msabah1 — Ahmed Riadh Baba-Ali2


1
Department of Computer Science, University of Science and Technology Houari
Boumedienne, USTHB, Bab Ezzouar, Algiers, Algeria, slmalg@yahoo.com
2
Department of Electronics, University of Science and Technology Houari Boumedienne,
USTHB, Bab Ezzouar, Algiers, Algeria, riadhbabaali@yahoo.fr

Abstract. The orthogonal cutting problem consists in finding an optimal ar-


rangement of n items on identical dimension bins. Several placement heuristics
are used to realize this task. The constraint of guillotine complicates more the
problem. In our article, we are interested in the orthogonal cutting problem, tak-
ing into account the guillotine constraint. To do it, we propose a new placement
heuristic inspired by the BLF routine, and which tries to place the items on le-
vels, to verify the guillotine constraint, while exploiting intra-levels residues, in
two directions vertically then horizontally. Our heuristic named BLF2G will be
combined with a guided genetic algorithm, to be compared with the other heu-
ristics and metaheuristics found in literature, on made test sets and known test
sets.

KEYWORDS: orthogonal cutting, guillotine constraint, combinatorial optimi-


zation, heuristics, genetic algorithm.

1 Introduction
The problem of cutting or placement is an optimization problem, whose objective is to
determine a suitable arrangement of various items in others that are wider. The main
objective is to maximize the use of the raw material, and thus to minimize the losses.
The orthogonal cutting problem pulls its interest of the fact that it is applicable on
several fields such as the cut of sheet steel, paper, fabrics, etc. … This is important for
the industries of mass production where the optimization of the material plays an im-
portant role in the cost of manufacturing.
In our work, we will propose a new guillotine placement routine, which aims the
exploitation of the raw material. The development of such routine has to consider
several parameters, such as the shape of the treated objects and the constraints im-
posed by the production system.
In our case, we considered an orthogonal cutting problem, which treats a strip of
fixed width and supposed infinite height, to generate items of rectangular shape. The
used material can be a steel sheet and the machines of production are typically shears
guillotine, which impose the cut from edge-to-edge (guillotine constraint). Items keep
their original orientations to be cut on decorated or textured plates.

* Proceedings of the 18th RCRA workshop on Experimental Evaluation of Algorithms for


Solving Problems with Combinatorial Explosion (RCRA 2011).
In conjunction with IJCAI 2011, Barcelona, Spain, July 17-18, 2011.
Full paper can be found in : http://slmalg.unblog.fr/

1
The routine, which we propose, pulls its profile of the BLF (Bottom Left Fill) rou-
tine; the layout of items is made in levels, to insure the guillotine constraint. Two oth-
er mechanisms are setup to exploit intra-levels residues, vertically and horizontally.
After this introduction to the problem, we lists in section 2 of this article the place-
ment routines found in literature. In section 3 we expose our routine, and we show its
utility in exploitations of residues on test sets made to fit the bin. The results of our
method are explained in section 4, which will be compared with the other heuristics,
on test sets found in literature. We end our article with a conclusion, which will be
presented in section 5.

2 State of the art


We are interested in this section to investigate the placement heuristics found in the
literature from the existing methods to propose a routine which fits to our case. In
order to take advantage.
Baker Coffman and Rivest, present the BL (Bottom Left) heuristics to place the
items in the lowest place to the left. They test several sequences of appearance of
items and find that the sorting of the list of items according to the diminution of the
widths gives better results [1].
Jakobs uses the BL heuristics, which tries to place the item in the highest place to
the right, then he slides the item successively in the lowest place then most to the left
possible. He uses afterward a genetic algorithm to find the sequence, which gives best
result [9].
Lui and Teng perfected the BL heuristics by favouring the sliding of the item from
top to bottom, the sliding of the item is made of right to the left if no movement below
is possible [10].
The BLF Routine tries to place items in the lowest place to the left by exploiting
the internal residues. Ramesh Babu and Ramesh Babu backs up the points which form
the left lower corners of residues in a list. For every item to be placed an algorithm
goes through the list, respecting the lowest order at the left and place the item in the
first suitable place. So the list is modified according to the dimension of the placed
item. [13].
Lodi et al. presented an approach named "Floor-Ceiling" (FC) who spreads the way
of placing items on the levels. The approach FC places the items from left to right at
the bottom of level and also place items from right to left at the top of the level [11].
Lodi et al. present a new variant of the FC routine to verify the guillotine con-
straint. This variant realizes the cuttings from edge to edge.
Ben Messaoud et al. present a new placement heuristic, based on levels (SHF) to
apply the FC algorithm by injecting items placed in the ceiling below [2].
Burke et al. present their Best Fit heuristic by the introduction of the notion of
neighborhood. They build their format of cut as one goes along. Initially the list of
items is sorted by decreasing height. They try to fill the lowest residue by the items
which suits, if no item is suitable, they fill this space by an irrecoverable scrap until
the next level, and so on; as a bricklayer who builds a wall. The item is moved either
to the left or to the right according to the height of neighboring items. They also intro-
duce a mechanism of orientation of the long items to reduce the losses [4].

3 Our contribution
The placement policy is the crucial point in the laying out process, to maximize the
exploitation of residues. We can divide the placement heuristics into two categories:

2
a) The direct heuristics, which place directly the current item in the first suitable place
according to the applied policy, we can quote the routines Finite First Fit, Finite Nest
Fit, Bottom Left, Bottom left Fill [3] … b) The heuristics which chooses, they choose
a suitable place among various according to their laying out policies, to place the cur-
rent item, such as Finite Best Fit, Finite Worst Fit …
The first category of heuristics realizes the laying out at short time, and it depends
totally on the order of appearance of the items. These heuristics are favorable to be
combined with stochastic algorithms and metaheuristics. The second category requires
an additional time to realize the laying out and gives, generally better results than the
first category.
Our placement routine tries to place items directly to the suitable place, according
to a laying out policy which verifies the guillotine constraint. It belongs to the first
category.

3.1 Our Placement policy


The use of an adequate placement policy gives better results. The guillotine constraint
makes the problem more complicated. To find a good placement policy, which fits
better to this established fact, we took advantage from placement methods which ex-
ists in the literature. The routines which base on the laying out in level are the best
adapted to our scenario, the cut from edge to edge required by shears-guillotines
adapts itself better to the layout in levels.
The laying out of an item on the strip can be made according to three possible cas-
es:

Placement in levels. The strip is structured in levels; every level is characterized by a


height and an available width. If the width of the item is lower or equal to the availa-
ble width the item is thus placed. The available width is updated, and if the height of
the item is superior to the level, the height of the level is redefined by the height of the
item. We named this phase BLG, because it is a question of applying the routine of
placement Bottom Left to levels to verify the constraint of Guillotine. The application
of this routine engenders BLGsub-levels (Fig.1).

Height of the Level i, as high as the


Level i longest item.

The BLGsub-levels Available width in the level i.

Level 1

Fig. 1. The layout on the levels

Placement in BLGsub-level:. A BLGsub-level is characterized by a width, equal to


the width of the item placed below, and by the available height. Items are ordered on
these residues vertically. If the width of the item is lower or equal to the width of
BLGsub-level and the height of the item is lower or equal to the available height of
BLGsub-level, the item is thus placed in this BLGsub-level and the available height is
updated. This phase is named BLFG (Bottom Left Fill Guillotine). The application of
this routine engenders BLFGsub-levels (Fig. 2).

3
Available height of
BLGsub-levels
The BLFGsub-levels

Width of BLGsub-levels

Fig. 2. The layout on BLGsub-levels

Placement in BLFGsub-level :. A BLFGsub-level is characterized by a height and


an available width. If the height of the current item is lower or equals the height of the
level and the width of the current item is lower or equal to the available width of
BLFGsub-level, the item is thus placed in this BLFGsub-level and the available width
of BLFGsub-level is updated. This phase is named BLF2G (Bottom Left Fill 2
(second exploitation of residues) Guillotine) (Fig. 3.).

Height of Available width of


BLFGsub-levels BLFGsub-levels

Fig. 3. The layout on BLFGsub-levels

After this stage, the residues are too small to be exploited, but if we have a data set of
smallest items, we can continue the exploitation using BLF3G, BLF4G, etc…

3.2 The placement algorithm


Read the dimension of the plates Then place item
Load list of item Update BLFGsub-level
For all items break End if
For all levels Pass at the following
If current item can be BLFGsub-level
placed in the level End For
then place item Pass at the following
Update the level BLGsub-level
Break End if End For
For all BLGsub-levels Pass at the following level
If current item can be End For
placed in the BLGsub-level If item not placed
Then place item then place item in a new level
Update the BLGsub- update the new level
level End if
Break End if Pass at the following item
For all BLFGsub-levels End for
If current item can be End
placed in the BLFGsub-level

3.3 The genetic Algorithm


To show the power of our policy of placement BLF2G, we will combine it with a
classic GA. We used a real codification[7] where the chromosome is defined by the
order of items. The order of appearance of items in the process of laying out, accord-

4
ing to our policy BLF2G, determines the quality of every individual. We implemented
a genetic algorithm approach with a population size of 100 and we fixed the number
of generations by 20 times the number of items. Initially we generate a random popu-
lation with a random ordering item in each individual. At each generation, our BLF2G
policy gives the quality of each individual. The genetic operators are so defined:

The crossover operator. We used the partially matched cross-over with 1-point
cross-over (PMX1). We make then a correction to make the children’s valid. We are
going to correct the child 1 with the missing genes according to their order of appear-
ance in parent 2, and replace the double genes in child 2 by the missing genes accord-
ing to their order of appearance in parent 1.
Parent 1 : 123|456 123321 Child 1 : 123654
PMX1  Correction
Parent 2 : 654|321 654456 Child 2 : 654123

The Mutation operator. It is about a permutation between two sites chosen random-
ly:
Child : 1|2365|4  143652

4 Experimental results
To estimate the performance of our heuristic we made a test sets which offer a max-
imal exploitation of the material (0 % of scraps) by applying the BLF2G policy.
name # of Item Plates dimension Optimal height
Msa17a, Msa17b, Msa17c 17 200 x 200 200
Msa35a, Msa35b, Msa35c 35 200 x 200 200
Msa75a, Msa75b, Msa75c 75 200 x 200 200
Msa150a, Msa150b, Msa150c 150 200 x 200 200
Table 1. Our made test sets
To estimate our heuristics we combined it with a genetic algorithm. The obtained
results are compared with our policy BLF2G by applying a sorting according to the
Decreasing Heights to the list of items.
Msa17 Msa35 Msa75 Msa150
a b c a b c a b c a b c
BLF2G DH 240 245 263 220 225 229 214 210 210 205 205 218
BLF2G GA 200 200 200 220 215 219 215 210 218 205 205 219

Table 2. Results of the routine BLF2G+DH and BLF2G+GA.


We notice that the GA combined with our policy of placement BLF2G, reached the
optimal for the test sets of 17 items (small size), and has given comparable results
with the BLF2G+DH heuristic for test sets of medium size. But for the test sets of big
size the BLF2G+DH heuristic is better.

4.1 Improvement
The GA failed in front of the DH heuristic for the test sets of large-size. To remedy
that we suggest injecting the individual sorted out according to DH policy in the initial
population of the evolutionary process, which we name GAguided. The following table
shows the results.

5
Msa17 Msa35 Msa75 Msa150
a b c a B c a b c a b c
BLF 2G DH 240 245 263 220 225 229 214 210 210 205 205 218
BLF2G GA 200 200 200 220 215 219 215 210 218 205 205 219
BLF2G GAguided 200 200 200 215 210 219 207 205 210 205 205 212

Table 3. Result of the BLF2G+DH routine, BLF2G+GA and BLF2G+GAguided


With this improvement the GAguided keeps its superiority with regard to the DH heuris-
tic and gives results equal and even better than the DH heuristic for the test sets of any
size.

4.2 Test sets found in the literature


To estimate better our method we are going to use the test sets of Hopper and Turton
(2001) [8], Burke et al. (2004) [4], which are the most used:
Test Size Optimal Test Size Optimal
Name box Size Name box Size
set # of items Height set # of items Height
C1:P1,P2,P3 16/17 20 20x20 N4 40 80 80x80
Burke et Hopper & Turton

C2:P1,P2,P3 25 15 40x15 N5 50 100 100x100


Burke et al. 2004

C3:P1,P2,P3 28/29 30 60x30 N6 60 100 50x100


2001

C4:P1,P2,P3 49 60 60x60 N7 70 100 80x100


C5:P1,P2,P3 73 90 60x90 N8 80 80 100x80
C6:P1,P2,P3 97 120 80x120 N9 100 150 50x150
C7:P1,P2,P3 196/197 240 160x240 N10 200 150 70x200
al. 2004

N1 10 40 40x40 N11 300 150 70x200


N2 20 50 30x50 N12 500 300 100x300
N3 30 50 30x50 N13 3152 960 640x960

Table 4. Test sets found in literature.


The following table presents the obtained results by applying our GAguided+BLF2G
policy with regard to GA+BLF, SA+BLF, New Best-Fit find in Burke et al. (2004).
Test Name GA+ SA+ New GA- Test Name GA+ SA+ New GA-
sets BLF BLF Best-Fit guided+BLF2G sets BLF BLF Best-Fit guided+BLF2G
C1P1 20 20 21 20 C6P3 126 126 124 126
C1P2 21 21 22 22 C7P1 255 255 247 250
C1P3 20 20 24 21 C7P2 251 253 244 248
C2P1 16 16 16 16 C7P3 254 255 245 249
C2P2 16 16 16 16 N1 40 40 45 40
Hopper & Turton 2001

C2P3 16 16 16 15 N2 51 52 53 50
C3P1 32 32 32 32 N3 52 52 52 53
C3P2 32 32 34 33 N4 83 83 83 87
Burke et al. 2004

C3P3 32 32 33 31 N5 106 106 105 105


C4P1 64 64 63 66 N6 103 103 103 106
C4P2 63 64 62 65 N7 106 106 107 116
C4P3 62 63 62 62 N8 85 85 84 85
C5P1 95 94 93 95 N9 155 155 152 153
C5P2 95 95 92 97 N10 154 154 152 154
C5P3 95 95 92 95 N11 155 155 152 153
C6P1 127 127 123 125 N12 313 312 306 309
C6P2 126 126 122 128 N13 - - 964 Out of service

Table 5. comparison of the GAguided+BLF2G heuristic to GA+BLF, SA+BLF and New Best-
Fit heuristics.
We notice that for the test sets of small-size the method GA+BLF and SA+BLF gave
results better than the heuristics New Best Fit. Our method gave comparable and
sometimes better results than the other methods, such thing is explained by the care of
the constraint of guillotine. For the test sets C2P3 and N2 our method reached the

6
optimal, while the other methods did not manage to reach it, although they are free of
the guillotine constraint. And that shows the power of our method in the exploitation
of the material.
For the test sets of medium size our method kept its position with regard to the oth-
er methods. But for the test sets of big size our method shows a better score with re-
gard to the methods GA+BLF, and SA+BLF. And that confirms the good adopted
laying out policy. The New Best-Fit heuristic marked a score better than our method
for the test sets of large-size. And that confirms the failure of the GA in front of the
test sets of large-size.
The following graph compares the four methods according to the percentage of fall.
GA+BLF
25
% over optimal

SA+BLF
20
New Best-Fit
15 GAguided+BLF2G
10
5
0
N1

C1P3

N2

C2P2

C3P1

C3P2

N4

C4P2

N5

N7

C5P2

N8

C6P2

N9

C7P3

N10

N12
Tests sets

Fig. 4. Comparison of the GAguided+BLF2G heuristic to GA+BLF, SA+BLF and New


Best-Fit heuristics (% over optimal)
According to the graph, we notice that the New Best-Fit heuristic has badly started for
the test sets of small-size. But for the test sets of medium and big size it took over and
shows better results. Both heuristics {GA, SA}+BLF, gave comparable results be-
tween them, with a light superiority of the genetic algorithm, whatever the size of
problem.
For the test sets of small-size, our method gave comparable results to GA+BLF and
SA+BLF heuristics, and sometimes better. For C1P3 and C1P2 of small size, our
method had a bad results.
For the test sets of medium-size our method diverge with regard to the other heuris-
tics, with some comparable results.
For the test sets of large-size, our method gave better results with regard to
{GA,SA}+BLF, but it stays below the New Best-Fit heuristics.

5 Conclusion
Our contribution, in the problem of rectangular cut, showed its efficiency. First of all
we developed a powerful routine in exploitation of residues named BLF2G. This rou-
tine takes into account the constraint of cut from edge to edge. Secondly, we guided
the genetic algorithm with the greedy heuristics DH, by the introduction of the DH
individual in the initial population.
The use of the GAguided combined with our routine BLF2G, has allowed to reach the
optimal for Msa17 test set (a, b, and c) of small-size. But for the test sets of medium
and large-size, the GA failed to investigate the space of search to find the sequence of
items which offers the optimal solution.
The comparisons made with the other methods which do not take the guillotine
constraint into account on test sets made to insure the optimum without considering

7
this constraint of guillotine, are very encouraging. The optimal is reached by our me-
thod repeatedly, especially for C2P3 and N2, such result is not reached by the other
methods. Such thing explained the legitimacy of our heuristics of placement, in ex-
ploitation of residues.
The Guided GA combined with our heuristic BLF2G allowed us to show the quali-
ties of our method of placement for the test sets of small and mid-size. For the test sets
of big size, our method BLF2G+GAguided was out of service for the test set N13 of
spatial size. Almost the same thing occurred for the test sets MT01, … MT10, of
Burke et al. [5].
Our BLF2G routine depends totally on the GA to find the optimal order of items.
In perspective, we intend in the near future to give more intelligence to our heuristics
BLF2G, by applying new policies of placement, of Best-Fit type, to escape from de-
fects met by the GA.

References
1. B. S. Baker, E. G. Coffman, and R. L. Rivest, “Orthogonal packings in two dimensions,”
SIAM Journal of Computing, 9, 4, pp. 846-855, 1980.
2. S. Ben Messaoud, C. Chu, and M. L. Espinouse, “An approach to solve cutting stock
sheets,” IEEE International Conference on Systems, Man and Cybernetics, pp. 5109-5113,
2004.
3. J. O. Berkey, and P. Y. Wang, “Two dimensional Finite bin packing algorithms.” Journal
of the Operational Research Society, 38, pp. 423-429, 1987.
4. E. K. Burke, G. Kendall and G. Whitwell, “A New Placement Heuristic for the Orthogonal
Stock Cutting Problem,” Operations Research, vol. 52, no. 4, pp. 655–671, 2004.
5. E. K. Burke, G. Kendall, and G. Whitwell, “A Simulated Annealing Enhancement of the
Best-Fit Heuristic for the Orthogonal Stock-Cutting Problem,” INFORMS Journal on
Computing, Vol. 21, No. 3, pp. 505-516, 2009.
6. D. E. Goldberg, “Genetic algorithms in search, optimization, and machine learning,”
Reading, MA: Addison-Wesley, 1989.
7. E. Hopper, and B. Turton, “A genetic algorithm for a 2D industrial packing problem,”
Computers and Industrial Engineering, vol. 37/1-2, pp. 375-378, 1999.
8. E. Hopper, and B. Turton, “An empirical investigation of metaheuristic and heuristic algo-
rithms for a 2D packing problem.” Eur. J. Oper. Res., 128, pp. 34-57, 2001.
9. S. Jakobs, “On genetic algorithms for the packing of polygons,” European Journal of Op-
erations Research, n° 88, pp. 165-181, 1996.
10. D. Liu, and H. Teng, “An improved BL-algorithm for genetic algorithm of the orthogonal
packing of rectangles,” Eur. J. Operational Research, 112, pp. 413-420, 1999.
11. A. Lodi, S. Martello, and D. Vigo, “Heuristic and metaheuristic approches for a class of
two-dimensional bin packing problems,” INFORMS journal on computing, Vol 11, pp.
345-357, 1999.
12. Z. Michalewics, “Genetic Algorithms + Data Structures = Evolution Programs,” Third,
Revised and Extended Edition, Springer, 1996.
13. A. Ramesh Babu, and N. Ramesh Babu, “Effective nesting of rectangular parts inmultiple
rectangular sheets using genetic and heuristic algorithms,” International Journal of Pro-
duction Research, Vol. 37, n°7, pp. 1625-1643, 1999.

You might also like