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

net/publication/220492665

Algorithm 864: General and robot-packable variants of the three-dimensional

bin packing problem.

Article · January 2007

Source: DBLP

CITATIONS READS
47 903

5 authors, including:

David Pisinger Daniele Vigo

Technical University of Denmark University of Bologna
152 PUBLICATIONS   8,917 CITATIONS    175 PUBLICATIONS   9,781 CITATIONS   

SEE PROFILE SEE PROFILE

Edgar den Boef Jan Korst

14 PUBLICATIONS   141 CITATIONS   
Philips
139 PUBLICATIONS   3,734 CITATIONS   
SEE PROFILE
SEE PROFILE

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

Competitive Liner Shipping Network Design View project

EURO, the Association of Operational Research Societies within IFORS View project

All content following this page was uploaded by Jan Korst on 21 May 2014.

The user has requested enhancement of the downloaded file.

Algorithm 864: General and Robot-Packable
Variants of the Three-Dimensional Bin
Packing Problem
SILVANO MARTELLO
University of Bologna
DAVID PISINGER
University of Copenhagen
DANIELE VIGO
University of Bologna
and
EDGAR DEN BOEF and JAN KORST
Philips Research Laboratories

We consider the problem of orthogonally packing a given set of rectangular-shaped boxes into the
minimum number of three-dimensional rectangular bins. The problem is NP-hard in the strong
sense and extremely difficult to solve in practice. We characterize relevant subclasses of packing
and present an algorithm which is able to solve moderately large instances to optimality. Extensive
computational experiments compare the algorithm for the three-dimensional bin packing when
solving general orthogonal packings and when restricted to robot packings.
Categories and Subject Descriptors: G.4 [Mathematical Software]: Algorithm design and anal-
ysis; G.2.1 [Discrete Mathematics]: Combinatorics—Combinatorial algorithms
General Terms: Algorithms, Experimentation
Additional Key Words and Phrases: Bin packing, exact algorithms, constraint programming,
branch-and-bound

S. Martello and D. Vigo were supported by the Italian Ministry of Education, University and Re-
search (MIUR).
D. Pisinger thanks The Danish Natural Science Research Council (SNF) for the support of this
project.
Authors’ current addresses: S. Martello and D. Vigo, DEIS, University of Bologna, Viale
Risorgimento 2, 40136 Bologna, Italy; email: {smartello,dvigo}@deis.unibo.it; D. Pisinger,
DIKU, University of Copenhagen, Universitetsparken 1, DK-2100 Copenhagen, Denmark; email:
[email protected]; E. den Boef, Quintiq, Goudsbloemvallei 12-28 5237, MJs-Hertogenbosch, The
Netherlands; email: [email protected]; J. Korst, Philips Research Laboratories Eind-
hoven, Building WY 2.23, Prof. Holstlaan 4, 5656 AA Eindhoven, The Netherlands; email: jan.
[email protected].
Permission to make digital or hard copies of part or all of this work for personal or classroom use is
granted without fee provided that copies are not made or distributed for profit or direct commercial
advantage and that copies show this notice on the first page or initial screen of a display along
with the full citation. Copyrights for components of this work owned by others than ACM must be
honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers,
to redistribute to lists, or to use any component of this work in other works requires prior specific
permission and/or a fee. Permissions may be requested from Publications Dept., ACM, Inc., 2 Penn
Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481, or [email protected].

C 2007 ACM 0098-3500/2007/03-ART7 $5.00. DOI 10.1145/1206040.1206047 http://doi.acm.org/
10.1145/1206040.1206047

ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
2 • S. Martello et al.

ACM Reference Format:

Martello, S., Pisinger, D., Vigo, D., den Boef, E., and Korst, J. 2007. Algorithm 864: General and
robot-packable variants of the three-dimensional bin packing problem. ACM Trans. Math. Softw.
33, 1, Article 7 (March 2007), 12 pages DOI = 10.1145/1206040.1206047 http://doi.acm.org/10.1145/
1206040.1206047

1. INTRODUCTION
The Three-Dimensional Bin Packing Problem asks for an orthogonal packing
of a given set of rectangular-shaped boxes into a minimum number of three-
dimensional rectangular bins. Each box j ( j = 1, . . . , n) is characterized by a
width w j , height h j , and depth d j . An unlimited number of identical three-
dimensional bins, having width W , height H, and depth D is available. The
boxes have fixed orientation, that is, they may not be rotated, but no further
restriction is imposed. Such packings will be hereafter denoted as general pack-
ings. Without loss of generality, we assume in the following that all input data
is positive integers. Coordinates originate from the bottom-left-behind corner
of a bin, and (x j , y j , z j ) is the point where the bottom-left-behind corner of box
j is positioned.
The problem has several industrial applications in cutting and loading con-
texts. The reader is referred to Dyckhoff et al. [1997] for an annotated bibliog-
raphy on this subject, and to Lodi et al. [2002a, 2002b] for recent surveys.
In some relevant practical applications, it is demanded that the boxes can be
obtained from the bin through a sequence of guillotine cuts, that is, face-to-face
cuts parallel to the faces of the bin: these packings are said to be guillotine cut-
table. In other industrial contexts, boxes have to be packed by robots equipped
with a “hand” parallel to the base of the bins, which is covered with vacuum
cells for lifting the boxes. When the boxes have to be palletized or packed into
a container that is open from three sides while loading, the robot is usually
located at one of the corners. To simplify the packing operations, and to avoid
collisions between the hand and the boxes, it can be demanded that no al-
ready packed box be positioned in front of, right of, or above the destination
of the current box: a robot packing is a packing which can be achieved by suc-
cessively placing boxes starting from the bottom-left-behind corner, and such
that each box is in front of, right of, or above each of the previously placed
boxes.
Note that each guillotine-cuttable packing is also a robot packing. Consider
a guillotine-cuttable packing and a feasible sequence (tree) of cuts: by first
packing the boxes in the bottom, left, or back part (depending on the cutting
direction) of each cut, a feasible robot packing is obtained. The converse is not
true, that is, not all robot packings are guillotine cuttable, and some sets of
items can be robot packable even though they are not guillotine cuttable, as
shown by the example depicted in Figure 1.
Martello et al. [2000] proposed a branch-and-bound algorithm for the three-
dimensional bin packing problem. den Boef et al. [2005] have shown that this
algorithm correctly solves the robot packing variant of the problem, but does
not generate all orthogonal packings, thus possibly failing the detection of an
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
 Algorithm 864: Variants of Three-Dimensional Bin Packing Problem • 3

Fig. 1. A two-dimensional robot packing that is not guillotine cuttable. The four items cannot be
packed in the given square bin so as to be guillotine cuttable.

optimal general packing. This behavior is due to an inner procedure used to

determine an optimal packing for a single bin.
We present here an extension of the algorithm in Martello et al. [2000],
obtained through a new single-bin packing procedure that combines the orig-
inal enumerative approach with a new constraint programming approach, so
that general as well as robot-packable versions of the problem can be solved.
The corresponding C language implementation, binpack3d is provided. In
Section 2 we give an overview of the main algorithm, while Section 3 describes
the single-bin filling subproblem. Section 4 presents the overall structure of the
algorithm. Finally, Section 5 reports computational experiments to evaluate the
average performance of the code, as an algorithm for solving both general and
robot-packable variants of the problem. Section 6 presents our conclusions and
discusses future research topics.

2. THE ALGORITHM
The algorithm is based on the two-level decomposition approach presented by
Martello and Vigo [1998] for the two-dimensional bin packing problem. A main
branching tree assigns the boxes to the bins, without specifying their actual
position. The actual feasibility of this assignment is then checked by an inner
level. The boxes are initially sorted by nonincreasing volume.
The lower bound used by the algorithm is L2 , whose detailed description is
given in Martello et al. [2000]. Bound L2 is determined through three lower
bound computations, each relative to a different dimension of bins and boxes.
Consider, say, the depth, and let p ≤ W/2 and q ≤ H/2 be two given threshold
values. The corresponding bound LWH 2 ( p, q) is obtained by extracting from the
box set three subsets of boxes: (i) very large (those with w j > W − p and
h j > H − q); (ii) large (among the remaining boxes, those with w j > W/2
and h j > H/2); (iii) small (among the remaining boxes, those with w j ≥ p and
h j ≥ q). The other (smaller) boxes are disregarded. The bound is then computed,
on the basis of the depths of the considered boxes, by determining, through
a simpler bound LWH 1 , the minimum number of bins needed to allocate the

ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
4 • S. Martello et al.

large and very large boxes, as well as the minimum number of additional bins
needed for the small ones. Lower bound LWH 2 is then the maximum LWH 2 ( p, q)
value, computed over all ( p, q) pairs. The two other bounds, LWD 2 (respectively
LHD
2 ) are obtained in the same way, by interchanging h j (respectively w j ) with
d j and H (respectively W ) with D, thus providing the overall bound L2 =
max{LWH 2 , L2 , L2 }.
WD HD

At the root node, we execute in sequence two heuristic algorithms, H1 and H2,
whose detailed descriptions are given in Martello et al. [2000]. Algorithm H1 is
based on a level approach where a number of layers of dimension W × H × d ,
(d ≤ D) are constructed: the resulting layers are then combined into full bins
of depth D by solving a one-dimensional bin packing problem defined on the
depths of the layers. Algorithm H2 is a greedy approach which fills every bin
as far as possible, by heuristically solving a single-bin filling problem in which
the objective is to maximize the volume filled. These algorithms can produce
different solutions by rotating the dimensions of bins and boxes; hence the
computations are performed three times, and the best solution is selected as
the incumbent. If lower bound L2 does not prove that this solution is optimal,
the enumeration begins.
The search is recursively performed, following a depth-first strategy, by an
algorithm called rec binpack. At each decision node, let u denote the incumbent
solution value and M = {1, . . . , m} the set of bins used to allocate boxes in the
ascendant nodes. A bin of M is called open if further boxes may be possibly
placed into it; otherwise it is called closed. The branching phase assigns the
next free box, in turn, to all the open bins; in addition, if |M | < u − 1 the box is
also assigned to a new bin (hence opening it).
As previously mentioned, the exploration of a generated node requires
checking the actual feasibility of the associated assignment. This is performed
by an algorithm called onebin decision. Assume that the node has assigned
a box k to a bin i already containing a nonempty subset J of boxes such that

j ∈J ∪{k} w j h j d j ≤ W H D. The subproblem of establishing whether J ∪ {k}
can be packed into a single bin is in itself NP-hard. We solve it through the
following steps.
(1) Compute lower bound L2 for the subproblem;
(2) if L2 > 1 then fathom the node (no feasible packing exists); otherwise,
(3) if robot packing is desired then
(3.1) execute, for the subproblem, a branch-and-bound search to enumer-
ate all single-bin robot packings through algorithm onebin robot, de-
scribed in detail in Section 3.1; otherwise, that is, if general packing
is desired,
(3.2) execute, for the subproblem, algorithm onebin general, a constraint
programming-based approach described in detail in Section 3.2, to
enumerate all possible packings;
(4) if a single-bin packing is found, accept the assignment; otherwise, fathom
the node.
When the assignment of box k to bin i is accepted, an attempt is made to
close bin i. To this end, for each unassigned box k  , we compute lower bound
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
 Algorithm 864: Variants of Three-Dimensional Bin Packing Problem • 5

L2 for the subinstance defined by the boxes of J ∪ {k} ∪ {k  }. If L2 > 1 for each
k  , then we know that no further box can be placed into i and we close the
bin. If this fails, let K be the subset of those unassigned boxes k  for which
we have obtained L2 = 1: if heuristic H1 or H2 finds a single-bin solution for
the subinstance defined by the boxes in J ∪ {k} ∪ K , then we know that no
better placing is possible for these boxes, so we assign all of them to bin i and
close it.
Whenever a bin is closed, lower bound L2 is computed for the instance defined
by all the boxes not currently assigned to closed bins: if u ≤ L2 + c, where c is
the number of closed bins, we backtrack.

3. THE SINGLE-BIN FILLING SUBPROBLEM

When the exploration of a node is not concluded by Steps 1–2 above, the feasibil-
ity of the assignment of the given subset of boxes to a single bin is determined,
at Step 3, through one out of two complete enumeration searches.
Computational experiments showed that in many cases the algorithm is ex-
ecuted for a very small set of boxes. Hence, we found it fruitful to implement
specialized routines for instances with up to three boxes, based on direct enu-
meration of all distinct placings of the boxes, avoiding symmetric or dominated
solutions like those described by Scheithauer [1997].

3.1 Search for a Feasible Robot Packing

The onebin robot algorithm establishes whether the set of boxes currently as-
signed to a bin can actually be packed into it, by restricting the search to robot
packings. This is obtained by considering the problem of finding a solution that
packs the largest possible volume into the bin. In Martello et al. [2000], it was
shown that only solutions where the boxes are moved backward, downward,
and leftward as far as possible need be considered.
Within such settings, a new box may be placed at any intersection be-
tween three boxes (or bin sides); hence a naive test would demand O(n3 )
time. An O(n2 ) test can be obtained by observing that every subsolution de-
fines an envelope below which no new box can be placed, and where a set of
corner points defines the undominated positions for placing a new box (see
Figure 2). At any decision node, let V p be the volume of the already placed
boxes, and Ve the volume below the envelope. Observe that V p ≤ Ve since
the envelope may identify some encapsulated volume that cannot be used
(like that within the dotted lines in Figure 2). Since no solution that packs
a volume larger than WHD − (Ve − V p ) can be obtained by the descendant
nodes, if this upper bound does not exceed the current best solution, we may
backtrack. The branching operation is easily performed by recursively plac-
ing each unpacked box, in turn, into all the feasible corner points. The in-
cumbent solution is updated whenever a pattern is found which fills a larger
volume.
At each decision node, the envelope and the set of feasible corner points are
obtained, in O(n2 ) time, by determining a sequence of O(n) two-dimensional
envelopes and corner points, as shown in Martello et al. [2000].
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
6 • S. Martello et al.

Fig. 2. Envelope for the robot-packing single-bin filling. Black dots show the corner points for
future box insertions. Dotted lines show an encapsulated volume that cannot be used.

3.2 Search for a Feasible General Packing

The second packing algorithm, onebin general, establishes whether a given set
of boxes can be packed into the bin using general packing patterns. Unlike robot
packing, we cannot fill the bin from one corner. Instead, a solution approach
based on constraint programming has been implemented.
Constraint programming has recently been recognized as a promising tool
for solving decision problems which are difficult to formulate in ILP form (see,
e.g., Brailsford et al. [1999]). A packing is feasible if all boxes are placed within
the bin, and no two boxes overlap. The latter criteria is satisfied if for each pair
of boxes i, j , box i is left of, right of, under, above, behind, or in front of box j .
This means that for each pair of boxes i, j we may introduce the set of relative
positions M ij = { l, r, u, a, b, f }. To avoid symmetries, boxes 1 and 2 have
the restricted domain M 12 = {l, u, b}.
The problem is solved recursively by an algorithm, onebin general, where
at each iteration two boxes i and j are considered, and one of the relative po-
sitions in M ij is imposed. The feasibility of the currently imposed relations is
then checked, and if it can be proved that no solution exists, the algorithm back-
tracks. Otherwise the algorithm calls itself recursively. If all pairs of boxes i, j
have had imposed on them a relation, and a feasible assignment of coordinates
exists such that no box exceeds the bin, the algorithm terminates with the value
true. If the recursive search did not find a feasible packing, onebin general re-
turns the value false.
The feasibility of a partially ordered problem is checked separately for each
dimension. For symmetry reasons, we will only describe the algorithm for the
x-coordinates. Hence, assume that those boxes that are forced to be left or right
of each other are sorted in topological order as follows: if M ij = { l } then i ≺ j

ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
 Algorithm 864: Variants of Three-Dimensional Bin Packing Problem • 7

and if M ij = { r } then j ≺ i. Next, for a given box j , its position x j is found

iteratively as
x j = max xi + wi
i : i≺ j

while x j is set to zero if we maximize over the empty set. The algorithm runs in
O(n2 ) time since the topological ordering takes O(n2 ) time, and the assignment
of coordinates can be done in O(n2 ) time as there are O(n2 ) relations l, r. If
x j + w j > W for some j then box j exceeds the bin, and we backtrack.
To further speed up the solution algorithm, domain reduction and constraint
propagation are applied. Domain reduction is obtained by the Forward Check
(FC) approach proposed by Haralick and Elliot [1980]. Having chosen a relation
between boxes i and j , we may remove all other relations from M ij as we know
that only one relation needs to be chosen.
We use the Maintaining Arc Consistency (MAC) approach proposed by
Gaschnig [1979] to propagate the chosen constraints. The MAC algorithm is
a more sophisticated FC algorithm which at any step keeps the problem arc
consistent. The present problem is arc consistent if for all possible assign-
ments of a relation between two boxes i and j there is a feasible assign-
ment of coordinates to the boxes such that all constraints are satisfied. If this
is not possible then the relation is removed from the domain of M ij . In the
case where a domain becomes empty, the problem cannot be satisfied, and we
backtrack.
Although the use of MAC means that the time complexity of each node in
the search tree grows considerably, the number of nodes investigated decreases
correspondingly. Sabin and Freuder [1994] showed that the additional effort of
MAC pays off when dealing with hard problems.
The MAC strategy is implemented as follows. For each recursive call of al-
gorithm onebin general (i.e., for each new assignment of a relation) we run
through all pairs of boxes i and j with |M ij | ≥ 2 and test the feasibility of each
relation R ∈ M ij . If the packing becomes infeasible with the additional relation
imposed, the relation is removed from the domain. If M ij in this way becomes
empty, we may conclude that no feasible packing can be obtained by following
this branch, and thus we may backtrack in onebin general. If only one relation
is left in a domain M ij then this relation is fixed. Having fixed a relation, in
order to maintain arc consistency it may be necessary to run through an ad-
ditional reduction. However, preliminary computational experiments showed
that this does not pay off. All reductions in the domain achieved by the for-
ward propagation are pushed to a stack, such that the domains can be quickly
restored upon backtracking from onebin general.
The best performance of algorithm onebin general was obtained by sorting
the boxes according to nonincreasing volume, and then branching on pairs of
boxes (1, 2), (1, 3), (2, 3), (1, 4), (2, 4), (3, 4), (1, 5) . . .. In this way the, relative
position of the largest boxes was settled at an early stage of the algorithm,
and infeasibility could quickly be detected. This complies with the First Fail
Strategy proposed by Barták [1998].

ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
8 • S. Martello et al.

Fig. 3. Structure of algorithm binpack3d .

4. PSEUDOCODE VERSION OF BINPACK3D

For a better understanding of algorithm binpack3d, we report its general struc-
ture in Figure 3. Initial lower bounds and heuristic solutions are computed at
the root node (see Section 2). If the problem is not solved to optimality, the
tree enumeration is recursively performed through algorithm rec binpack (see
Sections 2 and 3).
As observed in Section 2, different upper- and lower-bound values may be
obtained by considering the instance according to the three different orienta-
tions. To this purpose, the computations are performed for a single orientation:
the overall lower- and upper-bound values are then obtained making use of a
routine that rotates the dimensions as w ← h ← d ← w and W ← H ←
D ← W.

5. COMPUTATIONAL EXPERIMENTS
For each class, we considered instances of different size, with n ranging be-
tween 10 and 50. The code was implemented in ANSI-C, and the experiments
were run on Pentium 4 running at 3 GHz. A time limit of 3600 s was given
to each instance, and 10 instances were solved for each class and size of a
problem.
The experiments were performed on the nine classes of random instances
described in Martello et al. [2000]. For the first five classes, the bin size was
W = H = D = 100 and five types of boxes were considered by uniformly
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
 Algorithm 864: Variants of Three-Dimensional Bin Packing Problem • 9

Table I. Number of Instances, out of 10, Solved to Proved Optimality

in 3600 s
Class
Packing n 1 2 3 4 5 6 7 8 9
10 10 10 10 10 10 10 10 10 10
15 10 10 10 10 10 10 10 10 10
20 10 10 10 10 10 10 10 10 10
25 10 10 10 10 10 10 10 10 10
General 30 10 10 10 10 10 10 10 9 9
35 10 10 10 10 10 10 8 9 9
40 10 9 8 10 9 10 7 9 3
45 8 6 4 10 6 9 5 9 0
50 7 3 7 10 5 10 6 9 0
Total: 723 85 78 79 90 80 89 76 85 61
10 10 10 10 10 10 10 10 10 10
15 10 10 10 10 10 10 10 10 10
20 10 10 10 10 10 10 10 10 10
25 10 10 10 10 8 10 7 9 10
Robot 30 10 10 10 10 9 10 7 7 10
35 10 10 10 10 5 10 5 8 10
40 10 9 8 10 4 10 2 10 9
45 10 7 4 10 4 10 1 8 2
50 7 4 9 10 3 10 3 5 0
Total: 694 87 80 81 90 63 90 55 77 71

randomly generating the box sizes in different intervals, namely:

— Type 1: w j ∈ [1, 12 W ], h j ∈ [ 23 H, H], d j ∈ [ 23 D, D];
— Type 2: w j ∈ [ 23 W, W ], h j ∈ [1, 12 H], d j ∈ [ 23 D, D];
— Type 3: w j ∈ [ 23 W, W ], h j ∈ [ 23 H, H], d j ∈ [1, 12 D];
— Type 4: w j ∈ [ 12 W, W ], h j ∈ [ 12 H, H], d j ∈ [ 12 D, D];
— Type 5: w j ∈ [1, 12 W ], h j ∈ [1, 12 H], d j ∈ [1, 12 D].
Classes k (k = 1, . . . , 5) were then obtained by generating each box according
to type k with probability 60%, and according to the other four types with
probability 10% each. Classes 6 to 8 may be described as follows:
— Class 6: W = H = D = 10; w j , h j , d j uniformly random in [1, 10];
— Class 7: W = H = D = 40; w j , h j , d j uniformly random in [1, 35];
— Class 8: W = H = D = 100; w j , h j , d j uniformly random in [1, 100].
Finally, Class 9 consisted of difficult all-fill instances having a known solution
with three bins: the boxes were generated by randomly cutting the bins into
smaller parts as follows. Bins 1 and 2 were cut into
n/3 boxes each, while bin 3
was cut into n−2
n/3 boxes. The cutting was made using a recursive approach
which repeatedly cuts the bin into two parts using a guillotine cut until five
boxes remain. These boxes were then cut using a nonguillotine pattern.
The entries in Table I give the number of instances, out of 10, solved to
optimality using a time limit of 3600 s. The bottom line reports the total number
of instances solved to optimality for each class and packing version. The general
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
10 • S. Martello et al.

Table II. Average Solution Times in Seconds, as Average of 10 Instances Using a Time Limit
of 3600 s
Class
Packing n 1 2 3 4 5 6 7 8 9
10 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
15 0.00 0.00 0.00 0.00 0.01 0.00 0.02 0.00 0.00
20 0.01 0.01 0.01 0.00 0.01 0.01 0.16 0.00 0.01
25 0.05 0.04 0.10 0.00 0.09 0.19 1.76 0.91 0.11
General 30 15.18 1.09 3.50 0.00 1.41 7.34 93.73 360.35 468.81
35 61.07 102.39 26.30 0.00 3.01 2.90 764.97 390.11 546.46
40 217.87 614.06 861.15 0.00 448.32 15.79 1151.20 365.25 2859.09
45 1049.05 1493.49 2163.92 0.01 1464.79 363.07 1952.55 450.54 3600.01
50 1391.89 2696.83 1712.14 4.00 1971.17 4.22 1604.58 434.98 3600.02
10 0.00 0.00 0.00 0.00 0.00 0.00 7.81 0.00 0.00
15 0.00 0.00 0.00 0.00 0.04 0.01 4.85 0.02 0.00
20 0.01 0.01 0.01 0.00 3.96 0.04 207.76 4.31 0.00
25 0.11 0.03 0.18 0.00 843.37 0.09 1425.30 361.05 0.05
Robot 30 10.41 1.01 1.90 0.00 381.42 5.00 1308.55 1080.11 19.78
35 36.67 35.75 11.03 0.00 1819.74 1.53 2033.36 950.86 74.75
40 114.53 449.87 766.76 0.00 2184.86 12.84 2881.79 136.94 1250.15
45 748.34 1299.86 2162.28 0.01 2161.50 11.28 3240.01 751.35 3247.02
50 1262.65 2523.43 1333.56 3.52 2532.84 5.49 2742.58 1875.95 3600.01

Table III. Average Number of Branch-and-Bound Nodes (in Thousands)

Class
Packing n 1 2 3 4 5 6 7 8 9
10 0 0 0 0 0 0 0 0 0
15 0 0 0 0 0 0 0 0 0
20 0 0 0 0 0 0 0 0 0
25 0 0 1 0 0 0 0 0 0
General 30 419 6 103 0 0 1 0 0 0
35 4179 1829 1125 0 0 0 0 0 0
40 17163 4146 21409 0 0 4 0 0 0
45 31083 38449 60636 1 0 19 0 0 0
50 45604 84678 60629 674 0 20 0 3 0
10 0 0 0 0 0 0 0 0 0
15 0 0 0 0 0 0 0 0 0
20 0 0 0 0 0 0 0 0 0
25 0 0 1 0 0 0 0 0 0
Robot 30 420 6 103 0 0 1 0 0 0
35 4179 1863 1125 0 0 0 0 0 0
40 17163 9402 40950 0 0 4 0 0 0
45 51837 59719 115812 1 0 19 0 0 0
50 65470 192349 80268 674 0 20 0 3 0

algorithm was able to solve 723 out of 810 instances, while the robot-packable
algorithm was able to solve 694 of them. The difference may be explained by
the fact that the general version has more freedom in arranging the boxes and
thus may find solutions using fewer bins. In this way, having tighter upper
bounds, the algorithm may terminate faster as the lower bounds used by both
algorithms are the same. For Class 9, the picture changed as the robot-packable
algorithm was able to solve 71 instances while the general algorithm only solved
61 instances (out of 90). This behavior can be explained by the fact that the
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
 Algorithm 864: Variants of Three-Dimensional Bin Packing Problem • 11

Table IV. Average Number of Iterations in Procedure onebin decision (in Thousands)
Class
Packing n 1 2 3 4 56 7 8 9
10 0 0 0 0 0 0 0 0 0
15 2 3 9 0 0 0 1 0 1
20 10 22 9 0 0 1 5 0 8
25 39 42 94 0 2 16 36 68 59
General 30 12224 820 3876 0 5 1627 42912 149919 193643
35 27063 102944 31566 1 7 1715 309928 102222 363543
40 206458 598374 720682 8 166802 9161 430270 167427 1267382
45 1098733 1621442 1795338 24 674970 205247 816829 193217 1579066
50 1676059 3062527 2299310 17432 861892 3718 672255 152052 1266272
10 0 0 0 0 0 3 4123 0 0
15 1 2 3 0 16 4 2467 11 0
20 23 14 6 0 2098 27 100877 2087 4
25 127 35 223 0 285958 55 513293 174724 45
Robot 30 13651 1455 2742 0 129160 3024 400023 401562 14725
35 23214 54554 20954 2 548005 1168 895675 511179 28287
40 160077 644205 785467 12 1387156 13591 1436016 94393 493016
45 1137480 2139452 2439088 27 1576640 8260 1781172 509878 1343846
50 1966457 4570412 1990263 17027 989142 3821 975509 878014 1380453

generated instances were all robot packable, so the robot packing version had
the advantage that it only needed to search the relevant solution space. As a
whole, Classes 4 and 6 were generally the easiest to solve while Class 9 tended
to be the most difficult.
Table II gives the average CPU times with a time limit of 3600 s. If an
instance was not solved to proved optimality within the time limit, the time
spent is included in the computation. It can be seen that most instances up to
30 items were solved within a reasonable time. Classes 4 and 6 were particularly
easy to solve, and the solution times seldom exceeded a few seconds.
Table III reports the number of branch-and-bound nodes in the main search
tree, measured in thousands. Table IV shows the corresponding number of it-
erations in procedure onebin decision, measured in thousands. It can be seen
that Classes 1 to 3 demanded quite a lot of branch-and-bound nodes, while
Classes 5 to 9 seldom branched but mainly spent time in the onebin decision
procedure. Class 4 used neither many branch-and-bound nodes nor many iter-
ations in onebin decision.

6. CONCLUSIONS AND FUTURE RESEARCH TOPICS

To our knowledge, the present code is the first implementation of an algorithm
for the three-dimensional bin-packing problem. The code may be used as a whole
for solving this kind of problem, or the exact algorithm for filling a single bin may
be used as a subproblem in other applications. The computational experiments
in the present article, as well as those in Martello et al. [2000], have demon-
strated that the algorithm finds good solutions within a short computing time.
A closer study of the algorithm shows that more than 95% of the time is
spent in routines related to the single-bin filling problem. Hence future research
should be focused on solving this decision problem.
ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.
 12 • S. Martello et al.

REFERENCES

BARTÁK, R. 1998. Online guide to constraint programming. Available online at http://kti.mff.

cuni.cz/∼bartak/constraints/.
BRAILSFORD, S., POTTS, C., AND SMITH, B. 1999. Constraint satisfaction problems: Algorithms and
applications. European J. Operat. Res. 119, 557–581.
DEN BOEF, E., KORST, J., MARTELLO, S., PISINGER, D., AND VIGO, D. 2005. Erratum to the three-
dimensional bin packing problem: Robot-packable and orthogonal variants of packing problems.
Operat. Res. 53, 735–736.
DYCKHOFF, H., SCHEITHAUER, G., AND TERNO, J. 1997. Cutting and Packing. In Annotated Bibliogra-
phies in Combinatorial Optimization, M. Dell’Amico, F. Maffioli, and S. Martello, Eds. John Wiley
& Sons, Chichester, England, 393–413.
GASCHNIG, J. 1979. Performance measurement and analysis of certain search algorithms. Tech.
rep. CMU-CS-79-124. Carnegie-Mellon University, Pittsburgh, PA.
HARALICK, R. AND ELLIOT, G. 1980. Increasing tree search efficiency for constraint satisfaction
problems. Artific. Intell. 14, 263–313.
LODI, A., MARTELLO, S., AND MONACI, M. 2002a. Two-dimensional packing problems: A survey.
European J. Operat. Res. 141, 3–13.
LODI, A., MARTELLO, S., AND VIGO, D. 2002b. Recent advances on two-dimensional bin packing
problems. Discr. Appl. Math. 123, 379–396.
MARTELLO, S., PISINGER, D., AND VIGO, D. 2000. The three-dimensional bin packing problem. Op-
erat. Res. 48, 256–267.
MARTELLO, S. AND VIGO, D. 1998. Exact solution of the two-dimensional finite bin packing problem.
Manage. Sci. 44, 388–399.
SABIN, D. AND FREUDER, E. 1994. Contradicting conventional wisdom in constraint satisfaction.
In Proceedings of the Second International Workshop on Principles and Practice of Constraint
Programming, PPCP’94, Rosario, Orcas Island, WA, USA, May 1994, A. Borning, Ed. Lecture
Notes in Computer Science, vol. 874. Springer-Verlag, Berlin and Heidelberg, Germany, 10–20.
SCHEITHAUER, G. 1997. Equivalence and dominance for problems of optimal packing of rectangles.
Ricerca Operat. 83, 3–34.

Received May 2003; revised January 2006; accepted April 2006

ACM Transactions on Mathematical Software, Vol. 33, No. 1, Article 7, Publication date: March 2007.

View publication stats