A Constraint for Bin Packing

Paul Shaw

ILOG S.A., Les Taissounieres HB2

2681 Route des Dolines, 06560 Valbonne, France
pshaw@ilog.fr

Abstract. We introduce a constraint for one-dimensional bin packing.

This constraint uses propagation rules incorporating knapsack-based rea-
soning, as well as a lower bound on the number of bins needed. We show
that this constraint can significantly reduce search on bin packing prob-
lems. We also demonstrate that when coupled with a standard bin pack-
ing search strategy, our constraint can be a competitive alternative to
established operations research bin packing algorithms.

1 Introduction

The one-dimensional bin packing problem is described as follows: Given n indi-

visible items, each of a known non-negative size si , and m bins, each of capacity
C, can we pack all n items into the m bins such that the sum of the sizes
of the items in any bin is not greater than C? This problem is an important
NP-complete problem having various applications (a good review paper [2] cites
stock cutting and television commercial break scheduling as well as the obvious
application of physical packing).
In recent years, good exact algorithms for the bin packing problem have
been put forward (for example, see [10, 11, 15]), so why examine a constraint-
based approach? The main reason is that most real bin packing problems are
not pure ones, but form a component of a larger system. For example, almost
all resource allocation problems have a bin packing component; this set includes
timetabling, rostering, scheduling, facility location, line balancing, and so on.
When problems become less pure, standard algorithms often become inapplica-
ble, making constraint programming more attractive. Additionally, the use of
a dedicated constraint (in place of a collection of constraints defining the same
solution set) normally allows for a significant increase in constraint propagation,
the all-different constraint [14] being the classic example.
In this paper, we introduce a dedicated constraint P for the one-dimensional
bin packing problem. Section 2 introduces some notation as well as a typical bin
packing model. Section 3 describes new pruning and propagation rules based on
reasoning over single bins. Section 4 then introduces a lower bounding method on
the number of bins used. Section 5 describes experiments on the new constraint.
Finally, section 6 describes the most relevant work and section 7 concludes.

M. Wallace (Ed.): CP 2004, LNCS 3258, pp. 648–662, 2004.

c Springer-Verlag Berlin Heidelberg 2004
 A Constraint for Bin Packing 649

2 Preliminaries
2.1 General Notation
All constrained variables we consider in this paper are non-negative integer.
Associated with each variable x is an initial domain D0 (x) ⊂ {0 . . . ∞}. The
goal is to assign each variable an element from its domain without violating any
constraints: such an assignment is referred to as a solution. We assume that this
assignment procedure proceeds constructively, building the solution one piece at
a time. At each point in this construction, we have a partial assignment which
we define to be the set of current domains of all variables. The current domain
D(x) of variable x is always a (non-strict) subset of its initial domain D0 (x).
We also denote the minimum and maximum of the domain of x in the current
partial assignment as x and x respectively.
When performing a domain reduction on a variable x, assume that D (x) is
the domain of x after the reduction. We use x ← a to denote D (x) = D(x) ∩ a,
x ← a to denote D (x) = D(x) ∩ {a . . . ∞}, x ← a to denote D (x) = D(x) ∩
/ a to denote D (x) = D(x)\a.
{0 . . . a}, and x ←
If at any time, for any variable x, D(x) = ∅, then the constraint system prunes
the search. What happens then depends on the search strategy employed, but
normally the search will backtrack to a previous choice point (if one exists),
changing an earlier decision.

2.2 Bin Packing Problem

Any instance of the packing constraint P takes three parameters which are a
vector of m constrained variables l = l1 . . . lm  representing the load of each
bin, a vector of n constrained variables b = b1 . . . bn  indicating, for each item,
the index of the bin into which it will be placed, and a vector of n non-negative
integers s = s1 . . . sn  representing the size of each item to be packed. Without
loss of generality, we assume that the sizes are
sorted according to si ≥ si+1 .
n
The total size to be packed is denoted by S = i=1 si . The set of item indices
is represented by I = {1 . . . n} and the set of bin indices by B = {1 . . . m}.
Note that, because variables in b and l can have arbitrary domains when
given to the constraint, problems more general that pure bin packing problems
can be specified. These include variable-sized bin packing [4], problems where
bins have a minimum load requirement, and those where not all items can be
packed in all bins.
The semantics of the
bin packing problem dictate that P must ensure for
each bin j ∈ B that lj = {i | i∈I∧bi =j} si and for each item i ∈ I that bi ∈ B.

2.3 Bin Packing Notation

We introduce some notation to specify the states of bins in partial assignments.
We define as the possible set Pj of a bin j as the set of items that are packed
or may potentially be packed in it: Pj = {i | i ∈ I ∧ j ∈ D(bi )}. We define the
650 Paul Shaw

required set Rj of bin j as the set of items packed in the bin: Rj = {i | i ∈

Pj ∧ |D(bi )| = 1}. We refer to Cj = Pj − Rj as the candidate

set of bin j. We
refer to the total size of items packed in bin j as pj = i∈Rj si , and to the set
of unpacked items as U = {i | i ∈ I ∧ |D(bi )| > 1}.

2.4 Typical Bin Packing Model

We present what we consider to be a typical constraint programming model of
the bin packing problem, which will be referred to as a comparison base in the
rest of the paper. We introduce intermediate 0–1 variables xi,j that determine if
item i has been placed in bin j. These variables are maintained by nm constraints
as below. We have, for each item i:

∀j ∈ B xi,j = 1 ⇔ bi = j

remove the b variables from the model and add n

Alternatively, we could
constraints of the form j∈B xi,j = 1. The x variables would then become the
decision variables.
The loads on the bins are maintained by m scalar products. For bin j we
have:

lj = wi xi,j
i∈I
All items must be packed, and for each item i:

bi ∈ B
Finally, it is often useful to add the redundant constraint specifying that the
sum of the bin loads is equal to the sum of the item sizes.

lj = S
j∈B

2.5 Propagations Performed Both by the Typical Model and by P

When the typical model in the previous section is implemented in ILOG Solver [7]
it carries out certain “basic” constraint propagations. These propagations are
also carried out by P, and each one of them is detailed in this section.

Pack All. All items must be packed. This is enforced for each item i ∈ I via:
bi ← 1 bi ← m

Load Maintenance. The minimum and maximum load of each bin is main-

domains of the bin assignment variables b. For brevity,
tained according to the
we denote Sum(X) = i∈X si , where X ⊆ I is any set of items under consider-
ation. For each bin j ∈ B:
lj ← Sum(Rj ) lj ← Sum(Pj )
 A Constraint for Bin Packing 651

Load and Size Coherence. We perform the following propagations which are
derived from the fact that the sum of the items sizes must be equal to the sum
of bin loads. This means that for any bin j ∈ B, its load is equal to the total size
to be packed, minus the loads of all other bins. This translates into the following
propagation rules:
 
lj ← S − lk lj ← S − lk
k∈B\j k∈B\j

This propagation rule is very important as it is only rule which communicates

information between different bins (aside from the implicit rule disallowing item
i from all bins other than j when item i is packed in bin j.) Especially important
is that it can increase the lower bound of the bin load. This property is analogous
to the notion of spare capacity used in [5]. The spare capacity sc in a bin packing
problem is sc = mC−S (the capacity available less the total size to be packed). In
any solution, each bin must be filled to at least a level of C − sc, otherwise more
space would be wasted than spare capacity available. The above propagation
rules dynamically maintain this information, propagating bounds on load and
implicitly reducing spare capacity when a bin is packed to less than its full
capacity.
The commitment rule below makes use of the lower bound on bin load, as
do the additional pruning and propagation rules introduced in section 3.

Single Item Elimination and Commitment. An item is eliminated as a

candidate for packing in a bin if it cannot be added to the bin without the
maximum load being exceeded. For bin j:

if ∃i i ∈ Cj ∧ pj + si > lj then bi ←
/ j
An item is committed to a bin if packing all candidates into the bin except
that item would not increase the packed quantity to the required minimum load
of the bin. For bin j:

if ∃i i ∈ Cj ∧ Sum(Pj ) − si < lj then bi ← j

3 Additional Propagation Rules

The idea of constraint propagation is to eliminate domain values when it can be
ascertained that these values can never appear in a solution to the constraint. (A
solution to a constraint is an assignment of all variables involved in the constraint
which does not violate the constraint.) Constraint propagation algorithms often
try to achieve generalized arc consistency (GAC) which means that all values
which cannot be involved in a solution to the constraint are removed. Unfortu-
nately, for the packing constraint P, achieving GAC is NP-complete. So, instead
of trying to achieve GAC, we set ourselves the more humble goal of increasing
propagation strength over that of the typical model.
652 Paul Shaw

The additional constraint propagation rules that we introduce are all based
upon treating the simpler problem of packing a single bin. That is, given a bin
j, can we find a subset of Cj that when packed in the bin would bring the load
the range [lj , lj ]? This problem is a type of knapsack or subset sum problem [9].
Proving that there is no solution to this problem for any bin j would mean that
search could be pruned. However, as shown by Trick [17], we can go further and
even achieve GAC for this reduced single-bin problem.
For bin j we define Mj = {m | m ⊆ Cj ∧ lj ≤ pj + Sum(m) ≤ lj }. Mj
is the set of sets of additional items which can be packed while respecting the
constraints on load. If Mj is empty, there is no legal packing of bin j, and we
can prune search. For Mj = ∅, we make deductions on the candidate items:

if ∀m ∈ Mj i ∈ m then bi ← j if ∀m ∈ Mj i ∈ m then bi ←
/ j
That is, if an item appears in every set of items that can be placed in the bin,
we can commit it to the bin. Conversely, if the item never appears in such a set,
we can eliminate it as a candidate item. Trick did not treat the case where the
load was a variable, but it is nevertheless possible to deduce illegal bin loads:

if ∃v ∈ D(lj ) ∀m ∈ Mj v ∈ m then lj ←
/ v

So, if no legal packing can attain load v, v cannot be a legal load for the bin.
In [17], a pseudo-polynomial dynamic programming algorithm is used to
2
achieve consistency. The time complexity of the algorithm is O(|Cj |lj ), mean-
ing that it can become inefficient when items are large or many items can be
packed in a bin. Sellmann [16] proposes an approach where efficiency can be
traded for propagation strength by dividing down values; in our case, item sizes
and bin capacities. By selecting an appropriate divisor, the resulting algorithm
can produce a trade-off between propagation strength and time invested. Here,
we take a different approach, using efficient algorithms which do not depend on
item size or bin capacity, but which like [16], do not in general achieve GAC on
the subset sum subproblem.

3.1 Detecting Non-packable Bins

Here, we describe a method which can detect non-packable bins in time O(|Cj |).
The advantage of the method is its efficiency; the drawback is that it is not
complete, and some non-packable bins may not be identified as such.
We try to find a proof that ∀m ⊆ Cj pj + Sum(m) < lj ∨ pj + Sum(m) > lj .
That is, there is no subset of candidate items whose sizes – together with the
already packed items – sum to a legal load. We search for this proof in a particular
way based on identifying what we refer to as neighboring subsets1 of Cj . Assume
two sets Cj1 , Cj2 ⊆ Cj . These sets are neighboring if there is no other subset of
Cj whose items sum to a value strictly between Sum(Cj1 ) and Sum(Cj2 ). (This
1
The author has found no pre-existing definition of the proposed notion in the liter-
ature.
 A Constraint for Bin Packing 653

implies that Cj1 and Cj2 are neighboring if |Sum(Cj2 ) − Sum(Cj1 )| ≤ 1.) We
describe a neighboring predicate N (j, Cj1 , Cj2 ) as follows:

N (Cj1 , Cj2 ) ⇔ N (Cj2 , Cj1 )

Sum(Cj1 ) ≤ Sum(Cj2 ) ⇒ (N (Cj1 , Cj2 ) ∨ Between(j, Cj1 , Cj2 ))

Between(j, Cj1 , Cj2 ) ⇔ ∃m ⊆ Cj Sum(Cj1 ) < Sum(m) < Sum(Cj2 )

Generation of Neighboring Subsets. Neighboring subsets are generated

which conform to a particular structure. We use the terms low-set and high-
set for two subsets which we consider as candidates for being neighbors. The
understanding is that when the low-set has a sum not greater than the high-set,
then the two subsets are neighbors. This will become clearer in what follows.
Figure 1 shows a set of candidate items X sorted by non-increasing size, and
three subsets marked A, B and C. Subset A comprises the k largest candidate
items. Subset C comprises the k  smallest candidate items. Subset B comprises
the k + 1 smallest items outside subset C. Subsets A and B may contain the
same item, but subset C is disjoint from the other two. We choose the low-set
Lk to be the union of subsets A and C and the high-set Hk to be the subset B.

B C

Fig. 1. Structure of neighboring subsets

We now show that for this particular structure, if Sum(Lk ) ≤ Sum(Hk ), then
Lk and Hk are neighboring. When Sum(Lk ) = Sum(Hk ), the proof is trivial.
We therefore concentrate on the case where Sum(Lk ) < Sum(Hk ). Lk is formed
from the k largest items and the k  smallest items. We can see that for any
m ⊆ X for which Sum(m) > Sum(Lk ), |m| ≥ k + 1. This must be the case as
Lk already contains the k largest items. Moreover m cannot contain any of the
k  smallest items as then Sum(Lk ) ≥ Sum(Hk ) in contradiction to our initial
assumption. To see this, imagine i ∈ m and that i is one of the smallest k  items
in X. In this case m is composed of k items not in the smallest k  and item i,
while Lk is composed of the largest k items, plus item i (and possibly some other
items). Since the largest k items must have a total size not less than k items
chosen more freely, it follows that if m contains one of the k  smallest items then
654 Paul Shaw

boolean function NoSum(X, α, β)

if α ≤ 0 ∨ β ≥ Sum(X) then
return false
ΣA , ΣB , ΣC := 0 {See figure 1 for meaning of A, B, C}
k, k
:= 0 {k largest items, k
smallest items}
while ΣC + s|X|−k
< α do
ΣC := ΣC + s|X|−k

k
:= k
+ 1
end while
ΣB := s|X|−k

while ΣA < α ∧ ΣB ≤ β do
k := k + 1
ΣA := ΣA + sk
if ΣA < α then
k
:= k
− 1
ΣB := ΣB + s|X|−k

ΣC := ΣC − s|X|−k

while ΣA + ΣC ≥ α do
k
:= k
− 1
ΣC := ΣC − s|X|−k

ΣB := ΣB + s|X|−k
− s|X|−k
−k−1
end while
end if
end while
return ΣA < α
end function

Fig. 2. Procedure for detecting non-existence of sums in [α, β]

Sum(m) ≤ Sum(Lk ). Having now discounted the smallest k  items from m, the
smallest Sum(m) is then obtained when m is made up of the k + 1 smallest items
outside of the smallest k  , i.e. when m = Hk .
As an example of this reasoning, we consider a bin j with pj = 0, lj = 34,
lj = 35 and candidates items of sizes 10, 10, 10, 9, 9, 9, 9, 2, 1. The bin can
be shown non-packable by considering the neighboring subsets produced when
k = 3 and k  = 2. In this case, the low-set sums to 10 + 10 + 10 + 2 + 1 = 33 and
the high-set to 9 + 9 + 9 + 9 = 36.

Implementation. Here we describe a procedure NoSum which determines if

a subset of a set of items X cannot sum to a value in the range [α, β]. Being an
incomplete method, NoSum can generate false negatives but this is not prob-
lematic as only the positive result is actively used in pruning. We remind the
reader that item sizes obey the relation si ≥ si+1 . NoSum is detailed in figure 2.
The method is reasonably simple despite its appearance. The variables ΣA ,
ΣB and ΣC hold the total sizes of items in the subsets A, B, and C shown in
figure 1. The variables k and k  keep their meaning from the previous section.
 A Constraint for Bin Packing 655

Essentially, k is increased from 0 until the k largest items are not less than α,
the lower bound on the required sum. For each k, k  is chosen such that ΣA + ΣC
is maximized while being less than α. This is done initially by adding up the
k  smallest item sizes. Thereafter, each time k is increased, k  is reduced again
until ΣA + ΣC < α. During this reduction phase, ΣB is maintained by ‘sliding’
the window of the B subset to the right so that it remains adjacent to subset C
(again, see figure 1). If we find that ΣB > β, then we have found a proof that
no subset of the items can sum to a value in [α, β].
The time complexity of NoSum is linear in |X|, but is usually much better
as it is bounded by the value of k  after the initial loop. That is, the maximum
number of candidate items that can be added together while maintaining their
sum strictly less than α. We refer to this quantity as Njα . We assume that
Sum(X) can be computed in constant time by incrementally maintaining the
total item size of any candidate set when it is reduced.

3.2 Pruning Rule

Given the function NoSum, it is trivial to derive a pruning rule. The only ma-
nipulation needed is to remove the effect of items already packed in the bin from
the load target. For any bin j:

if NoSum(Cj , lj − pj , lj − pj ) then prune

3.3 Tightening Bounds on Bin Load

The reasoning used for detecting non-packable bins can also be used to tighten
bounds on the bin load. For this, NoSum needs slight modification to deliver on
return the values of α = ΣA + ΣC and β  = ΣB . These are the total sizes of the
two neighboring subsets when NoSum answers in the affirmative. We can then
specify the tightening rules as:

if NoSum(Cj , lj − pj , lj − pj ) then lj ← pj + β 

if NoSum(Cj , lj − pj , lj − pj ) then lj ← pj + α

3.4 Elimination and Commitment of Items

As proposed in [5], we construct from the pruning rule propagation rules to
eliminate and commit items. We make a proposition, then ask the pruning rule
if this proposition leads to a contradiction; if so, we can assert the negation of
the proposition. We commit item i to bin j and then ask if bin j is packable. If it
is not, then i can be eliminated from the candidates of bin j. A similar argument
holds for committing items; we assume that i is eliminated, then ask if the bin
is packable. If not, we can commit the item to the bin. Given a candidate item
i ∈ Cj :
if NoSum(Cj \i, lj − pj − si , lj − pj − si ) then bi ←
/ j

if NoSum(Cj \i, lj − pj , lj − pj ) then bi ← j

656 Paul Shaw

In both of these rules, item i is eliminated as a candidate, but in the first, the
bounds passed to NoSum are reduced to reflect the increased size of the packed
items due to the inclusion of i in Rj . In the second rule, the proposition is to
eliminate an item and so the packed size does not change.
Examining a bin j for all item eliminations and commitments takes time
lj
O(|Cj |Nj ), which can be up to O(|Cj |2 ) but is typically much less as normally
lj
Nj  |Cj |. When the pruning rule is restricted to consider only neighboring
subsets for which either k = 0 or k  = 0, we know of algorithms which run in time
O(|Cj |). These algorithms combine the logic of the pruning and propagation in
one procedure and are thus more complex than the use of the pruning rule as
a subordinate procedure. In this paper, these linear algorithms are not used do
to their increased complexity, reduced propagation strength (as the pruning rule
used is less general), and marginal reduction in calculation time per search node.

4 Using a Lower Bound

The usual simple lower bound for the fixed capacity bin packing problem is ar-
rived at by relaxing the constraint that each item is indivisible. Given a fixed ca-
pacity C, the bound, which we shall term L1 (following Martello and Toth [13]) is:
 
1 
L1 (C, s) = si
C
i∈I

Martello and Toth showed L1 to have an asymptotic performance ratio of 12 ,

which means that the the bound L1 can be arbitrarily close to one half of the
number of bins used in an optimal solution.
However there are better bounds which also are efficient to compute. In [13],
Martello and Toth introduced bound L2 which dominates L1 and has an asymp-
totic performance ratio of 23 . In their experiments, Martello and Toth show this
bound to be typically much tighter than L1 . Korf [10] also found this. In his tests
on problems with random sizes from 0 up to the bin capacity, 90 item problems
require an average of 47.68 bins. For these problems, the L1 bound averaged
45.497, whereas the L2 bound averaged 47.428. The L2 bound can be calculated
in linear time if the item sizes are sorted.
The bound works by splitting the items into four subsets using a parameter
K. N1 contains all items that are strictly larger than C − K. N2 contains all
items not in N1 , but which are strictly larger than half the bin capacity. N3
contains all items not in N1 or N2 of size at least K. The items of size strictly
less than K are not considered. Now, no items from N2 or N3 can be placed
with items from N1 , so |N1 | forms a term of the lower bound. Moreover, each
item in |N2 | needs a bin of its own and so |N2 | forms another term. Finally, the
free space in the bins of N2 is subtracted from the total size of items in N3 (as
items in N3 can be placed with those in N2 ). If the result is positive, this is an
overspill which will consume more bins. The result of this extra bin calculation
is added to |N1 | + |N2 | to arrive at the bound for a given K. The maximum of
this bound calculation over all 0 ≤ K ≤ C/2 is the lower bound L2 .
 A Constraint for Bin Packing 657

4.1 Adapting the Lower Bound to Partial Solutions

One simple way to apply the lower bound for a fixed capacity bin packing prob-
lem is to do so before search begins. If L2 returns a value greater than the
number of bins available, then we can declare the problem infeasible. However,
we would like to benefit from the lower bound test when bins are of different
sizes, meaning that it could be applied during search as well as treating problems
that have bins of variable sizes from the outset.
The idea is that after each extension of the partial assignment of items to
bins, we transform the current partial assignment into one which can be handled
by the bounding function L2 . We then prune the search if the lower bound
calculated exceeds the number of bins available.
The transformation carried out is relatively straightforward. First, we find the
bin with the maximum potential load; this will then become the fixed capacity
in the transformed problem: C = maxj∈B lj . Then, in order to account for items
already packed as well as bin capacities less than C, we create a new vector z of
item sizes that contains the sizes of all item in U , the unpacked items, plus one
additional item aj for each bin j. We define:

aj = p j + C − l j
This requires some explanation as it involves two manipulations. First, we
are introducing items into the packing problem representing the already packed
items. We could just add the items themselves, but that would miss the oppor-
tunity to group items that have already been packed together. When we group
these items, we have the potential for a better bound. Second, we are introducing
items to ‘top up’ each bin with capacity less than C. That is, we have overes-
timated the capacity of bin j. So, to emulate the fact that a bin has a lesser
capacity, we introduce an item into the transformed packing problem which says
that an additional C − lj must be packed. Finally, we take the further opportu-
nity to fuse these two items into one as we know that all the packed items and
the ‘top up’ item must be packed in the same bin.
We denote by a the vector a1 . . . am  and assume a function SortDec(a)
which sorts it non-increasing order. Likewise, we assume a function Merge(x, y)
which merges two such sorted vectors x and y, maintaining the ordering prop-
erty. We now define the vector z of items sizes to be passed to L2 as z =
Merge(si | i ∈ U , SortDec(a)). Search is pruned whenever L2 (C, z) > m.
The time complexity of the bounding procedure is O(n + m log m).

5 Experiments

We conduct experiments to compare P to the typical model presented in sec-

tion 2.4. Throughout these tests we use a machine with a Pentium IV processor
running at 2GHz and 1GB of memory. Tests are performed with ILOG Solver.
We compare the typical model, which we call the basic level of propagation
described in section 2.5, with the enhanced model incorporating the additional
658 Paul Shaw

rules described in sections 3 and 4. We decompose these extra rules into use of
the lower bound described in section 4 (which we term “+LB”), and use of the
pruning and propagation rules described in section 3 (which we term “+P”).
Use of both of these sets of rules together is indicated by “+LB +P”.
We use a standard search procedure, complete decreasing best fit (Cdbf) [5],
which packs items in order of non-increasing size, packing each item in the first
bin with least free space that will accommodate it. On backtracking, the search
states that the chosen item cannot be placed in the selected bin. A symmetry
breaking rule is also used which states that on backtracking, all “equivalent”
bins are also eliminated as candidates for the current item and indeed all other
unpacked items of the same size. An equivalent bin is one which carries the
same load as the one being eliminated from consideration. In addition, no choice
point is created if all available bins for a particular item are equivalent: the item
is packed into the first of them. Finally, we added the additional dominance
rule which states that if an item can fill a partially filled bin to capacity, then
it is immediately packed in this bin. This rule subsumes Gent’s pair packing
preprocessing rule [6], and is the simplest special case of a rule described in [11].
In Cdbf, to find the minimal number of bins, we solve a succession of decision
problems starting with the maximum number of bins set to the simple lower
bound L1 . Each time the packing problem is proven to be insoluble, the number
of bins is increased until a solution is found, which is necessarily optimal.
In the first instance, we chose a set of known benchmarks which could be
comfortably solved by all models, from basic to +LB +P, so that comparisons
could be made without extensive CPU times being expended. In this regard, we
examined the smallest instances of series 1 of [15]. These problems have 50 items
with sizes chosen from {[1, 100], [20, 100], [30, 100]} and bin capacities chosen
from {100, 120, 150}. The suite comprises 20 instances for each combination of
capacity and size distribution, making 180 instances in total.
Figure 3 plots the number of choice points taken for the 180 benchmarks,
where the benchmarks have been sorted according to the number of choice points
needed to solve the problem using all additional propagations (+LB +P). In
accordance with observations in [5, 10], there is great variation in the effort
needed to solve the problems. For some problems, the additional propagation
reduces the number of choice points by over three orders of magnitude. One
problem took over three million choice points using the basic level, but was
solved without branching using +LB +P.
Figure 4 plots two different views of the problems where additional propa-
gations (+P) and lower bound pruning (+LB) are activated independently. We
can see that the use of the additional propagations (+P) is more important than
the use of the lower bound (+LB). In fact, the lower bound often performs no
additional pruning over +P. However, on some problems, the search is signifi-
cantly cut by the used of the lower bound. We argue that the small additional
lower bound cost is paid back by increased robustness.
Table 1 shows all instances that were not solved in under 100 choice points.
As mentioned, we can see that the lower bound calculation often performs no or
 A Constraint for Bin Packing 659

1e+08
Basic
+LB +P
1e+07

1e+06

100000
Choice points

10000

1000

100

10

1
0 20 40 60 80 100 120 140 160 180
Sorted problem number

Fig. 3. Sorted 50 item instances

1e+08 1e+08
+LB +P
1e+07 +LB +P 1e+07 +LB +P
1e+06 1e+06
Choice points

Choice points

100000 100000
10000 10000
1000 1000
100 100
10 10
1 1
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Sorted problem number Sorted problem number

Fig. 4. Sorted 50 item instances, broken down into different propagation types

Table 1. Hardest instances, sorted by choice points needed by +LB +P

Choice Points Run Time (s)
Name Basic +LB +P +LB +P Basic +LB +P +LB +P
N1C2W1 G 4485 138 4133 136 0.38 0.02 0.48 0.02
N1C3W2 G 802 742 187 187 0.05 0.05 0.03 0.03
N1C3W2 J 16101 16035 225 225 0.88 0.76 0.04 0.04
N1C2W1 C 3412 435 2639 435 0.22 0.04 0.24 0.05
N1C3W1 R 3756 2902 609 586 0.25 0.18 0.10 0.09
N1C1W2 A 966684 966684 4154 4154 42.29 38.00 0.40 0.44
N1C3W2 H 382812 302637 4562 4562 15.74 10.89 0.62 0.66
N1C3W2 F 13971619 11671895 7491 7491 651.95 483.19 1.45 1.53
N1C3W4 I 2354291 2342765 9281 9281 105.67 92.60 0.74 0.80
660 Paul Shaw

little pruning, but it did reduce the number of choice points by over a factor of
thirty on one of these problems. The additional propagations (+P) can result in
massive speed increases however, reducing run times by orders of magnitude.
We now look at another benchmark set in order to compare P with other
other bin packing algorithms. We examine the benchmarks of Falkenauer [3].
These benchmarks are well-known, but have been criticized by Gent [6] as being
too easy. The problems can certainly be solved by simple methods. However,
Gent’s methods are geared towards consideration of bins, and finding items which
fill them to completely to capacity; they do not resemble Cdbf, which might
be applied as a first try at solving a problem with a bin packing component,
before exploring less well-known methods. Falkenauer gives the performance of
Martello and Toth’s branch and bound algorithm [12] on these problems which
was, until more recently, the best exact algorithm for bin packing. This method
is also based on the decreasing best fit strategy, but makes heavy use of quite
complex reduction procedures, lower bounds and dominance criteria.
Table 2 shows the results of runs on the smallest sizes of the “uniform”
and “triplets” benchmarks using all propagations (+LB +P). Comparison of
run times should be done with care, as the results of Martello and Toth’s al-
gorithm (MTP) are reproduced from [3] and are likely to be around one order
of magnitude greater than those of P. However, the number of choice points is
a fairly reliable measure. Where MTP is marked with a > sign, it means that
the optimum solution was not found. All problems are solved fairly easily by
the combination of P and Cdbf except for u120 08 and u120 19. The former is
solved in under 7 minutes, whereas the latter takes 15 hours. Martello and Toth’s
procedure falls foul of problems u120 08 and u120 19 as does ours, but performs
much worse on the triplets set, finding optima for only 6 of the problems. By
contrast, P plus Cdbf solves all of these problems quickly.
Korf [11] also tested his algorithm on these problem sets, finding that his
algorithm could solve all problems quickly. However, one interesting phenomenon
is that for the two problems that his algorithm found the most difficult, our
method found solutions instantly. Conversely, for the two most difficult problems
we report, his algorithm found solutions instantly. The fact that MTP and Cdbf
find uniform problems 08 and 19 difficult, while Korf’s algorithm does not, seems
to indicate that a change of search strategy could be useful. This is supported
by the fact that Gent’s methods [6] were quickly successful on these problems.
Korf also tested on the triplets set with 120 items and reported finding solutions
instantly to all but four of the twenty problems. We do not report our results
on these problems for reasons of space, but interestingly, our algorithm found
solutions to all of these four troublesome problems instantly.
What is perhaps surprising about these experiments is that although con-
straint programming is a general technology which is rarely the best method
to apply to a pure problem, by a combination of a dedicated constraint and a
standard search procedure, we manage to produce a competitive algorithm.
 A Constraint for Bin Packing 661

Table 2. Falkenauer’s problems solved with +LB +P, comparison with MTP

Choice Points Run Time (s) Choice Points Run Time (s)
Name MTP P MTP P Name MTP P MTP P
u120 00 56 39 0.1 0.02 t60 00 36254 62 9.5 0.03
u120 01 0 36 0.1 0.01 t60 01 28451 173 12.6 0.05
u120 02 124935 38 29.0 0.02 t60 02 > 1.5M 116 > 564.2 0.02
u120 03 74 31 0.0 0.02 t60 03 > 1.5M 195 > 444.7 0.04
u120 04 0 38 0.0 0.02 t60 04 > 1.5M 12 > 404.6 0.01
u120 05 43 32 0.1 0.02 t60 05 > 1.5M 176 > 415.2 0.04
u120 06 69 32 0.0 0.04 t60 06 > 1.5M 77 > 485.7 0.02
u120 07 54 38 0.0 0.03 t60 07 > 1.5M 193 > 395.9 0.04
u120 08 > 10M 2.63M > 3681.4 398.34 t60 08 > 1.5M 359 > 451.6 0.06
u120 09 103 35 0.1 0.03 t60 09 26983 201 9.6 0.06
u120 10 0 34 0.1 0.01 t60 10 1783 16 0.9 0.01
u120 11 64 32 0.1 0.02 t60 11 13325 36 6.3 0.01
u120 12 88 25 0.0 0.03 t60 12 6450 24 1.5 0.01
u120 13 0 34 0.0 0.02 t60 13 > 1.5M 30 > 385.0 0.01
u120 14 0 33 0.0 0.02 t60 14 > 1.5M 14 > 400.8 0.01
u120 15 36 36 0.1 0.02 t60 15 > 1.5M 90 > 537.4 0.02
u120 16 0 33 0.0 0.02 t60 16 > 1.5M 30 > 528.3 0.01
u120 17 48 30 0.0 0.02 t60 17 > 1.5M 50 > 429.9 0.01
u120 18 24 35 0.0 0.01 t60 18 > 1.5M 146 > 385.6 0.03
u120 19 > 7.5M 321.89M > 3679.4 15H08 t60 19 > 1.5M 140 > 399.5 0.03

6 Related Work
Johnson’s initial work [8] brought bin packing algorithms to the fore, but ex-
act algorithms were lacking until the Martello and Toth’s reduction and branch
and bound procedures [12, 13]. However, in the last five years there has been
significant interest and progress from the OR community (for example [1, 15])
using branch and bound and mathematical programming approaches. Korf’s bin
completion algorithm [10, 11], which makes heavy use of dominance properties,
and Gent’s work [6] have shown that the AI field has much to offer the domain.
Trick [17] has shown that GAC can be achieved for knapsack constraints – a key
subproblem of bin packing – using dynamic programming. Sellmann [16] has fur-
ther proposed approximating this consistency to accelerate solving procedures.

7 Conclusion
This paper has introduced a constraint for bin packing. The constraint uses
pruning and propagation rules based on a notion of neighboring subsets in sub-
set sum problems as well as a lower bound on the number of bins used. We
have demonstrated that this new constraint can cut search by orders of magni-
tude. Additional comparisons on established benchmarks showed that the new
constraint coupled with a simple standard packing algorithm can significantly
outperform Martello and Toth’s procedure.
662 Paul Shaw

The packing constraint P has been available in ILOG Solver since version
6.0. All propagations described here will be available in Solver 6.1 to be released
in autumn 2004.

References
1. J. Valerio de Carvalho. Exact solution of bin-packing problems using column gen-
eration and branch-and-bound. Annals of Operations Research, 86:629–659, 1999.
2. E. G. Coffman, Jr, M. R. Garey, and D. S. Johnson. Approximation algorithms
for bin packing: A survery. In D. Hochbaum, editor, Appoximation algorithms for
NP-Hard Problems, pages 46–93. PWS Publishing, Boston, 1996.
3. E. Falkenauer. A hybrid grouping genetic algorithm for bin packing. Journal of
Heuristics, 2:5–30, 1996.
4. D. K. Friesen and M. A. Langston. Variable sized bin packing. SIAM Journal on
Computing, 15:222–230, 1986.
5. I. Gent and T. Walsh. From approximate to optimal solutions: Constructing prun-
ing and propagation rules. In Proceedings of the 15th IJCAI, 1997.
6. I.P. Gent. Heuristic solution of open bin packing problems. Journal of Heuristics,
3:299–304, 1998.
7. ILOG S.A., Gentilly, France. ILOG Solver 6.0. User Manual, September 2003.
8. D. S. Johnson. Fast algorithms for bin packing. Journal of Computer and System
Sciences, 8:272–314, 1974.
9. H. Kellerer, U. Pferschy, and D. Pisinger. Knapsack Problems. Springer, 2004.
10. R. Korf. A new algorithm for optimal bin packing. In Proceedings of 18th AAAI,
pages 731–736, 2002.
11. R. Korf. An improved algorithm for optimal bin packing. In Proceedings of the 18th
IJCAI, pages 1252–1258, 2003.
12. S. Martello and P. Toth. Knapsack problems. Wiley, Chichester, 1990.
13. S. Martello and P. Toth. Lower bounds and reduction procedures for the bin pack-
ing problem. Discrete and Applied Mathematics, 28(1):59–70, 1990.
14. J.-C. Régin. A filtering algorithm for constraints of difference in CSPs. In Pro-
ceedings of the 12th AAAI, pages 362–367. American Association for Artificial
Intelligence, AAAI Press / The MIT Press, 1994.
15. A. Scholl, R. Klein, and C. Jürgens. BISON: A fast hybrid procedure for exactly
solving the one-dimensional bin packing problem. Computers & Operations Re-
search, 24:627–645, 1997.
16. M. Sellmann. Approximated consistency for knapsack constraints. In Proceedings
of the CP’ 03, pages 679–693. Springer, 2003.
17. M. Trick. A dynamic programming approach for consistency and propagation for
knapsack constraints. In Proceedings of CP-AI-OR ’01, 2001.