Activities Planning and Resource Assignment on Multi-place
Hospital System
Exact and Approach Methods Adapted from the Bin Packing Problem
Michel Gourgand, Nathalie Grangeon and Nathalie Klement
LIMOS CNRS UMR 6158, Universit
´
e Blaise Pascal, Complexe Scientifique des C
´
ezeaux, 63173 Aubi
`
ere Cedex, France
Keywords:
Bin Packing, Resources Assignment, Exact Method, Metaheuristics.
Abstract:
In France, the Hospital Community of Territory (HCT) has been defined since the settlement of the pricing
by activity (T2A) in 2004, and the new governance. This new community allows the pooling of the hospital
resources between any places in the same territory. It aims at increasing the continuity of healthcare. A medi-
cal exam needs human resources and one material resource, both of them compatible with the exam. Human
resources are supposed available all the time. In this paper, material resources are the critical resources. The
objective is to create a decision support tool which plans the exams on each material resource. Some con-
straints have to be respected: compatibilities between exams and resources, resources opentime. We propose a
mathematical model. It deals with an exams planning taking into account the assignment of material resources
in the community over a given horizon planning. The problem is solved in an exact way and with approach
methods applying bin packing heuristics and single solution based metaheuristics. Experimentation tackles
with randomly generated instances.
1 INTRODUCTION
Nowadays, the main objective of hospital system is
to improve its efficiency. Indeed, some hospitals can
have patients waiting a few months or more to get an
appointment. On the other hand, other hospitals have
almost no waiting time. These two kinds of places
can be very close geographically speaking. In France,
a new organization was created a few years ago, the
Hospital Community of Territory (HCT). It is a group
of distinct places which aim at improving their ef-
ficiency by putting means from different places to-
gether. The idea developed in this paper is that by dis-
patching medical exams on several places, the global
efficiency of all the places will be improved.
After the specification of our problem, the formu-
lation of the mathematical model is detailed. This
problem can be compared to the bin packing prob-
lem. This comparison is explained in the state of
the art. This problem is first solved in an exact way,
but because it is NP-hard, approach methods are also
considered by applying some single solution based
metaheuristics (kangaroo algorithm and simulated an-
nealing algorithm) combined to a bin packing heuris-
tic. The last section compares the results of the exact
method to those of the approach methods. We will
conclude with some further works.
2 SPECIFICATION
First, the considered problem is detailed in a generic
way and then the analogy with the field of medical
imaging is done.
2.1 Problem Statement
The problem studied in this paper can be compared
to a tasks assignment problem over a multi-place sys-
tem.
Each task needs a resource to be treated and a time
slot, a period when it will be done. We make the
hypothesis that human resources are available all the
time, only the material resources are discussed, they
are the critical resources. Each task has a known pro-
cess time. A task has to be completed before a due
date. Each task has got a reference place where it has
to be done, if possible.
The system is composed of several places. There
is a known distance between these places. In each
place, there are one to several resources. Every re-
117
Gourgand M., Grangeon N. and Klement N..
Activities Planning and Resource Assignment on Multi-place Hospital System - Exact and Approach Methods Adapted from the Bin Packing Problem.
DOI: 10.5220/0004803901170124
In Proceedings of the International Conference on Health Informatics (HEALTHINF-2014), pages 117-124
ISBN: 978-989-758-010-9
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
sources cannot treat all the tasks, there is a list of com-
patibilities between tasks and resources. Resources
are from one type. Resources have a planning defin-
ing their available time, we call it resources opentime.
The considered horizon planning is known. This
horizon is divided into periods. Each task starts in one
period and finishes in the same period.
The aim is to assign each task to one resource and
one period, respecting the compatibilities and the dif-
ferent times. Tasks have to be done as soon as pos-
sible, so the assigned periods have to be the smallest
possible.
The model has to respect some constraints which
are the following:
Resources have to be able to process their as-
signed tasks (resource compatibility).
Tasks have to be assigned to a resource during the
opentime of this resource.
A task has to be assigned to one resource and one
period.
2.2 Objective
The aim of this model is not to make a precise sched-
ule of the tasks but to attribute a period to each task.
The criteria that can be studied are the following:
The completion time of the last planned task. It
is the biggest period assigned to the tasks. All the
tasks are done before or during this period. By
misnomer, we call it in the following makespan.
The sum of the completion times of all the tasks
planned in the system, that is to say the sum of
each period assigned to all the tasks.
The number of tasks done before their due date.
The number of tasks done at their reference place,
or the distance between the reference place and
the effective one for each task.
The criteria which are the most relevant are about
the economic aspect of the problem. The earlier the
tasks will be planned, the earlier they will be com-
pleted. So resources will be available earlier to prac-
tice the next tasks. With the two first criteria, we can
assume that the tasks will be done as soon as possible.
2.3 Medical Imaging Example
For the hospital system, a task is an exam. The
considered system is the HCT, composed of several
places. In each place, there are one to several material
resources. A material resource is from one type: MRI,
x-ray, or scanner. Each material resource has a plan-
ning with its available time per period. For example,
we can assume that an MRI engine is only available
ninety minutes on Monday morning because it needs
a maintenance operation or because an external physi-
cian reserved it. Its opentime will span over ninety
minutes during this period. The considered horizon
planning is divided into periods. Each period can rep-
resent one half-day, one day or one week, the horizon
planning can represent one week, one month or one
year.
Depending on the used horizon planing, the de-
scribed model can be used at three levels:
Strategic: to determine the system dimensional,
the number of places needed in the system, how
many hospitals are part of the HCT.
Tactical: to determine the number of resources
needed in each place, how many material re-
sources of each type (MRI, x-ray, scanner, etc.)
are needed in each hospital.
Operational: to plan exams and assign them one
resource and one period.
In this paper, we are dealing with instances in
which the horizon planning is not greater than one
week, we are focusing on the operational level.
In the following, we will used the word exam and
not task.
3 STATE OF THE ART: BIN
PACKING
The bin packing problem considers N objects and
some bins. The aim is to put all the objects in bins
by minimizing the number of bins, with respect of the
size of objects and bins (Garey and Johnson, 1979).
It is a NP-hard problem and has been widely studied
(Beigel and Fu, 2012).
Our problem consists in the assignment of exams
to a material resource and a period. If we consider
that the horizon planning is composed of couples (l, s)
with material resource l (1, L) and period s (1, T ),
the aim of the problem is to assign exams to these cou-
ples (l, s). Exams have to be planed as soon as pos-
sible, the aim is to minimize the number of different
couples (l, s), that is to say to minimize the number
of bins. Table 1 summarizes the links between the bin
packing problem and the problem described in this
paper.
Several heuristics have been developed to solve
this problem (Coffman Jr et al., 1996).
The typical one is NextFit. Objects are put in the
current bin. If the current bin is full, objects are
put in the next one. We cannot come back to the
current bin anymore.
HEALTHINF2014-InternationalConferenceonHealthInformatics
118
Table 1: Analogies between the bin packing and the consid-
ered problem.
Bin packing problem Considered problem
Object Exam
Bin Couple (l, s)
Size of an object Process time
Due date
Reference place
To assign exams to
To assign objects in bins a period and a
material resource
Constraint of the Constraint of material
capacity of the bins resources opentime
Compatibility constraint
To minimize the number To minimize the number
of used bins of used couples (l, s)
To minimize the last
assigned period
To minimize the sum
of the number of bins
FirstFit is different from NextFit because objects
are put in the first available bin. A new bin is taken
into account if and only if all the previous bins do
not have enough space to accept the current ob-
ject. Bins are not definitively closed. If a smaller
object fits in a previous bin, it will be put in this
one. In the worst case, the quadratic complexity
in time of this algorithm is O(N log(N)), with N
the number of objects (Corcoran and Wainwright,
1995).
BestFit consists in putting objects in the best bin
which has some available space. The best bin is
the one which will have the least available space
after this object will be put in it.
FirstFit heuristic is adapted to our problem. This
will be explained in Section 5.2.
4 MATHEMATICAL MODEL
The used data in this problem are the following. Let
be:
N the number of exams to plan.
K the number of distinct places.
L the number of available material resources.
T the horizon planning. The planning will be done
for s(s = 1, T ) periods.
t
i
(i = 1, N) the process time of exam i.
d
i
(i = 1, N) the due date of exam i.
r
i,k
(i = 1, N and k = 1, K) = 1 if place k is the
reference place of exam i, 0 otherwise.
c
i,l
(i = 1, N and l = 1, L) = 1 if exam i can be
handled by material resource l, 0 otherwise.
s
l,k
(l = 1, L and k = 1, K) = 1 if material resource
l is located on place k, 0 otherwise.
p
l,s
(l = 1, L and s = 1, T ) the opentime of material
resource l during period s.
The considered variables are the following:
x
i,l,s
(i = 1, N; l = 1, L and s = 1, T ) = 1 if exam i
is assigned to material resource l during period s,
0 otherwise.
C
i
is the period assigned to exam i.
The constraints are written as follows:
Material resources opentimes have to be respected
(1).
N
i=1
t
i
.x
i,l,s
p
l,s
, l {1, L}, s {1, T } (1)
An exam is assigned to one material resource and
one period (2).
L
l=1
T
s=1
x
i,l,s
= 1, i {1, N} (2)
An exam is assigned to a compatible material re-
source (3).
L
l=1
(1 c
i,l
)
T
s=1
x
i,l,s
= 0, i {1, N} (3)
(4) defines C
i
, which is the period assigned to
exam i.
C
i
=
T
s=1
L
l=1
s.x
i,l,s
, i {1, N} (4)
(5) and (6) integrity and non negativity con-
straints.
x
i,l,s
{0, 1}, i {1, N}, l {1, L}, s {1, T } (5)
C
i
0, i {1, N} (6)
Four criteria can be studied:
The sum of the completion times, i.e. of the as-
signed periods to the planned exams (7).
H
1
=
N
i=1
C
i
(7)
The makespan, i.e. the period of the last planned
exam (8).
H
2
= max
i
C
i
(8)
ActivitiesPlanningandResourceAssignmentonMulti-placeHospitalSystem-ExactandApproachMethodsAdapted
fromtheBinPackingProblem
119
The number of exams done after their due date
(9).
H
3
= N
N
i=1
L
l=1
d
i
s=1
x
i,l,s
(9)
The number of exams performed out of their ref-
erence place (10).
H
4
= N
N
i=1
K
k=1
L
l=1
T
s=1
x
i,l,s
.s
l,k
.r
i,k
(10)
In the following, we are taking into consideration only
the economics aspects, that is to say to assign the
maximum number of exams during the smaller num-
ber of periods and material resources. The studied
objective function is the minimization of the first two
criteria.
5 APPROACH METHOD
5.1 Coding of the Set of Solutions
A solution X is the assignment of exam i to mate-
rial resource l during period s. It is coded by a three-
dimensional matrix, each dimension defines each set
of subscript. The dimension of the matrix is N L T .
X = (x
i,l,s
), i {1, N}, l {1, L}, s {1, T }
Let be the set of all the solutions. Let be C
the set of admissible solutions.
Size of is given in (11).
card() = L
N
T
(11)
For each solution X , all the exams are assigned
to one period and one material resource (2). If the
material resources opentimes (1) or the compatibili-
ties between exams and material resources (3) are not
respected, the solution is called non admissible solu-
tion.
5.2 Initial Solution
Two ways are used to compute an initial solution.
An admissible initial solution can be built by a
heuristic based on bin packing: FirstFit heuristic. It
is described in Algorithm 1. The maximum number
of exams is assigned to the first material resource dur-
ing the first period. The exam is assigned to the ma-
terial resource if they are compatible. When there is
no free time anymore on this material resource dur-
ing this period, exams are assigned during the same
period to the second material resource, then the third
one, etc. When there is no free time anymore on all
the material resources during the first period, exams
are assigned to the second period in the same way.
Algorithm 1: Heuristic for an initial solution.
Require:
temp
l,s
= 0, l {1, L}, s {1, T }
begin
for each exam i = 1, N do
OK = f alse, l = 1, s = 1
while s T.and.OK = f alse do
while l L.and.OK = f alse do
if exam i is compatible with resource l then
if p
l,s
temp
l,s
t
i
then
x
i,l,s
= 1
temp
l,s
= temp
l,s
+t
i
OK = true
end if
end if
l = l + 1
end while
s = s + 1
end while
end for
end
We can also randomly build an initial solution. It
consists in assigning a random material resource to
each exam during a random period. In this case, the
initial solution is admissible or not, compatibilities
between exams and material resources and resources
opentimes are not taken into account.
5.3 Metaheuristics
The used metaheuristics are single solution based.
Two stochastic algorithms are detailed and used: sim-
ulated annealing and kangaroo algorithm (iterated lo-
cal search). Under some hypotheses, inhomogeneous
algorithm for simulated annealing and kangaroo algo-
rithm converge in probability to an optimal solution
(Aarts and van Laarhoven, 1987) and (Fleury, 1993).
Principle algorithms are given, using the following
writing:
X current solution,
Y candidate solution (or neighbour),
V,W two neighbourhood systems,
H cost function.
HEALTHINF2014-InternationalConferenceonHealthInformatics
120
5.3.1 Simulated Annealing
Originally, the inhomogeneous simulated annealing
was used by Metropolis (Metropolis et al., 1953) to
simulate the physical annealing in metallurgy. Sim-
ulated annealing converges in probability to the set
of optimal solutions if neighbourhood system V sat-
isfies the accessibility and reversibility property. The
algorithm of simulated annealing is described in Al-
gorithm 2.
Algorithm 2: Simulated Annealing.
Require: temperature T
0
, initial solution X, record
solution RX := X , decreasing factor α, maximum
number of iterations IterMax
begin
iter := 0, T := T
0
while iter < IterMax do
Choose randomly and uniformly Y V (X )
if H(Y ) < H(RX) then
RX := Y
end if
if H(Y ) H(X) then
X := Y
else
X := Y with probability e
H(Y)H(X )
T
end if
iter = iter + 1
Generate a new temperature T = α T
end while
end
Two parameters have to be chosen: the initial tem-
perature and the decreasing factor. The initial temper-
ature T
0
is chosen so as all the transitions are autho-
rized at the beginning i.e. e
H(Y)H(X )
T
0
' 1, (X,Y ).
(Aarts and van Laarhoven, 1987) proposes an algo-
rithm to compute the initial temperature.
The used formula to compute the decreasing fac-
tor α is defined in Equation (12). T
a
is the latest tem-
perature computed in the simulated annealing, close
to 0.
α =
IterMax
p
(
T
a
T
0
) (12)
5.3.2 Kangaroo Algorithm
Kangaroo algorithm consists in a stochastic descent,
but if there is no improvement of the current solution
after a number A of iterations, a jump is made. To
make this jump, a solution is chosen in a neighbour-
hood system W different from V . Kangaroo algorithm
converges in probability to the set of optimal solutions
if neighbourhood system W satisfies the accessibility
property. Algorithm 3 describes kangaroo algorithm.
Algorithm 3: Kangaroo Algorithm.
Require: number of iterations A before a jump, ini-
tial solution X, record solution RX := X , maximum
number of iterations IterMax
begin
iter := 0, c := 0
while iter < IterMax do
if c < A then
Choose randomly and uniformly Y V (X )
if H(Y ) H(X) then
if H(Y ) < H(X) then
c := 0
if H(Y ) < H(RX) then
RX := Y
end if
end if
X := Y
end if
else
Choose randomly and uniformly Y W (X )
if H(Y ) 6= H(X) then
if H(Y ) < H(RX) then
RX := Y
end if
c := 0
end if
X := Y
end if
c := c + 1
iter = iter + 1
end while
end
The used formula to compute A is given in (13)
(Fleury, 1993).
A card(V ).ln(2) (13)
5.3.3 Neighbourhood System
The used neighbourhood system V is defined in Algo-
rithm 4. An exam i is randomly chosen. Its current as-
signment is deleted. A material resource is randomly
chosen among the compatible ones: l {1, L} such
as c
i,l
= 1. A period is randomly chosen: s {1, T }.
Exam i is assigned to material resource l and period
s. Equation (14) gives the size of V .
card(V ) N.L.T (14)
ActivitiesPlanningandResourceAssignmentonMulti-placeHospitalSystem-ExactandApproachMethodsAdapted
fromtheBinPackingProblem
121
Algorithm 4: Neighbourhood System.
Choose randomly and uniformly i {1, N}
begin
for l and s such as x
i,l,s
= 1 do
x
i,l,s
= 0
end for
Choose randomly and uniformly l {1, L} such as
l is compatible with i
Choose randomly and uniformly s {1, T }
x
i,l,s
= 1
end
In the kangaroo algorithm, a second neighbour-
hood system W is used. Its principle is to choose ran-
domly and uniformly an exam and to assign it to an-
other period and another material resource randomly
and uniformly chosen. W is used eight times each
time it is called. W satisfies the accessibility and re-
versibility property.
5.3.4 Studied Criteria
According to the mathematical model, two constraints
have to be respected:
Material resources opentimes (1),
Compatibility between assigned material resource
and the considered exam (3).
To check these constraints, overruns constraints
are computed. So a new criterion Cont is defined. It
is equal to the sum of differences between the sum
of the process times of the assigned exams to a mate-
rial resource during period and the opentime resource
over this period, plus the number of exams which are
assigned to non compatible material resources. It is
computed thanks to the constraints (1) and (3). If
Cont is equal to 0, all the constraints are respected.
Cont(X) =
l,s
max(0,
i
t
i
x
i,l,s
p
l,s
)+
i,l,s
x
i,l,s
(1 c
i,l
) (15)
A weight sum of the three criteria, H
1
, H
2
and
Cont can be used. The general form of the objective
function is given in Equation (16), with α
1
, α
2
, α
3
the
weight coefficient.
H(X) = α
1
.H
1
(X) + α
2
.H
2
(X) + α
3
.Cont(X) (16)
6 EXPERIMENTS AND RESULTS
The HCT is composed of three places. There are four
or eight material resources, each one located on one
place. The horizon planning is composed of several
periods, each period represents one half-day. The aim
is to plan N exams in the HCT, that is to say to assign
one period and one compatible material resource to
each exam. Exams have to be planned during material
resources opentimes. The objective is to plan all the
exams at the beginning of the horizon planning, the
last exam must be done over the smallest period.
6.1 Dataset
The data are randomly generated but the size of the
data represents real instances. Two kinds of data are
created:
For instances 50A to 500A, each process time is
between 5 and 45, with 5 minutes steps. Each ma-
terial resource opentime is equal to 300 minutes.
For instances 50B to 500B, each process time is
between 1 and 100, and each material resources
opentime is equal to 100.
Table 2 summarizes the data.
6.2 Exact Method
To solve this model, we use the version 12.4.0 of
CPLEX. The host machine is powered by an Intel
Xeon X5687 quad-core CPU running at 3.6 GHz. For
each size of problem, results about makespans and
sums of the different periods assigned to each exam
are summarized in Table 3. Computation times are
given in seconds.
The computations which are stopped after thirty
minutes are written with a *. In this case, results about
lower and upper bounds are given. The bigger the
problem is, the longer the computation times are.
H
1
is the sum of the completion times, of the as-
signed periods to all the planned exams in the HCT.
H
2
is the makespan, the period of the last planned
exam. The used objective function is given in Equa-
tion (17). α
2
equals 10.000 because in all the in-
stances, H
1
is always smaller than 10.000. Thus, each
criterion is easily readable.
H(X) = H
1
(X) + 10.000.H
2
(X) (17)
For the smallest size of problem, fifty exams have
to be planned. The makespan is equal to two, that
means that at least two periods are used to plan all
of them. The available time in one period is equal to
HEALTHINF2014-InternationalConferenceonHealthInformatics
122
Table 2: HCT Data.
Exams N 50A 50B 100A 100B 200A 200B 300A 400A 500A 500B
Resources L 4 4 4 4 8 8 8 8 8 8
Periods T 8 8 8 15 10 20 15 20 25 40
Table 3: Results with the exact method.
Exams 50A 50B* 100A 100B* 200A 200B* 300A 400A* 500A* 500B*
H
1
51 - 128 - 262 - 486 - - -
H
2
2 - 2 - 2 - 3 - - -
Time 0.22 - 0.16 - 0.98 - 4.41 - - -
Lower bound - 70135 - 120483.1 - 91056.3 - 40774.2 51114.8 162329.9
Upper bound - 70137 - 120520 - 141074 - 40776 51118 327231
material resources opentime during this period, three
hundreds minutes in this case. So one period is more
or less equal to half-a-day. At least, one day is used
to plan fifty exams over the four material resources in
the HCT.
6.3 Approach Methods
Results of the approach methods are compared to the
exact method’s ones. Two cases are possible:
[X ]: we can accept non admissible solutions.
Initial solution is randomly and uniformly gener-
ated. The criteria Cont is used to compute over-
runs constraints. In the objective function, α
3
is
set to 10
6
. Thus Cont is the first criterion to min-
imize. At the beginning of the computation, it is
greater than zero but at the end, it is always equal
to zero. So the best solution will be admissible.
[X C]: we only accept admissible solutions. Ini-
tial solution is created by the proposed heuristic.
Results about initial solution are given in table
4. The generated neighbour has to be admissible,
that is to say Cont = 0.
Table 4 summarizes results about the makespan
(H
1
) and the sum of the assigned periods to each exam
(H
2
).
For each replication, the number of iterations is
equal to 1.000.000. The gold standard are the results
of the exact method given by CPLEX in the previous
section. CPLEX optimal solution is written if found.
If not, upper bounds are written because they are the
best found integer solutions (*). Mostly, methaheuris-
tics find the optimal makespan.
For big instances (**), the makespan is not the
same in approach and exact methods. In these cases,
H
2
in the approach way can be better than in the exact
one.
Moreover, the heuristic can find a better makespan
than the metaheuristics for instances B. In these in-
stances, size of a ”bin” is equal to 100 and some ex-
ams have got a process time which is equal to 100.
With the heuristic, exams are sorted by decreasing
processing times. Thus, the biggest exams are pro-
cessed in the first periods. With the metaheuristics,
once a big exam is put in one of the last period, it can-
not be moved to another ”bin” otherwise it will not
respect resources opentimes. But every time, meta-
heuristics give better results than the heuristic about
H
2
. Most of the times, simulated annealing gives the
best results.
Results about H
2
are optimal if the number of ex-
ams is small. For large instances, there is no opti-
mal solution. The exact method gives an approxi-
mation after thirty minutes of computation. With the
approach methods, optimal makespan, if existing, is
found in some seconds and a good result of H
2
in
some minutes. Almost all the exams are assigned to
the optimal period, only some of them are assigned
during the next period of the optimal one.
7 CONCLUSIONS
In the current economic context, solutions have to be
found to improve hospital efficiency. By dividing up
exams over an Hospital Community of Territory, ex-
ams could be done more quickly. In a same horizon
planning, more exams could be planned.
In this paper, a formulation of the problem is
given. Exact and approach methods are compared.
For large instances, exact method does not return the
optimal solution while approach method find a good
solution in some minutes. Optimal makespan, if ex-
isting, is found in a few seconds and a good result
for all assignments is found in a few minutes. Other
heuristics and other neighbourhood systems have to
be tested to improve the best solution.
At the operational level, in this paper, assignment
are made over seven days. In real life, planner needs
to anticipate the planning one month in advance. By
ActivitiesPlanningandResourceAssignmentonMulti-placeHospitalSystem-ExactandApproachMethodsAdapted
fromtheBinPackingProblem
123
Table 4: Results with the metaheuristics.
Exams 50A 50B* 100A 100B** 200A 200B* 300A 400A 500A 500B**
CPLEX (upper bound)
H
1
2 7 2 12 2 14 3 4 5 32
H
2
51 137 128 520 262 1074 486 776 1118 7231
Solution generated by the heuristic
H
1
2 7 3 12 3 14 4 4 6 31
H
2
54 203 142 684 295 1435 579 957 1408 8228
Kangaroo algorithm, x
H
1
2 7 2 14 2 16 3 4 5 36
H
2
51 154 129 554 266 1220 504 815 1185 6520
Kangaroo algorithm, x C
H
1
2 7 2 13 2 15 3 4 5 33
H
2
51 140 128 508 264 1126 496 800 1161 6465
Simulated annealing, x
H
1
2 7 2 13 2 15 3 4 5 35
H
2
51 138 128 503 263 1154 491 796 1164 6326
Simulated annealing, x C
H
1
2 7 2 13 2 15 3 4 5 34
H
2
51 139 128 488 262 1127 494 799 1172 6228
increasing the number of periods, or by increasing
the size of the periods, we could apply this study to
the tactical level or the strategic one. This could be
used to determine the number of resources or places
needed in a HCT.
The made hypothesis here is that human resources
are always available. In further research, we could in-
clude them to our model. Human resources can work
or not overtime and they can be allowed to move over
the HCT. So the load between the places can be bal-
anced.
REFERENCES
Aarts, E. and van Laarhoven, P. (1987). Simulated Anneal-
ing: Theory and Applications. Kluwer Academic Pub-
lishers.
Beigel, R. and Fu, B. (2012). A dense hierarchy of sublinear
time approximation schemes for bin packing. In Fron-
tiers in Algorithmics and Algorithmic Aspects in Infor-
mation and Management, pages 172–181. Springer.
Coffman Jr, E. G., Garey, M. R., and Johnson, D. S. (1996).
Approximation algorithms for bin packing: A survey.
In Approximation algorithms for NP-hard problems,
pages 46–93. PWS Publishing Co.
Corcoran, A. L. and Wainwright, R. L. (1995). Using libga
to develop genetic algorithms for solving combinato-
rial optimization problems. Practical Handbook of
Genetic Algorithms, 1:143–172.
Fleury, G. (1993). M
´
ethodes stochastiques et d
´
eterministes
pour les probl
`
emes NP-difficiles. Ph.D., Universit
´
e
Blaise Pascal, Clermont-Ferrand II.
Garey, M. R. and Johnson, D. S. (1979). Computers
and Intractability: A Guide to the Theory of NP-
completeness. WH Freeman and Company, New
York.
Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N.,
Teller, A. H., and Teller, E. (1953). Equation of state
calculations by fast computing machines. The Journal
of Chemical Physics, 21:1087–1092.
HEALTHINF2014-InternationalConferenceonHealthInformatics
124