Journal Pre-proofs

A modified biogeography-based optimization algorithm with guided bed selec-

tion mechanism for patient admission scheduling problems

Abdelaziz I. Hammouri

PII: S1319-1578(19)31543-5
DOI: https://doi.org/10.1016/j.jksuci.2020.01.013
Reference: JKSUCI 740

To appear in: Journal of King Saud University - Computer and In-

formation Sciences

Received Date: 25 November 2019

Accepted Date: 26 January 2020

Please cite this article as: Hammouri, A.I., A modified biogeography-based optimization algorithm with guided bed
selection mechanism for patient admission scheduling problems, Journal of King Saud University - Computer and
Information Sciences (2020), doi: https://doi.org/10.1016/j.jksuci.2020.01.013

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover
page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will
undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing
this version to give early visibility of the article. Please note that, during the production process, errors may be
discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2020 Production and hosting by Elsevier B.V. on behalf of King Saud University.
 A modified biogeography-based optimization algorithm
with guided bed selection mechanism for patient admission
scheduling problems

Abstract
One of the complex combinatorial optimization problems is the Patient admission
scheduling problem (PASP), which is concerned with assigning the patients arriving
into a hospital to available beds to get medical services. The objective of PASP is
to maximize the patients’ comfort, medical treatment effectiveness, and hospital uti-
lization. This research proposes a new approach based on Biogeography-Based Op-
timization (BBO) algorithm for tacking the PASP. BBO was inspired from the idea
of species migration between different habitats . Due to the complexity of the search
space in PASP, the original BBO has been equipped with a guided bed selection (GBS)
mechanism in order to improve its results and performance, as well as the operator
capabilities of BBO which are modified to improve its diversity . These three variants
of BBO are compared with each other using six de facto data sets that are widely used
in the literature with varying sizes and complexity. The modified BBO is able to yield
better results than the other variants. In a nutshell, this paper provides a new PASP
method that can be considered an efficient alternative for the scheduling domain to be
used by other researchers.
Keywords: Biogeography Based Optimization, Evolutionary Algorithms,
Meta-Heuristic, Healthcare, Patient Admission scheduling.

1. Introduction

Nowadays, health-care organizations such as hospitals, face the issue of an in-

creased demand for maximizing the usage of resources and minimizing the duration
of patients stay [1]. The unit that is responsible for achieving this demand is called
the Central Admission Unit (CAU). It allocates beds to patients arriving to the hospi-
tal. The process of using beds is not only restricted to the situation of the patient but
also depends on the patient’s room preference, medical equipment in the room (e.g.,
oxygen, telemetry, etc.) and the medical skills of the department staff [2]. There-
fore, managing and controlling the task of finding suitable beds for arriving patients
is difficult and requires substantial knowledge and experience [3]. Thus, an efficient
admission scheduling becomes increasingly important.
In optimization terms, PASP is considered a combinatorial optimization problem
(COP) that belongs to the NP-hard optimization problem in almost all of its instances
and variations [4]. It has been rigorously formulated as an optimization scheduling

Preprint submitted to Journal of LATEX Templates January 25, 2020

problem. Therefore, several optimization methods have been adapted or adjusted to
deal with the complex nature of the PASP search space. Demeester et al. [5] for-
mulated the PASP and generated thirteen benchmark instances based on real hospital
situations. They hybridized tabu search algorithm with local search for tackling the
PASP. Bilgin et al. [6], proposed a framework based on the hyper-heuristic method
combined with local search neighborhoods algorithm for improving the obtained so-
lutions for several PASP instances. A multi-neighborhood local search procedure was
suggested by Ceschia and Schaerf [7] and that was used for adapting the relations be-
tween neighborhood and considered different penalty weights of the objective function
components, which leads to further solution improvements. Later on, Ceschia and
Schaerf [8] proposed and solved a new formulation of PASP, and extended the PASP
by adding new features, such as overcrowding of the rooms, the possibility of delayed
admissions, uncertainty in the length of stay, and the presence of emergency patients.
Ceschia and Schaerf made the new instances available to the public. In addition, they
developed a new metaheuristic method based on complex neighborhood structure and
simulated annealing for solving both of the static and dynamic versions of the problem,
the proposed method is able to solve large problem instances in a reasonable amount
of time. Guido et al. [9] defined an efficient model formulation for solving the new
instances of PASP that were proposed by Ceschia and Schaerf [8], Guido et al. de-
fined admission dates and assigned patients to rooms, they suggested how to determine
weight values, and how to use the penalise constraint violations. The results obtained
from the suggested model showed an improvement in scheduling quality.
In 2016, Ceschia and Schaerf [10] made a new extension to the PASP (called
Dynamic-PASP) in order to make it suitable for practical applications, as well as to
make it a challenging case study. Ceschia and Schaerf proposed an exhaustive model
that includes new components of the objective function, a complex notion of patient
delay and a flexible planning horizon. Moreover, Ceschia and Schaerf developed the
infrastructure and user interface of the prospective application, which included the in-
stance generator and the solution validator (both are available on the web along with
obtained solutions). They performed a comprehensive experimental evaluation on the
proposed method. This new extension of the PASP has been resolved by Guido et
al. [11] and Zhu et al. [12]. Where Guido et al. proposed a new integer programming
model for tackling the Dynamic-PASP. They applied a matheuristic that combines exact
methods with meta-heuristic approaches. The proposed approach showed promising
results. While Zhu et al. proposed a new method based on the full Mixed Integer Pro-
gramming for solving the Dynamic-PASP, they also studied the compatibility of short
and long terms objectives in solving the Dynamic-PASP. The results obtained from the
proposed approach are significantly better than the available results in the literature for
26 out of the 30 instances.
Dorgham et al. [13] proposed a new approach for tackling the PASP by hybridizing
the simulated annealing algorithm with genetic algorithm in order to obtain a global
search method. Guido et al. [14] developed a new matheuristic method that is classified
as cooperative between a meta-heuristic and an exact method. The proposed method
allows getting good results for all benchmark instances. Bastos et al. [15] addressed
PASP using a parameter-free exact method, the proposed method obtained new best
results for 9 out of 13 benchmark instances.

2
 The first evolutionary method based on employing the biogeography-based opti-
mization (BBO) algorithm for tackling the PASP was proposed in [16] .Results of the
proposed approach were modest. Later on, a new modification of the BBO algorithm
has been made in [17], where the authors proposed a new mechanism for selecting a bed
for a patient and called it the guided bed selection (GBS) mechanism. This mechanism
improved the results obtained by the BBO algorithm and it became closer and more
competitive to several results in the literature. Another heuristic based technique was
proposed in [18] that is based on dynamic constraint aggregation, column generation
and branch-and-bound techniques. This approach achieved new results for a subset of
the PASP instances. A non-linear adaptive great deluge method was proposed in [19]
for tackling the PASP, where a subset of the PASP instances has been solved in less
computational time comparing to other methods in the literature. Also in [20], the au-
thors reduced the required computational time for finding a high quality solution, where
two heuristics based on mixed integer programming were employed in order to break
down the PASP to sub-problems. The standard versions of optimization methods can
struggle in mastering the complexity of the PASP search space. Therefore, researchers
in this domain either modify or hybridize their methods for PASP [20, 21, 22, 10].
BBO is an efficient version of evolutionary algorithms derived from GA which has
been widely used in different research areas. It is a population based meta-heuristic al-
gorithm that was created in 2008 by [23], and was inspired by the migration of species
amongst different homeland. It has characteristics in common with other evolution-
ary algorithms such as Particle Swarm Optimization (PSO) and Genetic Algorithms
(GA) [23]. Moreover, the performance of BBO has been tested on the sensor selec-
tion problem and the 13 benchmarks and also compared to the performance of other
well-known algorithms such as Ant Colony Optimization (ACO) algorithm, Evaluation
Strategy (ES) algorithm, PSO, GA, . . . [23]. The results obtained from BBO show the
outstanding performance of BBO and provide evidence that the theory of BBO can be
successfully applied on different types of optimization problems such as the data clas-
sification problem [24], image segmentation [25], flow shop scheduling [26], global
optimization problems [27], and economic load dispatch problem [28]. However, for
complex optimization problems, BBO may have a population with low diversity, which
degrades the performance of BBO and can lead to poor convergence precision [29, 30].
Technically, BBO is started with a random population. Generation after genera-
tion, the BBO employs migration and mutation operators controlled by emigration and
immigration rates and mutation rate, respectively. Thereafter, the replacement operator
has to take over to achieve the survival of the fittest principle. This process proceeds
until a steady-state generation is reached. Quite recently, two versions of BBO have
been adapted for PASP. BBO [16] and BBO-GBS [17]. A new Guided Bed Selection
(GBS) mechanism was employed with BBO in order to improve its results and perfor-
mance [17]. These two versions were unable to obtain the optimal solutions in PASP
and therefore, can be further investigated.
The main objective of this paper is to propose a new modified version of BBO that
is able to tackle the PASP to cover the gaps of the previous BBO versions dealing with
the same problem. The new version is called MBBO-GBS which modifies the standard
replacement scheme of BBO to stress the diversity aspect more rigorously. In order to
evaluate the proposed method, six de facto problem instances with various complexity

3
and different sizes are used. Initially, the modified BBO is compared with BBO [16]
and BBO-GBS [17]. The proposed method is able to outperform the results obtained
by the previous BBO versions. Furthermore, the results produced by MBBO-GBS are
also compared to the results produced by nine well-known algorithms. Interestingly,
MBBO-GBS is able to produce comparable results to others and ranks sixth on average.
The remaining parts of this paper are organized as follows: Section 2 describes
the definition and problem formulation for the patient admission-scheduling problem,
While Section 3 presents the original BBO. The proposed method is discussed in Sec-
tion 4. The experimental settings and the obtained results are discussed in Section 5,
followed by Section 6 which represents the conclusion and future works.

2. Patient Admission Scheduling Problem (PASP)

The PASP is the process of assigning patients to beds with regards to rooms/departments
to achieve high proficiency medical treatment and to increase the comfort of patients
[2]. Each patient needs to spend a specific period of time in the hospital. The patient
should be assigned a room for each night between admission and discharge dates. A
time slot represents a night in the hospital schedule. Length-of-stay represents a set of
consecutive nights that the patient needs to spend in the hospital.

2.1. Constraints
The assignment of patient to a bed should respect two types of constraints: hard and
soft constraints. Satisfying all hard constraints is required to yield a feasible solution,
while the violations of soft constraints is allowable but not preferable. The main goal
of PASP is to produce a feasible solution with the least soft constraint violations. Hard
and soft constraints can be described as shown bellow [2]:

• Hard constraints
1. HC1: Beds cannot be double-booked at the same time slot.
2. HC2: The dates of admission and discharge must be predefined.
3. HC3: Time span must be connected with the length-of-stay for each patient.
4. HC4: The number of the assigned patients to a room should not exceed
room capacity per night.
5. HC5: Patient must be admitted to a free bed for each night of his/her stay
in the hospital.

• Soft constraints
1. Room Transfer (SC1); patient should not be moved from one room to
another during his/her stay in the hospital.
2. Room Gender (SC2); All patients admitted to the same room in a given
night should have the same gender property, unless the room property al-
lows for having different genders at the same time. The room genders are
Female (F), Male (M), Mix (N) which means that patients of both genders
can be admitted to the room at the same time, or Depending (D) on the
gender of the first patient admitted to the room each night.

4
 3. Age Policy (SC3); Patient has to be admitted to a room in a department
that is suitable for his/her age.
4. department Specialism (SC4); Patient has to be admitted to a room in a
department with suitable specialism for his/her pathology.
5. Room Mandatory Requirements (SC5); Patient has to be admitted to a
room with the mandatory requirements of his/her pathology.
6. Room Preferred Requirements (SC6); Patient has to be admitted to a
room with the preferred requirements of his/her pathology.
7. Room Size (SC6); The size of room (single, double or ward) should be
based on patient’s preference.

In this research, the first soft constraint (Room transfer) is considered a hard con-
straint. In which case the search space of the problem becomes smaller. The problem
is also tackled in a similar way as in [7, 16, 17, 31]. In [6, 5, 2] the authors keep the soft
and hard constraints as in the original problem. In both cases, the researches that tackle
the Room Transfer as a hard constraint can compare their results to the researches that
tackle the Room Transfer as a soft constraint. Since tackling the Room Transfer as hard
constraint means it is not allowed to violate this constraint. Thus, there is no penalty
to be added to the solution cost. Whereas if the search method violates this constraint
during the search process, it needs to repair the solution to become feasible. On the
other hand, if the search method tackles the Room Transfer as a soft constraint, it is
allowed to violate it but a penalty will be added to solution cost.
The original data set of PASP that is introduced in [2] was used in this research.
This data set consists of 6 instances. Kindly note that the patients that are admitted after
the planning horizon and the patients who leave the hospital in the same admission date
are omitted from the data.
The data set instances of PASP that we use in this research include different inputs,
such as departments, rooms, time slots (Nights) and patients [2]. The objective func-
tion of PASP is to fulfill all hard constraints for all patients, as well as minimize the
violations of soft constraints. In this research, the pre-processing steps given by [31]
have been adapted to simplify the objective function, where the penalties of violating
SC3, SC4, SC5, SC6, SC7, and partially SC2 have been merged in the patient-room
cost (PRC) matrix. In the pre-processing step after reading the input file, the PRC ma-
trix is calculated only once for all patient to room assignments. The SC2 (room gender
policy) is considered in PRC matrix if the gender policy of the room is M or F. How-
ever, if the gender policy of the room is N or D, the penalty of SC2 is not considered in
PRC matrix. Table 1 presents the weights of soft constraints in case of violation.

2.2. Problem Formulation

To evaluate the objective function of PASP, this research uses the same mathemat-
ical model of the problem that is defined in [32], which is identified as follows: the
symbols that are used in the problem are defined First as: 1) P is the set of patients,
2) P M is the set of male patients, 3) PF is the set of female patients, 4) R is the set
of rooms, 5) B is the set of beds, 6) cr is the capacity of room r, and 7) T is the set
of the time slots. The decision variables which are used in the development of the
mathematical model are defined as:

5
 Table 1: the weights of the soft constraints

Constraint Penalty
Room transfer (SC1) 11.0
Room gender (SC2) 5.0
Age Policy (SC3) 10.0
Department Specialism (SC4) 1.0
Room Mandatory Requirements (SC5) 5.0
Room Preferred Requirements (SC6) 2.0
Room Size (SC7) 0.8

• X p,r,t : if the patient p is assigned on the time slot t to the room r, then the value
of X p,r,t is 1, otherwise the value will be 0.
• g p : represents the gender of the patient; 1 if the patient is male and 0 for female.

g p .X p,r,t at the time slot t in the

P
• mr,t : the number of male patients is equal to
p∈P
room r.
P
• fr,t : the number of female patients is equal to (1 − g p ).X p,r,t at the time slot t
p∈P
in the room r.
• PS P : represents the length of stay, which is defined as:
{ t ∈ T |(t < t s,p + d p ) t >= t s,p }
V

where t s,p is the admission day and d p is the length of stay p.

• T r p,t represents the room transfer, which is 1 if ∃r∈R|X p,r,t ,X p,r,t+1 otherwise 0,
where t and t + 1 ∈ PS P

The mathematical formulation of the hard and soft constraints

• The following equation is used for selecting a room for a patient for all nights
during his/her stay: X
X p,r,t = 1, ∀t ∈ T, p ∈ P. (1)
r∈R

• the following equation is used to ensure that room capacity is not exceeded:
X
X p,r,t ≤ cr , ∀t ∈ T, r ∈ R. (2)
p∈P

• The following equation is used for calculating the total violation penalties of soft
constraint SC1 (room gender policy):
X
FRG = WRG .min(mr,t , fr,t ). (3)
t∈T,r∈R

where WRG represents the weight of SC1 for room R and gender G.

6
 • the following equation is used for calculating the total penalties of room transfer
constraints (SC6): X
FTr = WTr .T r p,t . (4)
r∈R,p∈P,t∈T

where WTr represents the transfer weight of the room r.

• The following equation is used to calculate the total room penalty of SC3, SC4,
SC5, SC6, SC7 and partially SC2 constraints:
X
F PRC = C p,r .X p,r,t . (5)
r∈R,p∈P,t∈PS P

where C p,r is the penalty of reserving room r to assigning patient p.

Objective function: When solving the PASP, the quality of the generated solution
can be assessed by computing the sum of penalties of violating the soft constraints
which represents the objective function of the PASP. The objective function can be
calculated as follows:
F = F PRC + FRG + FTr . (6)

3. Biogeography Based Optimisation

Biogeography is a natural way of distributing different types of species, where the

geographical areas can be considered good or bad homeland based on their features,
such as: diversity of vegetation, rainfall, temperature and diversity of topographic fea-
tures. These features are considered as independent variables or "Suitability Index
Variables" (SIVs). The quality of the homeland is computed based on dependent vari-
ables known as the "Homeland Suitability Indexes" (HSI) which reflect how good the
solution is. If the homeland has a high HSI, it will be considered a good homeland, and
if it has a low HSI then it will be considered as bad (poor) homeland.
In 2008, Dan Simon introduced the BBO algorithm [23] as a new population based
meta-heuristic algorithm that was inspired from the idea of moving species between
different islands as they looked for a good homeland. BBO algorithm employs the
HSI to indicate how good the homeland (solution) is; where the solution with high
HSI is considered a good solution, while the solution with low HSI is considered a
bad (poor) solution. BBO algorithm has two main operations which are: the migration
and mutation operations. The migration operation is used to improve poor solutions
by copying features from good solutions (high quality solutions) into the poor ones.
While the mutation operation is used for randomly modifying one or more features in
the solution based on a pre-computed probability. The mutation operation can lead to
improving the diversity in the population, as well as reducing the algorithm speed in
reaching the stagnant state.
In [23], the author shows how to design and implement the principle of biogeog-
raphy in the BBO algorithm. The author applies the BBO algorithm on 14 benchmark
functions, as well as in the problem of sensor selection for estimating aircraft engine
health. As shown in the paper, the BBO algorithm has shown outstanding performance
in comparison to other well-known methods. The success story of BBO encourages

7
researchers to employ the BBO in different domains, such as reliability optimization
problems [33], camera calibration problem [34], transportation network problem [35],
and global optimization problems [27].

Algorithm 1 The pseudo-code of the original BBO algorithm.

maxGeneration ← the maximum number of generations (Iteration).
populationS ize ← the size of the population.
Population ← {} // create empty population
BestS olution ← φ
for i from 1 to populationS ize do
S i ← new random solution
HS Ii ← fitness(S i )
Population ← Population ∪ S i .
Sort the Population based on the HS I for each solution
BestS olution ← Population(1)
for gIndx from 1 to maxGeneration do
for i from 1 to populationS ize do
Calculate λi and µi as shown in Eq. 8 and Eq. 7 respectively.
for i from 1 to populationS ize do
S i ← Population(i)
Execute the migration operation on S i
Execute mutation operation on S i
HS Ii ← fitness(S i )
Population(i) ← S i
Sort the Population based on the HS I for each solution
if Population(popS ize) worst than BestS olution then
Population(popS ize) ← BestS olution
return BestS olution

Alg. 1 shows the pseudo-code of the BBO algorithm. In the initialization state
of the BBO algorithm, a set of solutions (population) is randomly generated. Next,
the population is sorted based on the quality of its solutions. The improvement loop
in the BBO algorithm is then started. In each iteration in the improvement loop, the
emigration and immigration rates (µ and λ, respectively) are calculated according to
the following equations:

µi = E.(Ki /n) (7)

λi = I.(1 − Ki /n) (8)
where E and I are the maximum emigration and immigration rates, respectively.
Based-on the parameter tuning that performed in [23], the value for both of E and I is
set to 1. Ki represents the rank of solution i, where the worst solution has the lowest
value of the rank and the best solution has the highest value of the rank. n represents
the size of the population. Based on the computed values of µ and λ, the BBO al-
gorithm performs the migration operation in order to improve the quality of the poor
solutions by sharing features (SIVs) of the good ones. After the migration operation
is completed, the mutation operation is performed with respect to the mutation rate.
Finally, fitness function is calculated and the population is updated. The worst solution
in the population is then replaced with the best solution found so far.

8
4. BBO versions for PASP

In the literature, two version of BBO have been used for tackling PASP and these
are

• Original BBO with random bed selection (BBO-RBS) [16].

• Original BBO with guided bed selection (BBO-GBS) [17].
In BBO-RBS, the original BBO algorithm has been employed for tackling PASP
(as discussed in Section 4.1). Later on, a guided bed selection mechanism (in BBO-
GBS) has been introduced in order to improve the quality of the solutions (as discussed
in Section 4.2). In this research, we introduced a new mechanism for updating the
population during the search process in order to maintain the diversity of the population
(as explained in Section 4.3).

4.1. BBO-RBS for PASP

The original BBO algorithm (As shown in Alg. 1) has been employed for solving
the PASP. BBO starts with the feasible “initial solutions”. These solutions are then
improved by invoking the migration and mutation operations iteratively. The initial
solutions are generated as follows: one of the patients (who has not been assigned to
a bed) is selected randomly and assigned to any available bed for the length of his/her
stay. If there are no available beds, a bed is randomly chosen, freed and then assigned
to that patient. These steps are repeated until all patients are assigned to beds. The
details of the BBO-RBS are presented in the followed subsections.

4.1.1. Solution Representation

A two dimensional matrix is used to represent the solution (as shown in Fig. 1).
The index of the row in the matrix represents the bed ID. The index of the column in
the matrix represents the night ID and the entries of the matrix represent the patient ID.
This solution representation prevents the violation of two hard constraints, which are:
(i) maximum one patient per bed-time slot (HD1), and (ii) the number of patients per
night in the room cannot exceed its capacity (HD2).

Figure 1: The solution representation matrix.

9
4.1.2. Migration Operation
In BBO, the migration operation is used to improve the quality of poor solutions by
sharing the features of good ones. In PASP, the migration operation is used to migrate
the beds of specific patients from good solutions to poor ones based on the value of
λ and µ, which is calculated as in Eq. 8 and Eq. 7, respectively. Alg. 2 shows the
migration operation of BBO.

Algorithm 2 The pseudo-code of the migration operation.

numO f S IV ← the number of patient
S i ← current solution
for k from 1 to numO f S IV do
r ← random value between 0 and 1
if r < λi then
S j ← random solution with a probability proportional to µ j
B ← the bed of the Patientk in S j
move Patientk in S i to bed B

4.1.3. Mutation Operation

In the BBO algorithm, the mutation operation is used to increase the diversity of
the population, as well as reduce the convergence speed by randomly adding new fea-
tures to each solution in the population. Without the mutation operation, high quality
solutions tend to be more dominant than other solutions in the population which can
lead to premature convergence. Three different neighbourhood structures have been
employed in the mutation operation, which are:

• Nbs1 (move): randomly select a patient then move him/her to a new random bed.
• Nbs2 (swap): randomly select two patients then swap their beds.
• Nbs3 (swap and move): randomly select two patients then swap their beds. If
we cannot reserve one of the two beds to corresponding patients for the whole
duration of stay, then patient will be moved to a new random bed.

In the mutation operation, a random number is generated. If this number is less than
0.5 then the first neighbourhood structure (Nbs1) is executed. If the randomly gener-
ated number is greater than or equal to 0.5, then the second neighbourhood structure
is executed (Nbs2). If Nbs2 fails and cannot be completed, the third neighbourhood
structure (Nbs3) is executed.

4.1.4. Repair Method

In BBO, the migration operation can lead in some cases to unfeasible solutions.
Therefore a repair method is employed after the migration operation to maintain the
feasibility of the solution. The repair method is used to release the beds of any patients
who violate any hard constraints, then move him/her to a new bed that satisfies all hard
constraints. This process is repeated until the solution becomes feasible (all patients
assigned to beds that satisfy all hard constraints).

10
4.2. BBO-GDS for PASP
The guided bed selection (GBS) mechanism is used within the neighbourhood
structure in the BBO algorithm. For each patient, a list of three groups of beds will
be created. To do this, beds are sorted in ascending order according to the penalty of
assigning the patient to the bed. The first group of beds contains the most suitable
beds (beds with low penalty), while the second group of beds contains the next suitable
beds (beds with medium penalty) and the last group of beds contains the worst list of
beds (beds with high penalty). GBS mechanism is used by the neighbourhood structure
when selecting a bed for a new patient. The GBS mechanism tries to find a bed in the
first group of beds. If this fails then it tries to find a bed from the next group and so on.

4.3. MBBO-GBS for PASP: proposed method

In this paper, a modified version of BBO algorithm with GBS mechanism (called
MBBO-GBS) is proposed to update the population of the BBO during the search pro-
cess, in order to maintain the diversity of the population.
In the original BBO, the worst solution in the population is replaced by the best
one at the end of each iteration. This leads to the loss of diversity of the population by
creating a duplicate copy (or with high similarities) of the solutions, i.e., the replace-
ment makes two solutions which are almost similar to each other (following the best
solution, see Alg. 1). Moreover, these duplicate copies of the best solutions have very
low probability to accept changes during the migration operation. Accordingly and due
to this low probability, the high quality solutions (in general) do not accept new fea-
tures (SIV) from other solutions during the migration operation, unlike poor solutions
that accept many features from the good solutions. However, in the MBBO-GBS, the
best solution found in previous iterations replaces the same solution in the population
(referring to the same solution id). This replacement is only executed if the quality of
the best solution in the previous iteration is better than the quality of the best solution in
the current iteration. This mechanism is illustrated in the following scenario (as shown
in Fig. 2):
Assume that there are 50 solutions (coded as S1, S2, S3, . . . , S50), and S2 is treated
as the best solution. Let’s further assume that after the first two iterations, S2 remains
the best solution. Now consider the following steps in both the original and MBBO-
GBS:
Iteration 1:
• Step 1: Execute the migration and mutation operations then sort the population.

• Step 2: In the original BBO, replace the worst solution (S’9) with the previous
best (S2). While in the MBBO-GBS, the previous best (S2) replaces the solution
with the same id (S’2), we prevent the duplication of solution S2.
• Step 3: Sort the population before completing the first iteration.
Iteration 2:

• Step 1: Then execute the migration and mutation operations, then sort the popu-
lation.

11
 • Step 2: In the original BBO, replace the worst solution (S”18) with the previous
best (S2). While in the MBBO-GBS, the previous best (S2) replaces the solution
with the same id (S’2)
• Step 3: Sort the population before completing the second iteration, and continue
the same way for the third iteration.

%HVW6ROXWLRQ¼ 6 6
 6
 6 6
 6
 6
6 6
 6
 6
 6

 6

 6



2ULJLQDO%%2
6 6
 6
 6
 6

 6

 6

6 6
 6
 6
 6

 6

 6















6 6
 6
 6
 6

 6

 6


:RUVW6ROXWLRQ¼ 6

6
 6 6
 6

 6
6


%HVW6ROXWLRQ¼ 6 6
 6 6 6
 6 6
6 6
 6
 6
 6

 6

 6



0RGLILHG%%2
6 6
 6
 6
 6

 6

 6


6 6
 6
 6
 6

 6

 6















6 6
 6
 6
 6

 6

 6


:RUVW6ROXWLRQ¼ 6

6
 6
 6
 6

 6

 6


,QLWLDO 6WHS 6WHS 6WHS 6WHS 6WHS


6WHS 1H[W
3RS LWHUDWLRQV
,WHUDWLRQ ,WHUDWLRQ

Figure 2: Iteration steps in maintaining the diversity of the population.

In this scenario, the original BBO replaces the worst solution (in each of the it-
erations) by S2. As a result (after the second iteration), we get three versions of the
solution S2 (i.e., S2, S’2, S”2) with high similarity. However, in our MBBO-GBS, the
best solution does not cause any solution duplication in the population, which main-
tains the diversity of the population for a longer time. Furthermore, in the original
BBO, if we assume that the S2 remains the best solution for 30 iterations, we will
have 30 copies of the best solutions (S2).

5. Experimental Results

The MBBO-GBS was implemented using Java and performed on core i5 (2.53GHz)
PC with 6GB RAM. The experiments were executed for 10 independent runs. The
termination criterion is set as the maximum number of iterations which is equal to
2000 that caused the running time about 900 seconds (for the first instance), and 1200
seconds for the other instances. The parameter settings of the proposed algorithm are
Population size is set to 50 solution and the mutation rate is set to 0.001.

12
5.1. Comparisons between variants of BBO
Table 2 shows the results of minimum (Min), maximum (Max), average (Avg),
standard deviation (Std) and CPU time of the best result (Time) in seconds out of 10
runs. The best results are presented in bold.

Table 2: Experimental results of the comparisons between variants of BBO.

BBO-RBS [16] BBO-GBS [17] MBBO-GBS (proposed approach)

Inst.No.
Min Avg Std Time Min Avg Std Time Min Avg Std Time
1 1233.4 1312.8 59.5 686.3 850.6 878.3 35.2 672.4 800.0 823.8 15.0 578.4
2 2027.0 2088.9 60.4 945.7 1418.4 1460.6 43.3 1210.0 1324.4 1378.2 29.2 935.7
3 1385.2 1433.7 41.7 790.0 962.6 976.4 10.4 1047.8 897.0 917.2 16.0 939.1
4 2211.0 2301.5 69.9 902.5 1610.6 1636.2 16.2 1201.2 1556.0 1598.6 34.1 1052.9
5 800.8 828.9 28.3 592.6 714.3 729.2 12.7 774.4 687.2 710.7 10.4 754.6
6 1283.2 1317.1 37.9 684.7 1004.4 1027.8 27.2 946.3 937.6 970.8 18.8 838.5

As shown in Table 2, we can deduce that the MBBO-GBS outperforms the two
other variants of BBO algorithm in comparison to all instances in terms of penalty.
The comparisons in terms of the computational time, show that the BBO-RBS takes
less computational time compared to the others except for the first two instances. This
can be due to the fact that the GBS mechanism enforces the selection of the bed with
low penalty, which will of course take longer time than the RBS mechanism. Whereas
in the RBS mechanism, any available bed can be selected regardless of its penalty,
unlike the in the GBS mechanism which tries to choose a bed from the most suitable
subset of beds. If there are no available beds, it tries to find a bed from the next subset
of beds, and so on. However, this approach leads to better solutions.
Fig. 3 is a clear evidence of how good the proposed approach is as it shows the
convergence curve of the best solutions found during the search for each version of
BBO. Fig. 3 indicating that the convergence speed in the MBBO-GBS was reduced
and became slower than BBO-RBS and BBO-GBS, which delayed the stagnant stage
during the search process and therefore led to better results in all data set instances.
In most cases, the BBO algorithm reaches the stagnant state after the first 500 itera-
tions as shown clearly in the Fig. 3. About 95% of improvements in BBO are obtained
within the first 500 iterations and about 98% of the improvements are achieved within
the first 1000 iterations. Additionally, the guided bed selection mechanism improves
the quality of the solution from 25% to 30% for all instances (except for Instance5,
where it is about 12%). It can be clearly observed that the MBBO-GBS is much better
than both BBO-RBS and BBO-GBS.
We executed a statistical T-test to examine whether there are any significant dif-
ferences between the algorithms in the comparison with the significance interval of
95% (α = 0.05). We executed a pair of comparisons as follows: 1) BBO-RBS versus
BBO-GBS. 2) BBO-RBS versus MBBO-GBS. 3) BBO-GBS versus MBBO-GBS. Ta-
ble 3 shows the p-values of the tested instances, where the presented p-values indicate
enough evidence to conclude that there are significant differences (p-values < 0.05)
between algorithms in comparison for all data set instances.

5.2. Comparisons with the State-of-the-Art

In this subsection, the results obtained by MBBO-GBS are compared to other avail-
able approaches in the literature. Table 4 shows the methods names and abbreviations

13
  ,QVWDQFH 
,QVWDQFH
 %%25%6

KͲZ^ 
%%2*%6
 KͲ'^

DKͲ'^ 0%%2*%6









6ROXWLRQ&RVW

6ROXWLRQ&RVW




 





 





 




















































































,WHUDWLRQ1R ,WHUDWLRQ1R

,QVWDQFH ,QVWDQFH


%%25%6 %%25%6


 %%2*%6
%%2*%6
 

0%%2*%6  0%%2*%6




 




6ROXWLRQ&RVW



6ROXWLRQ&RVW





 





 





 














































          
,WHUDWLRQ1R ,WHUDWLRQ1R

 ,QVWDQFH 

,QVWDQFH
 
%%25%6
 %%25%6

%%2*%6
 %%2*%6
 0%%2*%6
 0%%2*%6






6ROXWLRQ&RVW

6ROXWLRQ&RVW













 

 

 

 


















































































,WHUDWLRQ1R ,WHUWLRQ1R

Figure 3: The best solution found in each of the iterations (for instances from 1 to 6).

Table 3: p-values obtained between different version of BBO.

Instance BBO-RBS vs. BBO-RBS vs. BBO-GBS vs.

.
No BBO-GBS MBBO-GBS MBBO-GBS
1 0.0007 0.0000 0.0088
2 0.0002 0.0003 0.0007
3 0.0003 0.0000 0.0000
4 0.0000 0.0000 0.0001
5 0.0000 0.0039 0.0012
6 0.0000 0.0066 0.0057

of the methods used in the comparison. While Table 5 shows the best, average, stan-
dard deviation (Stdev.), and the rank based on the average of the obtained results for

14
 Table 4: Key to comparative methods for solving the PAS problem

Key Method name Reference

SA Simulated annealing [31]
NLGD Non-linear great deluge [19]
DHS-GD Simple random great deluge with dynamic heuristic [6]
DHS-SA Simple random simulated annealing with dynamic heuristic [6]
HH-SA Simulated annealing with hyper-heuristic [2]
HH-IE Improved with equal hyper-heuristic [2]
DHS-OI Simple random improvement with dynamic heuristic [6]
TS Tabu Search [2]

Table 5: The results of the algorithms.

Instance# Criteron BBO SA NLGD DHS-GD DHS-SA HS HH-SA HH-IE DHS-OI TS

1 Best 800 659.2 664.6 674.8 680 742 NA NA 884.4 NA
Average 823.76 665.61 678.61 685.68 709.24 774.62 830.36 893.69 959.38 1051.02
Stdev. 15 3.2 6.09 7.43 26.65 23.6 18.83 46.1 51.25 48.17
Rank 6 1 2 3 4 5 7 8 9 10
2 Best 1324.4 1143.6 1162.4 1185.2 1191.6 1273.6 NA NA 1584.8 NA
Average 1378.24 1150.96 1175.43 1202.64 1213.3 1328.21 1382.28 1525.18 1622.82 2211.88
Stdev. 29.2 3.55 6.2 15.85 17.56 27.74 14.21 51.92 39.9 87.25
Rank 6 1 2 3 4 5 7 8 9 10
3 Best 897 776.6 792.4 803.8 823.6 908.4 NA NA 1126 NA
Average 917.18 786.67 804.72 822.62 847.28 939.46 923.5 991.14 1185.1 1184.24
Stdev. 16 5.31 6.63 22.53 15.12 16.75 15.3 30.71 35 60.31
Rank 5 1 2 3 4 7 6 8 10 9
4 Best 1556 1176 1187.4 1228.2 1274.6 1334.4 NA NA 1633.8 NA
Average 1598.58 1190.58 1219.9 1251 1310.4 1392.37 1624.04 1742.42 1749.8 3663.42
Stdev. 34.1 4.34 17.69 12.28 22.09 41.69 38.18 34.43 67.58 163.45
Rank 6 1 2 3 4 5 7 8 9 10
5 Best 687.2 625.6 634.4 639.2 640.8 668 NA NA 749.4
Average 710.66 631.87 635.81 642.88 648.08 680.59 661.6 684.48 771.44 714.16
Stdev. 10.4 1.78 1.2 4.77 5.08 8.4 4.45 9.11 16.87 6.23
Rank 8 1 2 3 4 6 5 7 10 9
6 Best 937.6 801.2 806.8 825.8 835.8 916.6 NA NA 1073.6 NA
Average 970.76 811.18 815.41 836 855.56 935.24 955.04 1001.8 1137.34 1188.1
Stdev. 18.8 3.07 5.35 8.87 14.11 17.19 20.1 32.99 43.89 58.1
Rank 7 1 2 3 4 5 6 8 9 10
Average (Rank) 6.33 1.00 2.00 3.00 4.00 5.50 6.33 7.83 9.33 9.67

each algorithm of the 10 independent runs for each problem instance. Note that the
Best results obtained by some comparative methods are not recorded, namely: HH-IE,
HH-SA, and TS. These are marked in Table 5 as "NA" (Not Available). In addition,
the smallest value for the best and average results is highlighted in bold font. The re-
sults presented in Table 5 prove that the proposed method is able to obtain competitive
results when compared to all other state-of-the-art algorithms in terms of the best and
average. Based on the average criterion, the proposed method is almost ranked in the
sixth place according to average ranking for all problem instances. Results prove that
the proposed method (MBBO-GBS) can be considered as an alternative algorithm for
PAS which is able to produce comparable results to the ones available in the literature.

6. Conclusion and Future Work

The patient admission scheduling problem deals with building up the hospital plan
schedule that share the night and the room (bed) for each patient over a given plan
horizon. Admission units at hospitals spend a lot of time and effort to achieve feasi-
ble resource utilization in a specific time slot. Therefore, finding a good solution by
developing a new algorithm is very helpful and beneficial. BBO is a population-based
meta-heuristic algorithm that has been created in 2008. BBO is inspired by the science

15
of biogeography, which uses the idea of the migration strategy of species for solving
optimization problems. In this research, we proposed a new modification on the BBO
algorithm (called MBBO-GBS). At the end of each iteration in the original BBO, the
best found solution replaces the worst one. This operation can lead to the loss of pop-
ulation diversity. In this research, we have modified the updating of solutions in the
population where the best solution replaces the solution with same ID rather than the
worst solution. This updating mechanism prevents having multiple copies of the best
solution during the search process as well as maintains the diversity in the population.
The performance of the approach is rigorously tested on the original PASP data set.
Experimental results show that MBBO-GBS outperformed other versions of BBO and
was able to produce favourable results in comparison to state-of-the-art. However, the
proposed approach shows that it has reached the stagnant state at an early point of the
search. We believe that this needs further investigations so this will be the subject of
our future work.

References

[1] J. M. Vissers, I. J. Adan, N. P. Dellaert, Developing a platform for comparison

of hospital admission systems: An illustration, European Journal of Operational
Research 180 (3) (2007) 1290–1301 (2007).
[2] B. Bilgin, P. Demeester, G. Vanden Berghe, A hyperheuristic approach to the
patient admission scheduling problem, Techical Report, KaHo Sint-Lieven, Gent
(2008).

[3] N. Ayvaz, W. T. Huh, Allocation of hospital capacity to multiple types of patients,

Journal of Revenue and Pricing Management 9 (5) (2010) 386–398 (2010).
[4] A. L. Bolaji, A. F. Bamigbola, P. B. Shola, Late acceptance hill climbing algo-
rithm for solving patient admission scheduling problem, Knowledge-Based Sys-
tems 145 (2018) 197–206 (2018).

[5] P. Demeester, W. Souffriau, P. De Causmaecker, G. V. Berghe, A hybrid tabu

search algorithm for automatically assigning patients to beds, Artificial Intelli-
gence in Medicine 48 (1) (2010) 61–70 (2010).
[6] B. Bilgin, P. Demeester, M. Misir, W. Vancroonenburg, G. V. Berghe, One hyper-
heuristic approach to two timetabling problems in health care, Journal of Heuris-
tics 18 (3) (2012) 401–434 (2012).
[7] S. Ceschia, A. Schaerf, Multi-neighborhood local search for the patient admission
problem, in: International Workshop on Hybrid Metaheuristics, Springer, 2009,
pp. 156–170 (2009).

[8] S. Ceschia, A. Schaerf, Modeling and solving the dynamic patient admission
scheduling problem under uncertainty, Artificial intelligence in medicine 56 (3)
(2012) 199–205 (2012).

16
 [9] R. Guido, V. Solina, D. Conforti, Offline patient admission scheduling problems,
in: International Conference on Optimization and Decision Science, Springer,
2017, pp. 129–137 (2017).
[10] S. Ceschia, A. Schaerf, Dynamic patient admission scheduling with operating
room constraints, flexible horizons, and patient delays, Journal of Scheduling
19 (4) (2016) 377–389 (2016).
[11] R. Guido, V. Solina, G. Mirabelli, D. Conforti, Offline patient admission, room
and surgery scheduling problems, in: New Trends in Emerging Complex Real
Life Problems, Springer, 2018, pp. 275–283 (2018).

[12] Y.-H. Zhu, T. A. Toffolo, W. Vancroonenburg, G. V. Berghe, Compatibility of

short and long term objectives for dynamic patient admission scheduling, Com-
puters & Operations Research 104 (2019) 98–112 (2019).
[13] K. Dorgham, I. Nouaouri, H. Ben-Romdhane, S. Krichen, A hybrid simulated
annealing approach for the patient bed assignment problem, Procedia Computer
Science 159 (2019) 408 – 417, knowledge-Based and Intelligent Information En-
gineering Systems: Proceedings of the 23rd International Conference KES2019
(2019). doi:https://doi.org/10.1016/j.procs.2019.09.195.
URL http://www.sciencedirect.com/science/article/pii/
S1877050919313778
[14] R. Guido, M. C. Groccia, D. Conforti, An efficient matheuristic for
offline patient-to-bed assignment problems, European Journal of Op-
erational Research 268 (2) (2018) 486 – 503 (2018). doi:https:
//doi.org/10.1016/j.ejor.2018.02.007.
URL http://www.sciencedirect.com/science/article/pii/
S0377221718301218

[15] L. S. Bastos, J. F. Marchesi, S. Hamacher, J. L. Fleck, A mixed integer program-

ming approach to the patient admission scheduling problem, European Journal of
Operational Research 273 (3) (2019) 831–840 (2019).
[16] A. I. HAMMOURI, B. ALRIFAI, Investigating biogeography-based optimisation
for patient admission scheduling problems, Journal of Theoretical & Applied In-
formation Technology 70 (3) (2014).
[17] A. Hammouri, M. Alweshah, E. Hezzam, M. Asmaran, Biogeography based op-
timization with guided bed selection mechanism for patient admission scheduling
problems, International Journal of Soft Computing 12 (2) (2017).

[18] T. M. Range, R. M. Lusby, J. Larsen, A column generation approach for solv-

ing the patient admission scheduling problem, European Journal of Operational
Research 235 (1) (2014) 252–264 (2014).
[19] S. Kifah, S. Abdullah, An adaptive non-linear great deluge algorithm for the
patient-admission problem, Information Sciences 295 (2015) 573 – 585 (2015).

17
[20] A. M. Turhan, B. Bilgen, Mixed integer programming based heuristics for the
patient admission scheduling problem, Computers & Operations Research 80
(2017) 38–49 (2017).
[21] R. M. Lusby, M. Schwierz, T. M. Range, J. Larsen, An adaptive large neigh-
borhood search procedure applied to the dynamic patient admission scheduling
problem, Artificial intelligence in medicine 74 (2016) 21–31 (2016).
[22] K. Anwar, M. A. Awadallah, A. T. Khader, M. A. Al-Betar, Hyper-heuristic ap-
proach for solving nurse rostering problem, in: 2014 IEEE symposium on com-
putational intelligence in ensemble learning (CIEL), IEEE, 2014, pp. 1–6 (2014).
[23] D. Simon, Biogeography-based optimization, IEEE transactions on evolutionary
computation 12 (6) (2008) 702–713 (2008).
[24] M. Alweshah, A. I. Hammouri, S. Tedmori, Biogeography-based optimisation for
data classification problems, International Journal of Data Mining, Modelling and
Management 9 (2) (2017) 142–162 (2017).
[25] M. Zhang, W. Jiang, X. Zhou, Y. Xue, S. Chen, A hybrid biogeography-based op-
timization and fuzzy c-means algorithm for image segmentation, Soft computing
23 (6) (2019) 2033–2046 (2019).
[26] F. Zhao, S. Qin, Y. Zhang, W. Ma, C. Zhang, H. Song, A hybrid biogeography-
based optimization with variable neighborhood search mechanism for no-wait
flow shop scheduling problem, Expert Systems with Applications 126 (2019)
321–339 (2019).
[27] X. Zhang, Q. Kang, Q. Tu, J. Cheng, X. Wang, Efficient and merged
biogeography-based optimization algorithm for global optimization problems,
Soft Computing 23 (12) (2019) 4483–4502 (2019).
[28] G. Xiong, D. Shi, Hybrid biogeography-based optimization with brain storm op-
timization for non-convex dynamic economic dispatch with valve-point effects,
Energy 157 (2018) 424–435 (2018).
[29] X.-w. Zheng, D.-j. Lu, X.-g. Wang, H. Liu, A cooperative coevolution-
ary biogeography-based optimizer, Applied Intelligence 43 (1) (2015) 95–111
(2015).
[30] R. Zi-wu, Z. Qiu-guo, Hybrid algorithm based on biogeography-based optimiza-
tion and differential evolution for global optimization, in: 2014 9th IEEE Confer-
ence on Industrial Electronics and Applications, IEEE, 2014, pp. 754–758 (2014).
[31] S. Ceschia, A. Schaerf, Local search and lower bounds for the patient admission
scheduling problem, Computers & Operations Research 38 (10) (2011) 1452–
1463 (2011).
[32] I. A. Doush, M. A. Al-Betar, M. A. Awadallah, A. I. Hammouri, M. Ra’ed,
S. ElMustafa, H. ALkhraisat, Harmony search algorithm for patient admission
scheduling problem, Journal of Intelligent Systems (2018).

18
[33] H. Garg, An efficient biogeography based optimization algorithm for solving re-
liability optimization problems, Swarm and Evolutionary Computation 24 (2015)
1–10 (2015).
[34] V. Garg, K. Deep, Performance of laplacian biogeography-based optimization al-
gorithm on cec 2014 continuous optimization benchmarks and camera calibration
problem, Swarm and Evolutionary Computation 27 (2016) 132–144 (2016).
[35] A. H. Niknamfar, S. T. A. Niaki, S. A. A. Niaki, Opposition-based learning for
competitive hub location: A bi-objective biogeography-based optimization algo-
rithm, Knowledge-Based Systems 128 (2017) 1–19 (2017).

19