School Bus Routing using Harmony Search

Zong Woo Geem

Johns Hopkins University
729 Fallsgrove Drive #6133
Rockville, MD 20850
+1-301-294-3893

[email protected]

ABSTRACT
A recently-developed nature-inspired algorithm, Harmony Search,
mimicking music improvisation, is introduced and applied to
transportation problem (school bus routing), and compared with
popular evolutionary algorithm (genetic algorithm). The Harmony
Search is conceptualized using the musical process of searching
for a perfect state of harmony. This algorithm was applied to a
test network consisting of one bus depot, one school and ten bus
stops with demand by commuting students. This school bus
routing example is a multi-objective problem to minimize the
number of operating buses and also the travel time of all buses,
with bus capacity and time window constraints that are considered
as penalty costs. Harmony Search could find good solution within
the reasonable time with other advantages such as no derivative
requirement and no initial value assumption. The presented
routing model is expected to be applied to large-scale real
networks in the future.

Categories and Subject Descriptors

J.2 [Computer Applications]: Physical Science s and Engineering
Engineering

General Terms
Algorithms, Design, Economics

Keywords
Harmony Search, School Bus Routing, Optimization

1. INTRODUCTION
Transportation researchers and professionals are sometimes coped
with situations where optimal decisions need to be made.
Traditionally various mathematical techniques have been used for
supporting these optimal decisions. However, their computational
disadvantages such as requiring derivative information, initial
value assumption, or huge amount of computation and memory,
made them rely on another type of methodology, that is,
evolutionary or meta-heuristic algorithms.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
GECCO 2005, June 2529, 2005, Washington, DC, USA.
Copyright 2004 ACM 1-58113-000-0/00/0004$5.00.

The basic ideas of existing evolutionary and meta-heuristic

algorithms are motivated by natural phenomena. For example,
the evolutionary algorithms [1-3] and the genetic algorithm [4, 5]
are inspired by biological evolutionary process; tabu search [6]
and ant algorithm [7] from animal's behavior; and simulated
annealing [8] from physical annealing process.
Harmony Search (HS) algorithm was recently developed in an
analogy with music improvisation process where musicians in an
ensemble adjust the pitches of their instruments in order to obtain
perfect harmony [9, 10]. The HS algorithm has been successfully
applied to both benchmarking problems and real-world problems
such as the traveling salesperson problem [9], Rosenbrock's
function [10], optimal design of pipeline network [11], parameter
calibration of hydrologic model [12], and optimal design of truss
structures [13, 14]. Consequently, the HS algorithm provides a
possibility of success in an optimization problem in transportation
research field.

2. HARMONY SEARCH ALGORITHM

Adopting the idea that existing evolutionary or meta-heuristic
algorithms are found in the paradigm of natural processes, a new
algorithm can be conceptualized from a musical performance
process (say, a jazz trio) involving searching for a better harmony.
Musical performance seeks a best state (fantastic harmony)
determined by aesthetic estimation, as the optimization process
seeks a best state (global optimum: minimum cost; minimum
error; maximum benefit; or maximum efficiency) determined by
objective function evaluation. Aesthetic estimation is done by the
set of the pitches sounded by joined instruments, as objective
function evaluation is done by the set of the values produced by
composed variables; the aesthetic sounds can be improved
practice after practice, as the objective function values can be
improved iteration by iteration.
Figure 1 shows the structure of the Harmony Memory (HM) that
is the core part of the HS algorithm. Consider a jazz trio
composed of saxophone, double bass, and guitar. There exist
certain amount of preferable pitches in each musician's memory:
saxophonist, {Do, Fa, Mi, Sol, Re}; double bassist, {Si, Do, Si,
Re, Sol}; and guitarist, {La, Sol, Fa, Mi, Do}. If saxophonist
randomly plays {Sol} out of its memory {Do, Fa, Mi, Sol, Re},
double bassist {Si} out of {Si, Do, Si, Re, Sol}, and guitarist
{Do} out of {La, Sol, Fa, Mi, Do}, the new harmony (Sol, Si, Do)
becomes another harmony (musically C-7 chord). And if this new
harmony is better than existing worst harmony in the HM, the
new harmony is included in the HM and the worst harmony is

excluded from the HM. This procedure is repeated until fantastic

harmony is found.

Subject to

xi X i , i = 1,2,..., N

f (x)

is an objective function;

Where

decision variable
for

each

xi ; X i

(2)

is the set of each

is the set of possible range of values

decision

variable,

that

is,

X i = {xi (1), xi (2),..., xi ( K )} for discrete decision variables

(xi (1) < xi (2) < ... < xi ( K ) ) ; N is the number of
decision variables (number of music instruments); and K is the

number of possible values for the discrete variables (pitch range

of each instrument).
The HS algorithm parameters are also specified in this step:
Harmony Memory Size (HMS) (= number of solution vectors),
Harmony Memory Considering Rate (HMCR), Pitch Adjusting
Rate (PAR), and Stopping Criteria (= number of improvisation).
Here, HMCR and PAR are the parameters of HS algorithm
explained in Step 3.
Figure 1. Structure of Harmony Memory

2.2 Initialize the Harmony Memory

In real optimization, each musician can be replaced with each
decision variable, and its preferred sound pitches can be replaced
with each variable's preferred values. Let us set that each decision
variable represents pipe diameter between two nodes and the
music pitches {Do, Re, Mi, Fa, Sol, La, Si} correspond to pipe
diameters {100mm, 200mm, 300mm, 400mm, 500mm, 600mm,
700mm}. And if first variable chooses {500mm} out of {100mm,
400mm, 300mm, 500mm, 200mm}, second one {700mm} out of
{700mm, 100mm, 700mm, 200mm, 500mm}, and third one
{100mm} out of {600mm, 500mm, 400mm, 300mm, 100mm},
those values (500mm, 700mm, 100mm) make another solution
vector. And if this new vector is better than existing worst vector
in the HM, the new vector is included in the HM and the worst
vector is excluded from the HM. This procedure is repeated until
certain stopping criterion is satisfied.
According to the above algorithm concept, the steps in the
procedure of HS for the school bus routing problem are as
follows:

In Step 2, the Harmony Memory (HM) matrix, as shown in

Equation 3, is filled with as many randomly generated solution
vectors as the size of the HM (i.e., HMS) and sorted by the values
of the objective function, f (x) .

x1
2
x

HM =
...

HMS
x

(3)

2.3 Improvise a New Harmony

A new harmony vector, x = ( x1 , x 2 ,..., x N ) is generated by

anyone of following three actions: HM consideration; Pitch
adjustment; or totally random generation. For instance, the value

x1 )

Step 1. Initialize the Problem and Algorithm Parameters.

of the first decision variable (

Step 2. Initialize the Harmony Memory (HM).

chosen from anyone of values stored in HM (

for the new vector can be

x11 ~ x1HMS ).

Step 3. Improvise a new harmony from the HM.

Values of other variables ( xi ) can be chosen in the same manner.

Step 4. Update the HM.

Here, there is a possibility that totally random value can be

chosen using the HMCR parameter, which varies between 0 and 1
as follows:

Step 5. Repeat Step 3 and 4 under the stopping criterion.

2.1 Initialize the Parameters

In Step 1, the optimization problem is specified as follows:
Minimize

f (x)

(1)

x {xi1 , xi2 , ..., xiHMS } w.p.

HMCR
xi i
xi X i
w.p. (1 - HMCR)

(4)

The HMCR sets the rate of choosing one value from the historical
values stored in the HM, and (1-HMCR) sets the rate of randomly
choosing one value from the possible range of values.
Next,

each

component

of

the

new

harmony

vector,

x = ( x1 , x 2 ,..., x N ) is examined whether it should be pitch-

adjusted. This procedure uses the PAR parameter that sets the rate
of adjustment for the pitch from the HM as follows:

Pitch adjusting for

PAR
Yes w.p.
xi
No w.p. (1 PAR)

(5)

The pitch adjusting process is performed only after a value is

chosen from the HM. The value (1-PAR) sets the rate of doing
nothing. If the pitch adjustment decision for

xi (k ) (the
adjusted value of xi ( k ) is
assumed to be

xi xi (k + m)

xi

is Yes and

xi

is

Xi

), the pitch-

for discrete decision variables

(6)

kth element in

m
is
the
m {..., 2, 1, 1, 2, ...} .

Where

neighboring

index,

Pattnaik et al. [19] proposed an optimization model of minimizing

overall cost (both the operator's and user's) while determining a
route configuration with a set of urban bus transit routes and
associated frequencies using GA. In their approach, a set of
candidate routes is first generated, and then GA is employed to
find the optimal one. Chien et al. [20] introduced a GA model to
the optimization of bus route and the corresponding headway
while minimizing total cost, subject to geography, capacity, and
budget constraints. Their results were validated by comparing
with those obtained from exhaustive search algorithms. Jung and
Haghani [21] used GA to solve a multi-vehicle pickup and
delivery problem with time window. They formulated the
problem as a mixed-integer linear program, with an objective
function to minimize the total cost, consisting of the fixed cost of
the vehicles, routing cost, and customer inconvenient cost. The
proposed GA scheme can solve a pickup and delivery problem
within an extremely short time compared with the B&B method.

4. PROBLEM FORMULATION

The HMCR and PAR parameters introduced in the Harmony

Search help the algorithm find globally and locally improved
solution, respectively.

2.4 Update the Harmony Memory

If the new harmony vector, x = ( x1 , x 2 ,..., x N )

The intractability of the SBRP depends on how the problem is

formulated and how many constraints are incorporated.
Generally, the wider the problem scope is, the harder the problem
becomes, especially to the exact methods such as the branch and
bound method (B&B). Because an exact method may take a very
long time to obtain the optimal solution, people have turned to
evolutionary or meta-heuristic algorithms such as genetic
algorithms (GA) that do not necessarily find the optimal solution,
but tend to find good solutions within a reasonable amount of
time [18]. The VRPs that the evolutionary or meta-heuristic
algorithms had success in applying include transit routing
problem and pickup and delivery problem.

is better than

the worst harmony in the HM, judged in terms of the objective

function value, the new harmony is included in the HM and the
existing worst harmony is excluded from the HM. The HM is then
sorted by the objective function value.

School bus routing in this study is a simple multi-objective

problem to minimize the number of operating buses and the travel
time of all buses, with two major constraints (bus capacity and
time window). The study network to be optimized consists of one
bus depot, one school, and ten bus stops as shown in Figure 2.
Each stop is demanded by certain amount of commuting students,
and travel time is specified between two stops.

2.5 Repeat Step 3 and 4

Step 3 and 4 are repeated until the termination criterion (for
example, the number of maximum improvisation) is satisfied.

3. SCHOOL BUS ROUTING PROBLEM

In order to demonstrate the searching ability of the Harmony
Search, the HS is applied to a school bus routing problem
(SBRP). The SBRP is a practical optimization problem that is
closely related to many people's daily lives. From a school's
perspective, the SBRP aims to supply students with an efficient
and equitable transportation service [15]. Many approaches [1517] have been developed to solve the SBRP. This problem falls
into a large class of problems called vehicle routing problems
(VRPs), in which a set of vehicles provides pickup, delivery or
simply a service to customers dispersed in an area.

Figure 2. Study Network of School Bus Routing

DNED = union
( DN ED )

4.1 Decision Variables

xi

= decision variable having served bus

for demand node i ,

VS

set of demand nodes and ending node

= vehicle set

i DN , k VS
x = vector of decision variables xi ,

nbus(x)
lkijk

i DN

= number of operating buses

4.4 Formulation
Minimize

1 if bus k travels node i and node j

,
otherwise
0

f (x) = fc nbus (x) + rc

k iSTDN jDNED

k VS , i STDN , j DNED

1 time window violation in bus k

, k VS
otherwise
0

4.2 Parameters

fc = fixed cost per school bus

rc = routing cost per moving time
spij

(7)
subject to

nbus(x) nset (VS )

(8)

DM

(9)

k
i

BC k , k VS

iSTDN jDNED

pc1 = penalty cost for bus capacity violation

pc 2

spij lk ijk + DM ik bt TW k , k VS
i

(10)

= penalty cost for time window violation

nset (VS )
k
i

= number of elements in set

VS

= number of boarding students in node

i by bus k ,

BC k = bus capacity of bus k , k VS

bt = boarding time per student
= time window of bus k ,

xi to each bus stop

bus k ( k VS ) that

By assigning a discrete decision variable

i DN , k VS

TW k

k
ij

= shortest path between node i and node

DM

vtmk

ij

+ pc1 vcp k + pc 2 vtm k

1 bus capacity violation in bus k

, k VS
otherwise
0

vcp k

sp lk

k VS

(demand node) i to denote the specific

serves the stop, the objective function of the problem is to
minimize both the number of operating buses and the moving
time of the buses as first and second terms in Equation 7. The
third and fourth terms represent penalty costs for the violation of
bus capacity and time window, respectively. Where the fixed cost
per school bus fc is assumed as $100,000/bus; routing cost per

rc is $105/min; shortest path between node i and

j , spij (in minutes), is calculated by Floyd and Warshall's
algorithm; connection status between node i and node j for bus
k , lkijk , has 1 when xi ( i STDN ) and x j ( j DNED )
have bus k ; penalty cost for bus capacity violation pc1 is
$100,000; and penalty cost for time window violation pc 2 is
moving time
node

4.3 Sets
DN
ST

= demand nodes (= bus stops)

= starting node (= bus depot)

ED = ending node (= school)

STDN = union set of starting
( ST DN )

$100,000 for the computation.

node and demand nodes

Equation 8 is the maximum operating bus number constraint. The

number of operating buses is less than or equal to the number of
candidate buses (= 4 buses). Equation 9 shows that the number of
boarding students in a bus is less than or equal to the bus capacity

(= 45 students for each bus). Equation 10 shows that the travel

time of a bus is less than or equal to the time window (= 32
minutes for each bus). Travel (in-vehicle) time in this work
consists of moving time and boarding time, where each student's
barding time bt is 6 seconds.

routing costs and computing times are $399,870 and 6.6 seconds
in HS, and $409,597 and 6.7 seconds in GA on Pentium II 233
MHz.

Table 1. Computational Results of Harmony Search

5. COMPUTATIONAL RESULTS
In order to apply the HS algorithm to the school bus routing
problem, parameters of HS algorithm are specified such as the
number of musical instruments (= 10 of demand nodes), pitch
range of each instrument (= {bus 1, bus 2, bus 3, bus 4}),
Harmony Memory Size (HMS) (= 10 ~ 100), Harmony Memory
Considering Rate (HMCR) (= 0.3 ~ 0.95), and stopping criteria (=
1000 improvisation). Next, harmonies (solution vectors) are
randomly generated from the possible range as many as HMS and
sorted by objective function value.
In Step 3, a new harmony is generated from the HM. For instance,
the bus of the first demand node in the new vector can be chosen
any bus out of the stored buses (for example, {bus3, bus1, bus 2,
bus 2, bus 3, bus 1, bus 4, bus 2, bus 4, bus 2}) of the first
demand node in HM. The buses of other demand nodes can be
chosen in the same manner. On the other hand, in smaller
possibility (1-HMCR), a bus can be chosen from all the possible
range {bus 1, bus 2, bus 3, bus 4}.
The new harmony x is put in the objective function to obtain
total cost which consists of fixed bus cost, bus moving cost, and
two penalty costs. If bus k violates equation 9, the variable for

HMCR

0.3

0.5

0.7

0.9

0.95

10

410185

410290

407665

410185

410500

20

410185

410185

410185

307980

307980

40

410185

410395

410185

410185

410185

100

410185

410185

410185

410185

410185

HMS

Table 2. Computational Results of Genetic Algorithm

MR

0.01

0.03

0.05

0.07

0.1

10

410395

509240

410185

410185

407350

20

307980

410185

406930

410185

410185

40

410290

410185

410185

410185

406930

100

410395

410290

410185

410185

410290

PS

bus capacity violation, vcp , becomes 1, and penalty cost for the
capacity violation is added. If bus k violates equation 10, the
k

variable for time window violation, vtm , becomes 1, and

penalty cost for the time window violation is added.
In Step 4, if the cost of the new harmony is better than the worst
cost of any harmony in the HM, the new harmony is included in
the HM, and the existing worst cost harmony is excluded from the
HM. After that, the HM is sorted.
Finally, in Step 5, the computation is terminated when the
stopping criteria is satisfied. If not, Step 3 and 4 are repeated.
The HS is computed for the school bus routing problem and its
results are examined by comparing with those of popular
evolutionary technique, genetic algorithm (GA). In order to fairly
compare HS with GA, the number of objective function
evaluations and the number of computational runs are same in
both algorithms: in HS (Table 1), the number of function
evaluations is 1,000 (= number of improvisation) and the number
of runs is 20 with various HMS (10 ~ 100) and HMCR (0.3 ~
0.95) which parameter values are frequently used in previous HS
applications; and in GA (Table 2), the number of evaluation is
1,000 (= population size number of generations) and the number
of runs is 20 with various population size (PS, 10 ~ 100), and
mutation rate (MR, 0.01 ~ 0.1), recommended by Koumousis and
Georgiou [22].
From the results presented in Table 1 and 2, both algorithms could
find the best solution ($307,980) that is the optimal solution of the
study network, demonstrated by an exact algorithm. The average

Table 3 shows the route, number of commuting students, and

travel time of each bus in optimal solution ($307,980) and nearoptimal solution ($410,185). "Do Nothing" in 5th row of the table
means that the bus is not operated while satisfying all the
constraints.

Table 3. Results of School Bus Routing Problem

Routing

Bus

Cost ($)

Routes

# of
Students

Travel
Time
(min)

Depot 8 9 10
School

45

31.5

Depot 4 5 6 School

45

28.5

Depot 1 2 3 7
School

40

29.0

307,980

Do Nothing

Depot 2 6 School

35

25.5

Depot 1 3 7 School

25

27.5

Depot 5 9 10 School

45

27.5

Depot 4 8 School

25

29.5

410,185

6. CONCLUSIONS
A newly developed algorithm, Harmony Search is modeled for
solving the school bus routing problem, and the results of HS
computation are compared with those of genetic algorithm. HS
mimics musician's behaviors in music improvisation process.
Musician's behaviors such as memory considering, pitch adjusting,
and random choosing are effectively translated as local and global
solution search schemes.

[9] Geem, Z. W., Kim, J. H., and Loganathan, G. V. A New

Heuristic Optimization Algorithm: Harmony Search.
Simulation, 76, 2 (2001), 60-68.
[10] Geem, Z. W., and Tseng, C. L. New Methodology,
Harmony Search and its Robustness. In Late-Breaking
Papers of Genetic and Evolutionary Computation
Conference (GECCO 02) (New York City, USA, July 9-13,
2002), 174-178.

The proposed HS model for the school bus routing problem is to

minimize the multi-objective function, consisting of the number
of operating buses, the travel time of all buses, and penalty costs
related with bus capacity and time window violations. HS could
find the optimal solution or near-optimal solutions within
reasonable time with advantages including no derivative
information requirement, no initial value assumption, no huge
memory requirement, and alternative solutions.

[11] Geem, Z. W., Kim, J. H., and Loganathan. G. V. Harmony

Search Optimization: Application to Pipe Network Design.
International Journal of Modelling and Simulation, 22, 2
(2002), 125-133.

From the above-mentioned advantages, the HS algorithm appears

to be successfully applied to optimization problems in
transportation engineering field. Especially, the presented school
bus routing model is expected to be applied to large-scale real
networks interfaced with other data-supporting packages such as
geographical information system in the future.

[13] Kang, S. L, and Geem, Z. W. A New Structural Optimization

Method Based on the Harmony Search Algorithm.
Computers and Structures, 82, 9-10 (2004), 781-798.

7. REFERENCES

[15] Bowerman, R., Hall, B., and Calamai, P. A Multi-Objective

Optimization Approach to Urban School Bus Routing:
Formulation and Solution Method. Transportation Research
A, 29A, 2 (1995), 107-123.

[1] Fogel, L. J., Owens, A. J. and Walsh. M. J. Artificial

Intelligence Though Simulated Evolution. John Wiley,
Chichester, UK, 1966.
[2] De Jong, K. Analysis of the Behavior of a Class of Genetic
Adaptive Systems. Ph.D. Thesis, University of Michigan,
Ann Arbor, 1975.
[3] Koza, J. R. Genetic Programming: A Paradigm for
Genetically Breeding Populations of Computer Programs to
Solve Problems. Report No. STA-CS-90-1314, Stanford
University, 1990.
[4] Holland, J. H. Adaptation in Natural and Artificial Systems.
University of Michigan Press, Ann Arbor, 1975.
[5] Goldberg, D. E. Genetic Algorithms in Search Optimization
and Machine Learning. Addison Wesley, MA, 1989.
[6] Glover, F. Heuristic for Integer Programming using
Surrogate Constraints. Decision Sciences, 8, 1 (1977), 156166.
[7] Dorigo, M., Maniezzo, V., and Colorni, A. The Ant System:
Optimization by a Colony of Cooperating Agents. IEEE
Transactions on Systems, Man, and Cybernetics-Part B, 26,
1 (1996), 29-41.
[8] Kirkpatrick, S., Gelatt, C., and Vecchi, M. Optimization by
Simulated Annealing. Science, 220, 4598 (1983), 671-680.

[12] Kim, J. H., Geem, Z. W., and Kim, E. S. Parameter

Estimation of the Nonlinear Muskingum Model using
Harmony Search. Journal of the American Water Resources
Association, 37, 5 (2001), 1131-1138.

[14] Kang, S. L, and Geem, Z. W. A New Structural Optimization

Method for Structures with Discrete Sizing Variables. In
Proceedings of 2004 Conference of the Structural
Engineering Institute of ASCE (STRUCTURES 2004)
(Nashville, USA, May 22-26, 2004), CD-ROM.

[16] Bennett, B. T., and Gazis, D. C. School Bus Routing By

Computer. Transportation Research, 6, (1972), 317-325.
[17] Tsay, H. S., and Fricker, J. D. Practical Approach for
Solving School Bus Routing. Transportation Research
Record, TRB, 1202, (1988), 44-56.
[18] Bodin, L., Golden, B., Assad, A., and Ball, M. The State of
Art in the Routing and Scheduling of Vehicle and Crews.
Computers and Operations Research, 10, (1983), 63-211.
[19] Pattnaik, S. B., Mohan, S., and Tom, V. M. Urban Bus
Transit Route Network Design using Genetic Algorithm.
Journal of Transportation Engineering, ASCE, 124, 4 (1998),
368-375.
[20] Chien, S., Yang, Z., and Hou, E. Genetic Algorithm
Approach for Transit Route Planning and Design. Journal of
Transportation Engineering, ASCE, 127, 3 (2001), 200-207.
[21] Jung, S., and Haghani, A. Genetic Algorithm for a Pickup
and Delivery Problem with Time Windows. Transportation
Research Record, TRB, 1733, (2000), 1- 7.
[22] Koumousis, V. K., and Georgiou, P. G. Genetic Algorithms
in Discrete Optimization of Steel Truss Roofs. Journal of
Computing in Civil Engineering, ASCE, 8, 3 (1994), 309325.