IEEE—ICET 2006

2nd International Conference on Emerging Technologies

Peshawar, Pakistan 13-14 November 2006

An Intelligent Search Technique to Train

Scheduling Problem Based on Genetic Algorithm
1
Muhammad Babar Khan, 2 Dianye Zhang, 3Ming shi Jun, 4Zhu Jiang Li
[1] [2] [3] [4]
College of Traffic and Transportation, Southwest Jiaotong University,
Chengdu 610031, P. R. of China
Email: 1 [email protected]

Abstract: Genetic algorithms are generally used to occur and on which track each train should be at
solve NP-complete problems, which include the train any given time to achieve certain predetermined
scheduling problem. Train scheduling consists of objectives such as minimum overall train delay,
finding optimal arrival, departure, and dwelling and safety of train movement. Given the above
times at the train stations. The objective is to
stated difficulties, we develop a genetic
minimize the associated delay that is the waiting
times at the station. In this paper, we present an algorithm, which can solve dispatching problem
application of genetic algorithms to a generic train by producing “good enough” solutions.
scheduling problem. The necessary algorithmic This paper presents the application of genetic
solution is described together with some empirical
algorithm (GA) to minimize delays at the
observations about the performance of our train
scheduling model. It is concluded that our model is intermediate and final train stations. The term
able to produce useful results in terms of minimizing delay describes the deviation of trains from its
delays and producing optimal schedules. scheduled departure and arrival times. The train
dispatching model is formulated subject to
1. INTRODUCTION operational constraints for safe operations of
trains between the stations. It is concluded that
One of the most challenging and difficult the model is able to produce useful results in
problems in dispatching trains is to resolve train terms of optimal schedules in reasonable time
conflicts. Conflicts arise due to deviation from for the test applications shown here. The article
the pre-established time schedules. Most of the is organized as follows.
railway systems rely on the established
schedules, which are made by the planners after First the problems in the optimal train
every two to three years. Any disruption of dispatching are presented. Section 2 focuses on
railway traffic in such an environment may the recent and old work done to solve train
cause the system to be unstable with a high scheduling problems. The basic formulation of
operating cost due to increasing interference our model is highlighted in section 3 and its GA
delays. Therefore, conflict has to be resolved based solution is given in section 4.
timely and effectively to minimize the losses. Subsequently, section 5 and 6 discuss the
Optimizing train schedules is a combinatorial implementation of results and conclusions
problem, which is NP-complete with time respectively.
complexity of O(2n), where n is the number of
conflicts [8]. For example, a problem instance 2. LITERATURE REVIEW
with 30 meets has 230 approximately 1 billion
possible solutions while an instance with 40 In the recent years, the complexities in the
meets has 240 approximately 1 trillion possible railway system are growing and numbers of
solutions. Therefore, emphasis is given finding researchers are focusing on optimal train
efficient heuristics in order to find the solution scheduling problems. The effort is mainly spent
in polynomial times. This means that the on developing mathematical models and to
computation time and memory requirement to obtain the solution in the desired time limits.
solve such a problem increases exponentially The techniques used to realize the models can be
with the problem size [2]. The optimum characterized into three groups: Mathematical
dispatching problem determines, by solving programming, Simulation techniques and
these conflicts, where meets and passes should Knowledge-based systems. In early seventies,

1-4244-0502-5/06/$20.00©2006 IEEE 593

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on June 01,2021 at 01:05:10 UTC from IEEE Xplore. Restrictions apply.
 Szpigel [10] developed the linear programming
model to determine the over taking and crossing , (1)
positions of trains provided the departure times Minimum dwell time at stations: The minimum
and upper velocities. Suader et. al, [9] proposed time for train q staying at stations is to
a decision support system for train dispatching allow for boarding and alighting, etc. Dwell time
that enumerates all feasible meet locations and for every station may not be the same. Thus if
selects the one, which minimizes total delay.
Peterson and Taylor [7] developed a dispatch train q arrives at station s at time it can
algorithm which calculates the crosses, transit
times of segments and determines which trains depart after the time .
takes the siding in order to minimize the total
travel times. Kraft et al [9] used different
approach and gave optimal dispatching of trains (2)
based on train priorities, delay penalty and track
running times of the trains. Branch and bound
approach was used to resolve the conflicts.
and ,
Kraay et al [6] proposed mathematical model
based on the conservation of fuel and to
Minimum headways: The safety time interval
minimize the delay of trains. Heuristics were
required for safe train operation, i.e. headways
used to find the solution of realistic size
has to be not less than the minimum headways.
problems. Jovanoic and Harker [4] formulated a
Headways are ensured between the trains during
model “SCAN” named as schedule analysis
meet and pass on entering or exiting the links.
based on combinatorial optimization and
We set out 0-1 integer variables.
simulation.

3. THE BASIC MODEL hpqs = 1 if train p immediately precedes

train q on link s,
This section deals with the definition and
derivation of the model. The set of trains is hpqs = 0 otherwise
given by T in which p and q denote trains. Let
l denote the link joining the stations S and Tl Entering link headways: Minimum headways
be the set of trains use link l , shown in Fig. 1. for the trains p and q while entering the link s
is described. V Aenter is the headway required for
the two trains. The entry departure times
( Aps and Aqs ) of trains p and q on link s are
given as,
Fig.1. Illustration of Railway Line with Stations

Constraints are formulated for the safe trains M (1 − hpqs ) + Aqs ≥ Aps + V Aenter (3)
operation on single track and to avoid the
conflicts between the trains traversing in the Constant M is chosen large enough so that
same direction and in opposite direction. equation constraints are satisfied. The term
Running time constraints: The running time of M (1 − hpqs ) = 0, if hpqs = 1 which makes sure
train q on link l is adjusted as the time of exit
that the constraint is necessary if and only if,
form the link ( X qs ) must be greater or equal to p follows q on link s.
the time of entry (arrival) to the link ( Aqs ) plus
the minimum time needed to traverse the link Exiting link headways: As above, to ensure the
minimum headway between train p and q on
( V t min ).
exiting from link s .

594

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on June 01,2021 at 01:05:10 UTC from IEEE Xplore. Restrictions apply.
 M (1 − hpqs ) + X qs ≥ X ps + V X exit (4) ∑∑ ⎡⎣δ
q∈T s∈S
A
qs ( Aqs ) +δ X qs ( X qs ) + δ t qs (rqs ) + δ w qs ( wqs ) ⎤⎦
Usually there are bounds on the desirable
departure and arrival times of some trains or all (7)
trains at initial and final stations. Considering all the constraints and cost terms
defined above the train scheduling problem can
be stated as,
Aqs ≥ Aqs ≥ Aqs (5) Minimize P, equation (7), by choosing the values
( )
of X qs , Aq , s , hpqs , subject to (1) - (7), hpqs = 0 or
X qs ≥ X qs ≥ X qs (6)
1, for all p, q ∈ Tl , s ∈ S .

Where Aqs and Aqs are the earliest and the latest 4. SOLVING THE OPTIMAL TRAIN
arrival time, X qs and X qs are the earliest and the SCHEDULING PROBLEM
In literature, train scheduling problem has been
latest exit time of train q on link s respectively.
formulated in different ways, one is a job-shop
Then the earliest exit time from link s scheduling. For example, each segment of the
is X qs = Aq , s + V t min , and the earliest departure train paths is regarded as working procedure in
job-shop and each section of working line as
time on link s + 1 is Aq , s +1 = X qs + V w dwell , and
working machine. Train scheduling is also
so on. Similarly, we can work forward to represented as a network flow problem and
compute bounds for all later stations. The latest mainly as a mixed integer-programming
exit X qs and latest Aqs arrival time is calculated problem. For the details refer to the survey paper
[1]. The later formulation drew much attention
in the same manner explained above.
and has been thoroughly studied by many
researchers. All the solutions to this mixed
Objective function: The objective function
integer-programming problem are the variants of
includes solving the conflicts between trains in
branch and bound method, usually combined
order to minimize the travel times, dwell time at
with simulations. A small example is given to
stations. Trains are scheduled with a
illustrate the procedure. Four trains ( p, q , m, n )
prespecified departure time given with some
flexibility in the time table. The deviation of traversing between station 1 and 4, (see Fig. 2).
trains from its scheduled times causes a cost, or Three conflicts are resolved at the locations
revenue loss. This cost is written as designed for trains to meet or overtake known as
sidings or passing loops. Sidings are provided at
δ d qs ( Aqs ) and is assumed convex with minimum stations or between the lines segments to
at the scheduled departure time. Similarly there optimize the railway schedules.
are usually costs associated with the arrival
To find the optimal solution, using branch and
times ( X qs ), running times on links (rqs ) and
bound means the traversal of the solution tree of
dwelling times at stations ( wqs ) . all the possible solutions in train diagram, shown
in Figure 2. Finding an initial solution and an
upper bound to total cost function and then
Where rqs = X qs − Aqs
exploration of other branches of the tree until the
wqs = Aq , s +1 − X qs branches are fathomed. This method is prone to
what is called “combinatorial explosion” a
These costs are also included in the objective situation where for slightly larger instances of a
function. problem, the effort required grows
disproportionately. Therefore, emphasis is given
The general objective function now to be finding efficient heuristics like genetic algorithm
minimized can be written as: (GA) in order to find the solution in polynomial
Minimize P, i.e. times.
4.2 Genetic Algorithm

595

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on June 01,2021 at 01:05:10 UTC from IEEE Xplore. Restrictions apply.
 Holland, [3] firstly developed genetic schedule as genes in the chromosome. It
algorithm (GA) as search algorithm based on contains mainly three attributes: the delayed
mechanics on natural selection and genetic rules train, the train passed by and the segment in
which train conflict occurs. This representation
will give the train schedule solution with
minimum chromosome length. There can be
different possibilities to make the feasible
solutions depending upon which train is given
priority over other (see Fig.3). Child train
schedule solutions are produced from the two
parents train schedule solutions.

Fig.2. Illustration of two possible solutions of Train Schedule

also known as “theory of evolution”. The use of

genetic algorithms and so-called “evolutionary
computing” has become a very popular method
of solving optimization problems in recent years.
Fig.3.Illustration of Encoding Style of Chromosomes
Genetic algorithm are typed of stochastic search
method and applied to NP-complete problems in
many areas ranging from scheduling Infeasible child solutions are replaced with one
optimization to designing optimal configuration of the parent solutions. When the cross over is
and layouts. performed duplicate conflicts must be removed
In genetic algorithms, a population is nothing from the chromosomes.
but a collection of chromosomes representing Crossover Operator
possible solutions. The chromosomes are altered One point crossover is selected for the train
or modified using genetic operators through scheduling problem so to keep the infeasible
which a new generation is created. New child solution minimum. The two parents (train
members of the population are generated by schedule solutions) are selected to mate based on
various operations; the two most frequently used their schedule fitness values. The first step of
being random mutation and genetic crossover crossover is to select two parents then crossover
between the parents. The fitter the parent, the point is randomly selected. Then next step is
better its chances of transmitting its genes to choosing the cutting point. The genes prior to
later members. cutting point are copied to one child solution and
the ones after the cutting point is copied to
4.3 Implementation of GA to optimal second child solution (see Figure 4).
train dispatching problem Mutation process
A generic algorithm basically consists of three Some of the new solutions in the next
operators: Reproduction, Crossover and generation, which are produced after the
Mutation. We explain the implementation of selection and crossover, are mutated using the
these operators for the optimal train-dispatching mutation operator. The most common mutation
problem. process is to take a gene from the child solution
and alter it with predetermined probability. For
Representation of a chromosome example, if a gene can take the value of 0 or 1,
It is important to keep the length of the when the child is born the gene value is 0 then
chromosomes as short as possible to achieve fast after mutation it will be changed to 1 (general
convergence. We used the conflicts in the

596

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on June 01,2021 at 01:05:10 UTC from IEEE Xplore. Restrictions apply.
 style is illustrated in Figure 4). In train There are variety of ways to choose the
scheduling problem it is desirable to have a coefficient Cmax .It may be chosen as a input
mutation which has good chance to improve the coefficient, as the largest g value observed thus
solution. far, as the largest g value in the current
population,

Fig.4. Crossover and Mutation Process

Fig.5. Conflict Delay of Trains p and m
Generation of initial solution
In the schedule, there are many options to or the largest of the k generations. We consider
resolve the conflicts on sidings (see Figure. 5). It the last possibility in our case i.e. the maximum
gives many initial solutions, which makes the value of objective function P for the solution k.
population size for GA. To have variability in The probability of selection function for the
the population, the conflicts are resolved solution k is given by:
randomly on the neighboring sidings. The P( selectk ) = f k / ∑ f .
conflicts are resolved systematically, starting
The stopping rule is implemented in the
from the earliest conflict to last conflict in the
algorithm when there is no improvement in the
schedule. First obtain the unresolved schedule
schedule of the child solutions when compared
and make out the list of conflicts in the order
with its parents. GA will terminate when the
they occur. The trains p and q are selected in
variance between the populations becomes too
the first conflict. The amount of delay d1 and small, it may not be the fittest solution.
d 2 are measured for the both trains as illustrated
in Fig.5. Using the random number generation, 5. IMPLEMENTATION OF RESULTS
delay train p and train m with the probability In GA implementation to the problem like
respectively. Update the schedule after delaying discussed here, one of the important parts is to
one the trains and move to the next conflict. select appropriate combination of parameters i.e.
P ( p ) = d1 / ( d1 +d 2 ) , P ( m ) = d 2 / ( d1 +d 2 ) crossover type, the mutation rate. It is not
desirable to allow fast convergence dominating
Repeat this process until there is no more the population too early or to have a population
conflicts left in the schedule. taking too long to converge. Therefore, it is
Mapping Objective function to Fitness difficult to predict a priori what combination of
evaluation settings will produce the best result for the
In many problems, the objective is more problem. Here we also investigate the effect of
naturally stated as the minimization of some cost different values of these parameters in order to
function rather than the maximization of utility choose the more adequate ones to use in our
or profit function. In our case, the fitness scheduling problem. After having tests on the
evaluation is based on the minimization of the varying GA parameters, finally crossover,
total delay of trains in the schedule. With GAs mutation and population size are set as 0.8, 0.01
the following cost-to-fitness transformation is and 50, respectively.
commonly used:
Conflicts are hypothetically and randomly
f ( x) = Cmax − g ( x) When g ( x) < Cmax , generated to test the efficiency of our model.
f ( x) = 0 Otherwise The departure and arrival times are varied at

597

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on June 01,2021 at 01:05:10 UTC from IEEE Xplore. Restrictions apply.
 initial and intermediate stations in the range of performance of the scheduling implementation.
10 to 30 trains in the schedule. Fig. 6 shows the It is concluded that the model is able to produce
number of violating constraints and the useful results in terms of minimizing total delays
generations. The running time, dwelling time and producing optimal schedules in reasonable
and the minimum headway constraints converge time.
to zero after 15 to 25 generations. It shows that
our model will prevent the violation of these
constraints and reduce the total delays.
The crossover types and the mutation rate are
varied to see the effect on schedule performance.
However, we found out that the single point
crossover produces good results faster. Results
for these tests were averaged over 10 runs for
each train problem instance.

Fig. 7. Illustration of Schedule Fitness Vs Generations

REFERENCES
[1] Cordeau, J.F., Toth, P., Vigo, D., 1998. A survey of
optimization models for train routing and scheduling.
Transportation Science 32, 380–404.
[2] Garey, M.R., Jonhson, D.S., 1979. Computers and
Intractability: A Guide to the Theory of NP-
completeness. W.H. Freeman, San Francisco, CA.
Fig. 6. Illustration of violating constraints [3] Holland J.H., Adaptation in natural and Artificial
Systems. An Arbor, MI:Uni. Michigan Press, (1975).
As the GA based solutions involve a random [4] Jovanovic, D., Harker, P.T., 1990. A decision support
component for which the solution can vary at system for train dispatching: an optimization-based
methodology. Journal of the Transportation Research
each run. This is demonstrated in the Fig.7. Our Forum XXXI, 25-37.
model produced satisfactory results in terms of [5] Komaya, K., 1992. An integrated framework of
minimizing the delays at the intermediate and simulation and scheduling in railway systems. In:
final train stations (see Fig.7). The schedule Murthy, T.K.S., Allan, J., Hill, R.J., Sciutto, G., Sone,
fitness indicates the delay in seconds for S. (Eds.), Computers in Railways III, Vol. 1:
Management. Computational Mechanics Publications,
different simulations. After 35 generations there Southampton, Boston, pp. 611-622.
is no improvement found in the schedule to [6] Kraft, E.R., 1987. A branch and bound procedure for
minimize the delays. In our tests the total delay optimal train dispatching. Transportation Research
at the intermediate and final station reduces from Forum 28, 263–267.
35 min (2100 sec.) to 12 min. (700 sec.) as well [7] Petersen, E.R., Taylor, A.J., Martland, C.D., 1986. An
the cost associated to these delays are introduction to computer-assisted train dispatch.
Journal of Advanced Transportation 20, 63-72.
minimized.
Peterson and Taylor (1986)
[8] Sahin, I., 1999. Railway traffic control and train
6. CONCLUSION scheduling based on inter-train conflict management.
This paper presents the application of genetic Transportation Research B 33, 511–534.
algorithm (GA) to minimize delay at the [9] Sauder, R.L., Westerman, W.W., 1983. Computer aided
intermediate and final train stations. The train train dispatching: decision support through
optimization. Interfaces 13, 24-37.
dispatching model is formulated subject to
[10] Szpigel B., 1973. Optimal train scheduling on a single
operational constraints for safe operations of track railway. In: Ross, M. (Ed.). OR'72, North
trains between the stations. Then we present Holland Publishing Co. pp. 343-352.
some empirical observations about the

598

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on June 01,2021 at 01:05:10 UTC from IEEE Xplore. Restrictions apply.