Using PSO and GA to Optimize Schedule Reliability in Container Terminal

Jack Xunjie. Luo1,2 , Defeng Wu1, Zi Ma1 Tianfei Chen1 and Aiguo Li1
1- Automation Research Center, 2 - Operation Department,
Dalian Maritime University (DLMU) Qingdao Qianwan Container Terminal Company (QQCT)
Dalian, China Qingdao, China
Email: [email protected] E-mail: [email protected] or [email protected]

Abstract—A schedule reliability problem (SRP) optimization In fact, the minimum hours between ship schedule
model for dynamic berth allocation in container terminal (CT) is departure time and actual departure time in a port which is
proposed. The model focuses on the minimum average schedule called schedule reliability (SR) is most important, since if the
missed hours of ships between the ship schedule departure time schedule departure time is delayed, it has to be compensated by
and the actual departure time to enhance the schedule reliability the container terminal (CT) company. At the same time, as
(SR) of ships in CT, and the quay crane allocation is considered logistics service sectors, more and more CT companies are
in this model. Particle swarm optimization (PSO) algorithm and concerning how to meet client’s demands, among of which the
genetic algorithm (GA) are employed to optimize the berth SR is one of the most important demands. So the SR is one of
allocation planning to improve the SR in CT. Both of PSO and
KPIs (key performance indicator) to assess the service level of
GA perform well and make average schedule missed hours
ports and shipping liner [7]. In this study, an improved model
reduce 40%. Experimental results also show that PSO has faster
convergence rate than GA in this case. The SRP model can also which focuses on the SR is proposed to overcome the
be used for related time windows problem optimization of the aforementioned disadvantages.
airport, railway station, bus station and other logistics industries. When the mathematical model is established, all that left is
to optimize the berth allocation to improve the SR of CT.
Keywords - PSO; GA; Container Terminal;Schedule Reliability Genetic algorithm (GA) was utilized to optimize the berth
Problem (SRP); Optimization; Berth Allocation Problem (BAP).
allocation problem in many studies. However, another
powerful optimization tool, particle swarm optimization (PSO),
I. INTRODUCTION has some advantages over other similar optimization
It has been witnessed that berth allocation problem (BAP) techniques such as GA, namely the following [8] [9].
has been widely studied due to the rapid development of 1) PSO is easier to implement and there are fewer
shipping and ports industry and the fierce competition between parameters to adjust.
ports. Edmond and Maggs [1] firstly proposed a model for
solving berth allocation and port cargo handling problem by 2) In PSO, every particle remembers its own previous best
using queuing theory. Lai and Shih [2] introduced a heuristic value as well as the neighborhood best; therefore, it has a more
algorithm for berth allocation, a first-come-first-served (FCFS) effective memory capability than the GA.
allocation strategy was employed for the problem. A solution 3) PSO is more efficient in maintaining the diversity of the
approach was presented by Imai et al. [3] for static berth swarm (more similar to the ideal social interaction in a
allocation problem (SBAP). For a given set of ships, the community), since all the particles use the information related
method can determine a collection of assignments to make the to the most successful particle in order to improve themselves,
sum of time they spend in berths minimum. Later on, both Imai whereas in GA, the worse solutions are discarded and only the
et al. [4] and Nishimura et al. [5] extended the SBAP model. good ones are saved; therefore, in GA the population revolves
Imai et al. [4] extended SBAP to dynamic BAP considering the around a subset of the best individuals.
fact that some ships arrive at berths while work is in process.
Nishimura et al. [5] extended the dynamic BAP to the multi- In our study, PSO and GA are successfully implemented to
water depth configuration and genetic heuristic algorithm to solve the dynamic BAP with time restriction so that the SR in
solve the problem. Tang and Dai [6] proposed a berth CT is improved. Comparisons between PSO and GA are also
allocation model with priority. They also developed genetic made.
algorithm technique to solve the above problem.
This paper is organized as follows. In section 2, the
However, there are two main shortcomings for the above schedule reliability problem (SRP) model is proposed briefly.
studies. One is that they all just consider the minimum hours In section 3, the implementation of PSO and GA is presented in
that ship spends in waiting for berths and handling cargos, detail. In section 4, real experimental data is used to validate
namely the service hours. The models just target to improve the the effectiveness of the presented PSO and GA methodology.
operation efficiency, not to improve the service level offered Finally, some conclusions are given to summarize the study.
by CT to the clients. The other is that the quay cranes assigned
for each ship are not included in the established model, so that II. ROBLEM FORMULATION
the practical operation process in berth is not sufficiently
described. The SR investigated in this study involves a fixed number
of berth allocation for a collection of ships within a period of
time. Figure 1 depictes the schedule reliability problem (SRP)
This work is supported by Department of Science and Technology of
Liaoning Province (2007219003) of China and Dalian Maritime University
Young Teacher Fund (DLMU-ZL-200805)

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on November 01,2024 at 02:26:58 UTC from IEEE Xplore. Restrictions apply.
978-1-4244-4994-1/09/$25.00 ©2009 IEEE
based on whole operation process for one ship serviced in a For ship j, its actual departure time could be the berthing
port. Each ship, which arrives at the port, firstly waits at an time bijk plus handling hour hij . In dynamic BAP, assume
anchorage area for the schedule berthing time, then berthes at
the assigned position in one of the berths, after ship j is the kth ship arriving at berth i, its berthing time could
loading/unloading the cargos, departs from the berth. The be two parts, one is the total service hour for preceding k-1
schedule missed hour (could be either positive or negative) is ships, and the other one is the leisure time before ship j arrived
the absolute time error between ship schedule departure time at berth i. That is
and actual departure time. The smaller is schedule missed
hour, the higher SR is berth allocation in CT. Therefore, the
bijk = s i + ¦ ¦h
m∈V n∈Ok
m x imn + ¦¦y
m∈V n∈Ok
imn x imn

goal is to minimize the schedule missed hour called SRP so Therefore, the mathematical model for dynamic schedule
that the SR could be improved. reliability problem optimization can be formulated as
Schedule missed hour
1
Arrive time
of ship Berth time
Schedule
departure
Actual
departure
Min z = ¦¦¦
L i∈B j∈V k∈O
bijk + hij − g j ⋅ x ijk (1)

time time
Constraint condition:

Waiting hour Handling hour st ⋅ ¦¦ x ijk = 1 ∀i ∈ B, j ∈ V (2)

i∈B k∈O

Service
hour ¦x ijk ≤ 1 ∀i ∈ B, k ∈ O (3)
j∈V
Figure1 The Schedule RELIABLITY PROBLEM
A. Model Assumptions xijk ∈ {0,1} ∀i ∈ B, j ∈ V , k ∈ O (4)
1) Each ship should call at one berth and only be served at
berth for once; Mj
hij = ∀i ∈ B, j ∈V (5)
2) Each berth should serve only one ship at the same time or F ⋅ Ni
just be waiting for arrived ship;
3) The length and depth of the berth can satisfy the ¦N i =N ∀i ∈ B (6)
requirement of ship berthing; i∈B
4) The allocation strategy is FCFS; In the above constraint conditions, condition (2) ensures
B. Symbol Definition every ship will be serviced at berth; condition (3) guarantees
each berth will serve only one ship at any time. Condition (5)
B — the berth aggregate; shows the higher operation efficiency of quay crane, the
V — the ship aggregate; shorter time will the ship be served at berth.
O — the aggregate of the ship sequence berthing at each
berth; III. PSO AND GA TO SOLVE THE OPTIMIZATION
k = (1......T ) ∈ O —the sequence of ship berthing;
A. PSO Solution
Ok = (1......k − 1) ∈ O ;
PSO is an evolutionary computation technique developed
F — the operation efficiency of quay crane; by Eberhart and Kennedy [9]. It simulates the food searching
N i — the number of quay cranes assigned for each berth; activities of a swarm of birds (particles). In the
multidimensional space in which each particle represents a
M j — the lay freight volume for jth ship; point at the intersection of all search dimensions, each particle
in the swarm is moved toward the optimal point by adding a
N — total number of quay cranes; velocity to its position. The velocity of a particle is updated by
hij — the handling hour for jth ship at ith berth; three components, which are called inertial, cognitive, and
social. The inertial component is the simulation of the inertial
si — the setup time of berth i; behavior of the bird to continue flying in the same direction.
C. Model Expression Meanwhile the memory of the bird about its previous best
position is modeled by the cognitive component model, and
For a CT, assume there are L (Number of V) ships in T the memory of the bird about the best position among the
(Number of B) berths. Each ship has its own arrival time a j , particles (interaction inside the swarm) is modeled by the
handling time hj and schedule departure time social component. The particles move around the
multidimensional search space until they find the food
g j ( j = 1, " , L ) . The model established in this study is based (optimal solution).
on the minimum between actual departure time and schedule The velocity update equation is given by
departure time so that the berths for ships can be reasonably vid = wvid + c1r1 ( Pbest id − xid )
allocated to enhance the SR of the CT. The number and + c 2 r2 (Gbest d − xid ) (7)
efficiency of quay crane are also considered in the model.
i = 1, 2, " , m ; d = 1,2, ", D

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on November 01,2024 at 02:26:58 UTC from IEEE Xplore. Restrictions apply.
 The position update equation is given by wmax − wmin
x id = xid + vid (8) w = wmax − × iter (9)
itermax
Where itermax is the maximum number of iterations
Where
(generations), and iter is the current number of iterations.
v id - dimension d of the velocity of particle i;
x id - dimension d of the velocity of particle i; B. GA Solution
w - inertia weight; Genetic algorithm uses a separate search space and solution
c1 , c 2 - acceleration coefficients; space. The search space is the space of coded solutions, i.e.
Pbest id - dimension d of the own best position of particle i; genotypes or chromosomes consisting of genes. Genetic
algorithm is widely used in optimization problem and is proved
Gbest d - dimension d of the best particle in the swarm; to be effective to solve NP-hard problem. The whole procedure
D - dimension of the optimization problem (number of of GA is shown in Figure 3.
decision variables);
Start
m - number of particles in the swarm;
r1 , r2 - two separately generated uniformly distributed
Generate intial ppulation
random numbers in the range [0, 1].

Comput ftness fnction vlue

Initialize particle
swarm
Yes
Finish Satisfy stopping
criterion?
Evaluate Update velocity
No
particles and position

Tracking Selection
No
thparticle
Crossover
Determine Gbest
and Pbest Maximum iteration
satisfied?
Mutation
Yes

Gbest is the optimized New population

centers and widths
FIGURE3 FLOW CHART of GA
FIGURE2 FLOW CHART of PSO
The chromosome coding and fitness function employed in
How to properly code the particles is one of the key parts in
this study are the same as [6] to make a comparison between
this study, a 2L dimension space is constructed and denotes
PSO and GA. An example of Whe chromosome coding is
the berth allocation for L ships. Each ship has two dimensions:
shown in Figure 4. The example shows ship 4, ship 6, ship 8
one for berth number i assigned to the ship and the other one
and ship 11 call at berth 1 in order. Ship 2, ship 1, ship 3 and
for serving number k in berth i. This coding method will ship 5 call at berth 2 and ship 7, ship 9 at berth 3 respectively.
guarantee every ship to be served and restrict each ship to be
laid only at one berth. Although this way makes the particle
representation dimension higher, however it guarantees the
constraint conditions (2) and (3) so that the computation time
will be saved.
The whole procedure of PSO is shown in Figure 2. The two
constants c1 , c2 represent the weighing of the stochastic FIGURE4 AN EXAMPLE of GA CHROMOSOME CODING
acceleration terms that pull each particle toward the best The fitness function both for PSO and GA is chosen as
positions. If low values are chosen, then the particles will follows:
roam far from the target regions before being tugged back. On
f ( x) = 1 /(1 + exp( y ( x) / 10))
the other hand, if high values are considered, it will result in
abrupt movement toward, or past, target regions. Therefore, Where y(x) is original objective function.
the acceleration constants c1 and c2 were set to be 2.0 in IV. EXPERIMENTAL RESULTS
accordance with past literature. In this study, 12 ships were chosen for experimental data.
Good selection of inertia weights in (7) will provide a The details of the real data are listed in Table 1. Selected
balance between global and local explorations, thus requiring parameters of PSO and GA are shown in Table2 and Table 3
less iteration on average to find a sufficiently optimal solution. respectively. After 200 epochs, both of the optimized results
The inertia weight generally decreases linearly from are illustrated in Table4. It can be seen that the average
approximately 0.9 to 0.4 during a run, it is set according to the schedule missed hour between ship schedule departure time
following equation: and ship actual departure time is reduced from 7.1382 to

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on November 01,2024 at 02:26:58 UTC from IEEE Xplore. Restrictions apply.
4.3042. The experimental result demonstrated in Figure 5 18
shows that both of PSO and GA is a convergent optimization PSO
algorithm while PSO has higher convergence rate than GA in 16 GA

Fitness Function Value

this case. 14
TABLE 1 SELECTED REAL DATA IN DETAIL – JULY,2009
Schedule 12
Arrival Berth Actual Schedule
Ship missed 10
time time dep. time dep. time
hour 8

S1 0700/05 0040/06 0350/06 0600/06 20.3300 6

S2 1300/05 1340/05 0130/06 0800/06 7.5000
4
0 50 100 150 200
S3 1830/05 1950/05 0050/06 0800/06 9.3300
Number of Iterations
S4 2000/05 2350/05 0920/06 0500/06 0.1700
FIGURE5 CONVERGENCE of PSO AND GA
S5 0700/06 1140/06 0320/07 2300/06 0.8300
1130/06 1230/06 0530/07 0
V. CONCLUSION
S6 0400/07
S7 0800/07 0900/07 1700/07 1900/07 3.5000 This study proposes a schedule reliability problem (SRP)
optimization model for berth allocation based on the analysis
S8 0500/07 0720/07 1810/07 1800/07 2.5000
of existed model and practical requirements to improve the SR
S9 0930/07 1020/07 0150/08 0600/08 6.5000 in CT and shipping industry. The model considers the
S10 1400/07 1450/07 0830/08 0000/07 32.6700 allocation of quay crane so that the operation in CT is
described in more detail. The schedule missed hour between
S11 0400/08 0440/08 1740/08 1800/08 1.5000
ship schedule departure time and actual departure time is
S12 0900/08 1140/08 2120/08 1900/08 0.8300 chosen as the objective function so that the SR in CT is
Average schedule missed hour 7.1383 improved. PSO and GA are used for dynamic berth allocation
optimization to enhance the SR in CT. Experimental results
demonstrate the effectiveness of these two techniques.
However, PSO can converge faster than GA. Furthermore, it
TABLE 2 PSO PARAMETERS
should be pointed out that the quay crane allocation is not
Number of Dimension 24
included in simulation. Future work should focus on the
Number of Particles 100
inclusion of quay crane model and optimization of the quay
Number of Iterations 200
Ȧmax 0.8 crane allocation, so that to improve the efficiency of the CT.
Ȧmin 0.2 The SRP model can also be used for the airport, railway station,
c1,c2 2.0 bus station and other related logistics industries.

REFERENCES
TABLE 3 GA PARAMETERS [1] Edmond E D and Maggs R P. How Useful Are Queue Models in Port
Investment Decision for Container Berths. Journal of the Operational
Number of Dimension 14 Research Society, vol.29, no.8, pp.741-750, 1978.
Number of chromosomes 100 [2] K.K. Lai and K. Shih, A study of container berth allocation. Journal of
Number of Iterations 200 Advanced Transportation, vol. 26, no.1, pp: 45-60, 1992.
Probability of mutation Pm 0.08 [3] A. Imai, K. Nagaiwa and W.T. Chan, Efficient planning of
Probability of crossover Pc 0.25 berthallocation for container terminals in Asia, Journal of Advanced
Transportation, vol. 31, no.1, pp: 75-94, 1997.
[4] A. Imai, E. Nishimura and S. Papadimitriou, The dynamic berth
TABLE 4 OPTIMIZED BERTH ALLOCATION RESULTS allocation problem for a container port, Transportation Research Part B,
Average vol.35, no.4, pp.401-417, 2001.
No. of [5] E. Nishimura, A. Imai and S. Papadimitriou, Berth allocation planning in
schedule Berth allocation results the public berth system by genetic algorithms, European Journal of
iterations
missed hour Operational Research, vol.131, no.2, pp.182-292, 2001.
[6] L. Tang and L. Dai, Berth Allocation with Service Priority for Container
Berth 1˖(4,6,8,11) Terminal of Hub Port, 4th International Conference on Wireless
PSO 200 4.3042 Berth 2˖(2,1,3,5,7,9,12) Communications, Networking and Mobile Computing, pp.1-4, 2008.
[7] Jack X.J. Luo and X.Y.Chen, Operations and Businesses Management
Berth 3˖(10) of Container Terminals. Dailian, China, Dailian Maritime University
Press, ISBN 978-7-5632-1217-0 , 1998
Berth 1˖(4,6,8,11)
[8] Del Valle Y, Venayagamoorthy, G.K., Mohagheghi, S., Hernandez, J.-C.
GA 200 4.3042 Berth 2˖(2,1,3,5,10,12) and Harley, R.G. Particle Swarm Optimization: Basic Concepts,
Variants and Applications in Power Systems, IEEE Transactions on
Berth 3˖(7,9) Evolutionary Computation, vol.12, no.2, pp.171-195 , 2008.
[9] Kennedy J. and Eberhart R., Particle swarm optimization, IEEE
International Conference on Neural Networks, pp.1942-1948, 1995.

Authorized licensed use limited to: UNIVERSIDADE DE SAO PAULO. Downloaded on November 01,2024 at 02:26:58 UTC from IEEE Xplore. Restrictions apply.