CHAPTER 4 MULTI-DEPOT VEHICLE ROUTING...
Transcript of CHAPTER 4 MULTI-DEPOT VEHICLE ROUTING...
189
CHAPTER 4
MULTI-DEPOT VEHICLE ROUTING PROBLEM
4.1 INTRODUCTION
The Vehicle Routing problem (VRP) is one of the most challenging
combinatorial optimization tasks in real time logistics applications. The VRP
can be defined as a problem of finding the optimal routes of delivery or
collection from one or several depots to a number of cities or customers,
while satisfying capacity and time constraints. In a real-world environment,
drivers choose the shortest path to reach a destination, due to this, the distance
travelled and the cost can be minimized. Collection of household garbage,
gasoline delivery trucks, goods distribution, and snow plough, street cleaning,
school bus routing, dial-a-ride systems, transportation for handicapped
persons, routing of salespeople and mail delivery are the most common
applications of the VRP. The VRP plays a vital role in distribution, logistics
and supply chain management. Huge research efforts have been devoted to
studying the VRP since 1959 where Dantzig and Ramser have described the
problem as a generalized problem of Travelling Salesman Problem (TSP).
The VRP is responsible in designing an optimal set of routes for a
fleet of vehicles in order to serve a given set of customers. The interest in
VRP is motivated by its practical relevance as well as by its considerable
difficulty. During the past few decades, an increasing number of optimization
techniques based on operations research are proposed, for the effective
management of the provision of goods and services in distribution systems.
190
The VRPs have several variants based on the operational mechanism and
mathematical modeling. Based on the constraints and characteristics of VRP,
they are classified into capacitated VRP, multi-depot VRP, periodic VRP,
split delivery VRP, stochastic VRP, VRP with backhauls, VRP with pick-up
and delivery and VRP with time window.
The Multi-Depot Vehicle Routing Problem (MDVRP), an
extension of classical VRP, is a NP-hard problem for simultaneously
determining the routes for several vehicles from multiple depots to a set of
customers and then returning to the same depot. The objective of the problem
is to find routes for vehicles such that all the customers are served at a
minimal cost in terms of total travel distance, without violating the capacity
and travel time constraints of the vehicles. During the past three decades,
classical VRPs are paid more attention in the research perspective,
comparatively, the number of research projects on MDVRP is fewer. Salhi
and Sari (1997) addressed a multi-level composite heuristic with two
reduction tests. The initial feasible solutions are constructed in the first level,
while the intra-depot and the inter-depot routes are improved in the second
and third levels. Wu et al (2002) reported a simulated annealing (SA)
heuristic for solving the multi-depot location routing problem (MDLRP). To
solve the problem on a larger scale, the original problem is divided into two
sub-problems, namely, the location-allocation problem, and the general
vehicle routing problem, respectively. Each sub-problem is then solved in a
sequential and iterative manner by the simulated annealing algorithm
embedded in the general framework for the problem-solving procedure.
Giosa et al (2002) developed a “cluster first, route second” strategy for the
MDVRP with Time Windows (MDVRPTW), an extension of the MDVRP.
Considering the operational nature of the MDVRPTW, the cluster first route
second technique, focuses more on minimizing the computational time.
191
Haghani and Jung (2005) presented a formulation for solving the dynamic
vehicle routing problem with time-dependent travel times using Genetic
Algorithm. The performance of the genetic algorithm is evaluated by
comparing its results with exact solutions and lower bounds for randomly
generated test problems. For small size problems with up to 10 demands, the
genetic algorithm provided same results as exact solutions. For the problems
with 30 demand nodes, the genetic algorithm results are found to have less
than 8% gap with lower bounds.
Lee et al (2006) handled the MDVRP by formulating the problem
as deterministic dynamic programming (DP) with finite-state and action
spaces, and then using a shortest path heuristic search algorithm. Creviera
et al (2007) proposed a heuristic combining tabu search method, and integer
programming for multi-depot vehicle routing problem in which vehicles may
be replenished at intermediate depots along their route. Jeon et al (2007)
suggested a hybrid genetic algorithm (HGA) for MDVRP, which considers
the improvement of generation for an initial solution, three different heuristic
processes, and a float mutation rate for escaping from the local solution in
order to find the best solution. Inorder to solve the MDVRP efficiently, two
hybrid genetic algorithms (HGA1 and HGA2) are developed by Ho et al
(2008). In their approach, the initial solutions are generated randomly in
HGA1. The Clarke and Wright saving method and the nearest neighbor
heuristic are incorporated into HGA2 for the initialization procedure. Results
proved that HGA2 is superior to HGA1 in terms of the total delivery time.
Chen and Xu (2008) developed a hybrid genetic algorithm (GA)
with simulated annealing for solving the MDVRP. Since the MDVRP
integrates hard optimization problems, three improvement heuristic
techniques are introduced by Mirabi et al (2010). Each hybrid heuristic
192
combines elements from both constructive heuristic search and improvement
techniques. The improvement techniques used in this method are
deterministic, stochastic and simulated annealing (SA) methods. These
techniques outperformed Giosa’s (2002) cluster first, route second method in
terms of minimum delivery time. Lau et al (2010) considered the cost due to
the total traveling distance, and the cost due to the total traveling time for
solving the MDVRP. They employed a stochastic search technique called
fuzzy logic guided genetic algorithms (FLGA) to solve the problem.
MDVRP solutions are provided by Renaud et al (1996) in three
phases: fast improvement, intensification, and diversification for 23
benchmark instances using the Tabu Search algorithm. They considered
MDVRPs with capacity and route length restrictions. GA for MDVRP
proposed by Ombuki-Berman and Hanshar (2009) employed an indirect
encoding and an adaptive inter-depot mutation exchange strategy for the
MDVRP with capacity and route-length restrictions. Thangiah and Salhi
(2001) presented a genetic clustering technique for MDVRP, where a
generalised clustering mechanism is applied to genetic algorithms as a
postoptimizer. Su (1999) presented a real time dynamic vehicle control and
scheduling system for multi-depot physical distribution. To perform the
system objectives effectively, their system had five major modules such as the
global information collection system, depot controller, route planner, vehicle
scheduler, vehicle route and time table feedback system. This method is more
suitable for scheduling based on real time status of the system.
Wenjing et al (2010) introduced a modified PSO algorithm with
mutation operator and improved inertia weight for solving MDVRP. The
simulation results showed that this modified method could not only avoid
premature automatically according to the convergence level but also get a
193
better optimal solution than the standard basic PSO. Wang et al (2008) used
particle position matrix based on goods for particle swarm optimization to
solve multi-depots single VRP model, and every matrix column corresponded
to one goods. Matrix elements are random number between 0 and 1, and
matrix elements sort rules are established to get single vehicle route thus
satisfying the objective function.
The Artificial Bee Colony algorithm (ABC) is a new
population-based meta-heuristic approach proposed by Karaboga (2005). This
approach is inspired by the intelligent foraging behavior of honeybee swarm.
Karaboga and Basturk (2007) demonstrated that ABC outperforms genetic
algorithms and particle swarm optimization in multivariable function
optimization. Brajevic (2011) presented the ABC algorithm for capacitated
vehicle routing problem. In general, ABC algorithms are applied for
continuous optimization problems. Since VRPs are combinatorial
optimization problems certain modifications are implemented by Ivona.
Though ABC is a fairly new approach introduced a few years ago, it has not
yet been applied to solve the MDVRP. Hence, it is worthwhile to evaluate the
performance of ABC algorithm for solving MDVRP.
In this thesis, the performance of the heuristics such as Genetic
Algorithms (GA) (Surekha and Sumathi Aug 2011), Modified Particle Swarm
Optimization (MPSO) (Surekha and Sumathi Oct 2011), Artificial Bee
Colony (ABC) (Surekha and Sumathi Aug 2012) optimization, Hybrid MPSO
with GA, the Genetic Swarm Optimization (GSO) and Improved Genetic
Swarm Optimization (IGSO) on MDVRP benchmark instances are addressed.
The solution to MDVRP is obtained in four stages namely grouping, routing,
scheduling and optimization. Customers are clustered based on distance
between customers and depots in the grouping phase. In routing, customers of
the same depot are assigned to several routes by Clarke and Wright saving
194
method and each route is sequenced in the scheduling phase. Better routing
and scheduling can result in shorter delivery distance, shorter time spent in
serving all customers, higher level of efficiency and lower delivery cost. The
scheduled routes are optimized using the optimization algorithms in order to
obtain a global optimum solution with optimal distance, improved robustness,
reduced computational time and better algorithmic efficiency.
A set of five different Cordeau’s benchmark instances (p01, p02,
p03, p04, p06) from the online resource of University of Malaga, Spain are
experimented using MATLAB R2008b software. Several investigations are
conducted on the benchmark instances using the proposed heuristics. The
results are evaluated in terms of depot’s route length, optimal route, optimal
distance, computational time, and number of vehicles.
The chapter is organized as follows: The basic concepts of MDVRP
along with the mathematical modeling and architecture of the proposed
approach are discussed in Section 4.2. Section 4.3 enlightens the
implementation strategies of the proposed bio-inspired intelligent heuristics
such as GA, MPSO, ABC, GSO and IGSO for solving the MDVRP. The
simulation results and experimental analysis of the MDVRP instances using
heuristic algorithms are described in Section 4.4. Section 4.5 compares and
discusses the outcome of this chapter while section 4.6 draws the conclusion
and future work of solving multi-depot vehicle routing problems.
4.2 FUNDAMENTAL CONCEPTS OF MDVRP
In MDVRP, the number and locations of the depots are
predetermined. Each depot is large enough to store all the products ordered by
the customers. Each vehicle starts and finishes at the same depot. The location
and demand of each customer is also known in advance and each customer is
195
Route 2
Route 1
Route 2
Route 1
A B
1
5 9
4
8
7
3 10
6
2
visited by a vehicle exactly once. Figure 4.1 shows an example of the
MDVRP with 2 depots and 10 customers. Since there are additional depots
for storing the products, the decision makers have to determine depots
through which the customers are served (Ho et al 2008). The decision making
stages are classified into grouping, routing, scheduling and optimization as
shown in Figure 4.2.
Figure 4.1 Example of an MDVRP with 2 depots and 10 customers
In grouping, customers are clustered based on distance between
customers and depots. In the example, customers 1,5,9,4,8 are assigned to
depot A while customers 7,10,3,6,2 are assigned to depot B. In depot A,
customers 1,5,9 are in the first route, while customers 4 and 8 are served in
the second route. The customers of the same depot are assigned to several
routes in the routing phase by Clarke and Wright saving method and each
route is sequenced in the scheduling phase. The aim of routing is to minimize
the number of routes without violating the capacity constraints. Since there
are two depots the minimum number of routes can be limited to two. More
number of routes increase the number of vehicles required thus reducing the
distance. In general, the objective of the MDVRP is to minimize the total
delivery distance or time spent in serving all customers thus utilizing efficient
amount of vehicles.
196
Figure 4.2 Decision making in MDVRP
4.2.1 Mathematical Formulation of MDVRP
The MDVRP is formulated with the objective of forming a
sequence of customers on each vehicle route. The time required to travel
between customers along with the depot and demands are known in advance.
It is assumed that all vehicles have the same capacity, and each vehicle starts
its travel from a depot, upon completion of service to customers, it has to
return to the depot. The notations used and the mathematical model are as
follows:
Sets
I – Set of all depots
J – Set of all customers
K – Set of all vehicles
Indices
i – depot index
j – customer index
k – route index
Grouping / Clustering
Routing
Scheduling and Optimization
197
Parameters
N – Number of vehicles
Cij – Distance between point i and j, JIji,
Vi – Maximum throughput at depot i
di – Demand of customer j
Qk – Capacity of vehicle (route) k
Decision variables
otherwisekrouteonjpreceedsyimmediateliif
xijk ,0,1
otherwiseidepottoallottedisjcustomerif
zij ,0,1
Ulk – auxiliary variable for sub-tour elimination constraints in route k
Mathematical model
The objective function is to minimize the total distance of all
vehicles given by Equation (4.1),
JIi JIj Kkijkij xCmin (4.1)
Each customer has to be assigned a single route according to
Equation (4.2),
Kk JIiijk Jjx ,1 (4.2)
The capacity constraint for a set of vehicles is given by
Equation (4.3),
198
Jj JIikijkj KkQxd , (4.3)
Equation (4.4) gives the new sub-tour elimination constraint set as,
KkJjlNNxUU ijkjklk ,,,1 (4.4)
The flow conservation constraints are expressed as in Equation (4.5),
JIiKkxx
JIj JIjjikijk ,,0 (4.5)
Each route can be served at most once according to Equation (4.6),
Ii Jjijk Kkx ,1 (4.6)
The capacity constraints for the depots are given in Equation (4.7) as,
IiVzd i
Jjiji , (4.7)
Constraints in Equation (4.8) specify that a customer can be
assigned to a depot only if there is a route from that depot going through that
customer,
KkJjIixxz
JIuujkiukij ,,,1)( (4.8)
The binary requirements on the decision variables are given by
Equations (4.9) and (4.10)
KkJjIixijk ,,},1,0{ (4.9)
JjIiz ij ,},1,0{ (4.10)
The positive values of the auxiliary variable is defined in
Equation (4.11) as,
199
KkJlUlk ,,0 (4.11)
The objective function according to Equation (4.1) minimizes the
total delivery distance. Constraint in Equation (4.2) ensures that each
customer is allotted with only one route while constraint in Equation (4.3)
guarantees the capacity limit of vehicles. Similarly, the sub-tour avoidance is
imposed by Equation (4.4) and the flow limits by Equation (4.5). The route to
be served and the limit on the depots are given by Equation (4.6) and
Equation (4.7) respectively. Thus the MDVRP aims at minimizing the total
delivery distance by satisfying the mentioned constraints.
4.2.2 Architecture of the Proposed Intelligent MDVRP System
Multi-depot vehicle routing problem is more challenging and
sophisticated than the single-depot VRPs. In addition, MDVRP is NP-hard,
which means that an efficient algorithm for solving the problem to optimality
is unavailable. Therefore, MDVRP is difficult to solve with exact procedures
such as branch and bound, branch and cut algorithms. To deal with the
problem efficiently and effectively, heuristic algorithms like GA, MPSO,
ABC, GSO and IGSO are applied in this work. Figure 4.3 illustrates the block
diagram of the proposed multi-depot vehicle routing problem.
The well known Cordeaux’s instances (p01, p02, p03, p04, p06) are
chosen as benchmark problems. The customers are clustered based on the
distance between customers and depots. The customers of the same depot are
assigned to several routes using the Clark and Wright saving method (Clarke
and Wright 1964). The routes obtained from Clarke and Wright saving
method is scheduled and optimized using heuristic algorithms such as genetic
algorithm, modified particle swarm optimization, artificial bee colony
algorithm, genetic swarm optimization and improved genetic swarm
optimization. The results of all optimization techniques such as GA, MPSO,
200
ABC, GSO and IGSO are evaluated with respect to number of customers
serviced, number of vehicles required, optimal route, optimal distance, and
computational time. A comparative analysis of the proposed heuristic
techniques is performed based on optimal distance, robustness, computational
competency, and algorithmic efficiency to identify the suitable optimization
algorithm for solving MDVRP.
Figure 4.3 Proposed block diagram for MDVRP
4.2.3 Grouping Assignment
The group assignment also referred as cyclic assignment requires
each depot to attract one customer every time until all customers are assigned.
The cyclic assignment assigns one customer at a time in a cycle based on the
location of the depot heads. First, the head of each depot is set as the depot
itself. Then, for each depot, the closest customer to the depot head is assigned
to it and the depot head is updated by the closest customer, if the vehicle of
the depot has enough serving capacity. The above procedure is repeated until
all customers are assigned.
The first step is to assign customers to each of n links, known as the
grouping problem. Because the objective here is to minimize the total delivery
Benchmark Input
Cordeaux’s instances
p01, p02, p03, p04, p06
Grouping
Grouped with respect to
distance between customers and depots
Routing
Clarke and Wright Saving method
Scheduling and Optimization
GA, MPSO, ABC, GSO and IGSO
Experimental results and comparative analysis
Optimal distance, computational
competency, algorithmic efficiency
201
time spent in distribution, customers are assigned to the nearest depot.
Grouping assignment procedure is given below:
Step 1 : Calculate the distance among each customer.
Step 2 : Calculate the distance between every customer and every depot.
Step 3 : Assign the customers to the nearest depot.
Step 4 : The customers are divided into M groups, where M is equal to the
number of depots.
For example, in a distribution network, a supplier owns two depots
(dA and dB) to deliver the products to a set of customers. Each customer, ci
should be assigned to a single depot exactly.
Equation (4.12) shows that, if the customer ci is located near to
depot A , then the customer ci is assigned to depot A .
If ),(),( dBciDdAciD , then assign ci to dA ; (4.12)
Equation (4.13) indicates that, if the customer ci is located near to
depot B , then the customer ci is assigned to depot B .
If ),(),( dBciDdAciD , then assign ci to dB ; (4.13)
Equation (4.14) ensures that, if the distance between customer ci
and depot A is equal to the distance between customer ci and depot B , then
customer ci is assigned to either depot Aor depot B .
In case ),(),( dBciDdAciD , select a depot arbitrarily (4.14)
Equation (4.15) represents the distance between customer i and
depot k .
202
22 )()(),( dkcidkci yyxxdkciD (4.15)
In the above Equation (4.15), xci and yci denote x and y coordinates of the customer ci , xdk and ydk indicates the x and y coordinates of depot k . The customers are grouped and served by either depot A or depot B .
4.2.4 Routing Algorithm
In recent years, since the evolution of supply chain management, the system design problem considering the location of distribution facilities has become more significant. In routing phase, customers in each group are divided into different routes. The aim of routing is to minimize the number of routes, or vehicles, used while not violating the vehicle capacity constraint. A key element of many distribution systems is the routing and scheduling of vehicles through a set of customers requiring service. Better routing decision can result in higher level of customer satisfaction. A wrong grouping assignment solution will result in routes of higher total cost (distance) than with a better grouping assignment. So the routing phase is strongly dependent on the grouping assignment.
In 1964 Clarke and Wright published an algorithm for solving the classical vehicle routing problem. The savings algorithm is a heuristic algorithm, and therefore it does not provide an optimal solution to the problem with certainty. The method does, however, often yield a relatively good solution, that deviates slightly from the best known solution. In the first step of the savings algorithm the savings for all pairs of customers are
calculated, and all pairs of customer points are sorted in descending order of the savings. Subsequently, from the top of the sorted list of point pairs one
pair of points is considered at a time. When a pair of points ji is
considered, the two routes that visit i and j are combined (such that j is
visited immediately after i on the resulting route), if this can be done without deleting a previously established direct connection between two customer
203
points, and if the total demand on the resulting route does not exceed the vehicle capacity. There are two versions of the savings algorithm, namely sequential and parallel. In the sequential version exactly one route is built at a time, while in the parallel version more than one route may be built at a time.
Figure 4.4 Illustration of the savings concept (a) Separate route (b) Same route
Initially in Figure 4.4(a), customers i and j are visited on separate
routes. An alternative to this is to visit the two customers on the same route, for example in the sequence ji as illustrated in Figure 4.4(b). Since the
transportation costs are given, the savings that result from driving the route in Figure 4.4(b) instead of the two routes in Figure 4.4(a) can be calculated.
Denoting the transportation cost between two given points i and j by cij, the
total transportation cost Da in Figure 4.4(a) is:
0000 jjiia ccccD (4.16)
Equivalently, the transportation cost Db in Figure 4.4(b) is:
00 jijib cccD (4.17)
By combining the two routes the savings Sij can be obtained as,
ijojibaij cccDDS 0 (4.18)
204
The customers with larger saving value are grouped in the same route without violating the vehicle capacity constraint.
4.3 BIO-INSPIRED TECHNIQUES FOR MDVRP
Due to the complexity of the problem, solving the MDVRP to optimality is extremely time-consuming. In order to tackle the problem efficiently, researchers have preferred heuristic methods over exact methods like branch and bound, branch and cut algorithms (Ho et al 2008). In this dissertation, the application of GA, MPSO, ABC, GSO and IGSO to solve MDVRPs are proposed. The step by step procedure required for implementing these intelligent algorithms is delineated in this section.
4.3.1 Solution Representation and Fitness Function
The solution representation and fitness function for all the proposed bio-inspired techniques is analogous and the feasible solution is generated based on three basic steps: Grouping, Routing and Scheduling. The individuals for the solution of the MDVRP are encoded using path representation, in which the customers are listed in the order in which they are visited. Consider an MDVRP instance with 6 customers designated 1-6. If the path representation for this instance is (0 2 4 1 0 3 6 5 0), then two routes are required by the vehicles to serve all the customers. The first route starts from the depot at 0 and travels to customers 2, 4 and 1, upon serving the vehicle
returns to the depot. Similarly, the second route starts from depot at 0, services customers 3, 6, 5 and returns to the depot. While applying heuristics based optimization techniques each individual in the initial population consists of n links for n depots in the MDVRP.
Grouping – In this stage, the customers are assigned to each of the n links. The objective of the MDVRP is to minimize the total delivery time and hence customers are assigned to the nearest depots. In the example, there are two depots A and B, each customer ci has to be assigned to a single depot exactly.
205
Depot B Depot A
Customers 0A 0B 1 2 3 4 5 6 7 8 9 10
Grouping
1 4 5 8 9 2 3 6 7 10
Routing
Depot B Depot A
0A 1 5 9 0A 0B 2 6 0B 7
Scheduling
0A 8 4 3 10 0B
Depot B Depot A
0A 9 5 1 0A 0B 2 6 0B 3 0A 8 4 7 0B 10
This process of grouping is done based on the distance computation according to the following rule:
If D(ci,A) < D(ci,B), then customer ci is assigned to depot A
If D(ci,A) > D(ci,B), then customer ci is assigned to depot B
If D(ci,A) = D(ci,B), then customer ci is assigned to a depot chosen arbitrarily between A and B
In the above cases, 22 )()(),( kckci yyxxkcDii
, represents the
distance between customer ci and depot k.
Figure 4.5 Solution representation and initial population
Routing – The customers in the same link are assigned to several routes using
Clarke and Wright Saving method. The routing is based on the distance
travelled by the vehicles for serving the customers. A saving matrix ),( ji ccS is
constructed for every two customers i and j in the same link. Further, the
206
customers with large saving value are grouped in the same route without
violating the capacity constraints. The saving matrix is constructed according
to ),(),(),(),( jijiji ccDckDckDccS .
Scheduling – Starting from the first customer, the delivery sequence is
chosen such that the next customer is as close as to the previous customer.
This process is repeated until all the unselected customers are sequenced. The
scheduling is performed by the respective optimization technique based on the
fitness function. The fitness function for the MDVRP is described in this
section. At the end of the scheduling phase, a feasible solution of the MDVRP
example problem (Figure 4.1) is constructed as shown in Figure 4.5.
Fitness Evaluation - For the MDVRP, the objective function is to minimize
the maximum delivery time spent among n depots. The delivery operations
start at the same time in every depot, while it takes different time to complete
serving the assigned customers. Some vehicles belonging to a depot may
complete the delivery faster while other vehicles may complete their task in a
longer duration. Let Dt be the total delivery time required by a depot k and let
)min( tD represent the minimum delivery time spent by all n depots, then
k cm
k
m
ict icicdcmcdD
1 1)](),1([)]0(),([ (4.19)
where V
yyxxbad abab
22 )()(),( is the travel time of a vehicle from
customer a to b. V is the speed of the vehicle, c(i) is the location of the ith
customer, c(0) is the initial position of the depot, mc is the number of
customers in route r, mk is the number of routes in depot k. Thus the fitness
function is defined as )min( tDF . This function acts as the objective
function to solve MDVRPs using the proposed intelligent techniques.
207
4.3.2 Implementation of MDVRP using GA
Genetic Algorithms (GA) is based on a parallel search mechanism
(Goldberg 1989), which makes it more efficient than other classical
optimization techniques such as branch and bound, tabu search method and
simulated annealing. The basic idea of GA is to maintain a population of
candidate solutions that evolves under selective pressure. GA imitates the
mechanism of natural selection and the survival of the fittest as witnessed in
natural evolution. The GA can avoid getting trapped in a local optimum by
tuning the genetic operators, crossover and mutation. Due to its high potential
for global optimization, GA has received great attention in solving multi-
depot vehicle routing problems. The general scheme of the genetic algorithm
for MDVRP is explained in this section.
Initial population
The initial population is formed based on the routes, depots and
customers according to the description available in section 4.3.1.
Selection
During each generation, the parents are selected for mating and
reproduction. In this MDVRP application, tournament selection is applied
(Renaud et al 1996) to generate new individuals in the population. This
selection strategy is based on fitness evaluation given in section 4.3.1. The
selection procedure is as follows:
Step 1 : Select a set of g individuals from the population in a random
manner to form the tournament set
Step 2 : Choose a random number rn in the range [0,1]
208
Step 3 : If rn < threshold select the fittest individual from the tournament
set for reproduction
Else choose any two chromosomes in random from the
tournament set for reproduction
Step 4 : Apply elitism to guarantee that the best individuals are selected
Crossover
A problem specific crossover technique, the Best Cost Route
Crossover (BCRC) developed by (Ombuki-Berman and Hanshar 2009), for
vehicle routing problem with time windows (VRPTW) is applied in this work
for MDVRP with slight improvements. The steps involved in BCRC are
shown below:
Step 1: Choose the parents from tournament selection
Step 2: Select a route from each parent in a random manner
Step 3: Remove all customers belonging to route 1 from parent 1
Step 4: For every customer belonging to route 1
Compute the cost of insertion of route 1 into each location of
parent 2 and store the costs in an ordered list.
For each insertion location, check whether the insertion is
feasible or not
Generate a random number ]1,0[nr
Choose the first feasible insertion location if rn < threshold
Else if rn > threshold choose the first entry in the ordered list,
despite of the feasibility
Step 5: Repeat Step 4 for customer belonging to route 2
209
Mutation
The flip bit mutation operator is used in finding an MDVRP
solution using GA. A substring is selected from the parent in a random
manner and flips to form an offspring. The route reversal mutation works on
only one chromosome at a time.
The design steps used for the implementation of MDVRP based on
GA are:
Step 1 : Generate randomly the initial population of GA based on routes,
depots and customers.
Step 2 : Evaluate the fitness function for each chromosome in the
population according to Equation (4.19).
Step 3 : Calculate the total fitness value, fi which is represented as the
summation of individual fitness of each chromosome
Step 4 : Repeat the following steps until new chromosomes have been
created
Select a pair chromosome from current population, based on
tournament selection.
With the crossover probability of 0.6, exchange the pair using
BCRC to form two new chromosomes.
Mutate the two new chromosomes at each row with the
mutation probability (0.02) and place the resulting
chromosomes in the new population.
Step 5 : Replace the current population with new population.
Step 6 : If the fixed number of generations is reached, stop and return the
best solution in current population else go to step 2.
210
The chromosomes in the population evolve through successive
iterations, called generations, thus evaluating the measure of fitness based on
Equation (4.19). The fitter the chromosomes, the higher the probabilities of
being selected to perform the genetic operations, such as crossover and
mutation. Selected individuals are chosen for reproduction (or crossover) at
each generation, with an appropriate mutation factor to randomly modify the
genes of an individual, in order to develop the new population. The current
population is replaced with the new population, until fixed number of
generations is reached.
4.3.3 Solving MDVRP Using MPSO
Particle Swarm Optimization (PSO) is a population based search
technique, consisting of potential solutions known as particles, duplicating a
flock of birds. The particles are initialized randomly based on the
optimization problem and they freely fly across the multidimensional search
space. Each particle updates its own velocity and position thus driving the
swarm towards the best fit regions. Eventually all the particles gather around
the point with the highest objective function. When compared with GAs, PSO
does not have any genetic operators like crossover and mutation. In GA, the
chromosomes share the information with each other and the whole population
moves towards the optimal solution, while in PSO, only the best particle
shares the information with others. Due to this one way information sharing
mechanism, the evolution of PSO is always towards the best solution.
Moreover, all particles tend to converge to the best solution quickly without
getting trapped in the local minima.
Solution representation and fitness evaluation
The particles for the solution of the MDVRP are encoded using
path representation and the fitness function is evaluated based on distance
according to the explanation given in section 4.3.1.
211
Updating Particles
In the evolutionary process, MPSO is used to generate new
offspring based on new position of the particle, thus updating the velocity
according to
itv
it
gt
it
it
itt
it
it xprcxprcvwxx
1
)()( 22111
(4.20)
The best position visited by the kth particle since the first time step
is known as the local best position denoted as pbest itp , while gbest g
tp
represents the best position that the kth particle and its neighbors have visited
from the beginning of the algorithm. In this work, since the distances are real
coded, the velocities and position update rules are slightly modified according to,
)}()(int{ 22111it
gt
it
it
itt
it xprcxprcvwv (4.21)
}{ 11it
it
it vxabsx (4.22)
where abs {} denotes the absolute value function, int {} is the integral
function, r1 and r2 are two independent random numbers uniformly distributed
in the range {0,1}, c1 and c2 are the weighting factors also known as cognitive
and social parameters respectively.
During each iteration, the pbest and gbest are updated based on the
objective (fitness) value. If the current value of pbest is better, pbest and its
objective value are updated with the current position and objective value.
Similarly, if the objective value is better than the objective value of gbest, then
gbest and its objective value are updated with the position and objective value
of the current best particle.
212
Algorithm
The step by step procedure for implementing the MDVRP using
MPSO is shown below:
Step 1 : Initialize the parameters of MPSO to suitable values. Set the
value of maximum number of iterations.
Step 2 : Initialize the population of particles according to the delivery
sequence of customers.
Step 3 : Evaluate the fitness function for all particles and choose the best
local and global positions pbest and gbest of the particles
Step 4 : Update the particles according to Equations (4.21) and (4.22)
Step 5 : Compute the fitness function of all the updated particles and
determine the best fit particle. Compare the fitness and update the
particles accordingly.
Step 6 : Test for stopping condition. If maximum number of iterations is
reached, then stop, else continue with Step 3.
4.3.4 Artificial Bee Colony Based MDVRP
The Artificial Bee Colony (ABC) optimization algorithm is
becoming more popular recently, due to the foraging behavior of honeybees.
ABC is a population based search technique, in which the individuals known
as the food positions are modified by the artificial bees during course of time.
The objective of the bees in turn is to discover the food sources with high
nectar concentration. The colony of artificial bees is grouped into employed
bees, onlooker bees and scout bees. During initialization phase, the objective
of the problem is defined along with the ABC algorithmic control parameters.
An employed bee is assigned for every food source available in the problem.
213
In employed bee phase, the employed bee stays on a food source and stores
the neighborhood of the source in its memory. During the onlooker phase,
onlooker bees watch the waggle dance of employed bees within the hive to
choose a food source. The employed bee whose food source has been
abandoned becomes the scout bee. Scout bees search for food sources
randomly during the scout phase. Thus the local search is carried out by the
employed bees and the onlooker bees while the global search is performed by
the onlooker and the scout bees, thus maintaining a balance between the
exploration and exploitation process. In this section, the step by step
procedure to implement ABC technique for MDVRP is discussed.
Generation of initial solutions
The initial solutions to the ABC problem are formed using the
grouping, and routing stages. In this mode, ABC algorithm generates
randomly distributed initial solutions, due to which the constraints satisfaction
may not be guaranteed. Therefore during initialization every solution has to
be checked according to each route as follows:
If the value of more than one position in the corresponding
positions of all routes in the solution is 1, one position is
selected in random and its value is set to 1, while the value of
others is set to 0.
If the values of the corresponding positions of all routes in the
solution are all 0, one position is selected in random and its
value is set to 1, and the others remain unchanged.
This kind of an initialization ensures for every initial solution that
each customer is served exactly once by exactly one vehicle. If the vehicle
starts and ends at different depots then the solution is infeasible.
214
Constraints handling
In order to handle the constraints, the ABC algorithm employs
Deb’s rules (Deb 2000), which are used in constrained optimization problems.
The method uses a tournament selection operator, where two solutions are
compared at a time by applying the following criteria:
Any feasible solution satisfying all constraints is preferred to
any infeasible solution violating any of the constraints.
Among two feasible solutions, the one having better fitness
value is preferred
Among two infeasible solutions, the one having the smaller
constraint violation is preferred
Because initialization with feasible solutions is a very time
consuming process and in some cases it is impossible to produce a feasible
solution randomly, the ABC algorithm does not consider the initial population
to be feasible. Structure of the algorithm already directs the solutions to
feasible region in the running process due to the Deb’s rules.
Neighborhood operators
In order to produce new solutions for the employed and onlooker
bees two neighborhood operators are applied (Eiben and Smith 2003). The
rst neighborhood operator is called SwapMutation. It is introduced as a
neighborhood operator for solving the travelling salesman problem (TSP) and
is called “2-change”. The idea of the mutation operation is to randomly
mutate the tour and hence produce a new solution g that is not very far from
the original one f. In this dissertation, the mutation operator is designed to
215
conduct customer exchanges in a random fashion. The steps for the
SwapMutation operator are as follows:
Randomly select two routes from the solution f and randomly
select two customers from each selected route.
Exchange the customers in the different routes and generate
the new solution g.
The second neighborhood operator, also based on random changes,
is called InsertMutation. The steps for the InsertMutation operator are as
follows:
Randomly select the routes from the solution f and randomly
select one customer from one selected route.
Remove the selected customer from one route to the other and
generate the new solution g.
After the new solution is produced selection process is applied
based on Deb’s method. If the new solution g is accepted instead of the
solution f, SwapMutation operator is applied in the mutated routes in order to
improve the new solution. SwapMutation is applied in the following way:
Compare all possible pairwise exchanges of customer
locations in the route to find the exchange that produces the
shortest distance.
When the pair of customers whose exchange produces the
shortest distance is found, the route is rearranged. If such pair
of customers is not determined, the route stays unchanged.
216
In employed bee phase SwapMutation operator is applied two
times. Onlooker bee phase has one difference from employed bee phase: it
applies one SwapMutation and one InsertMutation operator instead of two
SwapMutation operators.
The pseudo–code of the proposed ABC algorithm is given below:
Construct initial employed bee colony solutions;
Evaluate fitness value for each solution;
iteration = 1;
While iteration
EMPLOYED BEE PHASE
For each employed bee apply two times:
SwapMutation operator
Selection process based on Deb’s method
If the new solution is accepted, improve the new solution and
evaluate them
ONLOOKER BEE PHASE
Calculate the probability values for the solutions
For the onlookers selected depending on probabilities, the
following is applied two times:
First time Swap Mutation operator, second time Insert
Mutation operator
217
Selection process based on Deb’s method
If the new solution is accepted, improve the new solution and
evaluate them
SCOUT BEE PHASE
Determine every infeasible solution for the scout and replace them
with new produced solutions
iteration = iteration + 1
Memorize the best solution achieved so far
End While
4.3.5 GSO and IGSO for MDVRP
Hybrid algorithms are used as effective methods for obtaining
optimal solution to large scale constraint based optimization problems. In this
work, solution to MDVRP is achieved using hybrid combination of GA and
MPSO, termed as GSO. In GSO, GA and MPSO are hybridized to replace the
worst particles found by MPSO by the best particles obtained by GA and vice
versa. The GSO is modified with an enhancement step using MPSO, denoted
as Improved GSO (IGSO) in order to improve the optimal distance and
computational time of MDVRP instances obtained by GSO. The basic ideas
behind the implementation of IGSO are:
The infeasible solutions produced during the algorithmic run
are adjusted by enhancing the elites in the hybrid combination
of GA and MPSO.
218
The chances of the algorithm getting trapped into the local
optimization at a very early stage can be avoided by the IGSO
algorithm.
The step by step implementation of the proposed hybrid algorithm
for solving MDVRP is described as follows:
Solution coding and fitness evaluation
The structure of the initial solutions for the MDVRP based on
IGSO is framed based on the grouping and routing techniques. The routes
obtained using Clarke and Wright saving method is further scheduled based
on fitness function using IGSO. The description of solution representation and
fitness evaluation are provided in section 4.3.1.
Enhancement
In each generation, half of the best individuals termed as elites are
selected based on the fitness evaluation. The elites are enhanced, instead of
reproducing next generation offspring directly. The enhancement of the elites
is performed by updating the velocity and position of particles in MPSO
according to Equation 4.23 and 4.24 respectively.
)}()(int{ 22111it
gt
it
it
itt
it xprcxprcvwv (4.23)
}{ 11it
it
it vxabsx (4.24)
Selection
While the elites are enhanced using MPSO, the GA operations are
performed in parallel. In order to select parents from the selected best half
219
elites, tournament selection scheme is applied. Two enhanced elites are
selected randomly, and their fitness values are compared to select the one
with better fitness as a parent and placed in the mating pool.
Offspring generation
Parents are selected randomly from the mating pool in groups of
two and two offspring are created by performing crossover on the parent
solutions. Single individual is selected based on the mutation probability and
a new genetic offspring is generated to maintain diversity in the population. In
this work, best cost route crossover (BCRC) and route reversal mutation are
applied to increase the diversity in the population such that better elites are
produced in the next generation. The offspring generated from GA and the
enhanced elites from MPSO form a new population and their fitness is
evaluated and compared in order to select the elites for the next generation.
Stopping condition
The algorithm terminates with the number of iterations and
computation of the best solution.
The flowchart of the proposed IGSO approach is shown in
Figure 4.6. The initial swarm consists of the scheduled sequence based on the
fitness function of the MDVRP. The enhanced elites using MPSO and
offspring generated using GA form a new population for further generations
and evaluated for fitness. The process is repeated for a fixed number of
iterations and the optimal results for the MDVRP instances are determined.
220
Figure 4.6 Flowchart of the proposed IGSO algorithm
4.4 SIMULATION RESULTS
In this section, the simulation results of stages involved in solving
the Multi-Depot Vehicle Routing Problem (MDVRP) such as grouping,
routing and optimization is analyzed. The main objective of the multi-depot
vehicle routing problem is to minimize the total delivery distance, number of
vehicles required and computational time. The performance of the MDVRP
using GA, MPSO, ABC, GSO and IGSO is evaluated using a set of five
Cordeau’s instances namely p01, p02, p03, p04 and p06 taken from
http://neo.lcc.uma.es/radi-aeb/WebVRP/ online resource of University of
N
Y
Initial population
Fitness Evaluation
Stop?
Tournament selection
Best Cost Route Crossover and Route
reversal mutation
Offspring generation
New population
Enhancement using MPSO
Enhanced elites
Select best half elites and discard
others
Best Individuals
221
Malaga, Spain. The simulation is done using MATLAB R2008b on Intel i3
CPU, 2.53GHz, 4GB RAM PC.
4.4.1 Experimental Benchmark Instances
The specifications of five Cordeau’s instances such as p01
(4 depots and 50 customers), p02 (4 depots and 50 customers), p03 (5 depots
and 75 customers), p04 (2 depots and 100 customers) and p06 (3 depots and
100 customers) are shown in Table 4.1. In this work, the depots and
customers are represented in the form of coordinates and each depot has a
limited number of vehicles. Each vehicle has a constant loading capacity and
each vehicle starts and finishes at the same depot. The demand of each
customer is also known in advance and each customer is visited by a vehicle
exactly once.
Table 4.1 Specifications of benchmark instances
Parameters / Instances p01 p02 p03 p04 p06 Total No. of Customers 50 50 75 100 100
Total No. of Depots 4 4 5 2 3 Total No. of Vehicles 32 20 35 24 30
No. of Vehicles in each Depot 8 5 7 12 10 Capacity of each Vehicle 80 100 140 100 100
Best known distance in km 576.87 473.53 641.19 1001.59 876.5
4.4.2 Grouping and Routing
In grouping, the customers are assigned to the adjacent depots so
that the distance travelled by the vehicle is shorter. The customers are
clustered based on the minimum distance between customers and depots. In
the grouping phase, the Euclidean distance between the customer and the
depot is computed, and based on the minimum distance, the Cordeau’s
instances p01, p02, p03, p04, p05 are grouped and the results are shown in
Table 4.2.
222
Table 4.2 Assignment of customers to depot for benchmark instances
Depot Customers Allotted No. of customers Problem instance: p01
A(20,20) 4-13-17-18-19-25-40-41-42-44-45 11 B(30,40) 5-6-7-10-12-14-15-23-24-27-33-37-38-39-43-46-47-48-49 19 C(50,30) 1-2-8-9-11-16-21-22-26-28-29-30-31-32-34-50 16 D(60,50) 3-20-35-36 4
Problem instance: p02 A(20,20) 4-13-17-18-19-25-40-41-42-44-45 11 B(30,40) 5-6-7-10-12-14-15-23-24-27-33-37-38-39-43-46-47-48-49 19 C(50,30) 1-2-8-9-11-16-21-22-26-28-29-30-31-32-34-50 16 D(60,50) 3-20-35-36 4
Problem instance: p03
A(40,40) 3-4-5-6-9-12-15-17-18-20-25-26-27-29-30-32-34-37-39-40-44-45-47-48-50-
51-55-60-67-68-70-75
32 B(50,22) 7- 8-13-35-46-52-57-58-72 9 C(55,55) 10-11-14-19-31-38-53-54-59-65-66 11 D(25,45) 2-16-21-24-28-33-36-49-62-63-69-71-73-74 14 E(20,20) 1-22-23-41-42-43-56-61-64 9
Problem instance: p04
A(15,35)
2-5-6-7-8-13-14-15-16-17-18-21-22-23-36-37-38-40-41-42-43-44-45-46-47-48-52-53-56-57-58-59-60-61-72-73-74-75-82-83-84-85-86-87-89-91-92-93-
94-95-96-97-98-99-100
55
B(35,55)
1-3-4-9-10-11-12-19-20-24-25-26-27-28-29-30-31-32-33-34- 35-39-49-50-51-54-55-62-63-64-65-66-67-68-69-70-71-76-77-78-79-80-81-88-90
45
Problem instance: p06
A(15,20) 5-14-15-16-17-37-38-41-42-43-44-45-46-57-59-61-84-85-86-87- 91-92-93-
95-96-97-98-99-100
29
B(50,20) 1-3-9-10-11-12-20-24-25-29-30-32-33-34-35-50-51-54-55-63- 64-65-66-68-
70-71-76-77-78-79-80-81-90
33
C(35,35) 2-4-6-7-8-13-18-19-21-22-23-26-27-28-31-36-39-40-47-48-49-52-53-56-58-
60-62-67-69-72-73-74-75-82-83-88-89-94
38
In p01, 11 customers (customer IDs: 4, 13, 17, 18, 19, 25, 40, 41,
42, 44, 45) are assigned to depot A whose location is at (20,20) indicating the
x and y coordinates, 19 customers (customer IDs: 5, 6, 7, 10, 12, 14, 15, 23, 24,
27, 33, 37, 38, 39, 43, 46, 47, 48, 49) to be served by depot B located at (x-
coordinate:30, y-coordinate:40), 16 customers (customer Ids: 1, 2, 8, 9, 11, 16,
21, 22, 26, 28, 29, 30, 31, 32, 34, 50) are assigned to depot C located at (x-
coordinate:50, y-coordinate:30) and remaining customers (customer IDs: 3, 20,
35, 36) are assigned to depot D located at (x-coordinate:60, y-coordinate:50).
Similarly in the other benchmark problems such as p02, p03, p04 and p06, the
customers are grouped to the associated depots as shown in Table 4.2.
In routing phase, the customers in each group are divided into
different routes. The aim of routing is to minimize the number of routes, or
vehicles used, without violating the vehicle capacity constraint. The Clark and
223
Wright saving method is applied to solve the routing process. There are two
versions of the savings algorithm, a sequential and a parallel version. In the
sequential version exactly one route is built at a time (excluding routes with
only one customer), while in the parallel version more than one route may be
built at a time. In this work, the parallel version is used to compute the routes
at a faster rate. It is worth noting that the number of routes may be reduced
during the process of the parallel version of the savings algorithm. For
example, the two routes 0-1-2-0 and 0-3-4-0 will be combined into one route
if the connection from depot A to B is established; in that case the resulting
route becomes 0-1-2-3-4-0. Fully loaded percentage (FLP) is defined as the
number of products loaded in each vehicle in every depot and is calculated
according to Equation (4.26),
100
QLQQFL
(4.26)
where FL indicates the fully loaded percentage of each vehicle, Q stands for
capacity of each vehicle, LQ represents the loading quantity of each vehicle
during the distribution process.
Table 4.3 Route allocation for customers in instance p01
Depot Route FLP (%) D (km) NV
A(40,40)
A-4-13-17-A 85.85 28.825
5 A-18-19-A 78.69 21.21 A-25-40-A 86.82 20.65
A-41-42-44-A 92.5 34.28 A-45-A 89.65 16.25
B(50,22)
B-5-6-7-B 75.64 43.2
6
B-10-12-14-B 74.78 47.28 B-15-23-24-27-B 93.75 56.225
B-33-37-38-B 83.5 28.425 B-39-43-46-47-B 93.5 53.82
B-48-49-B 64.85 32.7
C(55,55)
C-1-2-C 78.51 28.46
5 C-8 -9-11-C 86.84 42.3
C-16-21-22-26- 28-29-C 92.4 73.3 C-30-31-32-C 86.5 27.405
C-34-50-C 72.65 19.325
D(25,45) D-3-20-D 76.56 16.42
2 D-35-36-D 85.6 14.88
224
In the routing phase, the route of vehicles, fully loaded percentage (FLP), distance (D) and the number of vehicles (NV) served by a depot are computed and the results are tabulated. The route allocation of customers for
instance p01 using the Clark and Wright saving method is shown in Table 4.3. The depots along with their coordinates are specified, and the set of routes followed by the customers with the distance is tabulated. It can be observed that for depot B located at (50, 22), there are 6 set of routes generated for 6 different vehicles. The total distance for the assigned vehicles to start from the depot B, serve all the customers and return to the depot B is 261.65 km.
Table 4.4 shows the results obtained for the benchmark instance p02. From the experiments carried out it is seen that 4 vehicles are used by depot A, 5 by depot B, 4 by depot C and 1 by depot D, respectively. The distances are computed using Euclidean distance and do not guarantee the optimal value.
Table 4.4 Route allocation for customers in instance p02
Depot Route FLP (%) D (km) NV
A(40,40)
A-4-13-17-A 85.85 28.825
4 A-18-19-A 78.69 21.21 A-25-40-41-A 86.82 20.65 A-42-44-45A 92.5 34.28
B(50,22)
B-5-6-7-10-B 75.64 43.2
5 B-12-14-15-B 74.78 47.28
B-23-24-27-33-B 93.75 56.225 B-37-38-39-43-46-B 83.5 28.425
B-47-48-49-B 93.5 53.82
C(55,55)
C-1-2-8-C 78.51 28.46
4 C-9-11-16-21-22-26-C 86.84 42.3 C-28-29-30-31-32C 92.4 73.3
C-34-50-C 86.5 27.405 D(25,45) D-3-20-35-36-D 76.56 16.42 1
The routing algorithm is run for the p03 benchmark instance consisting of 5 depots, 75 customers and the results are shown in Table 4.5.
The highest fully loaded percentage of depot A is 96.82%, which implies that the vehicle serves customers through route A-20-25-26-27-29-A in an efficient manner. Each route is followed by a vehicle, 7 vehicles serve the customer requirements in depot A, while 2 vehicles serve depot B, 3 vehicles in depot C and D, and 2 vehicles serve depot E following the routes as shown in the Table 4.5.
225
Table 4.5 Route allocation for customers in instance p03
Depot Route FLP (%) D (km) NV
A(40,40)
A-3-4-5-6-A 85.85 43.49
7
A-9-12-15-17-18-A 78.69 52.21 A-20-25-26-27-29-A 96.82 74.2 A-30-32-34-37-39-A 93.5 68.4
A-40-44-45-47-A 89.65 53.065 A-48-50-51-55-60-A 93.65 79.27
A-67-68-70-75-A 68.51 46.77
B(50,22) B-7-8-13-35-46-52-B 95.64 54.29 2 B-57-58-72-B 74.78 29.82
C(55,55) C-10-11-14-C 78.51 34.475
3 C-19-31-38-53-C 86.84 48.32 C-54-59-65-66-C 90.4 41.79
D(25,45) D-2-16-21-D 76.56 48.34
3 D-24-28-33-36-D 85.6 54.87 D-49-62-63-69-71-73-74-D 94.65 74.165
E(20,20) E-21- 22-23-41-42-43-E 91.84 50.78 2 E-56-61-64-E 81.35 23.445
The benchmark instance p04 consisting of 2 depots and 100
customers, considered as one of the large size problems is run using the
routing algorithm and the results are shown in Table 4.6. It is observed that
the total distance required by depot A with 11 vehicles to serve customers is
708.51 km, and similarly depot B used 10 vehicles with a total distance of
733.94 km, respectively such that all customers are served.
Table 4.6 Route allocation for customers in instance p04
Depot Route FLP (%) D (km) NV
A (15,35)
A-2-5-6-7-8-13-A 90.62 67.835
11
A-14-15-16-17-18-21-A 93.68 100.99 A-22-23-36-37-38-A 89.37 83.725
A-40-41-42-43-44-45-46-A 93.5 99.28 A-47-48-52-A 76.7 49.325
A-53-56-57-58-59-60-A 92.5 81.15 A-61-72-73-74-75-A 85.55 50.1
A-82-83-84-85-A 76.5 46.725 A-86-87-89-91-92-A 87.6 44.945
A-93-94-95-A 78.96 39.24 A-96-97-98-99-100-A 81.5 53.39
B (35,55)
B-1-3-4-9-10-B 90.5 48.965
10
B-11-12-19-20-24-25-B 94.5 94.885 B-26-27-28-29-30-B 93.5 104.395
B-31-32-33-34-B 88.6 72.38 B-35-39-49-B 87.6 43.325
B-50-51-54-55-62-63-B 92.5 117.46 B-64-65-66-67-B 76.5 52.435
B-68-69-70-71-76-B 96.61 78.085 B-77-78-79-80-81-B 95.5 86.585
B-88-90-B 58.67 34.37
226
The route allocation for customers in p06 instance with 3 depots
and 100 customers is shown in Table 4.7. Depot A is allocated with a set of 6
routes with a total distance of 216.49 km with loading capacities of 93.75%,
95%, 68.75%, 95%, 88.75%, and 73.75%, respectively. Similarly, depot B is
allocated with 7 routes whose total distance is 397.22 km and depot C is
allocated with 9 routes with a total distance of 476.36 km.
Table 4.7 Route allocation for customers in instance p06
Depot Route FLP (%) D (km) NV
A (15,20)
A-5-14-15-16-17-A 93.75 44.905
6
A-37-38-41-42-43-44-45-46-A 95 72.725 A-57-59-61-84-A 68.75 39.605
A-85-86-A 95 19.46 A-87-91-92-93-95-A 88.75 26.77
A-96-97-98-99-100-A 73.75 13.025
B (50,20)
B-1-3-9-10-11-B 83.75 54.055
7
B-12-20-24-25-29-30-B 83.75 88.05 B-32-33-34-35-50-51-B 98.75 54.06
B-54-55-63-64-65-B 73.75 83.555 B-66-68-70-B 82.5 37
B-71-76-77-78-79-80-B 92.5 58.5 B-81-90-B 56.25 22
C (35,35)
C-2-4-6-7-8-13-18-C 97.5 108.86
9
C-19-21-22-23-C 93.75 57.965 C-26-27-28-31-C 95 69.2 C-36-39-40-47-C 90 24.805
C-48-49-52-C 93.75 32.23 C-53-56-58-60-62-C 75 79.6 C-67-69-72-73-74-C 91.25 47.1 C-75-82-83-88-89-C 86.25 47.6
C-94-C 53.25 9
4.4.3 Impact of GA on MDVRP
In GA, each chromosome is represented by the route of each
vehicle. The set of chromosomes form the initial population thus forming the
search space. The population size decides the number of chromosomes in a
single generation. A larger population size slows down the GA run, while a
smaller value leads to exploration of a small search space. A reasonable range
of the population size is between [50,100]. Based on the encoding of
chromosomes in this work, the population size is set to 50. Genetic algorithms
generate a new route sequence by selecting two individuals in the population
227
to which the genetic operators crossover and mutation are applied. Best cost
route crossover technique is applied to exploit a better solution for MDVRP.
The choice of mutation is route reversal type, which explores a wider search
space for the MDVRP. The parameters used in solving MDVRP using GA for
five benchmark problems are shown in Table 4.8.
Table 4.8 Parameter settings for GA based MDVRP
Parameters Settings Population size Based on the number of customers
Selection Tournament selection Crossover Best Cost Route crossover Mutation Route reversal mutation
Crossover probability 0.6 Mutation probability 0.02
Elite 4
The routes are scheduled and optimized using GA to determine the
number of customers serviced (NC), number of vehicles required (NVR), and
optimal route. NVA indicates the total number of vehicles available in each
depot for the corresponding benchmark instance. These parameters are
computed and presented in Table 4.9 for benchmark instances p01, p02, p03,
p04 and p06, respectively. A total of 11 customers are serviced by depot A in
instance p01 with 4 optimal routes such as A-17-4-13-A, A-19-18-A, A-25-
41-40-A, A-44-42-45-A. Likewise depot B serves 19 customers, depot C
serves 16 customers and depot D serves 4 customers with the optimal routes
as shown in table. Instance p02 has four depots, each of which serves 11, 19,
16 and 4 customers with 3, 5, 3 and 1 optimal route respectively. A set of 16
optimal routes is generated by instance p03 using GA to serve 75 customers
from 5 depots. Though instance p04 has only 2 depots, the number of vehicles
required is too large since there are 100 customers to be served from these
depots. Depot A has 10 optimal routes to serve 55 customers while depot B with
8 optimal routes to supply 45 customers. Instance p06 has 3 depots each
generating 5, 5 and 7 optimal routes to serve 29, 33 and 38 customers respectively.
228
Table 4.9 Optimal route using GA
Depot NVA NC NVR Optimal Route Instance p01
A(20,20) 8 11 4 A-17-4-13-A, A-19-18-A, A-25-41-40-A, A-44-42-45-A B(30,40) 8 19 5 B-7-10-5-6-B, B-12-14-15-B, B-23-27-33-37-24, B-39-38-46-43-47-B, B-48-49-B C(50,30) 8 16 4 C- 8-2-9-1-C, C-28-26-11-21-16-22-C, C-29-31-32-30-C, C-50-34-C D(60,50) 8 4 1 D-35-20-3-36-D
Instance p02 A(20,20) 5 11 3 A-13-18-4-17-A, A-25-41-40-19-A, A-44-45-42-A B(30,40) 5 19 5 B-6-7-10-5-B, B- 15-12-23-14-B, B-27-24-37-33-38-B, B- 46-39-47-43-48-B, B-49-B C(50,30) 5 16 3 C-9-2-8-1-C, C- 11-16-21-29-28-26-22-C, C-34-50-31-32-30-C, D(60,50) 5 4 1 D-3-20-35-36-D
Instance p03
A(40,40)
7
32
6 A-9-3-6-5-4-A, A-18-17-20-15-26-12-25-A, A-34-27-37-29-30-32-A, A- 40-39-44-47-
45-A, A-48-60-67-55-50-51-68-A, A-75-70-A B(50,22) 7 9 2 B-52-46-7-35-8-13-57-B, B-72-58-B C(55,55) 7 11 3 C-19-14-11-10-C, C-59-65-31-38-53-54-C, C-66-C D(25,45) 7 14 3 D-2-24-16-21-D, D-49-36-71-69-28-62-73-33-63-D, D-74-D E(20,20) 7 9 2 E-61-22-42-41-56-23-43-1-E, E-64-E
Instance p04
A(15,35) 12 55 10 A-8-7-6-13-2-14-5-A, A-22-15-16-17-18-21-A, A-38-43-41-23-40-36-37-42-A, A-45-46-47-44-A, A-57-56-58-53-52-48-A, A-73-72-74-59-61-60-A, A-82-83-84-85-75-A,
A-91-86-89-87-92-A, A-97-95-94-96-93-A, A-99-98-100-A
B(35,55) 12 45 8 B-1-4-3-9-11-10-B, B-27-19-20-24-25-12-26-B, B-29-30-31-28-B, B-33-32-35-34-39-B, B-49-51-55-54-50-62-B, B-67-65-66-64-63-B, B-69-70-71-78-77-68-76-B, B-88-
90-81-79-80-B Instance p06
A(15,20) 10 29 5 A-17-5-37-15-14-38-16-A, A-41-59-45-46-44-43-42-57-A, A-61-84-86-85-A, A-96-95-87-97-92-91-93-A, A-100-98-99-A
B(50,20) 10 33 5 B-24-12-3-20-9-11-10-1-B, B-34-29-25-30-32-33-35-B, B-55-65-64-63-51-50-54-B, B-70-66-71-68-76-B, B-81-79-78-80-77-90-B
C(35,35) 10 38 7 C-7-19-8-18-6-13-4-2-C, C-26-23-22-21-27-C, C-31-36-40-39-28-C, C-49-47-48-C,
C-52-53-56-67-58-60-62-C, C-69-82-83-72-75-74-73-C, C-88-94-89-C
The computational results obtained through GA for the five
benchmark instances are shown in Table 4.10 in terms of computed optimal
distance, computational time and gap between the best known and optimal
distance. The gap between the best known distance and optimal
distances )( dP , is calculated using the following Equation (4.27),
100
bk
bkod V
VVP (4.27)
where oV is the optimal distance obtained by the optimization technique and
bkV is the best known distance of the instance. From the table, it is observed
that the optimal distances computed by GA for all the instances are near
229
optimal to the best known solution. These distances are computed based on a
total of 14 routes for instance p01, 12 routes for instance p02, 16 routes for
instance p03, 18 routes for instance p04 and 17 routes for instance p06. The
differences between the best known and optimal distances interpret near
optimal solutions for instances p04 (0.975449%), and p06 (0.682259%).
Relatively less difference is obtained for instances p01 and p02 with values
3.740877% and 1.081241% respectively, while for instance p03 GA is not
capable of finding the optimal distance since the difference
9.051919% is quite large.
Table 4.10 Computational results for benchmark instances using GA
Benchmark Instance Type
No. of customers
No. of Depots
Best known distance (km)
Optimal distance
(km)
Computational time (seconds)
Gap between best known and
optimal distance (%)
p01 50 4 576.87 598.45 4.0692 3.740877 p02 50 4 473.53 478.65 3.4207 1.081241 p03 75 5 641.19 699.23 6.8128 9.051919 p04 100 2 1001.59 1011.36 10.1081 0.975449 p06 100 3 876.50 882.48 9.3177 0.682259
4.4.4 Evaluation of MPSO for MDVRP
The parameters of MPSO are set as shown in Table 4.11 based on
empirical study and studies from literature. The acceleration constants c1 and
c2 are set to 1.28 and initial population of swarm varies with respect to the
customers and depots in each instance. Inertia weight, w, determines the
search behavior of the algorithm. Large values for w facilitate searching new
locations whereas small values provide a finer search in the current area. A
balance can be established between global and local exploration by decreasing
the inertia weight during the execution of the algorithm.
230
Table 4.11 Parameter settings for MPSO based MDVRP
Parameters Settings
Swarm size Based on the number of customers
Cognitive factor, c1 1.28
Social coefficient, c2 1.28
First inertia weight, wint 0.9
Last inertia weight, wend 0.4
No. of iterations 300
The routes for each vehicle are taken as the swarm particles and the
MPSO algorithm iterates to determine the pbest and gbest to optimize the
delivery distance. The scheduled routes for the five benchmark instances p01,
p02, p03, p04 and p06 are optimized using MPSO, and the route length,
optimal route, number of customers serviced and number of vehicles required
are evaluated and the results are shown in Table 4.12.
Instance p01 has generated a total of 13 routes from 4 depots thus
servicing all the 50 customers. In instance p02 with 4 depots, depot A
produces 3 optimal routes, depot B generates 4 optimal routes, while depot C
framed 3 optimal routes and depot D with one optimal route. A total of 75
customers are served in instance p03 with 5 depots requiring 14 vehicles
among the available 35 vehicles. Similarly, instance p04 has generated 15
optimal routes with 2 depots and instance p06 constructed 17 optimal routes
with 3 depots.
231
Table 4.12 Optimal route using MPSO
Depot NVA NC NVR Optimal Route Instance p01
A(20,20) 8 11 3 A-13-18-4-17-A, A-40-41-25-19-A, A-44-45-42-A B(30,40) 8 19 5 B-7-6-5-10-B, B-15-12-23-14-B, B-33-37-24-27-38-B, B-39-46-48-43-47-B, B-49-B C(50,30) 8 16 4 C-9-1-8-2-C, C-26-28-29-21-C, C-16-11-22-C, C-30-34-50-31-32-C D(60,50) 8 4 1 D-20-3-36-35-D
Instance p02 A(20,20) 5 11 3 A-13-19-17-4-18-A, A-25-41-40-42-A, A-45-44-A B(30,40) 5 19 4 B-6-7-10-5-B, B-12-15-14-24-23-B, B-39-38-27-43-37-33-B, A-47-48-46-49-A C(50,30) 5 16 3 C-9-1-8-2-C, C-26-28-29-21-16-11-22-C, C-30-34-50-31-32-C D(60,50) 5 4 1 D-3-20-35-36-D
Instance p03
A(40,40) 7 32 5 A-12-9-3-6-5-4, A-25-18-17-29-20-15-27-26-A, A-39-34-37-30-32-A,A- 40-44-47-48-45-A,A-67-75-70-60-68-51-50-55-A,
B(50,22) 7 9 2 B-52-46-7-35-8-13-57-B, B-72-58-B C(55,55) 7 11 3 C-11-14-19-10-C, C-59-54-53-38-31-65-C, C-66-C D(25,45) 7 14 2 D-21-2-24-16-28-D, D-74-36-71-69-62-73-33-63-49-D E(20,20) 7 9 2 E-56-41-42-22-61-1-43-23-E, E-64-E
Instance p04
A(15,35) 12 55 8 A-8-5-14-15-2-13-6-7-A,A-37-16-17-36-18-21-23-22-A,A-38-44-45-46-47-40-41-42-43-A, A-52-48-57-56-58-53-A, A-60-73-72-75-74-59-61-A, A-84-85-82-83-A, A-93-
92-87-89-86-91-A, A-94-95-97-100-98-99-96-A
B(35,55) 12 33 7 B-11-10-1-12-4-3-9-B, B-26-28-27-19-20-29-24-25-B, B-30-31-33-34-35-32-B, B-51-50-54-55-39-49-B, B-67-62-64-63-66-65-B, B-78-71-70-69-76-68-77-B, B-79-81-90-
88-80-B Instance p06
A(15,20) 10 29 5 A-5-37-15-14-38-16-17-A,A- 45-61-44-43-41-57-42-59-46-A, A-86-84-85-A,A- 91-92-97-87-95-96-93-A, A-100-98-99-A
B(50,20) 10 33 5 B-20-11-10-1-12-24-3-9-B, B-29-25-33-30-32-35-34-B, B-55-54-50-51-65-64-63-B, B-68-71-66-70-76-B,B- 80-77-79-78-81-90-B
C(35,35) 10 38 7 C-6-13-2-4-19-7-8-18-C, C-26-23-22-21-27-C, C-31-36-40-39-28-C, C-49-47-48-52-
C, C-67-56-58-60-62-69-53-C, C-88-82-83-73-74-75-72-C, C-94-89-C
The results obtained through MPSO are shown in Table 4.13 in
terms of optimal distance, computational time, and gap between best known
distance and optimal distance. The gap between the best known distance and
optimal distances obtained by MPSO is computed according to Equation
(4.27). It can be inferred from the results that the MPSO algorithm is capable
of obtaining near optimal solutions when compared with the best known
distance. For instance, the best known distance of p02 is 473.53 km while the
optimal value is 475.47 km with a gap of 0.41%. It can also be observed that
the computational time increase with the size of the problem.
232
Table 4.13 Computational results for benchmark instances using MPSO
Benchmark Instance Type
No. of customers
No. of Depots
Best known distance (km)
Optimal distance (km)
Computational time (seconds)
Gap between best known and optimal distance (%)
p01 50 4 576.87 580.35 3.6319 0.603255 p02 50 4 473.53 475.47 2.9434 0.409689 p03 75 5 641.19 687.65 5.1899 7.245902 p04 100 2 1001.59 1008.55 7.7211 0.694895 p06 100 3 876.50 880.35 8.1418 0.439247
4.4.5 Solution of MDVRP based on ABC
In ABC, a colony of artificial forager bees act as agents searching
for artificially rich food sources analogous to good solutions for a given
problem. In this work, the position of a food source represents a set of
possible routes to the multi-depot vehicle routing problem. The parameters
that govern the ABC algorithm are colony size, number of food sources, limit,
number of employed bees, number of onlooker bees and maximum number of
iterations. The colony size is based on the number of customers in each
MDVRP instance. Smaller the colony size faster the convergence rate and
vice versa. The number of employed bees and the onlooker bees are usually
set to one half of the colony size for best results. The number of food sources
is always chosen to be equal to the number of employed bees. The ABC
algorithm is run for a maximum of 800 iterations irrespective of the size of
the problem. The parameters and their values used for running the ABC
algorithm for MDVRP are shown in Table 4.14.
Table 4.14 ABC parameters for MDVRP
S.No Parameters Value
1 Colony size Based on No. of customers
2 No. of food sources No. of employed bees
3 Food source limit No. of onlooker bees x dimension of the problem
4 No. of employed bees No. of customers/2
5 No. of onlooker bees No. of customers/2
6 Maximum No. of iterations 800
233
The optimal route evaluated using ABC along with the number of
vehicles required and the number of customers serviced for the problem
instances is presented in Table 4.15. In instance p01, Depot A requires 3
vehicles for generating three optimal routes A-13-17-4-18-A, A-40-41-25-19-
A, and A-42-45-44-A, respectively. In the same way, from the table it can be
seen that depot B requires 5 vehicles for generating the optimal routes B-6-7-
10-5-B, B-15-12-23-14-B, B-33-37-24-27-38-B, B-39-46-48-43-47-B, and B-
49-B. Likewise depots C and D produce optimal routes with 3 and 1 vehicles
respectively. Similarly, for instance p02, ABC locates the optimal routes for
depots A, B, C and D each serving 11, 19, 16 and 4 customers with
constructive number of vehicles such as 3, 4, 3 and 1 respectively. ABC on
instance p03 with 5 depots produces 5 optimal routes from depot A, 2 from
depot B, 3 from depot C, 2 from depot D and 2 from depot E.
Table 4.15 Optimal route using ABC
Depot NVA NC NVR Optimal Route Instance p01
A(20,20) 8 11 3 A-13-17-4-18-A, A-40-41-25-19-A, A-42-45-44-A B(30,40) 8 19 5 B-6-7-10-5-B, B-15-12-23-14-B, B-33-37-24-27-38-B, B-39-46-48-43-47-B, B-49-B C(50,30) 8 16 3 C-2-9-1-8-C, C-11-22-26-28-29-21-16-C,C-30-32-31-50-34-C, D(60,50) 8 4 1 D-20-3-36-35-D
Instance p02 A(20,20) 5 11 3 A-19-13-18-4-17-A, A-42-40-41-25-44-A, A-45-A B(30,40) 5 19 4 B-7-5-10-12-6-B, B-27-33-15-14-24-23-B, B-47-43-48-46-38-39-37-B, B-49-B C(50,30) 5 16 3 C-1-11-9-2-8-C, C-28-31-26-22-16-30-21-29-C, C-50-32-34-C D(60,50) 5 4 1 D-36-35-20-3-D
Instance p03
A(40,40) 7 32 5 A-6-3-9-12-4-5-A, A-25-18-17-29-20-15-27-26-A, A-32-39-34-37-30-40-A, A-45-60-47-48-51-44-50-55-A, A-67-75-68-70-A
B(50,22) 7 9 2 B-52-46-7-35-8-13-57-B, B-72-58-B C(55,55) 7 11 3 C-10-19-14-11-C, C-65-31-38-53-54-59-C, C-66-C D(25,45) 7 14 2 D-2-24-16-28-21-D, D-49-74-36-71-69-62-73-33-63-D E(20,20) 7 9 2 E-22-42-41-56-23-43-1-61-E, E-64-E
Instance p04
A(15,35) 12 55 8 A-8-7-6-13-2-15-14-5-A,
A-22-23-21-18-36-17-16-37-A, A-46-45-44-38-43-42-41-40-47-A, A-52-48-57-56-58-53-A, A-73-72-75-74-59-61-60-A, A-85-84-83-82-A, A-89-86-91-93-92-87-A, A-96-99-98-100-97-95-94-A
B(35,55) 12 33 7 B-3-9-11-10-1-12-4-B, B-24-25-26-28-27-19-20-29-B, B-31-30-32-35-34-33-B, B-51-50-54-55-39-49-B,B- 67-65-66-63-64-62-B, B-78-77-68-76-69-70-71-B, B-80-88-90-81-79-B
Instance p06
A(15,20) 10 29 4 A-17-16-38-14-15-41-42-37-5-A, A-59-46-45-84-61-44-43-57-A, A-91-86-85-92-87-A, A-95-96-99-93-100-98-97-A
B(50,20) 10 33 5 B-20-11-10-1-12-25-24-3-9-B, B-29-50-33-51-30-32-35-34-B, B-63-55-54-65-66-64-B, B-68-76-70-71-78-79-77-B, B-90-81-80-B
C(35,35) 10 38 6 C-19-7-8-18-6-13-2-21-4-C, C-27-22-23-26-28-C, C-47-40-39-31-36-C,C- 52-48-49-56-53-C, C-67-69-62-60-58-73-72-C,C- 89-83-82-88-75-74-94-C
234
Instance p04 has a total of 24 vehicles available from 2 depots, and
the optimal routes framed require only 15 vehicles shared as 8 and 7 among
depots A and B correspondingly. It is also observed from the table that
problem instance p06 with 3 depots has 30 vehicles available, among which
50% of the vehicles are only used to serve all the customers.
The evaluated results in terms of optimal distance and
computational time are shown in Table 4.16. The optimal distance obtained
by instance p01 for serving all customers from 4 depots with a total of 12
vehicles is 584.72 km. Instance p02 has generated 11 optimal routes with near
optimal distance of 477.91 km. For instance p03 with 75 customers and 5
depots, the best known distance reported is 641.19 km while the proposed
ABC is capable of achieving an optimal distance of 668.84 km. For instances
p04 and p06, the gap between the best known distance and the obtained
optimal is found to be 0.9065586% and 0.635482% respectively.
Comparatively, ABC takes a longer time to converge in spite of the less
tuning parameters. This fact is evident from the very high computational time
tabulated for the instances considered in this work.
Table 4.16 Computational results for benchmark instances using ABC
Benchmark Instance Type
No. of customers
No. of Depots
Best known distance
(km)
Optimal distance
(km)
Computational time (seconds)
Gap between best known and optimal
distance (%) p01 50 4 576.87 584.72 8.1456 1.3607919 p02 50 4 473.53 477.91 8.9547 0.9249678 p03 75 5 641.19 668.84 12.6472 1.1930941 p04 100 2 1001.59 1010.67 17.5541 0.9065586 p06 100 3 876.50 882.07 18.0546 0.635482
4.4.6 MDVRP based on GSO and IGSO
The hybrid combination of GA and MPSO delivered better optimal
distance when compared to the individual performance of both GA and
MPSO for the p01, p02, p03, p04 and p06 benchmark instances. The
235
parameters of GSO and IGSO are set as shown in Table 4.17. The size of
population is based on the customers in each MDVRP instance. The social
and cognitive factors are set according to 40 21 cc . The inertia weight is
varied in order to maintain a proper balance between the exploration and
exploitation of the algorithm and is set such that 112
c+c 21 . In the GA
constituent of IGSO, best cost route crossover with a crossover rate of 0.7 and
route reversal mutation with a rate of 0.028 are chosen as the optimal setting.
The hybridization parameter applicable in GSO is set to 0.5 to ensure that
individuals of GA and MPSO participate equally in the hybrid combination.
Table 4.17 Parameter settings for GSO and IGSO
Parameters GSO IGSO Population size Instance dependant Instance dependant
Maximum inertia weight 0.9 0.9 Minimum inertia weight 0.4 0.4
Initial velocity 0 0 Initial position Random Random
Cognitive factor (c1) 1.28 1.24 Social factor (c2) 1.28 1.24
Error gradient 1e-25 1e-25 Type of Crossover Best cost route crossover Best cost route crossover Type of Mutation Flip bit Route reversal mutation
Crossover rate 0.6 0.7 Mutation rate 0.02 0.028
Maximum number of iterations Instance dependant Instance dependant Hybridization factor 0.5 -
For the problem instances, the number of vehicles required and the
optimal route are computed by GSO and IGSO and the results are presented
in Table 4.18 (a) and Table 4.18 (b). From Table 4.18 (a), for instance p03,
Depot A requires 5 vehicles for generating five optimal routes such as A-5-6-
3-9-12-4-15-A, A-20-29-30-17-18-25-26-27-A, A-40-44-32-39-34-37-A, A-
47-60-45-48-55-50-51-A, A-70-68-75-67-A respectively. In the same way,
from the table it can be seen that depot B requires one vehicle for generating
the optimal route B-13-8-35-7-58-72-46-52-57-B. Likewise, depot C, D and E
require 2 vehicles each for finding the optimal route.
236
Table 4.18(a) Optimal route using GSO
Depot NVA NC NVR Optimal Route Instance p01
A(20,20) 8 11 3 A-13-19-17-4-18-A, A-44-42-40-41-25-A, A-45-A B(30,40) 8 19 4 B-7-6-12-10-5-B, B-15-14-24-23-27-B,B-46-38-39-33-37-43-B,B- 48-47-49-B C(50,30) 8 16 3 C-9-11-1-8-2-C, C-28-31-26-22-16-30-21-29-C, C-32-34-50-C D(60,50) 8 4 1 D-20-3-36-35-D
Instance p02 A(20,20) 5 11 2 A-13-19-17-4-18-A, A-40-42-45-44-25-41-A B(30,40) 5 19 3 B-7-5-10-12-6-B, B-33-15-37-14-24-23-27-B, B-46-48-43-47-39-49-38-B C(50,30) 5 16 3 C-11-1-8-2-16-9-C, C-29-22-28-31-26-32-30-21-C, C-34-50-C D(60,50) 5 4 1 D-3-20-35-36-D
Instance p03
A(40,40) 7 32 5 A-5-6-3-9-12-4-15-A, A-20-29-30-17-18-25-26-27-A, A-40-44-32-39-34-37-A, A-47-
60-45-48-55-50-51-A, A-70-68-75-67-A B(50,22) 7 9 1 B-13-8-35-7-58-72-46-52-57-B C(55,55) 7 11 2 C-11-14-19-10-31-C, C-38-65-66-59-54-53-C D(25,45) 7 14 2 D-2-21-28-16-24-D, D-69-71-36-74-49-63-33-73-62-D E(20,20) 7 9 2 E-41-42-22-61-1-43-23-56-E, E-64-E
Instance p04
A(15,35) 12 55 8 A-13-6-7-8-5-14-15-2-A, A-18-21-23-22-37-16-17-36-A, A-43-42-41-40-47-46-45-44-38-A, A-57-56-58-53-52-48-59-A, A-72-75-74-61-84-83-82-60-73-A, A-87-85-91-86-
89-A, A-98-93-99-96-94-95-97-92-A, A-100-A
B(35,55) 12 45 6 B-3-9-11-10-1-12-4-B, B-30-19-27-28-26-25-24-29-20-B, B-33-34-35-32-31-39-B,B-51-
55-54-50-62-49-64-63-B, B-69-67-68-65-66-70-B, B-80-76-88-90-71-81-78-79-77-B Instance p06
A(15,20) 10 29 4 A-14-38-43-15-41-42-37-5-17-16-A, A-61-44-57-59-46-45-84-A, A-91-87-92-85-86-A
B(50,20) 10 33 4 B-11-10-1-12-25-24-29-3-9-20-B, B-32-30-51-33-50-54-34-35-B, B-65-68-55-70-63-
64-66-B, B-77-79-78-81-71-90-76-80-B
C(35,35) 10 38 6 C-19-7-8-18-6-13-2-21-4-C, C-27-22-23-26-28-C,C- 47-40-39-31-36-C,C-56-58-60-
48-49-52-53-C, C-67-69-62-73-72-74-75-C, C-83-94-89-88-82-C
Based on the results obtained in Table 4.18 (b), to serve all the
customers, instance p01 with 50 customers and 4 depots requires 11 vehicles,
while instance p02 requires 9 vehicles. Instance p03 has 5 depots and the
optimal number of routes from each depot A, B, C, D and E is 5, 1, 2, 2, and 1
respectively. The number of optimal routes generated by instance p04 is 7 and
5 corresponding to depots A and B. Instance p06 serves 29 customers from
depot A with 4 vehicles, 33 customers from depot B with 4 vehicles and 38
customers from depot C with 6 vehicles.
237
Table 4.18(b) Optimal route using IGSO
Depot NVA NC NVR Optimal Route Instance p01
A(20,20) 8 11 3 A-19-13-18-4-17-A, A-40-41-25-44-42-A, A-45-A B(30,40) 8 19 4 B-7-5-10-12-6-B, B-15-27-23-24-14-B, B-46-38-39-33-37-43-B, B-49-47-48-B C(50,30) 8 16 3 C-2-8-1-11-9-C, C-29-28-31-26-22-16-30-21-C, C32-34-50-C D(60,50) 8 4 1 D-3-36-35-20-D
Instance p02 A(20,20) 5 11 2 A-19-13-18-4-17-A, A-40-41-25-44-45-42-A B(30,40) 5 19 3 B-7-5-10-12-6-B, B-37-15-33-27-23-24-14-B, B-43-48-46-38-49-39-47-B C(50,30) 5 16 3 C-9-16-2-8-1-11-C, C-31-26-32-30-21-29-22-28-C, C-50-34-C D(60,50) 5 4 1 D-3-36-35-20-D
Instance p03
A(40,40) 7 32 5 A-12-9-3-6-5-15-4-A, A-17-18-25-26-27-20-29-30-A, A-40-44-32-39-34-37-A, A-45-60-47-48-51-50-55-A, A-75-68-70-67-A
B(50,22) 7 9 1 B-72-46-52-57-13-8-35-7-58-B C(55,55) 7 11 2 C-14-19-10-31-11-C, C-53-38-65-66-59-54-C D(25,45) 7 14 2 D-21-2-24-16-28-D, D-69-71-36-74-49-63-33-73-62-D E(20,20) 7 9 1 E-41-42-22-61-1-43-23-56-64-E
Instance p04
A(15,35) 12 55 7 A-5-8-7-6-13-2-15-14-A, A-21-23-22-37-16-17-36-18-A, A-45-44-38-43-42-41-40-47-46-A, A-57-56-58-53-52-48-59-A, A-74-75-72-73-60-82-83-84-61-A, A-86-91-85-87-
89-A, A-98-93-99-100-96-94-95-97-92-A
B(35,55) 12 45 5 B-11-9-3-4-12-1-10-25-26-28-27-19-30-20-29-24-B, B-33-34-35-32-31-39-B, B-51-55-54-50-62-49-64-63-B, B-67-69-70-66-65-68-B, B-79-78-81-71-90-88-76-80-77-B
Instance p06
A(15,20) 10 29 4 A-41-15-43-14-38-16-17-5-37-42-A, A-44-57-59-46-45-84-61-A, A-91-87-92-85-86-A, A-95-96-99-93-100-98-97-A
B(50,20) 10 33 4 B-20-11-10-1-12-25-24-29-3-9-B, B-32-30-51-33-50-54-34-35-B, B-65-68-55-70-63-64-66-B, B-76-90-71-81-78-79-77-80-B
C(35,35) 10 38 6 C-6-18-8-7-19-4-21-2-13-C, C-27-22-23-26-28-C, C-36-31-39-40-47-C, C-58-56-53-52-49-48-60-C, C-75-67-69-62-73-72-74-C, C-88-82-83-94-89- C
The experimental results obtained through GSO and IGSO for the
benchmark instances considered in this work are shown in Table 4.19 in terms
of optimal distance, computational time and gap between best known distance
and optimal distance. The gap between the best known distance (BKD) and
optimal distance obtained by IGSO is calculated according to Equation (4.27).
The optimal distances computed through IGSO for instances p01, p02, p03,
p04 and p06 are 576.94 km, 474.03 km, 644.29 km, 1001.94 km, and 876.68
km respectively. The gap between the best known distance and the computed
optimal distance for IGSO is 0.012134% in the case of instance p01,
0.10559% in p02, 0.483476% in p03, 0.034944% in p04, and 0.020536% in
p06. The reduction in gap for instances in IGSO show that the enhancement
step introduced in the hybrid combination has improved the optimal distance
when compared to GSO. The computational time obtained shows that the
238
IGSO algorithm is capable of converging at a faster rate towards the optimal
solution than the GSO algorithm.
Table 4.19 Computational results for MDVRP instances using GSO and IGSO
Instance No. of customers
No. of Depots
BKD (km) GSO IGSO
Optimal distance
(km) CT (s) Gap (%)
Optimal distance
(km) CT (s) Gap (%)
p01 50 4 576.87 578.52 1.5316 0.286026 576.94 1.3725 0.012134 p02 50 4 473.53 474.85 1.4259 0.278757 474.03 0.9750 0.10559 p03 75 5 641.19 653.78 2.3659 1.963537 644.29 2.2773 0.483476 p04 100 2 1001.59 1005.73 3.0191 0.413343 1001.94 2.3918 0.034944 p06 100 3 876.50 877.19 3.0976 0.078722 876.68 2.4094 0.020536
4.5 COMPARATIVE ANALYSIS AND DISCUSSIONS
The experimental results obtained on MDVRP instances using the
heuristic optimization techniques such as GA, MPSO, ABC, GSO and IGSO
are compared and discussed in terms of optimal distance, robustness,
computational efficiency and algorithmic efficiency as follows:.
4.5.1 Optimal Distance
The optimal distance obtained for the benchmark instances using GA,
MPSO, ABC, GSO and IGSO algorithms are compared with the existing
approaches in literature such as genetic clustering (GC) (Thangiah and Salhi
2001) and GA (Ombuki-Berman and Hanshar 2009). The optimal distance
obtained for the five MDVRP instances using the existing approaches and
proposed algorithms are presented in Table 4.20. For instance p03, the distance
obtained by proposed GA, MPSO, ABC, GSO and IGSO are comparatively less
than the GA (Ombuki-Berman and Hanshar 2009) in literature by 1.08%,
2.72%, 5.38%, 7.51% and 8.85% respectively. MPSO yields minimum distance
for instances p01, p02, p03, p04 and p06 when compared with GA with a
difference of 3.02%, 0.66%, 1.66%, 0.28% and 0.24% respectively. Though
239
ABC took more time to generate the optimal route, the distances are reduced
when compared with the proposed GA. For instance p01, the distance obtained
by ABC is 2.29% lesser than that obtained by GA. In the same way, the
optimal distance obtained by ABC is 0.15% (instance p02), 4.34% (instance
p03), 0.07% (instance p04), and 0.05% (instance p06) lesser than GA. The
improvement in optimal distance obtained by IGSO over GSO is given by
0.273% in p01, 0.173% in p02, 1.45% in p03, 0.377% in p04, and 0.058% in
p06. The enhancement step introduced in IGSO has shown a significant
improvement in terms of optimal distance specifically in instance p03. Thus
from the comparative analysis in terms of the optimal distance, it can be
revealed that IGSO is a better approach to solve the chosen MDVRP
instances.
Table 4.20 Comparative analysis in terms of optimal distance
Instance Type
No. of customers
No. of Depots
Best known
distance (km)
Distance reported in literature (km)
Distance obtained by the proposed approaches (km)
GC (2001)
GA (2009) GA MPSO ABC GSO IGSO
p01 50 4 576.87 591.73 622.18 598.45 580.35 584.72 578.52 576.94 p02 50 4 473.53 463.15 480.04 478.65 475.47 477.91 474.85 474.03 p03 75 5 641.19 694.49 706.88 699.23 687.65 668.84 653.78 644.29 p04 100 2 1001.59 1062.38 1024.78 1011.36 1008.55 1010.67 1005.73 1001.94 p06 100 3 876.50 976.02 908.88 882.48 880.35 882.07 877.19 876.68
4.5.2 Robustness
The robustness of the heuristic techniques applied in this work is
evaluated based on the effective vehicle management without violating the
capacity limit of the vehicles. Computational results in terms of vehicles
required (NVR) and total number of vehicles available (NVA) is shown in
Table 4.21. Among the proposed intelligent algorithms, it is found from the
table that GA utilized more number of vehicles for all the instances to
determine the optimal distance, thus increasing the vehicle utilization cost.
For test instances p02, p03, p04, it is observed that MPSO and ABC use the
240
required vehicles in a similar manner without any difference. ABC exploits
less number of vehicles in a more efficient way in the case of instance p01
and p06 with 37.5% and 50% respectively, while MPSO utilizes 40.63% and
56.67% of the vehicles available. This fact implies that ABC is robust against
MPSO, in spite of MPSO achieving the optimal distance.
While comparing GSO and IGSO, no significant difference in
vehicle utilization efficiency is found in instances p01, p02 and p06. IGSO
has shown significant improvement over GSO in utilizing vehicles
specifically for instances p03 (31.43%) and p04 (50%). Overall, for all the
test instances GSO and IGSO used less number of vehicles when compared
with GA, MPSO and ABC algorithms. Considerably, from the results based
on the vehicle utilization efficiency for instances p03 and p04, it can be
demonstrated that IGSO is robust over GA, MPSO, ABC and GSO
algorithms.
Table 4.21 Comparative analysis of vehicle utilization efficiency
Instance Type
No. of customers
No. of Depots NVA NVR (%)
GA MPSO ABC GSO IGSO p01 50 4 32 43.75 40.63 37.50 34.38 34.38 p02 50 4 20 60.00 55.00 55.00 45.00 45.00 p03 75 5 35 45.71 40.00 40.00 34.29 31.43 p04 100 2 24 75.00 62.50 62.50 58.33 50.00 p06 100 3 30 56.67 56.67 50.00 46.67 46.67
4.5.3 Computational Time
The computational efficiency of the bio-inspired heuristics is
evaluated in terms of the time taken for the algorithm to converge. The
computational time of the proposed techniques for the benchmark instances
are presented in Table 4.22. It can be observed from the table that
computational time of IGSO (2.3918s) for instance p04 is comparatively less
than GA (10.1081s), MPSO (7.7211s), ABC (17.5541s) and GSO (3.0191s).
241
Overall for all the instances it is found that ABC consumes more time to
determine the optimal route and distance. On the average, the computational
time of ABC is 48.39% higher than GA, 57.73% higher than MPSO, 85.58%
higher than GSO and 82.5% higher than IGSO. The longer time for
convergence required by ABC leads to a setback in applying the ABC
technique to solve practical MDVRP instances. Computational results
indicate that the time taken by IGSO is comparatively less since the hybrid
combination of GA and MPSO minimizes the chances of the algorithm
getting trapped in the local minima.
Table 4.22 Comparative analysis of computational time
Instance Type
No. of customers
No. of Depots
Computational time (s) GA MPSO ABC GSO IGSO
p01 50 4 4.0692 3.6319 8.1456 1.5316 1.3725 p02 50 4 3.4207 2.9434 8.9547 1.4259 0.9750 p03 75 5 6.8128 5.1899 12.6472 2.3659 2.2773 p04 100 2 10.1081 7.7211 17.5541 3.0191 2.3918 p06 100 3 9.3177 8.1418 18.0546 3.0976 2.4094
4.5.4 Algorithmic Efficiency
The algorithmic efficiency of the heuristic algorithms used to solve
MDVRP in this work is calculated using BigO notation in terms of the
estimated time and calculated time of the algorithm. The algorithmic
efficiency of the proposed intelligent heuristics is evaluated for the problem
instances and the results are presented in Table 4.23. Based on the size of the
code the efficiency varies significantly. High amount of programming lines in
the code leads to a decrease in the efficiency of the algorithm. The
algorithmic efficiency of instance p02 for GA is 88.21%, for MPSO is
90.27%, for ABC is 95.88%, for GSO is 93.11% and for IGSO is 92.81%. On
an average, the algorithmic efficiency for all the MDVRP instances in this
242
work, using GA is 89.15%, 90.84% for MPSO, 95.19% for ABC, 93.22% for
GSO and 92.71% for IGSO. It is observed that due to the less number of
control parameters in ABC, the algorithmic efficiency is improved over GA,
MPSO, GSO and IGSO algorithms.
Table 4.23 Comparison of algorithmic efficiency
Instance Type
No. of customers
No. of Depots
Algorithmic efficiency (%) GA MPSO ABC GSO IGSO
p01 50 4 89.77 90.48 94.52 92.56 92.27 p02 50 4 88.21 90.27 95.88 93.11 92.81 p03 75 5 89.09 91.12 94.37 93.71 92.64 p04 100 2 89.56 90.55 95.06 93.29 92.98 p06 100 3 89.14 91.79 96.11 93.43 92.86
4.6 SUMMARY
The intelligent heuristics such as GA, MPSO, ABC, GSO and
IGSO are proposed in this chapter for solving multi-depot vehicle routing
problem. Initially the customers are grouped based on its distance from the
depot. Further the customers of the same depot are assigned to routes using
Clarke and Wright saving method. This procedure also ensures that two routes
can be easily merged into a single route. The routes are later scheduled and
optimized by the proposed intelligent heuristics. The objective of multi-depot
vehicle routing problem is to find routes for vehicles to service all the
customers at a minimal cost in terms of number of routes and total travel
distance, without violating the capacity and depot limits.
The effectiveness of the proposed techniques is tested on a set of
five different Cordeau’s benchmark instances namely p01, p02, p03, p04 and
p06 in the MATLAB R2008b environment. The intelligent heuristics are
applied to evaluate the optimal routes, number of vehicles required, and
optimal distance of the MDVRP instances. In addition to the problem
243
dependant evaluations such as optimal distance and robustness, additional
assessments in terms of computational time and algorithmic efficiency of the
proposed algorithms are also estimated.
Considering the optimal distance, it is observed from Table 4.20
that IGSO showed improvements among the other proposed algorithms such
as GA, ABC, MPSO and GSO. From Table 4.21, it can be seen that IGSO
utilizes relatively less number of available vehicles when compared with GA,
MPSO, ABC and GSO, thus implying robustness. Likewise, the
computational efficiency of IGSO is improved thus demonstrating earlier
convergence rate, which is evident from the results presented in Table 4.22.
The algorithmic efficiency of ABC (Table 4.23) is reasonably higher due to
the less number of control parameters in the algorithm. Thus based on the
investigations carried out on the proposed intelligent algorithms, it is
observed that IGSO is an appropriate approach for solving practical MDVRP
instances based on the optimal solution, robustness and computational
efficiency.
In future, it would be of considerable interest to incorporate several
practical constraints such as urgency delivery, random serve time and non-
uniform vehicle capacity to the MDVRP. This work may be extended with
new optimization techniques like Bacterial Foraging (BFO), Biogeography-
based optimization (BBO) and Artificial Immune Systems (AIS). The
experimental results may be used to analyze, compare and determine better
optimization technique in forthcoming prospects.