HYBRID ANT COLONY ALGORITHM FOR THE MULTI-DEPOT PERIODIC OPEN CAPACITATED ARC ROUTING PROBLEM

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.11, No.1, January 2020
DOI : 10.5121/ijaia.2020.11105 53
HYBRID ANT COLONY ALGORITHM FOR THE
MULTI-DEPOT PERIODIC OPEN CAPACITATED ARC
ROUTING PROBLEM
Bilal Kanso
Department of Computer Science, Lebanese University, Beirut
ABSTRACT
In this paper we present a hybrid technique that applies an ant colony optimization algorithm followed by
simulated annealing local search approach to solving the Multi-Depot Periodic Open Capacitated Arc
Routing Problem (MDPOCARP). This problem is a new variant of OCARP that has never been studied in
the literature and consists of determining optimal routes in each period where each route starts from a
given depot, visits a list of required edges and finishes by the last one. The final edge of the route is not
required to be a depot. We developed a constructive heuristic, called Nearest Insertion Heuristic (NIH) to
build an initial solution. The proposed algorithm is evaluated on three different benchmarks sets and
numerical results show that the proposed approach achieves highly efficient results.
KEYWORDS
Arc routing problem, Periodic, Multi-Depot, Meta heuristic, Ant Colony Optimization, Simulated
Annealing algorithm.
1. INTRODUCTION
The Capacitated Arc Routing Problem (CARP) is a hard-combinatorial optimization problem
introduced which consists of a network (or graph) where edges have a positive length and some
of the edges (so-called required edges) are served [1]. The CARP aims to find an optimal set of
vehicles routes of total minimum length. Each route should start and end at a predefined depot
and each required edge should be serviced in exactly one route. A wide range of real word
applications can be modeled as arc routing problems such as transportation industry, collection of
household waste, mail delivery, meter reader, etc.
The periodic CARP (PCARP) is a generalization of CARP in which routes are built over M
periods that represent predefined schedules for visiting customers. A solution of the PCARP is a
set of optimal routes for each planning period so that the both period constraint and capacitated
vehicles along routes are satisfied. The Multi-Depot Open CARP (MDOCARP) is an extension of
CARP in which vehicles can depart at more than one depot for servicing customers. The vehicles
are not required to return back to a depot as in the CARP, instead they finish their route at the last
served task. The Multi-Depot Periodic Open CARP (MDPOCARP) is a variant of OCARP in
which the customer demands are served throughout planning periods. In MDPOCARP, there is a
frequency f for each customer c indicating how many times within the planning period associated
to c, the customer c must be visited.
In this paper, we intend to contribute to solve MDPOCARP. Then, we focus on three features of
the CARP problem at the same time: the periodic, open and multi-depot CARP. The motivation
for addressing this problem comes from the wide range of real-world applications that can be
seen as MDPOCARP. Indeed, the study of both periodic and multi-depot CARP is receiving

54
increasing attention, mostly in the last years, thanks to many real-world applications related to
recycling, periodic deliveries of products to customers, periodic visits for providing specific
services, Household waste collection, transportation of goods to supermarkets by manufacturing
companies using trucks, maintenance of equipments, quality inspectors or home health care. In all
these applications, vehicles start from a particular depot and finish at the last serviced customer.
There is no need to return back at the end of the working period to the depots.
This work proposes an Ant Colony Optimization (ACO) combined with the guided local search
meta-heuristic Simulated Annealing (SA) for the MDPOCARP. Computational experiments were
performed to evaluate the proposed method. The results are compared to a common initial
heuristic method. The obtained solutions are high quality and results reveal a good performance.
Furthermore, the use of multiple depots is addressed where we will show by experiments that the
number of depots and the total travelled distance are proportional. This means as the number of
involved depots increases, the generated solutions are significantly better, and the total cost
reached is decreased.
The remainder of this paper is organized as follows. Section 2 presents the related works
associated to the studied problem. Section 3 describes the MDPOCARP problem and introduces
an illustrative example. Section 4 presents a constructive initial solution and describes our
proposed meta heuristic for solving the MDPOCARP problem. Section 5 shows the numerical
experiments that we conducted on small, medium and large new instances. Finally, our
conclusions follow in Section 6.
2. RELATED WORKS
In the best of our knowledge a solution for the studied MDPOCARP problem has never been
developed in the literature, in spite of important practical problems could be easily modelled as an
MDPOCARP. In this section, we present a brief overview of contributions which are technically
close to our proposal, but which focus on various aspects of routing optimization problems. We
discuss the difference between the problematic addressed by those contributions and those
addressed by our approach. The CARP case with the characteristics studied in this work has not
been extensively studied. The periodic CARP (PCARP) has been introduced by Lacomme et al.
[2]. They proposed a memetic algorithm based on a linear sophisticated crossover. Recently, a
two Phased Hybrid Local Search (PHLS) algorithm was proposed by Chen and Hao [3] for the
PCARP. Usberti et al. [4] have addressed a variant of the classical open capacitated arc routing
problem (OCARP) in which the set of edges are extended by depots and routes are not required to
take the form of cycles. This variant of OCARP has been also solved by a hybrid genetic
algorithm in [5]. The classical OCARP studied in this work, where a departure from a depot is
mandatory by each vehicle, has been studied by Fung and Liu [6] . They proposed a mathematical
formulation and generated lower bounds. They also developed a memetic algorithm which turns
out superior than the classical genetic algorithm in solution quality for the classical OCARP. The
Multi-depot CARP problem was developed by Kansou et Yassine [7] where two ant colony
approaches and a memetic algorithm based on the splitting procedure and a special crossover
have been proposed. Regarding the vehicle routing problem VRP on nodes case, several
important works have been proposed. The periodic vehicle routing problem PVRP with
intermediate facilities problem was introduced by Angelelli and Speranza in [8]. They solved it
by a Tabu Search algorithm and presented computational results on a set of random generated
instances and on a set of PVRP instances taken from the literature. Archetti et al. [9] worked on
the flexible periodic VRP and a mathematical formulation has been proposed. Cornillier et al.
[10] developed heuristics for the multi-depot petrol station replenishment problem with time
windows while Cornillier et al. [11] developed an efficient heuristic for the same problem with
unique depot and multiple periods without time windows. Popovic et al. [12] proposed a VNS

55
algorithm to solve the multi-product multi-period inventory routing problem in fuel delivery case.
Vidal et al. [13] implemented a hybrid genetic algorithm for the multi-depot periodic vehicle
routing problem VRP, heterogeneous capacitated vehicles and constrained route duration. The
same last problem with identical vehicles was studied by Rahimi-Vahed et al. [14] where a
modular heuristic algorithm was proposed. The authors worked on three vehicle routing
problems, i.e., periodic, multi-depot and periodic with multi-depot problem. They conducted their
experiments using suitable instances that show the effectiveness of the proposed method in terms
of both quality and computational time. Nguyen et al. [15] proposed a genetic algorithm for the
periodic VRP with time windows. They introduced two new crossover to tackle simultaneously
the periodic and time windows cases. They found new best solutions for a number of large
instances. The periodic VRP for Retail Distribution of Fuel Oils studied by Carotenuto et al. [16],
a hybrid genetic algorithm was developed for solving it on a set of real case studies and on a set
of randomly generated instances. Carotenuto et al. [16] provided a mathematical formulation for
the multi-depot periodic VRP. They also proposed a hybrid genetic algorithm for solving the
problem due to the large size of the real instances which the company has to deal with. Ran Liu et
al. [17] studied for the first time the multi-depot open VRP (MDOVRP). The authors presented a
mixed integer programming mathematical formulation and an effective meta-heuristic solution
approach (hybrid genetic algorithm) for the resolution. Eduardo Lalla-Ruiz et al. [18] improved
the last formulation proposed by [17] for the same problem MDOVRP and they provided better
results in terms of solution quality and tighter linear bounds.
3. PROBLEM DESCRIPTION
The Multi-Depot Periodic Open problem (MDPOCARP) is an extension of Open CARP. It
consists of a fleet of identical vehicles. Each vehicle is assigned to one depot and cannot exceed
its capacity. The vehicle starts its route from one depot and it ends at the least served task. The
tasks must be served a given number of periods (which is equal to its frequency ft on a whole
multi period horizon H). H is composed of np periods and each task t has a predefined set of
possible service combinations of periods Ct such that 𝑓𝑡≤ 𝑛 𝑝. Each k belonging to a combination
Ct must contain ft periods where the task will be served, i.e. |k| = ft, ∀ k∈ Ct, ∀ t ∈ Er. The
number of tasks, or required edges, will be denoted by n and the cost of each task t is C(t). In the
experiments conducted in this work, we will use np= |H| = 5 periods (see Section 5). In order to
code the combinations of periods, we use the binary code and five sets of combinations
depending on the frequencies. For example, the combination of five periods 10100 means that the
task has the frequency 2 and will be served in period 1 and period 3.
Table 1: Information about an example with nine tasks
Table 1 presents the information about a concrete example of the studied problem. It consists of 9
tasks that should be visited in 5 periods according to given frequencies. For example, Task 5 must
be visited 4 times and Task 9 must be visited 3 times. Each task t has a given combination
indicating the periods in which the Task t should be visited. For example, Task 5 should be
visited in all periods except Period 1 and Task 9 should be visited in Period 1, Period 2 and
Period 4. Table 2 illustrates the routes representing the solutions of the studied problem
MDPOCARP. For example, the solution in period 1 consists of two routes (d1, task2, task7, task8)
and (d2, task1, task6, task9) while the solution in period 4 consists of three routes (d1, task4, task5,
task6), (d2, task3, task3, task1) and (d2, task8, task7, task9).

56
Table 2: example of a feasible MDPOCARP solution
4. ALGORITHM FOR THE MDPOCARP PROBLEM
We propose a Hybrid algorithm (that we call HACSA) based on both the Ant Colony
optimization (AC) and the Simulated Annealing (SA) local search algorithm to solve the
MDPOCARP. Ant colony algorithms are global search optimization techniques based on the
foraging behavior of real ants, which have been found to be very effective in solving the
travelling salesman problem TSP [19] [20] and the capacitated arc routing problem CARP [1]. In
this work, we construct an initial solution S0 by a constructive heuristic explained in the next
section. The AC algorithm uses a colony of Na artificial ants where each ant k represents a giant
route (order of all different tasks) without delimiters, i.e. the capacity constraint is not verified. At
each iteration of the AC, every ant starts its route from a random task. Moving of ants depends on
two amounts: (1) Visibility amount 𝜇𝑖𝑗 which is a constant and represents the inverse of the
distance Cij between tasks i and j (Cij = C(i, j) is the shortest distance calculated by Dijkstra's
algorithm from the end node of i to the start node of j); (2) Pheromone amount τij which is a
variable and represents initially the inverse of the cost of initial solution S0, i.e. τij =
1
C(SO)
, ∀ i, j ∈
set of tasks. The cost C(S0) represents the sum of distances traversed by all M vehicles on all
periods. Note that the amount of pheromone will be updated at the end of each iteration v by the
following rule τij(v+1) = ρτij(v) + (1 - ρ)Δτij, where ρ is the evaporation parameter in [0,1], Δτij =
1
C(best)
if (i,j) is used by the ant that gives the best solution best and it will be zero otherwise.
4.1. Initial Solution for MDPOCRP
To build the initial solution of the proposed Hybrid Ant Colony algorithm, we propose a
constructive heuristic, called Nearest Insertion Heuristic (NIH), which constructs the solution in
sequential manner and mainly relies on the insertion method of tasks into a current solution. This
heuristic consists in:
• Sorting in increasing order all tasks to be inserted in an empty solution according to the
distance from each task to its associated closet depot;
• Choosing the starting task s of the route as the one which has the minimal distance from
closet depot to it;

57
• Choosing other tasks of the route in such a way the next task after the last selected task
among the candidate tasks using the nearest neighbor rule i.e. the closet task to the last
selected one;
• For each task t in the constructed route, we choose the best combination ct for task t and
then we insert it into each day of ct. ct is the best combination that minimizes the insertion
cost of t in the current solution, i.e., into each day d of that combination. Note that, in a
given day, the insertion cost is the cost of inserting t after the last inserted task if the
capacity constraint and the maximal distance constraints are verified. For the other case,
it represents the cost to insert it at the beginning of a new route in the same day d.
4.2. Local Search Procedure
In order to improve the obtained solutions using our ACO-based algorithm, we integrate the
Simulated Annealing local search method (SA). The (SA) is a heuristic which is commonly used
to prevent the optimization algorithm to quickly falling into local minimum [21] [22]. It is based
on random acceptance strategy with certain probability. At each iteration of our proposed (SA)
algorithm, we will use one of the following four moves:
• Swap: this move consists in exchanging the positions i and j of two different tasks
belonging to the same feasible route;
• Relocate: this move consists in removing one task from a position i and putting it in
another position j chosen in the same route;
• 2-opt: this move consists in choosing two positions i and j in a given route R and then re-
organizes R by replacing the sequence of tasks (i, i+1, … ,j) by the sequence (inverse(j),
inverse(j-1), … , inverse(i)) where inverse(t) represents the inverse task for the task t;
• Cross-exchange: this move involves two routes R1 and R2 where it chooses a task in a
position i of R1 and a task in a position j of R2. It generates two new routes R1' and R2' as
follows: R1' takes the first part of R1 from first task to task located at position i completed
by another part of the route R2 starting with the task located at position j. R2' is built in
the same way and that by linking the remaining first part of R2 with the second remaining
part of R1.
The algorithm used in the local search procedure is a Simulated Annealing (SA) method
presented in Algorithm 1.

58
4.3. Ants Production
Suppose that Visitedk is the set of already visited tasks by ant k. An ant k moves from the last
visited task i to another task j by choosing the closest task j (j ∉ Visitedk) to i if the random
variable q is strictly less than q0. Otherwise, it is chosen by using the following probability
decision rule:
(1)
where α and β are respectively the pheromone and visibility parameters, q0 is a constant
parameter belongs to [0, 1] and q is a uniform random parameter in [0, 1]. This move of every ant
is called the transition rule to add a task at the end of every sequence of tasks already constructed
for every ant k. Once each ant is ready, then we will construct an associated feasible
MDPOCARP solution by apply the following insertion procedure.
4.4. Insertion Procedure
Consider an ant k = (t1
k
, t2
k
, …, tn
k
) where each ti
k
is the task of order i in k. We associate firstly
for k an empty solution Sk
= (S1
k
, S2
k
, …, Snd
k
) where each Sd
k
, d = 1, …, nd, is the associated
subsolution to the period d that contains firstly M empty routes. Secondly, we will insert each
task ti
k
indicated in the order of k into ftik sub-solutions (ft is the associated predefined frequency
of task t). For the task t1
k
we will find the best combination of days (from the associated set of
predefined day combinations Ct1k that minimizes the cost to insert it in the associated sub-
solutions. Note that the insertion cost of a task t1k
in a combination c ∈ Ct1k is Insertion = Σd ∈ c
C(CDt1k, t1
k
) where CDt represents the closest depot to task t. And the general insertion cost to
insert another task t in a combination c ∈ Ct is Insertion:
∑ 𝑚𝑖𝑛
𝑙∈𝑆 𝑑
𝑘
[𝐶(𝐶𝐷𝑡, 𝑡) + 𝐶(𝑡, 𝑓𝑖𝑟𝑠𝑡) − 𝐶(𝐶𝐷𝑓𝑖𝑟𝑠𝑡, 𝑓𝑖𝑟𝑠𝑡), 𝐶(𝑙𝑎𝑠𝑡, 𝑡), 𝐶(𝑙, 𝑡) + 𝐶(𝑡, 𝑙 + 1) − 𝐶(𝑙, 𝑙 + 1) ]
𝑑∈𝐶
With first and last denote respectively the first served task in the current route Sdk and its last
served task, l+1 denotes the task next to t and it can be nothing if l is the last served task (in this
case C(t, l+1)=C(l, l+1)=0). We will continue to insert the other tasks, one after the other but
always checking the capacity constraint of each route when trying to find the best possible
combination. It means that when a route in a such period (a sub-solution) can no longer serve the
tested task (selected to insert), we open another route in this period and again by checking that the
number of routes will not exceed M. In case if all M routes in a given period p of a day
combination c ∈ Ct are closed to a given task t due to the capacity Q and maximal number of
routes M constraints, we will study another day combination c' ∈ Ct that will not involve the
period p. The insertion procedure finishes when all tasks tik
(by the order i =1. …, n) will be
inserted in the solution Sk
that will be a MDPOCARP feasible solution.
Suppose that Rdmk
= (t1dm
, ..., t|Rdmk|dm
) is the mth route in the period d of the solution Sk, then the total
cost of a solution Sk is calculated by the following rule:

59
𝐶𝑜𝑠𝑡 = ∑ 𝑓𝑙 𝐶(𝑙)
𝑛
𝑙=1
+ ∑ ∑ 𝐶 (𝐶𝐷 𝑡1
𝑑𝑚, 𝑡1
𝑑𝑚
)
𝑛
𝑚=1
𝑛 𝑝
𝑑=1
+ ∑ 𝐶( 𝑡𝑗−1
𝑑𝑚
, 𝑡 𝑗
𝑑𝑚
)
𝑗=|𝑅 𝑑𝑚
𝑘
|
𝑚=1
4.5. The Proposed HACSA Algorithm
The proposed Hybrid Ant Colony with the Simulated Annealing algorithm HACSA is shown in
Algorithm 2.
5. NUMERICAL EXPERIMENTS
5.1. Benchmark instances
In [5] the authors tested the hybrid genetic on the OCARP instances derived from the benchmark
of CARP instances, which includes 23 gdb ( [23]) 34 val ( [24]), and 24 egl ( [25]) instances. The
authors in [2] developed a method to build the periodic instances for the original CARP problem
using a horizon of 5 periods on the gdb and egl instances. They assigned a random frequency
between 1 and 5 to each task and then built a set of combinations of periods for each frequency.
For the studied problem MDPOCARP, we use the same OCARP instances used in [5]. Similarly,
to [2] we generate our benchmark instances for three new sets: 23 mdpogdb, 34 mdpoval and 24
mdpoegl, totaling 81 instances. We constructed five sets of combinations depending on the
frequencies. The minimal value of the number of vehicles M used in our experiments is the
minimal number of vehicles necessary for finding a feasible solution and is obtained by carrying
out 2000 random solutions to have at least 50% of feasible solutions between them. We added
three depots to Golden’s instances, four depots to Benavant's instances and six depots to Eglese's
instance. For all instances, the first node represents a depot and the other depots are randomly
selected.

60
5.2. Results
The experiments were implemented in Java and run on a 2.6GHz Dual Core computer with a
memory of 16GB under Windows Ten. Tables 3, 4 and 5 show the results for all instances after
running 1000 iterations of our algorithm. For each table, the columns correspond respectively to:
(1) problem name instance; (2) the number of vehicles M; (3) cost obtained by NIH method; (4)
cost obtained with our HACSA method; (5) running time T in seconds for HACSA ; (6) gap in
percentage of HACSA cost compared to NIH. The gaps GG(%), GB(%) and GE(%) are
respectively the percent improvement of AS solution cost with respect to initial solution NIH.
They are computed as follows:
GX(%) =
NIH−HACSA
NIH
X 100 with X ∈ {G, B, E}
The average gaps GG, GB and GE are respectively 31.8, 45 and 42.4. The overall gap averages
39.7 between the obtained solution cost with HACSA method and that obtained with the initial
method NIH for the 81 test instances shows that the HACSA method improves significantly the
quality of the solutions and provides better solutions. The average running time 33.7 seconds
shows the performance of our approach.
Table 3: Results of HACSA algorithm on Golden's set instances
Instance M NIH HACSA T GG Instance M NIH HACSA T GG
mdpogdb1 5 1152 774 24 32.8 mdpogdb12 8 1559 1044 23 33
mdpogdb2 5 1372 898 36 34.5 mdpogdb13 7 1923 1552 19 19.3
mdpogdb3 5 955 668 23 30.1 mdpogdb14 5 424 275 14 35.1
mdpogdb4 4 1135 686 16 39.6 mdpogdb15 4 237 171 15 27.8
mdpogdb5 5 1388 953 45 31.3 mdpogdb16 6 495 357 41 27.9
mdpogdb6 5 1478 872 33 41 mdpogdb17 5 347 259 29 25.4
mdpogdb7 5 1154 804 32 30.3 mdpogdb18 5 691 489 68 29.2
mdpogdb8 10 1417 852 109 39.9 mdpogdb19 3 198 128 1 35.4
mdpogdb9 8 1410 860 82 39 mdpogdb20 5 471 326 17 30.8
mdpogdb10 4 1037 675 25 34.9 mdpogdb21 6 722 499 24 30.9
mdpogdb11 5 1865 1170 84 37.3 mdpogdb22 8 741 594 111 19.8
mdpogdb23 10 934 682 105 27
Table 4: Results of HACSA algorithm on Benavent's set instances
Instance M NIH HACSA T GB Instance M NIH HACSA T GB
mdpoval1A 2 1268 576 91 54.6 mdpoval6A 9 2549 2056 23 19.3
mdpoval1B 4 1004 535 108 46.7 mdpoval6B 4 1344 702 3 47.8
mdpoval1C 9 820 523 48 36.2 mdpoval6C 10 1160 760 2 34.5
mdpoval2B 3 1263 682 58 46 mdpoval7B 5 2280 1121 17 50.8
mdpoval3B 3 467 210 65 55 mdpoval8B 4 2339 1188 10 49.2
mdpoval4A 5 3600 1957 6 45.6 mdpoval9A 4 2174 978 24 55
mdpoval4D 17 3111 2102 14 32.4 mdpoval9D 10 2222 1194 15 46.3
mdpoval5D 9 2822 1473 6 47.8 mdpoval10D 9 2818 1489 20 47.2

61
Table 5: Results of HACSA algorithm on Eglese's set instances
Instance M NIH HACSA T GE Instance M NIH HACSA T GE
mdpoegle1A 5 12551 7481 8 40.4 mdpoegls1A 7 17829 9020 15 49.4
mdpoegle1B 7 12325 7201 6 41.6 mdpoegls1B 10 20362 10454 10 48.7
mdpoegle1C 10 10694 7464 6 30.2 mdpoegls1C 14 15091 9561 9 36.6
mdpoegle3A 8 21355 11315 18 47 mdpoegls3A 13 37034 18947 42 48.8
mdpoegle3B 12 22379 12081 14 46 mdpoegls3B 20 382112 22216 35 41.9
mdpoegle3C 17 20407 14080 14 31 mdpoegls3C 28 40054 25301 55 36.8
5.3. Parameters
The results in Tables 3, 4 and 5 are given by executing ten runs of the algorithm HACSA for each
instance. For all runs, the HA algorithm parameters were set as: Na =25, q0 = 0.5, pr =0.7, α = 1, β
= 1, ρ = 0.2 and iterH = 1000. We selected these parameters values after previous empirical tests.
5.4. The Effect of using Multi-Depots
Table 6 reports the results when executing our algorithm HACSA on Eglese's instances with 3
depots. The first column is the cost obtained by HACSA method while the second column is the
running time in seconds. The third column reports the gap G36 between HACSA3 (with 3 depots)
and HACSA6 (with 6 depots) which is computed as
G36=
HACSA3−HACSA6
HACSA3
X 100
The fourth column diffG63 = G6 - G3 reports the difference between the gap GE with 6 depots
and the gap GE with 3 depots.
Table 6: Comparison of HACSA algorithm with multiple depots on Eglese's instance set.
Instance HACSA3 T3 GE(%) G36(%) DiffG63
mdpoegle1A 7919 9 36.9 5.531 3.4
mdpoegle1B 7532 7 38.9 4.395 2.7
mdpoegle1C 8822 8 17.5 15.393 12.7
mdpoegle2A 10589 11 46 2.115 1.2
mdpoegle2B 10978 17 33.6 6.24 4.1
mdpoegle2C 12610 19 34.8 12.974 8.5
mdpoegle3A 11580 19 45.8 2.288 1.2
mdpoegle3B 13012 16 41.9 7.155 4.1
mdpoegle3C 17783 19 12.9 20.823 18.1
mdpoegle4A 14891 25 42.3 12.968 7.5
mdpoegle4B 14444 23 34.9 7.671 5
mdpoegle4C 16289 17 29.3 3.272 2.3

62
Table 7: Comparison of HACSA algorithm with multiple depots on Eglese's instance set.
Instance HACSA3 T3 GE(%) G36(%) DiffG63
mdpoegls1A 9583 13 46.3 5.875 3.1
mdpoegls1B 11575 15 43.2 9.685 5.5
mdpoegls1C 10934 11 27.5 12.557 9.1
mdpoegls2A 20951 52 49.8 1.785 0.9
mdpoegls2B 23148 50 36.2 11.703 7.4
mdpoegls2C 23083 56 34.1 8.049 5.3
mdpoegls3A 19658 59 46.9 3.617 1.9
mdpoegls3B 24212 49 36.6 8.244 5.3
mdpoegls3C 26903 57 32.8 5.955 4
mdpoegls4A 27050 64 48.1 10.429 5.4
mdpoegls4B 29510 66 38.4 10.051 6.2
mdpoegls4C 37810 96 36.9 2.05 1.3
Figure 1 and Figure 2 illustrates the importance of using multiple depots. It reports the difference
of results between the use of 3 depots and 6 depots for the Eglese's test instances. Figure 1
illustrates a histogram that represents the difference of costs obtained by using 3 depots and 6
depots and Figure 2 shows a curve that represents the difference between the gap GE when using
3 depots and 6 depots. Both figures emphasize that using 6 depots instead of 3 depots decreases
significantly the obtained cost. The average gap GE when using 3 depots is 37.15 while the
average gap GE when using 6 depots is 42.4. That yields that the average total distance
improvement is 5.25%. Consequently, as the number of depots increases, the obtained average
cost decreases.
Figure 1: HACSA3's costs vs. HASCSA6's costs

63
Figure 2: GE with 3 depots vs. GE with 6 depots
6. CONCLUSION
This work introduces for the first time the Multi-Depot Periodic Open Capacitated Arc Routing
Problem (MDPOCARP) which is a NP-hard problem optimization problem deriving from the
family of arc routing problem. Firstly, a constructive heuristic was developed for generating
initial solution for MDPOCARP which relies on finding nearest neighbor. Secondly, a hybrid ant
colony algorithm (HACSA) was proposed. It is composed by the ant colony optimization
algorithm combined with both insertion procedure and local search algorithm. The insertion
procedure is used to construct feasible ants. The local search uses swap, relocate, 2-opt and cross-
exchange moves.
The computational experiments, using three set of benchmarks mdpogld, mdpoval and mdpoegl
derived respectively from the CARP instances [23], [24], [25] has shown that HACSA generates
good and promising solutions for MDPOCARP.
REFERENCES
[1] B. L. G. a. R. T. Wong., «Capacitated arc routing problems,» 1981, p. 305–315.
[2] C. P. a. W. R.-C. P. Lacomme, «Evolutionary algorithms for periodic arc routing problems,»
European Journal of Operational Research, n° %1Project Management and Scheduling, p. 165(2):535
– 553, 2005.
[3] Y. C. a. J. Hao., «Two phased hybrid local search for the periodic capacitated arc routing problem,»
European Journal of Operational Research, 06 2017.
[4] P. F. a. A. F. F. Usberti, «The open capacitated arc routing problem,» Computers & Operations
Research, p. 38(11):1543 – 1555, 2011.
[5] R. A. a. F. Usberti, «Hybrid genetic algorithm for the open capacitated arc routing problem,»
Computers & Operations Research, p. 90:221 – 231, 2018.

64
[6] R. L. a. Z. J. R.Y.K. Fung, « A memetic algorithm for the open capacitated arc routing problem,»
Transportation Research Part E: Logistics and Transportation Review, p. 50:53 – 67, 2013.
[7] A. K. a. A. Yassine, «New upper bounds for the multi-depot capacitated arc routing problem,»
IJMHeur, p. 1(1):81–95, 2010.
[8] E. A. a. M. Speranza, «The periodic vehicle routing problem with,» European Journal of Operational
Research,, p. 137:233–247, 2002.
[9] E. F. a. D. H.-M. C. Archetti, «The flexible periodic vehicle routing problem,» Computers &
Operations Research, p. 85, 2017.
[10] F. B. a. J. R. F. Cornillier, «Heuristics for the multi-depot petrol station replenishment problem with
time windows,» European Journal of Operational Research, p. 220:361–369, 07 2012.
[11] F. B. G. L. a. J. R. F. Cornillier, «A heuristic for the multiperiod petrol station replenishment
problem.,» European Journal of Operational Research, p. 191(2):295 – 305, 2008.
[12] M. V. a. G. R. D. Popovic, «Variable neighborhood search heuristic for the inventory routing problem
in fuel delivery.,» Expert Systems with Applications, p. 39(18):13390 – 13398, 2012.
[13] T. C. M. G. N. L. a. W. R. T. Vidal, « A hybrid genetic algorithm for multidepot and periodic vehicle
routing problems>,» Operations Research, p. 60:611–624, 2012.
[14]T. C. M. G. a. W. R. A. Rahimi-Vahed, «Fleet-sizing for multidepot and periodic vehicle routing
problems using a modular heuristic algorithm.,» Computers & Operations Research, p. 53:9 – 23,
2015.
[15] T. C. a. M. T. P. Nguyen, «A hybrid generational genetic algorithm for the periodic vehicle routing
problem with time windows,» Journal of Heuristics, p. 20:383–416, 08 2014.
[16] S. G. S. M. a. F. V. P. Carotenuto, «Periodic capacitated vehicle routing for retail distribution of fuel
oils.,» vol. volume 10, p. 735–744, 2015.
[17] Z. J a. N. G. R. Liu, «A hybrid genetic algorithm for the multi-depot open vehicle routing problem,»
OR Spectrum, p. 36(2):401–421, Mar 2014.
[18] C. I. S. T. a. S. V. E. Lalla-Ruiz, «An improved formulation for the multi-depot open vehicle routing
problem,» OR Spectrum, p. 38:1–13, 07 2015.
[19] M. V. C. A. Dorigo M., «Ant system: optimization by a colony of cooperating agents.,» IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 26, pp. 29 - 41, 1996.
[20] T. D. M. Stutzle, «A short convergence proof for a class of ant colony optimization algorithms.,»
IEEE Transactions on Evolutionary Computation, vol. 6, pp. 358 - 65, 2002.
[21] P. J. M. A. E. H. L. L. J. van Laarhoven, «Job shop scheduling by simulated annealing.,» Operational
research, vol. 40, pp. 113-25, 1992.
[22] I. P. C. Osman, «Simulated annealing for permutation flow-shop scheduling,» Omega, vol. 17, pp.
551 - 17, 1989.
[23] B. D. J. B. E. Golden, « Computational experiments with algorithms for a class of routing problems,»
Computers & Operations Research, vol. 10, pp. 47 - 59, 1983.
[24] E. C. V. C. A. M. E. Benavent, «The capacitated arc routing problem: Lower bounds.,» Networks,
vol. 22, pp. 669 - 90, 1992.
[25] R. W. Eglese, «Routing winter gritting vehicles.,» Discrete Appl. Math., vol. 48, pp. 231 - 44 , 1994.

HYBRID ANT COLONY ALGORITHM FOR THE MULTI-DEPOT PERIODIC OPEN CAPACITATED ARC ROUTING PROBLEM

More Related Content

What's hot (15)

Similar to HYBRID ANT COLONY ALGORITHM FOR THE MULTI-DEPOT PERIODIC OPEN CAPACITATED ARC ROUTING PROBLEM (20)

Recently uploaded (20)

HYBRID ANT COLONY ALGORITHM FOR THE MULTI-DEPOT PERIODIC OPEN CAPACITATED ARC ROUTING PROBLEM