You are on page 1of 6

Solving the last mile delivery problem using

iterated local search approach


Zhongkai Cai Zizhen Zhang* Huang He
School of Data and Computer Science School of Data and Computer Science School of Data and Computer Science
Sun Yat-Sen University Sun Yat-Sen University Sun Yat-Sen University
Guangzhou, China Guangzhou, China Guangzhou, China
caizhk@mail2.sysu.edu.cn zhangzzh7@mail.sysu.edu.cn sserdoubleh@gmail.com

Abstract—This paper deals with a real-life vehicle routing


problem (VRP) called the last mile delivery problem (LMDP). We
formulate this problem into a mixed integer linear programming
model, which can be viewed as a combination of the capacitated
VRP, the multi-depot VRP, the open VRP and the pickup-
and-delivery problem with time windows (PDPTW). To solve
the LMDP, we develop a meta-heuristic algorithm based on
iterated local search, in which an adaptive large neighborhood
search is applied in the perturbation phase to enlarge the search
scope. Results from the computational experiments show that the
proposed approach performs well when applied to the benchmark
instances. Furthermore, we demonstrate the effectiveness of our
algorithm by applying it to some real cases of the LMDP.
Index Terms—Last mile delivery, vehicle routing, pickup-and- Depot Shop customer
delivery, time windows, iterated local search
Fig. 1. An example of three routes.

I. I NTRODUCTION
As e-commerce grows at an unprecedented fast pace, we 5 O2O shops are served. Note that a vehicle must start from a
have observed that a large portion of express orders is generated certain depot, visit several customers or O2O shops but has
by e-commerce retailers. In China, for example, it is estimated no need to return to the depot.
that over 60 percent of express delivery orders are related to The LMDP is one variant of the vehicle routing problems
the e-commerce trading. The most challenging issue and cost (VRPs). It combines the capacitated VRP (CVRP), the multi-
ineffective part in the express industry is the last mile delivery depot VRP, the open VRP (OVRP) and the pickup-and-delivery
where packages are delivered from local branches of express problem with time windows (PDPTW). Obviously, this problem
companies to individual customers. The last mile delivery also is an NP-hard problem and difficult to be optimally solved. In
includes the intra-city pickup and delivery requests, i.e., the literature, there are many approaches for CVRP, OVRP and
online and offline (O2O) service, such as online food delivery PDPTW. However, the models of both OVRP and PDPTW
service. The last mile deliveries in both e-commerce and intra- cannot be directly applied to the LMDP. Thus, we introduce a
city O2O services play a very important role in nowaday’s mathematical programming model for the LMDP and design
express industry in China. a meta-heuristic algorithm to solve it. In our algorithm, the
In this paper, we deal with a real-life last mile delivery iterated local search (ILS) framework [7] is used and an
problem (LMDP) appearing in Shanghai, China. In this adaptive large neighborhood search procedure is applied in
problem, more than one depot is used and all couriers have the perturbation phase. The effectiveness of our algorithm is
the same capacity. Two types of orders need to be delivered, evaluated through a set of randomly generated LMDP instances,
namely, e-commerce packages and intra-city O2O packages. a set of benchmark PDPTW instances and two real-life LMDP
For the e-commerce packages, couriers pick up packages at the instances from Shanghai provided by a top express company
local branches of the express companies and deliver them to in China.
individual customers. While for the intra-city O2O packages, The rest of this paper is organized as follows. Section
couriers need to pick up packages at O2O shops and deliver II introduces some literature relevant to the LMDP. Section
them to the end customers at specified time windows. Fig. 1 III gives problem description and formulation. Section IV
shows an example of three routes in which 25 customers and presents the details of the proposed ILS-based algorithm. The
computational results are reported and analyzed in Section V.
978-1-5386-5053-0/18/$31.00 2018
c IEEE Finally, the conclusions are provided in Section VI.
II. L ITERATURE REVIEW total amount of time spent by all the vehicles. The following
assumptions are made:
The LMDP we study is most related to a combination of
the multi-depot OVRP (MDOVRP) and the PDPTW in the 1) Each vehicle starts at one depot and ends at the last
existing literature. Thus, we briefly discuss the methodologies customer that receives its service.
that have been reported for these two problems as follows. 2) The vehicle capacity must be respected.
3) The vehicle performing an O2O order should visit the
A. MDOVRP: multi-depot open vehicle routing problem shop before its delivery customer vertex.
A number of heuristics and exact methods were proposed 4) The vehicle cannot travel directly between different
to solve the OVRP. Derigs and Reuter [3] solved the OVRP depots.
using an attribute-based hill-climbing heuristic. Repoussis et al. 5) Each customer can only be served by exactly one depot.
[12] developed a route-construction insertion-based sequential The notations related to the model is described as follows.
approach for the OVRP with time windows. A stochastic
A. Sets
multiple neighborhood search in a (1+1)-evolutionary strategy
was proposed for the OVRP by Reinholz and Schneider [11]. D : the set of depots.
Most previous works assumed that all vehicles start from C0 : the set of e-commerce customers.
the same central depot and end at a customer vertex. However, C1 : the pickup set of O2O customers (shops).
the situation is often different in real-life. Transportation C2 : the delivery set of O2O customers.
companies typically hire a large number of vehicles, which O : the set of pairwise O2O orders. If (i1 , i2 ) ∈ O, then
usually stationed at several different depots. This kind of i1 ∈ C1 and i2 ∈ C2 .
C : the set of customers, i.e., C = C0 ∪ C1 ∪ C2 .
problem is much more complex than the single-depot OVRP,
V : the vertex set, i.e., V = D ∪ C.
and is called multi-depot OVRP. To the best of our knowledge,
K : the vehicle set.
there are few research papers studied the MDOVRP in the
literature. Tarantilis and Kiranoudis [15] presented a list-based B. Input parameters
threshold-accepting (LBTA) algorithm for the MDOVRP. Liu
Q : the vehicle capacity.
et al. [6] presented a hybrid genetic algorithm to solve the
ETi : the earliest arrival time at vertex i.
MDOVRP.
LTi : the latest arrival time at vertex i.
B. PDPTW: pickup-and-delivery problem with time windows tij : the traveling time between vertex i and j. Because a
vehicle is not necessary to return to the depot, we simply
Over the last decade, several heuristics have been proposed
set tim = 0 for every i ∈ V, m ∈ D.
for the PDPTW. Ropke and Cordeau [13] devised a new branch- si : the service time of vertex i.
and-cut-and-price algorithm to solve the PDPTW. Nagata di : the demand of vertex i.
and Kobayashi [8] presented an efficient route minimization M : a sufficiently large number.
heuristic for the PDPTW, which is based on the guided ejection
search (GES).Qu and Bard [10] developed a greedy randomized C. Decision variables
adaptive search procedure (GRASP) with several novel features aki : the arrival time of vehicle k at vertex i.
for the PDPTW. Blocho and Nalepa [2] proposed a selective lik : the load of vehicle k after sevicing vertex i.
route exchange crossover, which is based on the longest xkij : a binary variable which is equal to 1 if vehicle k travels
common subsequence, and applied this operator in the memetic from vertex i to vertex j, and 0 otherwise.
k
algorithm (MA) for the PDPTW. ym : a binary variable which is equal to 1 if vehicle k starts
In practice, more than one depot is often required, which from depot m, and 0 otherwise.
necessitates to extend the PDPTW to the multi-depot PDPTW
D. Mathematical model
(MDPDPTW). Sombuntham and Kachitvichayanukul [14]
presented a particle swarm optimization algorithm with multiple The model of the LMDP is described as follows. It is
social learning structures for the MDPDPTW. Alaa et al. [1] modified from the PDPTW by considering multiple depots
proposed an approach based on the combination of genetic and open routes.
algorithm with the clustering algorithm for the optimization of   
min tij xkij + si (1)
the MDPDPTW. k∈K i∈V j∈V i∈C

III. P ROBLEM FORMULATION subject to:  k


ym = 1, ∀k ∈ K (2)
The LMDP is defined as follows. Let G = (V, E) be a m∈D
graph where V is the vertex set and E is the edge set. The 
vertex set V is partitioned into two subsets C and D, which xkji = 1, ∀i ∈ C (3)
represent the set of customers and the set of depots. A set K k∈K j∈V

of identical vehicles, each with a capacity Q, is available for  


xkij − xkji = 0, ∀ k ∈ K, i ∈ C (4)
transportation. The objective of the LMDP is to minimize the j∈V j∈V

xkmj = ym
k
, ∀ m ∈ D, k ∈ K (5) Algorithm 1 The proposed algorithm ILS ALNS.
j∈C0 ∪C1 1: Generate an initial solution S0 ;
 2: Perturbed solution Sp , Sp ← S0 ;
xkim = k
ym , ∀ m ∈ D, k ∈ K (6) 3: Current solution Sc , Sc ← S0 ;
i∈C0 ∪C2 4: Best solution Sb , Sb ← S0 ;
  5: Improved solution St ;
xki1 ,j − xkj,i2 = 0, ∀k ∈ K, (i1 , i2 ) ∈ O (7) 6: N1 and N2 are the preset numbers of iterations;
j∈C j∈C 7: iter1 ← 0;
8: iter2 ← 0;
aki + si + tij ≤ akj + M ∗ (1 − xkij ), ∀i, j ∈ C, k ∈ K (8) 9: while iter1 < N1 do
ETi ≤ aki ≤ LTi , ∀i ∈ C, k ∈ K (9) 10: Use an improvement neighborhood search to improve Sp ;
11: St ← Sp ;
aki1 + si1 + ti1 ,i2 ≤ aki2 , ∀k ∈ K, (i1 , i2 ) ∈ O (10) 12: if St satisfies the acceptance criterion then
13: Sc ← St ;
ljk + dj ≤ lik + M ∗ (1 − xkij ), ∀i ∈ C, j ∈ C0 ∪ C2 , k ∈ K (11) 14: end if
15: if Sc is better than Sb then
ljk − dj ≤ lik + M ∗ (1 − xkij ), ∀i ∈ C, j ∈ C1 , k ∈ K (12) 16: Sb ← Sc ;
17: end if
di ≤ lik ≤ Q, ∀i ∈ C0 ∪ C2 , k ∈ K (13) 18: Sp ← Sc ;
0≤ lik ≤ Q − di , ∀i ∈ C1 , k ∈ K (14) 19: if iter2 = N2 then
20: Sc ← Sb ;
xkij , ym
k
∈ {0, 1}, ∀m ∈ D, k ∈ K, i, j ∈ V (15) 21: Use ALNS to perturb Sc ;
22: Sp ← Sc ;
The objective (1) is to minimize the total time spent, 23: iter2 ← 0;
including the traveling time and the service time. Constraints 24: end if
(2) guarantee that each vehicle starts from only one depot. 25: iter1 ← iter1 + 1;
26: iter2 ← iter2 + 1;
Constraints (3) ensure that each customer is visited exactly 27: end while
once. Constraints (4) – (6) define the flow structure indicating 28: return Sb
that the vehicle starts from its designated depot, visits a number
of customers and finally returns to the depot. Note that the open
route consideration can be overcome by setting tim = 0 for In the rest of this section, we describe the main components
i ∈ V, m ∈ D. Constraints (7) require that the pickup customer of our algorithm, which include the construction of the
and delivery customer must be serviced by the same vehicle. initial solution, the improvement procedure, the perturbation
Constraints (8) give the transit time constraints. Constraints (9) procedure with the large neighborhood search mechanism, the
are time window constraints. Constraints (10) force the vehicle acceptance and termination criteria.
to visit the pickup customer before the delivery customer.
Constraints (11) – (12) present the vehicle load dependency A. Initial solution
constraints. Constraints (13) – (14) impose the initial vehicle
load. Constraints (15) are decision variable constraints. We modify the sequential insertion heuristic [4] to generate
an initial solution for the LMDP. This method is described
IV. T HE PROPOSED SOLUTION METHOD in Algorithm 2. Unlike the traditional insertion methods, our
We employ an iterated local search (ILS) as the algorithm method relies on the hill climbing algorithm to improve the
framework to solve the LMDP. The ILS consists of two quality of the current route, with no need to calculate the cost
procedures, namely, the improvement and the perturbation of each possible insertion position.
procedures. Furthermore, to enhance the solution quality,
several neighborhood operators are used in the improvement Algorithm 2 Generating the initial solution.
procedure, and the perturbation procedure is realized using an 1: The number of vehicles used k ← 0;
adaptive large neighborhood search (ALNS) mechanism. 2: repeat
The pseudocode of our ILS-based algorithm is given in 3: Initialize an empty route r;
4: k ← k + 1;
Algorithm 1. At first, we construct an initial solution S0 . Then, 5: for all unassigned orders do
the algorithm applies an improvement neighborhood search to 6: Get the next unassigned order i;
this solution to obtain an improved solution St . If St satisfies 7: Insert order i to route r;
an acceptance criterion, it replaces the current solution Sc . If 8: Call the hill climbing routing heuristic to improve r;
Sc can not be improved after N2 consecutive iterations, the 9: if r is a feasible route then
10: Mark i as an inserted order;
best solution Sb found during the search replaces the current 11: else
solution Sc . Then a perturbation procedure is applied to this 12: Remove i from r;
solution Sc and generates a new perturbed solution Sp . The 13: end if
next improvement procedure is applied to solution Sp . Note 14: end for
that in the perturbation procedure, Sp is required to be a new 15: until All order have been inserted.
solution that has not been visited.
R-relocate1,2 R-relocate1

Different depots R-exchange1,2 R-exchange1 relocate

R-mix R-relocate2 exchange


Inter-route
R-exchange2
R-relocate1,2 reverse
R-mix
Same depot R-exchange1,2
Neighborhood
structure
R-mix
Fig. 3. New Neighborhood operators.
relocate

Intra-route exchange Algorithm 3 The adaptive large neighborhood search process.


reverse 1: A solution S to be perturbed;
2: A set of removal operators R;
3: A set of insertion operators I;
Fig. 2. Neighborhood structure.
4: Sbest ← S;
5: Scurrent ← S;
6: while stopping criterion is not satisfied do
B. Improvement procedure 7: removel ← chooseRemovalOperator(R);
We use three types of neighborhood structures to explore the 8: reinsert ← chooseReinsertOperator(I);
solution space of the problem, which are inter-route operators 9: Snew ← removel(Scurrent );
within the same depot, inter-route operators among different 10: Snew ← reinsert(Snew );
11: if Snew satisfies the acceptance criterion then
depots and intra-route operators. The inter-route operators 12: Scurrent ← Snew ;
include R-relocate1, R-relocate2, R-exchange1, R-exchange2 13: end if
and R-mix. The intra-route operators include relocate, exchange 14: if f (Snew ) < f (Sbest ) then
and reverse. The neighborhood structure is given in Fig. 2. 15: Sbest ← Snew ;
The neighborhood operators used are as follows. 16: end if
17: updateOperatorsProbility(R, I);
• relocate: remove one vertex or one order (i.e., two vertices) 18: end while
first, and then reinsert it or them in another locations in 19: return Sbest .
the same route.
• exchange: exchange two orders in the same route.
• reverse: reverse a segment of a route. • Route Removal (ROR): randomly remove a route from the
• R-relocate1: remove one or two vertices from its current solution.
route and insert it or them in another route. • Proximity-Based Removal (PR): remove a set of orders
• R-relocate2: remove one or two orders from one route to that are close in terms of distance.
another route. • Shaw Removal (SR): remove orders that are similar
• R-exchange1: exchange a sequence of one or two consec- according to several criteria.
utive vertices in one route and a sequence of one or two • Worst Removal (WR): remove a number of orders with
vertices in another route. high costs in the current solution.
• R-exchange2: exchange an order in one route and an order
The insertion operators are defined as follows:
in another route.
• Basic Greedy Insertion (BGI): calculate the insertion costs
• R-mix: remove some selective orders from its route and
reinsert them in another route. for all removed orders, and then insert the order with the
smallest cost increment.
In order to improve the solution efficiently, we bond each
• Random Insertion (RI): seek the feasible positions avail-
inter-route operator and each intra-route operator into one.
able in the solution for a removed order and then randomly
Each combination is regard as a new neighbohood operator.
select one position for insertion.
These neighborhood operators in the improvement procedure
• Regret Insertion (RI): calculate for each unserved order
of ILS ALNS is shown in Fig. 3. Note that the movements
a regret value that is equal to the total cost difference
that lead to infeasible solutions are prohibited.
between the best insertion and the second best one. The
C. Perturbation procedure order with the biggest regret value is inserted.
We use the perturbation procedure to diversify the search A roulette-wheel mechanism selects a removal operator in
process. It is essentially an adaptive large neighborhood search set R and an insertion operator in set I. The adaptive layer is
(ALNS) process. The pseudocode of the ALNS process is similar to the one used in Pisinger and Ropke [9]. For each
presented in Algorithm 3. iteration, each removal (or insertion) operator i is assigned
The removal operators used in our implementation are as an initial value wi at the beginning of the algorithm, i ∈
follows: {1, 2,..., k }. Then the operator i is selected with a probability
• Random Removal (RR): randomly remove some orders wi /( i wi ). If operator i is selected and the resultant solution
from the current solution. improves the current solution, then update wi = wi + 1.
D. Acceptance and stopping criterion TABLE I
C OMPUTATIONAL RESULTS FOR THE SMALL - SIZE LMDP INSTANCES .
Once a new best solution is found in the improvement
procedure, this solution replaces the current solution. Moreover, Instance |C0 | |D| |O| |K| LINGO 11 ILS ALNS Gap (%)
when a solution is only better than the current solution and 1 3 1 1 2 417.25 417.25 0.00
the solution has not been accepted before, it will be accepted. 2 3 2 1 2 313.32 313.32 0.00
3 5 2 1 4 347.63 347.63 0.00
If the current solution (Sc ) has not been improved when the 4 6 3 2 4 431.87 436.92 1.17
preset number (N2 ) of consecutive iterations is reached, the Average 377.52 378.78 0.33
perturbation operator is run again. When a specified number Notes: |D|: the number of depots; |C0 |: the number of customers;
(N1 ) of iterations is reached, the whole algorithm stops. |O|: the number of O2O orders; |K|: the number of vehicles.

V. C OMPUTATIONAL RESULTS TABLE II


In this section, we report the results of our computational C OMPUTATIONAL RESULTS FOR L I & L IM ’ S PDPTW INSTANCES .
experiments for the LMDP. We applied LINGO to solve the
BKS PSO ILS-ALNS
the mathematical programming model introduced in Section Instance
NV Distance NV Distance NV Distance
III-D, on small-size problem instances to obtain their optimal
lc101 10 828.94 10 828.94 10 828.94
solutions. The ILS ALNS was encoded in C++, and run on a lc102 10 828.94 10 828.94 10 828.94
laptop computer with an Inter Xeon 3.5GHz processor with lc103 9 1035.35 10 827.86 10 827.86
lc104 9 860.01 10 1063.63 10 818.60
4 GB RAM and Windows 7 professional edition. Since there lc105 10 828.94 9 863.36 10 828.94
are no standard benchmark test instances for the LMDP, we lc106 10 828.94 10 828.94 10 828.94
lc107 10 828.94 10 828.94 10 828.94
used the benchmark test instances for the classical PDPTW to lc108 10 826.44 10 826.44 10 826.44
validate our proposed algorithm and compare the results with lc201 3 591.56 3 591.56 3 591.56
lc202 3 591.56 3 591.56 3 591.56
those generated by previous algorithms. At last, the algorithm lc203 3 585.56 3 591.17 3 591.17
was applied to two real-life LMDP instances obtained by an lc204 3 590.60 3 590.60 3 590.60
lc205 3 588.88 3 588.88 3 588.88
express company located in Shanghai, China. lc206 3 588.49 3 588.49 3 588.49
lc207 3 588.29 3 588.29 3 588.29
A. Small-size LMDP instances lc208 3 588.32 3 588.32 3 588.32
lr101 19 1650.80 19 1650.80 19 1650.80
Four test instances were randomly generated to evaluate lr102 17 1487.57 17 1512.25 17 1487.57
lr103 13 1292.68 13 1300.77 14 1340.60
the effectiveness of our algorithm. The number of customer lr104 9 1013.39 10 1050.90 9 1013.39
vertices are 4, 6 and 8, the number of depots are 1, 2 and 3, the lr105 14 1377.11 14 1389.43 14 1389.43
lr106 12 1252.62 12 1270.46 12 1252.62
number of shops are 1 or 2. The maximum number of vehicles lr107 10 1111.31 10 1147.12 10 1158.54
is 5 and the vehicles capacity is 50. The duration of the time lr108 9 968.97 9 968.97 9 968.97
window is 100. Service time for each vertex ranges from 50 to Notes: NV: the number of vehicles used. Bold value in this table
100. Customer demands are random numbers within [0, 50]. indicates that it has the same value as the best known solution.
These test instances were solved using both LINGO and
our ILS based algorithm whose results are compared and
Table II presents the results obtain by our ILS ALNS, PSO
analyzed. Gaps between optimal solutions and our algorithm
[14] and the best known solutions. The comparison results
were calculated by the following formulation:
show that our algorithm has an excellent convergence capability
ILS ALN S average value − LIN GO solution value
Gap = ×100% when solving the PDPTW instances. The performance of our
LIN GO solution value
(16) ILS-based algorithm is competitive compared to the existing
Note that the proposed ILS ALNS was executed ten times methods for the PDPTW.
for each instance and the average value was used to calculate
the gap. Table I shows the characteristics of instances and C. Real-life LMDP instances
the results of LINGO and ILS ALNS. As illustrated in this This section reports the experimental result of two real-life
table, our method solved 75% of instances to optimally (i.e., large-scale LMDP instances: one express company that owns
instances 1, 2, and 3) and the average gap between ILS ALNS 124 service branches in Shanghai and is responsible for all
and LINGO is 0.33%. e-commerce delivery requests in Shanghai. All e-commerce
packages arrive at service branches before 8:00 am. All
B. PDPTW instances couriers start working at 8:00 am and are required to deliver
We set the traveling cost from each customer vertex to each e-commerce packages to customers before 8:00 pm. At the
depot to zero and applied the ILS ALNS to a special case same time, these couriers also deliver intra-city O2O packages.
of the LMDP, namely, PDPTW. We tested our approach on They pick up orders from O2O shops at the designated time
the data set generated by Li and Lim [5]. All best known and need to deliver the orders to customers no later than the
solutions (BKS) are summarized in Table II. We compared our designated time. The problem tries to decide the ways of
results with the particle swarm optimization (PSO) algorithm sending the orders to the customers with the least total delivery
by Sombuntham and Kachitvichayanukul [14]. time. There are two real-life data sets about this problem
ACKNOWLEDGEMENTS
This research was partially supported by the National Natural
Science Foundation of China (No. 71601191), Guangdong
Natural Science Funds (No. 2016A030313264) and China
Scholarship Council (No. 201706385060).
R EFERENCES
[1] Essia Ben Alaa, Imen Harbaoui Dridi, Hanen Bouchriha,
and Pierre Borne. Insertion of new depot locations for the
optimization of multi-vehicles multi-depots pickup and delivery
problems using genetic algorithm. In International Conference
on Industrial Engineering and Systems Management, pages 695–
701, 2016.
Fig. 4. A part of final routes obtained for the Set 1 instance. [2] Miroslaw Blocho and Jakub Nalepa. Lcs-based selective route
exchange crossover for the pickup and delivery problem with
time windows. In European Conference on Evolutionary
available at: https://tianchi.aliyun.com/competition/information. Computation in Combinatorial Optimization, pages 124–140.
htm?spm=5176.100068.5678.2.36abe412b2McIG&raceId= Springer, 2017.
231581& lang=en US. [3] Ulrich Derigs and K Reuter. A simple and efficient tabu search
The results generated by the ILS ALNS for the Sets 1 and heuristic for solving the open vehicle routing problem. Journal
of the Operational Research Society, 60(12):1658–1669, 2009.
2 are 285343 and 533077, respectively. In order to demonstrate [4] Manar I Hosny and Christine L Mumford. Constructing initial
the effectiveness of our method, we show the best 10 (from solutions for the multiple vehicle pickup and delivery problem
1000+ teams) results of the two sets in Table III. From this with time windows. Journal of King Saud University-Computer
table, we can see that the results obtained by our method are and Information Sciences, 24(1):59–69, 2012.
promising. Specifically, our result ranks the third place in Set [5] Haibing Li and Andrew Lim. A metaheuristic for the pickup
and delivery problem with time windows. International Journal
1 and ranks seventh place in Set 2. Therefore, our algorithm on Artificial Intelligence Tools, 12(02):173–186, 2003.
can be considered as an effective method for the LMDP. For [6] Ran Liu, Zhibin Jiang, and Na Geng. A hybrid genetic algorithm
better visualization, a part of routes we obtained for the Set 1 for the multi-depot open vehicle routing problem. Or Spectrum,
instance is depicted in Fig. 4. 36(2):423–424, 2014.
[7] Helena R Lourenço, Olivier C Martin, and Thomas Stützle.
Iterated local search. In Handbook of metaheuristics, pages
TABLE III
320–353. Springer, 2003.
C OMPUTATIONAL RESULTS FOR THE REAL - LIFE LMDP INSTANCES .
[8] Yuichi Nagata and Shigenobu Kobayashi. Guided ejection
search for the pickup and delivery problem with time windows.
Rank Set 1 Set 2 In European Conference on Evolutionary Computation in
1 282874 514751 Combinatorial Optimization, pages 202–213, 2010.
2 283177 521024 [9] David Pisinger and Stefan Ropke. A general heuristic for vehicle
285343 ← our score
3 287312 525127
routing problems. Computers and Operations Research, 34(8):
4 288780 526337 2403–2435, 2007.
5 289543 528505 [10] Yuan Qu and Jonathan F Bard. A grasp with adaptive large
6 295793 531932 neighborhood search for pickup and delivery problems with
533077 ← our score
7 296534 535177
transshipment. Computers & Operations Research, 39(10):2439–
8 296622 538100 2456, 2012.
9 297171 540749 [11] Andreas Reinholz and Holger Schneider. A hybrid (1+1)-
10 299109 540958 evolutionary strategy for the open vehicle routing problem. In
Average 291691.5 530266 Advances in Metaheuristics, pages 127–141. Springer, 2013.
[12] Panagiotis P Repoussis, Christos D Tarantilis, Olli Bräysy, and
George Ioannou. A hybrid evolution strategy for the open
VI. C ONCLUSIONS vehicle routing problem. Computers & Operations Research, 37
(3):443–455, 2010.
In this paper, we investigate a real-life LMDP, which is [13] Stefan Ropke and Jean-François Cordeau. Branch and cut and
an extension of the classical MDOVRP and PDPTW. An price for the pickup and delivery problem with time windows.
iterated local search (ILS) based algorithm is proposed to Transportation Science, 43(3):267–286, 2009.
solve the problem. Our solution procedure combines the [14] Pandhapon Sombuntham and Voratas Kachitvichayanukul. A
iterated local search with an adaptive large neighborhood particle swarm optimization algorithm for multi-depot vehicle
routing problem with pickup and delivery requests. Lecture
search process. The algorithm is tested on various test instances. Notes in Engineering and Computer Science, 2182(1):965–972,
Computational results demonstrate that our proposed algorithm 2010.
is a promising method for solving real-life large-scale LMDP [15] C. D. Tarantilis and C. T. Kiranoudis. A list-based threshold ac-
instances. Meanwhile, our algorithm is also capable of dealing cepting method for job shop scheduling problems. International
with the classical PDPTW. Regarding the future work, we can Journal of Production Economics, 77(2):159–171, 2002.
impose more constraints and multiple objectives for a more
realistic LDMP.

You might also like