1

A Memetic Algorithm for Vehicle Routing with

Simultaneous Pickup and Delivery and Time
Windows
Zhenyu Lei and Jin-Kao Hao*

Accepted to IEEE Transactions on Evolutionary Computation, Aug. 2024.

Abstract—The Vehicle Routing Problem with Simultaneous and a service time si (the time the vehicle spends at node vi ).
Pickup and Delivery and Time Windows (VRPSPDTW) has a The depot time window [e0 , l0 ] specifies the earliest time the
number of real-world applications, especially in reverse logistics. vehicle can leave the depot and the latest time it must return
In this work, we propose an effective memetic algorithm that
integrates a lightweight feasible and infeasible route descent to the depot, with the service time s0 equal to 0. Furthermore,
search and a learning-based adaptive route-inheritance crossover each edge eij ∈ E is associated with a travel distance cij and
to solve this complex problem. We evaluate the effectiveness of the a travel time tij .
proposed algorithm on the set of 65 popular benchmark instances A fleet of M homogeneous vehicles with a capacity of Q is
as well as 20 real-world large-scale benchmark instances. We available to serve the given customers. VRPSPDTW aims to
provide a comprehensive analysis to better understand the design
and performance of the proposed algorithm. determine the best routes of the vehicles starting and ending at
Index Terms—Vehicle routing; Simultaneous pickup and deliv- depot v0 while satisfying the capacity and time window con-
ery with time windows; Combinatorial optimization; Heuristics; straints. Thus, a solution S of VRPSPDTW is a set of closed
Memetic algorithm. routes S = {R1 , . . . , RM }, where each route Ri consists of
a sequence of nodes {ni,0 , ni,1 , . . . , ni,L , ni,L+1 } visited by
the i-th vehicle. Here, ni,j indicates the j-th visiting node in
I. I NTRODUCTION route Ri , and L indicates the number of customers in route Ri .

T HE Vehicle Routing Problem with Simultaneous Pickup

and Delivery and Time Windows (VRPSPDTW) [1] is
a member of the large family of Vehicle Routing Problems
Notably, both the first and last nodes in route Ri are the depot
v0 (i.e., ni,0 = v0 and ni,L+1 = v0 ). The load of each vehicle
cannot exceed the capacity Q, thus qni,j < Q must be satisfied
(VRPs) [2]. Specifically, VRPSPDTW is a variant of the Ve- for all ni,j ∈ Ri , where qni,j is the load of the i-th vehicle
hicle Routing Problem with Simultaneous Pickup and Delivery after visiting node ni,j . Additionally, let ani,j be the arrival
(VRPSPD) [3] that incorporates time window constraints. This time at node ni,j . Arriving before eni,j (ani,j < eni,j ) results
is a computationally challenging problem because it can be in a wait time wni,j = max{eni,j − ani,j , 0}. Conversely,
trivially reduced to the NP-hard VRPSPD problem [1]. arriving after lni,j (ani,j > lni,j ) is considered infeasible.
VRPSPDTW has a wide range of applications in real world Let µ1 and µ2 be the costs or weights assigned to the
scenarios, especially in the field of logistics. Many logistics vehicles and the travel distance, respectively. A typical ob-
companies now offer pickup services in addition to delivery jective function of VRPSPDTW is then to minimize the total
services to improve transportation efficiency and reduce costs, cost, which is the weighted sum of the number of vehicles
a concept commonly referred to as reverse logistics [4]. and the travel distance, as shown in Equation (1). Another
VRPSPDTW can be described on a directed complete graph typical hierarchical objective function aims to minimize first
G = (V, E), where the vertices V = {v0 , v1 , . . . , vN } consist the number of vehicles used (primary objective) and then the
of the depot node v0 and the N customer nodes (v1 , . . . , vN ). travel distance of the vehicles (secondary objective). This can
The edges E = {eij |vi , vj ∈ V} represent connections be achieved by assigning a very high cost µ1 to vehicles and
between the nodes. Each customer node vi ∈ V (i 6= 0) is a relatively low cost µ2 to travel distances.
associated with a delivery demand di and a pickup demand
pi . This implies that the vehicle must deliver di units of goods Minimize f (S) = µ1 · M + µ2 · D(S) (1)
from the depot v0 to vi and pick up pi units of goods from vi to L
M X
X
the depot v0 . Additionally, each node vi has a time window [ei , Subject to D(S) = cni,j ni,j+1
i=1 j=0
li ] (the earliest and latest time to start the service at node vi )
S = {R1 , . . . , RM }
This work benefited from the computing resources of the Centre de Calcul Ri = {ni,0 , ni,1 , . . . , ni,L , ni,L+1 }, i = 1, . . . , M
Intensif des Pays de la Loire (CCIPL). The first author is supported by
a scholarship from the China Scholar Council (Grant No. 202206330014)
(Corresponding author: Jin-Kao Hao) Since the introduction of VRPSPDTW [1], many algo-
Z. Lei and J.K. Hao are with the Department of Computer Science, LERIA,
Université d’Angers, 2 Boulevard Lavoisier, 49045 Angers Cedex 01, France rithms have been developed to solve this problem. Section
(e-mails: zhenyu.lei@etud.univ-angers.fr, jin-kao.hao@univ-angers.fr). II provides a comprehensive review of related studies. The
 2

inherent complexity of the problem is such that no single encoding method to represent solutions and introduced a
method consistently excels across all benchmark instances. penalty function to deal with illegal solutions. Kassem et
Additionally, research on large instances and real-world appli- al. [14] presented the INST-SA heuristic, using an insertion
cations remains limited. Meanwhile, a notable trend in recent heuristic for solution initialization, and simulated annealing
algorithm design is the use of large neighborhood search and neighborhood search for solution improvement. Tang et
techniques with destroy and repair operators [5, 6, 7]. While al. [18] studied VRPSPDTW with the objective of minimizing
these strategies enhance search space exploration, they also the weighted total cost of the vehicle dispatching and the
increase algorithmic complexity. travel distance. They proposed a co-evolution of parameterized
To address these challenges, we propose a novel algorithm, search (CEPS) to achieve generalizable parallel algorithm
MA-FIRD, which is based on the memetic algorithm frame- porfolio (PAP) based on some training instances. Liang et
work. This framework has been proven effective for solving al. [12] studied a variant of VRPSPDTW with soft time
various VRPs, including VRPTW [8], split delivery VRP window constraints, allowing delayed arrivals beyond the time
[9], and multi-trip VRP [10]. One recent work [6] has also windows with associated penalty costs.
showed the effectiveness of the memetic algorithm approach Until the work of Wang and Chen [11], there was a lack
for VRPSPDTW. Following this framework, our proposed of benchmark instances for VRPSPDTW, making it difficult
algorithm integrates several novel components and strategies. to compare the solution algorithms for the problem. Wang
These include a lightweight feasible and infeasible route and Chen filled the gap by introducing a set of 65 bench-
descent search where a dedicated penalty function is embedded mark instances (denoted as WC) derived from the Solomon
to manage infeasible solutions, and a multi-parent route- VRPTW benchmark [21]. They also proposed a co-evolution
inheritance crossover supported by reinforcement learning. genetic algorithm (co-GA), which maintains two populations
Furthermore, we implement a max-min normalization-based for diversification and intensification, to minimize the hierar-
fitness-distance population management strategy to maintain chical objective of the number of vehicles first and the travel
population diversity. The proposed algorithm is evaluated on distance second. Following their study on the hierarchical
the popular WC benchmark [11] and the large-scale real- objective, a succession of research efforts using metaheuristics
world JD benchmark [6], demonstrating its superiority and [15, 16, 5, 17, 6] have contributed to advancing the state of
practicality compared to existing algorithms. the art in solving VRPSPDTW, particularly in the context of
The rest of this paper is organized as follows. Section the WC benchmark.
II provides a literature review on VRPSPDTW. Section III Wang et al. [15] introduced a simulated annealing approach
presents the proposed algorithm. Section IV shows computa- and later developed a parallel simulated annealing (p-SA)
tional results on benchmark instances compared to state-of- algorithm [16], utilizing multi-processor or multi-thread ca-
the-art methods, highlighting its ability to solve real-world pabilities to enhance and accelerate the search process. Shi
large-scale instances. Section V provides an in-depth analysis et al. [17] presented a two-stage algorithm VNS-BSTS. In
of the key components of the algorithm. Section VI concludes the first stage, a Variable Neighborhood Search (VNS) was
the paper and suggests possible avenues for future research. used to minimize the number of vehicles, featuring a novel
learning-based evaluation function to assess each move. The
II. L ITERATURE REVIEW
second stage employed a Bi-Structure based Tabu Search
TABLE I (BSTS) to intensify the optimization. Recently, Wu et al. [7]
S UMMARY OF THE ALGORITHMS FOR VRPSPDTW IN THE LITERATURE proposed an ant-colony optimization algorithm with destroy
Literature Year Method Objective and repair strategies (ACO-DR) for VRPSPDTW, focusing on
Angelelli et al. [1] 2002 Branch-and-Price travel cost
Liang et al. [12] 2009 GA total cost with penalty minimizing the weighted total cost of vehicle dispatching and
Lai et al. [13] 2010 DE travel cost
Wang et al. [11] 2012 co-GA hierarchical objective travel distance.
Kassem et al. [14] 2013 Insertion heuristic, SA travel cost
Wang et al. [15] 2013 SA hierarchical objective Hof et al. [5] proposed an Adaptive Large Neighborhood
Wang et al. [16]
Hof et al. [5]
2015
2019
parallel SA
ALNS, Path-relinking
hierarchical objective
hierarchical objective
Search with Path Relinking (ALNS-PR) algorithm for a class
Shi et al. [17] 2020 VNS, TS hierarchical objective of VRPSPD, including VRPSPDTW. The algorithm involves
Tang et al. [18] 2021 Co-evolution of parameterized search weighted total cost
Liu et al. [6] 2021 MA hierarchical objective 7 removal operators and 4 insertion operators, dynamically
Wu et al. [7] 2023 ACO weighted total cost
Praxedes et al. [19] 2024 Branch-Cut-and-Price travel cost selected through an adaptive mechanism to iteratively de-
stroy and repair the solution. Moreover, it incorporates a
As a variant of the well-studied VRPSPD problem [3, 20], path-relinking component to explore promising search spaces
the first study of VRPSPDTW can be traced back to the work between elite solutions. The algorithm temporarily accepts
of Angelelli et al. [1]. Their pioneering study introduced a infeasible solutions by incorporating penalty terms into the
Branch-and-Price algorithm based on a set covering formula- objective function. Results showed the robustness of the
tion to address the problem, with the objective of minimizing ALNS-PR algorithm in solving this group of problems with
the total travel cost of the vehicles. Table I summarizes the competitive results on VRPSPDTW.
most important solution approaches that have been carried out Liu et al. [6] proposed the Memetic Algorithm with Ex-
in the literature on VRPSPDTW since then. tended Neighborhoods (MATE). This innovative approach
Regarding the objective of minimizing the total travel cost employs local search for small-step exploration and removal-
only, Lai et al. [13] presented an improved differential evo- reinsertion of large neighborhoods for large-step exploration,
lution algorithm (IDE), where they devised a novel decimal effectively navigating the search space. To generate promising
 3

solutions, MATE incorporates a RARI crossover combined parent solutions (line 8). This new solution S 0 is then improved
with regret insertion. Experiments on the WC instances showed by the feasible and infeasible route descent search (line 9).
the efficacy of the MATE algorithm by discovering 12 new The improved solution S 0 is used to update the population
best solutions. The authors also generated a new set of 20 according to the population management strategy (line 10).
large-scale benchmark instances, denoted as JD benchmark, If a superior best solution is found, S ∗ is updated, and the
derived from the real-world JD logistics network. Testing on stagnation counter is reset to 0 (lines 12-14). Otherwise, the
the JD instances confirmed MATE’s effectiveness in solving stagnation generation counter ϕst is incremented (lines 15-
VRPSPDTW in real-world scenarios. 16). The algorithm terminates and returns the best solution
Finally, Praxedes et al. [19] introduced the exact Branch S ∗ found (line 20) when the termination condition is met
Cut and Price (BCP) algorithm to solve a broad class of (line 7), which includes reaching a maximum number of
VRPSPD, including VRPSPDTW. With a 12-hour time limit, generations ϕmax , surpassing a maximum patience threshold
BCP achieved 45 optimal solutions out of 65 WC instances ρ for stagnation generations ϕst , or exceeding a predetermined
for minimizing travel cost only. maximum running time τ . Below, we provide a detailed
Given their superior performance on popular benchmarks description of each component of the algorithm.
compared to other approaches, ALNS-PR [5] and MATE
[6] are considered state-of-the-art methods for solving VRP- A. Preprocessing
SPDTW.
To improve computational efficiency, a preprocessing step
III. M EMETIC A LGORITHM WITH F EASIBLE AND is employed before executing the main algorithm to prune the
I NFEASIBLE ROUTE D ESCENT S EARCH FOR VRPSPDTW search space. As outlined in Section I, VRPSPDTW can be
The memetic algorithm framework [22, 23] combines the defined on a directed complete graph G = (V, E), where the
advantages of genetic algorithms and local search methods, vertices V = {v0 , v1 , . . . , vN } represent the depot and cus-
providing an interesting way to balance search diversification tomer nodes, and the edges E = {eij |vi , vj ∈ V} represent the
and intensification. Leveraging this powerful framework, we connections between the nodes. The preprocessing procedure
design dedicated search operators and strategies to address the removes edges that violate the time window or vehicle capacity
diverse features and constraints of VRPSPDTW. Specifically, constraints of VRPSPDTW. For any pair of nodes vi and vj ,
we propose a lightweight feasible and infeasible route descent the directed edge eij is eliminated if one of the following rules
search, a learning-based adaptive route-inheritance crossover, is true.
and a max-min normalization-based fitness-distance popula- ei + si + tij > lj (2)
tion management strategy to enable efficient search. di + dj > Q (3)
Algorithm 1 Main framework of MA-FIRD pi + dj > Q (4)
1: Input: Instance I, Population size NP , Maximum number of generations pi + pj > Q (5)
ϕmax , Patience for stagnation generations ρ, Maximum running time τ .
2: Output: The best solution S ∗ .
3: ϕ ← 0, ϕst ← 0 /* Current generation and stagnation generation
counter */ Rule (2) allows to prune the connection between vi and
4: I ← P reprocessing(I) /* Section III-A */ vj if the arrival time at vj after serving vi is later than the
5: P ← Initialization(I, NP ) /* Section III-B */
6: S ∗ ← BestSolution(P) /* Record current best solution */
allowed latest time lj for vj . Rules (3), (4), and (5) prohibit
7: while ϕ ≤ ϕmax and ϕst ≤ ρ and time() ≤ τ do the connection between vi and vj if the capacity constraint is
8: S 0 ← ARIX(P) /* Section III-D */ violated before, between, and after consecutive visits to vi and
9: S 0 ← F IRDSearch(S 0 ) /* Section III-C */
10: P ← U pdateP opulation(S 0 , P) /* Section III-F */
vj .
11: Sbest ← BestSolution(P) /* The best solution in this generation
*/
12: if f (Sbest ) < f (S ∗ ) then
B. Initialization
13: S ∗ ← Sbest /* Update the best solution */ The initialization stage generates a group of candidate
14: ϕst ← 0
15: else
solutions to form the initial population. A high-quality initial
16: ϕst ← ϕst + 1 /* Stagnation counter is incremented */ solution, characterized by its feasibility and a relatively lower
17: end if number of vehicles, helps the algorithm in its subsequent
18: ϕ=ϕ+1
19: end while
search. We use the RCRS algorithm [4] to generate such
20: return S ∗ initial solutions. RCRS is an insertion-based heuristic that
starts with an empty solution and iteratively inserts customer
Algorithm 1 outlines the framework of MA-FIRD. For nodes until all nodes are accommodated. Notably, RCRS only
a given problem instance I, the algorithm starts with a accepts feasible insertions, and a new route is introduced if
preprocessing step to reduce the instance (line 4). Then, NP no feasible insertion is possible with the existing routes. In
candidate solutions are generated to form the initial population addition to considering travel cost, RCRS incorporates two
(line 5), and the best solution S ∗ is recorded (line 6). The other metrics: Residual Capacity (RC), which represents the
population is updated iteratively by generating new candidate insertion’s freedom in terms of capacity, and Radial Surcharge
solutions (lines 7-19). In each generation, the adaptive route- (RS), which is to prevent late and unfavorable insertions
inheritance crossover creates a new solution from multiple of remote customer nodes. Using these three metrics, the
 4

RCRS algorithm efficiently constructs a relatively high-quality search is performed on the best feasible solution S 0 (lines 19-
solution in a short time. 21), and the final solution S 0 is returned (line 22).

C. Feasible and infeasible route descent search

This section presents the feasible and infeasible route de-
scent search (FIRD), which is one key component of the
proposed algorithm.
1) General FIRD procedure: As shown in [24], allowing
controlled exploration of infeasible solutions facilitates the
transition between structurally different feasible solutions and
thus improves the performance of local search algorithms.
Indeed, this approach has proven successful in several difficult Fig. 1: Illustration of the route descent procedure. The solution has four routes.
combinatorial optimization problems with complex constraints First, the shortest route 4-12 is removed. Then, the nodes 4 and 12 are inserted
into other routes using the infeasible regret insertion strategy.
(see, e.g., [25, 26, 27]).
We adopt this feasible and infeasible search approach in
The feasible and infeasible search takes two inputs: the
the context of solving VRPSPDTW. This is basically achieved
solution to be improved and an optional target solution. If a
by introducing penalty terms for constraint violations in the
target solution (e.g., the best-found solution S 0 in lines 5 and
evaluation function (Section III-C2). Combined with a set
13) is provided, the search terminates prematurely if a better
of local search operators (Section III-C3), the feasible and
solution is found. If no target solution is provided, the search
infeasible route descent search is able to tunnel through
intensively improves the input solution (e.g., the best-found
infeasible regions of the search space to obtain high-quality
solution S 0 in line 20).
solutions that would otherwise be difficult to attain. 2) Evaluation function:
Algorithm 2 Feasible and infeasible route descent search Minimize feval (S) = D(S) + λ1 · V1 (S) + λ2 · V2 (S) (6)
1: Input: The solution S for improvement. L
M X
X
2: Output: Improved solution S 0 . Subject to D(S) = cni,j ni,j+1
3: S0 ← S /* Current best solution */ i=1 j=0
4: if S is infeasible then P L
5: S ← F easibleInf easibleSearch(S, S 0 ) /* Try to make S ck X X
V1 (S) = P max(ai,j − li,j , 0)
feasible */ tk i∈R j=0
r
6: S0 ← S
7: end if V2 (S) = (2 · max ck − min ck )
8: while S is feasible do L
P
i∈Rr max(maxj=0 (qni,j − Q), 0)
9: S ← RouteDescent(S) /* Remove a route of S */
10: if S is better than S 0 then Q
11: S0 ← S S = {R1 , . . . , RM }
12: else Ri = {ni,0 , ni,1 , . . . , ni,L , ni,L+1 }, i = 1, . . . , M
13: S ← F easibleInf easibleSearch(S, S 0 ) /* Optimize S and try
to find a better solution than S 0 */
14: if S is better than S 0 then
15: S0 ← S
16: end if
(
17: end if λi · κ if Vi = 0
λi = λi i = 1, 2, κ ∈ (0, 1) (7)
18: end while if Vi > 0
19: if S 0 is feasible then κ
20: S 0 ← F easibleInf easibleSearch(S 0 , ∅) /* Intensive search from
the best-found solution S 0 */
The feasible and infeasible search dynamically accepts
21: end if feasible and infeasible solutions to explore diverse areas of
22: return S 0 the search space. To achieve this, we define an evaluation
function that incorporates penalty terms of time window and
As shown in Algorithm 2, the FIRD procedure starts with capacity constraint violations, as shown in Equation (6). Since
an input solution S, which can be either feasible or infeasible. the number of vehicles (primary objective) is handled through
If S is infeasible, the feasible and infeasible search is applied the route descent strategy, the feasible and infeasible search
to ensure that the solution entering the route descent procedure only focuses on the travel distance objective D(S). V1 (S)
is feasible (lines 4-7). The route descent procedure removes and V2 (S) represent the normalized violation values of time
the shortest route from the solution S, and its nodes are window and capacity constraints, respectively. The time-warp
reinserted into other routes using the infeasible regret insertion technique [28, 8] is adopted to address violation of time
strategy presented in Section III-E (line 9), as illustrated in window constraints. During the search process, the coefficients
Figure 1. If the resulting solution of route descent is not λ1 and λ2 are dynamically adjusted at each move step based on
feasible or not better than the current best solution S 0 , the the feasibility of the constraints. This adjustment is facilitated
feasible and infeasible search is re-applied to try to find a by the coefficient adjustment factor κ, as depicted in Equation
better solution (line 13). The route descent procedure continues (7). If the time window or capacity constraint is violated, the
until it becomes impossible to find a feasible solution with the corresponding coefficient is increased to prioritize satisfying
current number of vehicles (lines 8-18). Finally, an intensive that constraint. If both constraints are violated, the coefficients
 5

bandit) adaptive mechanism is embedded to make the best

possible decisions (see Section III-E).

Algorithm 3 Adaptive Route-Inheritance Crossover

1: Input: Population P.
2: Output: Offspring solution S 0 .
3: S0 ← ∅ /* Initialize offspring solution */
4: Determine the number of parents Np , the insertion operator I based on
the adaptive strategy /* Section III-E */
5: Randomly select Np solutions from the population P as parents P
(a) Relocate (b) Swap (c) 2-opt*
6: Pd ← P [0] /* Dominant parent */
7: Rd ← ∀R ∈ P [0] /* Dominant routes */
Fig. 2: Illustration of operators and concatenation operation 8: Rp ← ∀R ∈ P [1 :] /* Routes from other parents form the route pool */
9: Determine the retention ratio γr
10: Nr ← bγr × |Pd |c /* Number of retained routes from the dominant
routes */
λ1 and λ2 are both increased. Conversely, if there is no 11: Ni ← |Pd | − Nr /* Number of introduced routes from the route pool */
constraint violation, the coefficients λ1 and λ2 are decreased 12: while route number of S 0 is smaller than the route number of Pd do
13: if Nr > 0 then
to encourage exploration of infeasible areas. 14: Select the route with the lowest conflict rate from Rd and add it
3) Move operators and move evaluation: The feasible and to S 0
infeasible search employs a set of local search move operators 15: Nr ← Nr − 1
16: end if
to explore the search space, including Relocate, Swap, and 2- 17: if Ni > 0 then
opt*. Relocate moves a sequence of nodes to the same or 18: Select the route with the lowest conflict rate from Rp and add it
another route. Swap exchanges two sequences of nodes in the to S 0
19: Ni ← Ni − 1
same route or different routes. 2-opt* takes place between two 20: end if
routes. It first breaks one edge in two candidate routes, and 21: end while
then exchanges and reconnects the sequences of nodes. Here, 22: Remove redundant nodes in S 0
23: U ← {n ∈ V|n ∈ / S0} /* Unrouted nodes */
the sequence length of nodes for Relocate and Swap operators 24: Insert the unrouted nodes U into S 0 using the insertion operator I
ranges from 1 to 3. Considering their similar complexity of the 25: return S 0

employed operators, a random order is employed to execute

these operators. The feasible and infeasible search stops under The ARIX crossover (see Algorithm 3) starts by determining
several conditions: when no improvement is detected after the number of parents Np (≥ 2) and the insertion operator I
applying all operators in the sequence, or when the maximum based on the adaptive strategy (line 4). Then, ARIX randomly
iteration limit is reached, or when a better solution than the selects Np parent solutions from the population P (line 5),
given target solution is found. designating the first parent as the dominant parent Pd , and
The search procedure adopts the best-improvement strat- gathering its routes in Rd (lines 6-7). Routes from the other
egy to apply each local search operator. To speed up move parents are pooled in Rp (line 8). The retention ratio γr
evaluation and accommodate infeasible solutions, we introduce is then determined (line 9), which is dynamically adjusted
an infeasibility-allowed constant-time move evaluation method during the search process and influences the number of routes
inspired by prior work [6, 8]. Since the used operators affect retained from Pd and the number of introduced routes Ni
only one or two routes, the move evaluation process can be (lines 10-11). ARIX then iteratively constructs the offspring
performed in constant time by computing the variation (called solution S 0 by selecting routes from Rd and Rp based on
move gain) of the evaluation function between the original Nr and Ni (lines 12-21). Routes with the lowest conflict rate,
solution and a new solution. We conceptualize moves as rear- representing the ratio of nodes in S 0 to the total nodes in
rangements of subsequences of routes, using a concatenation the route, are prioritized for selection. This process continues
operation denoted as ⊕ to symbolize the merging of two route until the number of routes in S 0 reaches that of Pd (line 12).
subsequences. Illustrative examples of various operators are Redundant nodes in S 0 are subsequently removed based on
depicted in Figure 2, while detailed concatenation operation selection order (line 22), and unrouted nodes are inserted using
calculations are provided in the online supplement. With this insertion operator I to obtain the final offspring (line 24). If the
method, the computation of the D(·), V1 (·), and V2 (·) values insertion leads to infeasibility, FIRD is employed to establish
in the evaluation function can be done in constant time, which the feasibility.
greatly speeds up the evaluation process. Five insertion operators are designed for the insertion of
unrouted nodes: Feasible best insertion (FBI) and infeasible
best insertion (IBI) prioritize the minimal cost increase, with
D. Adaptive route-inheritance crossover FBI creating a new route if no feasible insertion is possible,
We propose an adaptive route-inheritance crossover (ARIX) while IBI accepts infeasible insertions without introducing
for VRPSPDTW, which relies on inheriting routes from parent new routes. Feasible regret insertion (FRI) and infeasible
solutions and employs feasible and infeasible insertion op- regret insertion (IRI) evaluate each insertion based on regret
erators to complete the construction of each new offspring value, defined as the difference in travel cost between the best
solution. To decide the number of parent solutions and the and second-best insertions. Random insertion (RI) inserts an
insertion operators to apply, a learning-based (multi-armed unrouted customer randomly into a random route.
 6

In our context, we define the set of actions A as the

set {2, ..., |P|} of all possible Np parents, and the set
{F BI, IBI, F RI, IRI, RI} of all possible insertion oper-
ators I. The reward r of the action a is defined as the
improvement of the objective function value of the offspring
(a) Parent solutions and their routes. S 0 from the action a compared with the dominant parent Pd .

F. Population management
Following [31, 32], we use a fitness-distance based popula-
tion management strategy, which relies on the extended fitness
function ψ defined in Equation (9). This function is derived
from the max-min normalization of the objective value and
the distance within the population. The coefficient ξ adjusts
(b) Inheritance from the dominant routes Rd and the route pool Rp .
the balance between the effects of the objective value and the
distance, with higher fitness values indicating superior solution
fitness. Furthermore, Equation (10) quantifies the distance of a
given solution S from the population P, while Equation (11)
measures the dissimilarity between two solutions. ES1 and ES2
represent the sets of edges in solutions S1 and S2 , respectively.
(c) Redundant nodes removal and unrouted nodes insertion.
max f − f (S) DP (S) − min DP
ψ(S) = +ξ· (9)
Fig. 3: Illustration of the ARIX crossover. max f − min f max DP − min DP
Subject to max f = max f (Si ), min f = min f (Si )
Si ∈P Si ∈P
max DP = max DP (Si ), min DP = min DP (Si )
Figure 3 illustrates the ARIX process with three parent Si ∈P Si ∈P
solutions, P1 , P2 , and P3 (Figure 3a), where P1 is the DP (S) = min DS (S, Si ), Si 6= S (10)
Si ∈P
dominant parent Pd , whose routes form the dominant routes 
|ES1 ∩ ES2 |

Rd . The routes from P2 and P3 form the route pool Rp . Given DS (S1 , S2 ) = 100% 1 − , ES1 , ES2 ∈ E
max(|ES1 |, |ES2 |)
that Pd has four routes and the retention ratio γr is set to (11)
0.5 in this example, ARIX retains two routes from Rd and
selects two routes from Rp . Next (Figure 3b), the inheritance During the execution of the algorithm, the population is
process begins. Each route is marked with its conflict rate. The expanded for NP /2 generations before being reduced to its
offspring solution S 0 is constructed by alternately selecting original size NP . This resizing involves eliminating solutions
routes with the lowest conflict rates from Rd and Rp until the with the worst ψ values, thereby maintaining an adaptive
number of routes in S 0 matches that of the dominant parent Pd . balance in the population composition.
Finally (Figure 3c), the redundant node 4 is removed from S 0 ,
and the unrouted node 5 is inserted using the selected insertion G. Discussion
operator, resulting in the final offspring solution S 0 . The proposed algorithm has a number of novel features
compared to existing studies on VRPSPDTW, particularly in
E. Learning-based adaptive strategy for the ARIX crossover terms of local optimization and crossover.
For local optimization, unlike the complex large neighbor-
During the ARIX crossover, ARIX faces two main decision- hood search techniques used in recent studies [5, 6, 7], we em-
making problems to decide the number of parents Np used to ploy a lightweight feasible and infeasible search with simple
generate offspring solutions and the insertion operator I used local search operators. Key distinctions include the manage-
to insert the unrouted nodes in the last step of the crossover ment of penalty terms, the infeasibility-allowed constant-time
process. To help the ARIX crossover to make the best possible move evaluation, and the route descent strategy. Unlike penalty
decisions for these choices, we employ an adaptive strategy design in ALNS-PR [5], we normalize penalty terms and
based on multi-armed bandit (MAB) [29]. Specifically, we use dynamically adjust coefficients to balance the exploration of
the UCB1 algorithm from the Upper Confidence Bound (UCB) feasible and infeasible solutions for two constraints. Compared
family of algorithms [30]. For a set of actions ai ∈ A, UCB1 to the evaluation method in MATE [6], which focuses only on
calculates the upper confidence bound of the estimated reward feasible solutions, our method accommodates both feasible and
for each action ai at step t using Equation (8). R̄id is the infeasible solutions and facilitates the calculation of violation
empirical mean reward of action ai , Ni is the number of times values. Additionally, unlike VNS-BSTS [17], which handles
action ai has been selected so far. The action ai with the optimization objectives in two stages, our algorithm simultane-
maximum value of U CB1(i, t) is selected at each step. ously optimizes the number of vehicles and the travel distance
by integrating the route descent strategy into the feasible and
s
2 · ln(t)
U CB1(i, t) = R̄id + (8) infeasible search.
Ni
 7

For crossover, although our ARIX crossover follows the B. Experimental protocol and parameters tuning
similar idea of inheriting routes from parent solutions, ARIX The MA-FIRD algorithm was coded in C++3 and run on a
differs from those in [11, 6] by using multiple parents and computer with an AMD EPYC 7282 2.8 GHz processor and
insertion operators that are adaptively determined by a learning 4 GB RAM, running Linux.
mechanism to create more diversified and promising offspring To balance solution quality and computational efficiency,
solutions. In addition, ARIX strategically evaluates and selects we empirically determined several parameters. The population
the inherited routes based on the conflict rate instead of the size NP was set to 10, the maximum number of generations
random selection to improve the offspring quality. ϕmax was set to 5000, and the patience for stagnation gener-
In summary, we present a novel and effective approach to ations ρ was set to 500. Additionally, the algorithm relies on
solve VRPSPDTW. Its strategies and operators can also be three critical parameters: the coefficient of the fitness function
extended to other VRPs. The route descent strategy could be ξ, the coefficient adjustment factor κ, and the initial retention
useful for VRP variants with a flexible number of vehicles. ratio γr0 . To find a suitable setting for these parameters,
The management of penalty terms and the constant-time move we conducted an exhaustive parameter sensitivity analysis,
evaluation can be applied to VRPs with complex constraints. detailed in the online supplement. The resulting parameter
The idea of our adaptive route-inheritance crossover, leverag- values are summarized in Table II.
ing a learning-based strategy, could be extended to various
TABLE II
VRPs due to its general nature. PARAMETERS TUNING RESULTS
Parameter Value range Final value
IV. C OMPUTATIONAL RESULTS ξ : coefficient of the fitness function (0, 1) 0.8
κ : coefficient adjustment factor (0, 1) 0.5
γr0 : initial retention ratio (0, 1) 0.9
We show computational experiments to evaluate the MA-
FIRD algorithm and comparisons with the state-of-the-art
algorithms based on two VRPSPDTW benchmarks. To ensure a fair comparison, the maximum running time τ
for the JD benchmark is set to 7200 seconds following [6].
Each instance was solved with 30 independent runs of the
A. Benchmark instances algorithm.
The first benchmark consists of the popular 65 WC in-
stances1 from [11]. Among these instances, 9 are of small C. Reference algorithms
size with 10 to 50 customers, while the remaining 56 are of For the WC instances, we compare our results with the best-
medium size with 100 customers. This benchmark is divided known solutions (BKS) ever reported in the literature and with
into three types: C instances with clustered customers, R the top performing algorithms ALNS-PR [5] and MATE [6].
instances with customer locations generated uniformly and These two algorithms outperformed the other algorithms, and
randomly on a square, and RC instances with a combination together produced the current best-known solutions for the WC
of randomly placed and clustered customers. These instances instances. For the JD instances, we focus on a comparative
are further classified to two categories: instances with narrow study with MATE [6], which tested these 20 large-scale
time windows and small vehicle capacity, and instances with instances. Since the code of MATE is open source, we reran
large time windows and large vehicle capacity. Following it on our computer with the same experimental protocol as for
[11, 15, 16, 5, 17, 6], we adopt the hierarchical objective for our algorithm and the results are indicated by MATE*.
the WC benchmark with the number of vehicles and travel The experimental environments of the reference algorithms
distance serving as the primary and secondary objectives, are as follows. ALNS-PR [5] was programmed in Java and
respectively. The instances of the WC benchmark have been performed on a Windows 10 Professional desktop computer
extensively tested in the literature [11, 15, 16, 5, 17, 6, 7, 19] with an Inter Core i5-6600 3.30 GHz processor and 16 GB
The second benchmark consists of the 20 large JD in- RAM. MATE [6] was programmed in C++ and executed on
stances2 from [6], which was derived from real distribution a machine with an Intel Xeon E5-2699A V4 2.40 GHz and
system of JD Logistics company. This benchmark contains 128 GB RAM, running Centos 7.5. To account for different
4 groups of 5 instances with 200, 400, 600, 800, and 1000 experimental environments, we apply a time conversion ratio
customers. Unlike the WC benchmark designed for the hier- γ to standardize the computing time in following experiments.
archical objective (first the number of vehicles and then the Detailed information can be found in the online supplement.
travel distance), the JD benchmark was initially designed to
minimize the total cost, which is the weighted sum of the D. Computational results and comparisons
number of vehicles and the travel distance with predefined
weights. This benchmark allows us to verify the usefulness of The comparative results on the two benchmark sets are
the proposed algorithm in real-world applications. Since this summarized in Table III, which provides a global view of
benchmark set is very recent, it has only been tested in [6]. the performance of the proposed algorithm across different
3 The executable file and our solutions for benchmark instances are available
1 https://oz.nthu.edu.tw/∼d933810/test.htm
at https://github.com/leizy1008/MA-FIRD, the source code will be available
2 https://github.com/senshineL/VRPenstein upon publication of the paper.
 8

benchmark groups. These groups are identified as WC-S, WC-

C, WC-R, WC-RC, and JD, corresponding to the small WC, Gap = 100% ·
(2000 · M + D) − (2000 · MBKS + DBKS )
(12)
clustered WC, random WC, clustered-random WC instances, 2000 · MBKS + DBKS
and JD instances, respectively. The row #Instances gives the
number of instances within each group. The table contains
Table V shows that the MA-FIRD algorithm outperforms
information about the instances where the proposed algorithm
the reference algorithms. It reaches the BKS values for 48
shows better-same-worse results compared to the BKS and the
instances and improves the best-known results for the 8
reference algorithms. The ’-’ symbol indicates the unavail-
remaining instances (indicated in bold). Notably, within these
ability of the results. In addition, the Wilcoxon signed-rank test
8 instances, MA-FIRD obtains a new best upper bound for the
[33] with a confidence level of 0.05 was conducted to ascertain
instance Rdp104 with a smaller M value (with 9 vehicles in-
the statistical significance of the results, and the corresponding
stead of 10 for the BKS). MA-FIRD performs consistently and
p-values are given in the table.
stably with small, and even zero, standard deviations for the
TABLE III number of vehicles and the travel distance for most instances.
S UMMARY OF COMPARATIVE RESULTS ON BENCHMARK INSTANCES For the primary objective M (the number of vehicles), MA-
BETWEEN MA-FIRD AND THE REFERENCES .
FIRD has a better mean value of 7.27 against 7.30 for ALNS-
Group
#Instances
WC-S
9
WC-C
17
WC-R
23
WC-RC
16
JD
20
Total
85
p-value PR, 7.36 for MATE, and 7.45 for MATE*. In terms of the
MA-FIRD
MA-FIRD
vs
vs
BKS
ALNS-PR [5]
0-9-0
-
0-17-0
3-14-0
5-18-0
11-12-0
3-13-0
10-6-0
20-0-0
-
28-57-0
24-41-0
3.79E-06
1.82E-05
gap to the BKS, MA-FIRD holds the best gap of -0.19% (a
MA-FIRD
MA-FIRD
vs
vs
MATE [6]
MATE*
0-9-0
0-9-0
2-15-0
2-15-0
10-13-0
11-12-0
5-11-0
6-10-0
20-0-0
20-0-0
37-48-0
39-46-0
8.39E-08
9.56E-09 negative gap, indicating an improved result) against 0.14% for
ALNS-PR, 2.00% for MATE, and 2.79% for MATE*. In terms
Table III shows that the MA-FIRD algorithm consistently of the running time, all algorithms perform similarly, except
provides comparable or superior results compared to the BKS for MATE*, which requires a longer running time.
and the reference algorithms. For the WC benchmark, MA- 2) Comparison on the large JD instances: Table VI shows
FIRD achieved 8 new best upper bounds and reached the BKS the comparison results for the large JD instances between
values for all remaining instances. For the JD benchmark, MATE, MATE*, and MA-FIRD. Recall that for the JD in-
MA-FIRD established new best upper bounds for all 20 stances, the objective is to minimize the total cost, which is
instances. The associated p-values (<< 0.05) indicates that the weighted sum of the number of vehicles and the travel
the proposed algorithm statistically performs better than the distance. The columns fBest and fAvg ± std show the best
BKS and reference algorithms. and average total cost of each algorithm, including standard
1) Comparison on the small and medium WC instances: deviations. Using MATE’s results as a baseline, the columns
Tables IV and V show a detailed comparison of the proposed GapBest and GapAvg show the relative gap for the best and
algorithm with the BKS and the reference algorithms on an average values obtained by MATE* and MA-FIRD over 30
instance-by-instance basis for the small and medium WC in- independent runs. So a negative gap indicates an improved
stances, where the hierarchical objective function aims to first result. The number of vehicles M and the travel distance D
minimize the number of vehicles and then the travel distance. are also reported for reference.
The results of ALNS-PR and MATE are taken directly from It is evident that the MA-FIRD algorithm consistently
[5] and [6] respectively, while MATE* shows the results we outperforms both MATE and MATE* with significant im-
obtained by re-running the MATE code on our computer. provements in all instances in terms of the total cost, with
The Instance column indicates the instance names, and a negative mean gap of -2.28% for average results. Even the
BKS shows the best-known solutions in the literature. The average objective value is better than the best reference value
M and D columns display the number of vehicles and travel (144677.53 against 179464.90). If we examine the number
distance obtained by each algorithm. M ±std and D±std of vehicles and the travel distances separately, we observe
respectively show the average values and standard deviations that MA-FIRD dominates MATE and MATE* in both criteria.
of the number of vehicles and the travel distance obtained by These results demonstrate the effectiveness of the proposed
the corresponding algorithm. The t column shows the running algorithm in solving large instances in real-world scenarios.
time in seconds for each algorithm, while the γ· t column 3) Time-to-target analysis: To further compare the compu-
reports the converted time using the time conversion ratio γ. tational efficiency between the proposed MA-FIRD algorithm
For the 9 small WC instances, despite small differences in and the MATE algorithm, we perform a time-to-target (TTT)
computation time, all methods solve these instances consis- analysis [34, 35] on the medium WC instances. This analysis
tently and reliably within a short time. These instances do not involves solving each instance 100 times for both algorithms,
represent a challenge for the compared algorithms. measuring the probability distribution of the required time to
For the 56 medium WC instances, a weight value of 2000 achieve a predefined target objective value. In this analysis,
is assigned to the number of vehicles (M ), and 1 to the travel the maximum running time per run is set to 300 seconds.
distance (D) following [6]. The Gap column shows the relative Figure 4 visually depicts the results for four representative
gap between the solutions and the BKS, calculated using instances Cdp101, Rdp203, RCdp101, and RCdp204. The re-
Equation (12). A negative gap indicates that the algorithm has sults confirm that the proposed algorithm consistently achieves
found a better solution than the BKS, and these improved the given targets within a short running time, indicating its
values are highlighted in bold. remarkable convergence speed and computational efficiency.
 9

TABLE IV
C OMPARATIVE RESULTS ON THE 9 SMALL WC BENCHMARK INSTANCES BETWEEN MATE, MATE* AND MA-FIRD.
MATE (Avg) MATE* (Avg) MA-FIRD (Avg)
Instance
M ± std D ± std t γ·t M ± std D ± std t M ± std D ± std t
RCdp1001 3.00±0.00 348.98±0.00 1.00 1.24 3.00±0.00 348.98±0.00 0.06 3.00±0.00 348.98±0.00 0.16
RCdp1004 2.00±0.00 216.69±0.00 1.00 1.24 2.00±0.00 216.69±0.00 0.16 2.00±0.00 216.69±0.00 0.28
RCdp1007 2.00±0.00 310.81±0.00 1.00 1.24 2.00±0.00 310.81±0.00 0.19 2.00±0.00 310.81±0.00 0.26
RCdp2501 5.00±0.00 551.05±0.00 1.00 1.24 5.00±0.00 551.05±0.00 0.25 5.00±0.00 551.05±0.00 0.55
RCdp2504 4.00±0.00 473.46±0.00 1.00 1.24 4.00±0.00 473.46±0.00 1.28 4.00±0.00 473.46±0.00 1.42
RCdp2507 5.00±0.00 540.87±0.00 1.00 1.24 5.00±0.00 540.87±0.00 1.23 5.00±0.00 540.87±0.00 1.25
RCdp5001 9.00±0.00 994.18±0.00 1.00 1.24 9.00±0.00 994.18±0.00 1.38 9.00±0.00 994.18±0.00 2.56
RCdp5004 6.00±0.00 733.21±0.00 9.00 11.16 6.00±0.00 733.21±0.00 9.59 6.00±0.00 733.21±0.00 19.25
RCdp5007 7.00±0.00 809.72±0.00 9.00 11.16 7.00±0.00 809.72±0.00 9.00 7.00±0.00 809.72±0.00 6.51
Mean 4.78±0.00 553.22±0.00 2.78 3.44 4.78±0.00 553.22±0.00 2.57 4.78±0.00 553.22±0.00 3.58

TABLE V
C OMPARATIVE RESULTS ON THE 56 MEDIUM WC BENCHMARK INSTANCES BETWEEN BKS, ALNS-PR, MATE, MATE* AND MA-FIRD.
BKS ALNS-PR MATE MATE* MA-FIRD (Best) MA-FIRD (Avg)
Instance
M D M D t γ·t Gap M D t γ·t Gap M D t Gap M D t Gap M ± std D± std t Gap
Cdp101 11 976.04 11 976.04 19.30 23.93 0.00% 11 976.04 102.04 108.16 0.00% 11 976.04 10.08 0.00% 11 976.04 14.46 0.00%11.00±0.00 977.36±0.53 18.77 0.01%
Cdp102 10 941.49 10 941.49 28.11 34.86 0.00% 10 941.49 78.18 82.87 0.00% 10 941.49 32.23 0.00% 10 941.49 51.50 0.00%10.00±0.00 941.49±0.00 62.36 0.00%
Cdp103 10 892.98 10 892.98 48.03 59.56 0.00% 10 892.98 78.66 83.38 0.00% 10 892.98 100.54 0.00% 10 892.98 136.24 0.00%10.00±0.00 893.75±1.11 175.70 0.00%
Cdp104 10 871.40 10 871.40 46.51 57.67 0.00% 10 871.40 79.41 84.17 0.00% 10 871.40 111.53 0.00% 10 871.40 226.19 0.00%10.00±0.00 871.51±0.60 247.13 0.00%
Cdp105 10 1053.12 10 1053.12 15.97 19.80 0.00% 10 1074.51 67.36 71.40 0.10% 11 980.55 10.00 9.16% 10 1053.12 11.09 0.00%10.00±0.00 1053.12±0.00 15.67 0.00%
Cdp106 10 963.45 10 967.71 17.36 21.53 0.02% 10 963.45 100.87 106.92 0.00% 10 963.45 18.65 0.00% 10 963.45 19.56 0.00%10.00±0.00 963.45±0.00 25.72 0.00%
Cdp107 10 987.64 10 987.64 18.06 22.39 0.00% 10 988.60 88.37 93.67 0.00% 10 1072.35 18.49 0.40% 10 987.64 23.07 0.00%10.00±0.00 988.09±1.41 33.73 0.00%
Cdp108 10 932.50 10 932.88 18.22 22.59 0.00% 10 932.49 76.52 81.11 0.00% 10 932.50 90.64 0.00% 10 932.50 36.99 0.00%10.00±0.00 932.50±0.00 45.96 0.00%
Cdp109 10 909.27 10 910.95 38.45 47.68 0.01% 10 909.27 78.86 83.59 0.00% 10 909.27 104.62 0.00% 10 909.27 107.32 0.00%10.00±0.00 909.27±0.00 131.31 0.00%
Cdp201 3 591.56 3 591.56 24.37 30.22 0.00% 3 591.56 90.21 95.62 0.00% 3 591.56 24.27 0.00% 3 591.56 3.98 0.00% 3.00±0.00 591.56±0.00 4.43 0.00%
Cdp202 3 591.56 3 591.56 46.09 57.15 0.00% 3 591.56 158.59 168.11 0.00% 3 591.56 36.36 0.00% 3 591.56 10.09 0.00% 3.00±0.00 591.56±0.00 10.90 0.00%
Cdp203 3 591.17 3 591.17 44.19 54.80 0.00% 3 591.17 62.22 65.95 0.00% 3 591.17 54.60 0.00% 3 591.17 26.50 0.00% 3.00±0.00 591.17±0.00 29.62 0.00%
Cdp204 3 590.60 3 590.6 51.76 64.18 0.00% 3 590.60 78.93 83.67 0.00% 3 590.60 100.62 0.00% 3 590.60 44.40 0.00% 3.00±0.00 590.62±0.10 48.30 0.00%
Cdp205 3 588.88 3 588.88 36.35 45.07 0.00% 3 588.88 134.77 142.86 0.00% 3 588.88 30.74 0.00% 3 588.88 4.42 0.00% 3.00±0.00 588.88±0.00 4.87 0.00%
Cdp206 3 588.49 3 588.49 36.39 45.12 0.00% 3 588.49 176.76 187.37 0.00% 3 588.49 37.96 0.00% 3 588.49 7.29 0.00% 3.00±0.00 588.49±0.00 7.64 0.00%
Cdp207 3 588.29 3 588.29 39.83 49.39 0.00% 3 588.29 196.92 208.74 0.00% 3 588.29 39.85 0.00% 3 588.29 11.51 0.00% 3.00±0.00 588.29±0.00 12.49 0.00%
Cdp208 3 588.32 3 588.32 34.47 42.74 0.00% 3 588.32 215.96 228.92 0.00% 3 588.32 46.42 0.00% 3 588.32 10.33 0.00% 3.00±0.00 588.32±0.00 11.66 0.00%
Rdp101 19 1650.80 19 1650.80 22.86 28.35 0.00% 19 1650.80 53.82 57.05 0.00% 19 1650.80 6.56 0.00% 19 1650.80 10.27 0.00%19.00±0.00 1651.24±1.04 12.28 0.00%
Rdp102 17 1486.12 17 1486.12 20.89 25.90 0.00% 17 1486.12 49.49 52.46 0.00% 17 1486.12 20.62 0.00% 17 1486.12 22.18 0.00%17.00±0.00 1488.95±5.14 35.19 0.01%
Rdp103 13 1294.64 13 1297.01 17.83 22.11 0.01% 13 1294.64 78.82 83.55 0.00% 13 1294.64 72.16 0.00% 13 1294.64 131.76 0.00%13.77±0.43 1233.31±34.96 58.16 5.39%
Rdp104 10 984.81 10 984.81 28.40 35.22 0.00% 10 984.81 79.19 83.94 0.00% 10 984.81 116.38 0.00% 9 1026.42 73.55 -9.33% 9.97±0.18 987.83±7.95 75.87 -0.30%
Rdp105 14 1377.11 14 1377.11 17.58 21.80 0.00% 14 1377.11 105.43 111.76 0.00% 14 1377.11 12.08 0.00% 14 1377.11 10.68 0.00%14.00±0.00 1377.11±0.00 13.46 0.00%
Rdp106 12 1252.03 12 1252.03 27.54 34.15 0.00% 12 1252.03 78.81 83.54 0.00% 12 1252.03 44.62 0.00% 12 1252.03 103.97 0.00%12.00±0.00 1257.26±4.05 59.03 0.02%
Rdp107 10 1112.55 10 1121.86 18.79 23.30 0.04% 10 1124.90 78.99 83.73 0.06% 10 1126.94 180.03 0.07% 10 1112.55 107.73 0.00%10.00±0.00 1121.51±6.38 95.27 0.04%
Rdp108 9 965.22 9 965.54 20.60 25.54 0.00% 9 965.22 79.71 84.49 0.00% 9 965.22 603.82 0.00% 9 965.22 112.70 0.00% 9.00±0.00 970.27±4.84 119.47 0.03%
Rdp109 11 1194.73 11 1194.73 16.08 19.94 0.00% 11 1203.97 76.69 81.29 0.04% 12 1156.84 31.43 8.46% 11 1194.73 36.80 0.00%11.47±0.51 1183.27±27.94 41.06 3.97%
Rdp110 10 1121.46 10 1148.20 19.13 23.72 0.13% 10 1166.47 78.42 83.13 0.21% 11 1081.56 112.89 9.28% 10 1121.46 68.97 0.00%10.23±0.43 1131.49±34.63 87.89 2.26%
Rdp111 10 1098.84 10 1098.84 22.15 27.47 0.00% 10 1098.84 79.07 83.81 0.00% 10 1098.84 129.21 0.00% 10 1098.84 101.44 0.00%10.17±0.38 1096.63±19.15 96.34 1.57%
Rdp112 9 1010.42 9 1010.42 28.63 35.50 0.00% 10 953.63 78.32 83.02 10.22% 10 953.63 174.89 10.22% 9 997.27 101.98 -0.07% 9.93±0.25 961.87±11.65 38.03 9.56%
Rdp201 4 1252.37 4 1253.23 33.44 41.47 0.01% 4 1252.37 78.80 83.53 0.00% 4 1252.37 67.51 0.00% 4 1252.37 22.12 0.00% 4.00±0.00 1253.57±2.29 24.55 0.01%
Rdp202 3 1191.70 3 1191.70 46.71 57.92 0.00% 3 1223.69 77.46 82.11 0.44% 3 1244.54 165.95 0.73% 3 1191.70 39.56 0.00% 3.00±0.00 1193.36±2.53 54.76 0.02%
Rdp203 3 939.58 3 946.28 84.91 105.29 0.10% 3 939.58 73.94 78.38 0.00% 3 939.58 585.88 0.00% 3 939.50 88.73 0.00% 3.00±0.00 941.67±1.49 96.49 0.03%
Rdp204 2 833.09 2 833.09 111.52 138.28 0.00% 2 835.28 78.54 83.25 0.05% 2 860.16 520.65 0.56% 2 825.52 240.22 -0.16% 2.00±0.00 834.61±3.56 124.19 0.03%
Rdp205 3 994.43 3 994.43 80.35 99.63 0.00% 3 994.43 80.49 85.32 0.00% 3 994.43 289.25 0.00% 3 994.43 32.95 0.00% 3.00±0.00 997.59±6.63 48.02 0.05%
Rdp206 3 906.14 3 913.68 89.88 111.45 0.11% 3 906.14 78.23 82.92 0.00% 3 906.14 341.72 0.00% 3 906.14 106.05 0.00% 3.00±0.00 908.81±3.91 84.06 0.04%
Rdp207 2 890.61 2 890.61 82.51 102.31 0.00% 3 811.51 71.13 75.40 39.28% 3 811.51 579.05 39.28% 2 890.61 101.82 0.00% 2.03±0.18 896.76±18.39 134.66 1.49%
Rdp208 2 726.82 2 726.82 100.25 124.31 0.00% 2 726.82 81.23 86.10 0.00% 2 728.22 487.98 0.03% 2 726.82 118.93 0.00% 2.00±0.00 732.89±4.71 90.82 0.13%
Rdp209 3 909.16 3 909.16 86.59 107.37 0.00% 3 909.16 76.27 80.85 0.00% 3 909.16 1060.62 0.00% 3 909.16 102.72 0.00% 3.00±0.00 911.51±2.87 86.53 0.03%
Rdp210 3 939.37 3 939.37 82.84 102.72 0.00% 3 939.37 78.07 82.75 0.00% 3 939.37 449.32 0.00% 3 939.37 106.08 0.00% 3.00±0.00 945.45±5.99 86.71 0.09%
Rdp211 2 904.44 2 904.44 85.87 106.48 0.00% 3 767.82 68.64 72.76 37.99% 3 767.82 412.68 37.99% 2 885.71 55.53 -0.38% 2.03±0.18 905.64±29.15 79.89 1.38%
RCdp101 14 1708.21 14 1776.58 10.59 13.13 0.23% 14 1708.21 27.13 28.76 0.00% 14 1708.21 20.63 0.00% 14 1708.21 15.41 0.00%14.83±0.38 1647.08±27.81 23.14 5.40%
RCdp102 12 1570.28 12 1583.62 19.19 23.80 0.05% 12 1570.28 78.92 83.66 0.00% 12 1589.91 71.53 0.08% 12 1570.28 114.33 0.00%12.03±0.18 1575.69±20.05 61.87 0.28%
RCdp103 11 1282.53 11 1283.52 29.08 36.06 0.00% 11 1282.53 78.64 83.36 0.00% 11 1282.53 204.05 0.00% 11 1282.53 111.63 0.00%11.00±0.00 1285.74±5.18 99.93 0.01%
RCdp104 10 1171.37 10 1171.65 22.57 27.99 0.00% 10 1171.37 79.88 84.67 0.00% 10 1171.37 186.61 0.00% 10 1171.37 100.26 0.00%10.00±0.00 1171.56±0.51 146.17 0.00%
RCdp105 13 1632.29 14 1548.96 16.95 21.02 6.94% 13 1646.36 72.62 76.98 0.05% 14 1548.36 32.51 6.93% 13 1632.29 39.36 0.00%13.77±0.43 1570.50±39.01 51.43 5.33%
RCdp106 12 1392.47 12 1392.47 20.47 25.38 0.00% 12 1392.47 78.62 83.34 0.00% 12 1392.47 41.38 0.00% 12 1392.47 37.52 0.00%12.00±0.00 1393.42±3.32 46.37 0.00%
RCdp107 11 1252.79 11 1255.06 21.22 26.31 0.01% 11 1252.79 78.67 83.39 0.00% 11 1252.79 105.82 0.00% 11 1252.79 101.39 0.00%11.00±0.00 1256.00±2.84 101.35 0.01%
RCdp108 10 1194.40 10 1198.36 19.52 24.20 0.02% 10 1208.58 79.18 83.93 0.07% 11 1151.71 101.75 9.24% 10 1177.98 104.70 -0.08%10.00±0.00 1186.63±12.00 99.92 -0.04%
RCdp201 4 1406.94 4 1406.94 26.68 33.08 0.00% 4 1406.94 78.89 83.62 0.00% 4 1406.94 78.58 0.00% 4 1406.94 20.52 0.00% 4.00±0.00 1409.57±9.13 29.81 0.03%
RCdp202 3 1412.52 3 1414.55 39.94 49.53 0.03% 4 1161.29 79.10 83.85 23.59% 4 1161.29 73.45 23.59% 3 1365.65 107.51 -0.63% 3.00±0.00 1372.33±14.13 66.64 -0.54%
RCdp203 3 1050.64 3 1050.64 83.95 104.10 0.00% 3 1056.96 77.17 81.80 0.09% 3 1060.45 1152.32 0.14% 3 1049.62 60.83 -0.01% 3.00±0.00 1058.63±5.98 70.81 0.11%
RCdp204 3 798.46 3 798.46 94.85 117.61 0.00% 3 798.46 71.01 75.27 0.00% 3 798.46 334.63 0.00% 3 798.46 103.14 0.00% 3.00±0.00 799.33±2.76 85.43 0.01%
RCdp205 4 1297.65 4 1297.65 33.58 41.64 0.00% 4 1297.65 77.82 82.49 0.00% 4 1297.65 168.17 0.00% 4 1297.65 100.06 0.00% 4.00±0.00 1298.80±2.14 73.07 0.01%
RCdp206 3 1146.32 3 1146.32 59.36 73.61 0.00% 3 1146.32 78.54 83.25 0.00% 3 1146.32 392.20 0.00% 3 1146.32 37.67 0.00% 3.00±0.00 1152.35±5.86 46.23 0.08%
RCdp207 3 1061.14 3 1061.84 79.06 98.03 0.01% 3 1061.14 84.47 89.54 0.00% 3 1061.14 461.54 0.00% 3 1061.14 104.54 0.00% 3.00±0.00 1061.95±2.28 87.89 0.01%
RCdp208 3 828.14 3 828.14 72.78 90.25 0.00% 3 828.44 72.98 77.36 0.00% 3 829.00 506.23 0.01% 3 828.14 162.63 0.00% 3.00±0.00 830.39±3.36 112.32 0.03%
Mean 7.29 1044.45 7.30 1045.68 42.12 52.23 0.14% 7.36 1037.92 86.03 91.19 2.00% 7.45 1033.92 201.15 2.79% 7.27 1043.34 70.77 -0.19% 7.36±0.06 1041.11±7.13 66.63 0.65%

TABLE VI
C OMPARATIVE RESULTS ON THE 20 LARGE - SCALE JD BENCHMARK INSTANCES BETWEEN MATE, MATE* AND MA-FIRD.
MATE MATE* MA-FIRD
Instance
fBest fAvg ± std M D fBest GapBest M ± std D± std fAvg ± std GapAvg M D fBest GapBest M ± std D± std fAvg ± std GapAvg
F201 65106.00 66097.00±292.00 41 52364.90 64664.90 -0.68% 41.63±0.49 52647.61±213.67 65137.61±213.67 -1.45% 40 52083.40 64083.40 -1.57% 40.40±0.50 52420.04±374.59 64540.04±374.59 -2.36%
F202 65012.00 66038.00±422.00 43 51923.00 64823.00 -0.29% 43.00±0.00 52141.76±149.35 65041.76±149.35 -1.51% 42 51778.10 64378.10 -0.98% 42.20±0.48 52309.00±491.89 64969.00±491.89 -1.62%
F203 65980.00 67090.00±332.00 41 53293.30 65593.30 -0.59% 41.37±0.49 53636.95±196.48 66046.95±196.48 -1.55% 41 53247.40 65547.40 -0.66% 40.93±0.25 53762.45±375.53 66042.45±375.53 -1.56%
F204 64747.00 65851.00±326.00 42 51982.70 64582.70 -0.25% 42.13±0.35 52460.08±180.22 65100.08±180.22 -1.14% 41 51737.30 64037.30 -1.10% 41.30±0.47 52678.48±539.36 65068.48±539.36 -1.19%
F401 122319.00 123261.00±446.00 78 98157.70 121557.70 -0.62% 79.30±0.75 98636.59±256.89 122426.59±256.89 -0.68% 75 93376.00 115876.00 -5.27% 76.17±0.95 96172.30±1566.68 119022.30±1566.68 -3.44%
F402 126887.00 128091.00±410.00 83 101617.00 126517.00 -0.29% 82.47±0.90 102619.60±418.57 127359.60±418.57 -0.57% 79 98297.00 121997.00 -3.85% 79.57±0.77 100708.27±1260.18 124578.27±1260.18 -2.74%
F403 120130.00 122306.00±682.00 78 96729.80 120129.80 0.00% 79.30±0.60 97604.31±403.95 121394.31±403.95 -0.75% 74 93340.00 115540.00 -3.82% 75.80±0.85 95049.40±1398.61 117789.40±1398.61 -3.69%
F404 124517.00 125242.00±359.00 81 98897.50 123197.50 -1.06% 81.60±0.77 99879.39±452.51 124359.39±452.51 -0.70% 78 96051.00 119451.00 -4.07% 79.07±0.78 98081.70±1199.85 121801.70±1199.85 -2.75%
F601 182504.00 184119.00±608.00 116 147617.00 182417.00 -0.05% 115.92±0.89 148360.73±441.98 183137.65±441.98 -0.53% 109 141408.00 174108.00 -4.60% 110.97±1.10 144920.43±1783.13 178210.43±1783.13 -3.21%
F602 187236.00 188891.00±645.00 119 151059.00 186759.00 -0.25% 119.41±0.98 151888.00±591.75 187712.14±591.75 -0.62% 113 144417.00 178317.00 -4.76% 113.83±1.46 148246.53±2376.39 182396.53±2376.39 -3.44%
F603 186644.00 188050.00±621.00 118 149843.00 185243.00 -0.75% 118.63±1.01 151448.04±553.83 187036.93±553.83 -0.54% 110 144979.00 177979.00 -4.64% 113.63±1.25 148734.27±1812.42 182824.27±1812.42 -2.78%
F604 186289.00 188110.00±790.00 120 150449.00 186449.00 0.09% 119.67±1.12 151388.07±478.76 187288.07±478.76 -0.44% 114 144001.00 178201.00 -4.34% 115.00±1.55 147555.10±2227.07 182055.10±2227.07 -3.22%
F801 213661.00 214634.00±561.00 154 167407.00 213607.00 -0.03% 152.17±1.23 168743.57±610.66 214393.57±610.66 -0.11% 147 164225.00 208325.00 -2.50% 150.03±2.36 167414.03±2867.45 212424.03±2867.45 -1.03%
F802 212752.00 213276.00±292.00 150 166591.00 211591.00 -0.55% 151.00±1.05 167888.93±541.21 213188.93±541.21 -0.04% 145 162714.00 206214.00 -3.07% 148.57±1.76 165733.33±2238.98 210303.33±2238.98 -1.39%
F803 214126.00 214870.00±318.00 155 167790.00 214290.00 0.08% 153.60±1.40 168701.80±545.41 214781.80±545.41 -0.04% 150 165213.00 210213.00 -1.83% 152.93±2.23 168220.27±2641.75 214100.27±2641.75 -0.36%
F804 209431.00 210845.00±429.00 150 165302.00 210302.00 0.42% 148.93±1.46 165991.13±404.11 210671.13±404.11 -0.08% 145 160471.00 203971.00 -2.61% 147.57±2.31 164444.90±2467.06 208714.90±2467.06 -1.01%
F1001 312606.00 314914.00±1096.00 203 251707.00 312607.00 0.00% 203.41±1.74 253436.22±849.77 314458.44±849.77 -0.14% 195 244613.00 303113.00 -3.04% 197.27±2.32 248709.20±2547.42 307889.20±2547.42 -2.23%
F1002 309158.00 311718.00±1153.00 203 248731.00 309631.00 0.15% 202.41±1.50 250399.26±917.50 311121.48±917.50 -0.19% 193 241933.00 299833.00 -3.02% 195.73±2.83 245324.30±2909.40 304044.30±2909.40 -2.46%
F1003 311377.00 313989.00±981.00 204 249512.00 310712.00 -0.21% 203.87±1.58 251582.26±947.36 312743.13±947.36 -0.40% 195 242551.00 301051.00 -3.32% 196.90±2.72 246727.33±2729.98 305797.33±2729.98 -2.61%
F1004 308816.00 311415.00±943.00 204 248314.00 309514.00 0.23% 203.19±1.41 249761.69±840.35 310719.38±840.35 -0.22% 193 240655.00 298555.00 -3.32% 196.67±2.52 244625.53±2643.12 303625.53±2643.12 -2.50%
Mean 179464.90 180940.35±585.30119.15 143464.40 179209.40 -0.23% 119.15±0.99 144460.80±499.72 180205.95±499.72 -0.59%113.95 139354.51 141764.89 -3.12% 115.73±1.47 142091.84±1822.54 144677.53±1822.54 -2.28%
 10

(a) TTT value of Cdp101. (b) TTT value of Rdp203.

(a) Comparison of MA-FIRD with MA-FIRD-wo-norm and MA-FIRD-const-coef.

(c) TTT value of RCdp101. (d) TTT value of RCdp204.

Fig. 4: Examples of time-to-target analysis between MA-FIRD and MATE.

The x-axis represents the running time to reach the target value and the y-axis
indicates the cumulative probability of reaching the given target value.

V. P ERFORMANCE ANALYSIS
(b) Comparison of MA-FIRD with MA-F and MA-FI.
We now present experiments to study the influences of the
algorithmic components based on the medium WC instances Fig. 5: Comparison of MA-FIRD with its variants. The x-axis denotes the
instance names and the y-axis illustrates the relative gap between the solutions
and the large JD instances. obtained by variants compared to those of MA-FIRD. A positive gap indicates
that the variant performs worse than MA-FIRD, and vice versa.
A. Analysis on the feasible and infeasible route descent search
First, we investigate the impact of the penalty term man-
agement in the feasible and infeasible search, focusing on
normalization and dynamic coefficient adjustment. We com-
pare the MA-FIRD algorithm with two variants: MA-FIRD-
wo-norm, which removes the penalty term normalization, and
MA-FIRD-const-coef, which sets the penalty coefficients to a
constant value (λ1 = λ2 = 1.5). Figure 5a shows the relative
gap between the solutions obtained by these variants and
those of MA-FIRD. The results indicate that normalization of
penalty terms and dynamic coefficient adjustment significantly
enhance performance as the variants perform much worse than
the baseline MA-FIRD in most instances. (a) M = 15 and f (S) = 1634.85. (b) M = 14 and f (S) = 1708.21.
Next, we assess the role of the feasible and infeasible route Fig. 6: Two solutions on the instance RCdp101 with the distance of 32.17.
descent search by comparing it with the variant MA-F, which
eliminates the route descent strategy and accepts only feasible
solutions, and the variant MA-FI, which also removes the route
descent strategy but allows infeasible solutions. Figure 5b
compares MA-F and MA-FI with the baseline MA-FIRD. MA-
FI shows slightly better performance than MA-F, suggesting
that accepting infeasible solutions can help escape feasible
local optima and improve search capability. However, both
variants perform significantly worse than MA-FIRD in almost
all instances, highlighting the importance of the route descent
strategy.
To better understand the benefits of the route descent
strategy, Figures 6 and 7 illustrate high-quality solutions for
adjacent numbers of vehicles in two representative instances. (a) M = 3 and f (S) = 767.82. (b) M = 2 and f (S) = 885.71.
Different colors represent distinct routes. One notices that Fig. 7: Two solutions on the instance Rdp211 with the distance of 72.82.
high-quality solutions for adjacent numbers of vehicles are
 11

not necessarily close and may show significant differences.

For example, the distance between two solutions for RCdp101
is 32.17, and for Rdp211, it is 72.82. This indicates signif-
icant dissimilarity between solutions with adjacent numbers
of vehicles. This observation highlights that the distribution
of solutions in the search space is not uniform, and high-
quality solutions for different numbers of vehicles may not be
proximate. The route descent strategy enables rapid traversal
of the search space, helping to locate high-quality solutions
for varying numbers of vehicles.

B. Analysis on the adaptive route-inheritance crossover (a) Comparison of the MA-FIRD and its variants.

To study the impact of the ARIX crossover, we create the

following variants of the MA-FIRD algorithm:
• MA-FIRD-wo-X: ARIX is removed, and only one solu-
tion is maintained in the population. At each generation,
the correlated destroy and repair operator from [6] is used
to perturb the solution, which is then improved by the
FIRD search to generate a new solution.
• MA-FIRD-RARI: ARIX is replaced by the RARI
crossover from [6].
• MA-FIRD-RDM-P: The number of parents is randomly
chosen from {2, ..., |P|}. (b) Results with number of parents Np = 2, 4, 6, 8, 10.
• MA-FIRD-RDM-I: The insertion operator is randomly
selected from the set {FBI, IBI, FRI, IRI, RI} at each
generation.
Figure 8a provides a visual comparison of the variants
against MA-FIRD. The instances are grouped into WC-C,
WC-R, WC-RC, and JD. Note that the WC-C instances are
less sensitive to configuration variations, while instances in
the other groups are more sensitive. In general, the baseline
MA-FIRD algorithm outperforms MA-FIRD-wo-X and MA-
FIRD-RARI in most cases, which confirms the importance
of ARIX in the algorithm. Furthermore, the results of MA-
(c) Results with insertion operations FBI, IBI, FRI, IRI, and RI.
FIRD-RDM-P and MA-FIRD-RDM-I lag behind MA-FIRD,
suggesting that the learning-based adaptive mechanism helps Fig. 8: Comparison of the MA-FIRD algorithm and its variants. The x-
axis denotes the instance groups and the y-axis illustrates the relative gap
ARIX to select the appropriate number of parents and insertion between the solutions obtained by variants compared to those of MA-FIRD.
operators for different instances. The mean relative gap for each group is indicated by a solid line, while the
To gain deeper insights into the impact of different ARIX 0.95 confidence interval is represented by the corresponding shadow.
configurations, we conducted comprehensive experiments ana-
lyzing the influence of the number of parents and the insertion
operations. Figures 8b and 8c present the visualized results. VI. C ONCLUSION
Figure 8b shows the results for different numbers of parents We introduce an effective memetic algorithm to solve
(Np = 2, 4, 6, 8, 10), while Figure 8c presents the results for the challenging Vehicle Routing Problem with Simultane-
different insertion operators (FBI, IBI, FRI, IRI, and RI). ous Pickup and Delivery and Time Windows. Based on the
Overall, instances in the WC-C group still demonstrate memetic algorithm framework, our algorithm integrates sev-
less sensitivity to variations in setups. However, performance eral innovative strategies to effectively enhance the exploration
across the other three groups varies with different configura- of the solution space. These include a lightweight feasible
tions. Notably, no single configuration consistently performs and infeasible search mechanism featuring the route descent
well across all instances. For example, Np = 6, 8, 10 perform strategy, penalty term management, and infeasibility-allowed
well on JD but poorly on WC-R, while certain insertion constant-time move evaluation. We also introduce a learning-
operators (IBI, IRI, and RI) perform well on WC-R but poorly based adaptive route-inheritance crossover to enhance gen-
on JD. This suggests the necessity of an adaptive mechanism eralization and robustness, and a fitness-distance population
to dynamically determine configurations and values. Addition- management strategy based on max-min normalization to
ally, these configurations generally do not perform better than maintain population diversity.
the baseline MA-FIRD, indicating the efficacy of the learning- Extensive experiments have been conducted to evaluate the
based adaptive mechanism. algorithm’s performance. The results show that the proposed
 12

algorithm consistently outperforms the reference algorithms pickup–delivery and time windows,” Computers & Industrial Engineer-
with high computational efficiency. Remarkably, The algo- ing, vol. 83, pp. 111–122, 2015.
[17] Y. Shi, Y. Zhou, T. Boudouh, and O. Grunder, “A lexicographic-based
rithm attains the best solutions for all benchmark instances, two-stage algorithm for vehicle routing problem with simultaneous
reaching 57 best-known solutions (BKS) and achieving 8 new pickup–delivery and time window,” Engineering Applications of Arti-
best upper bounds for the 65 popular benchmark instances. ficial Intelligence, vol. 95, p. 103901, 2020.
[18] K. Tang, S. Liu, P. Yang, and X. Yao, “Few-shots parallel algorithm port-
Furthermore, the algorithm achieves new best upper bounds folio construction via co-evolution,” IEEE Transactions on Evolutionary
for all 20 large real-world instances, demonstrating its effec- Computation, vol. 25, no. 3, pp. 595–607, 2021.
tiveness and practicality in real-world scenarios. [19] R. Praxedes, T. Bulhões, A. Subramanian, and E. Uchoa, “A unified
exact approach for a broad class of vehicle routing problems with
The flexibility and versatility of innovative strategies in simultaneous pickup and delivery,” Computers & Operations Research,
the proposed algorithm lay the basis for future extensions to vol. 162, p. 106467, 2024.
tackle other multi-attribute Vehicle Routing Problems. Future [20] Ç. Koç, G. Laporte, and İ. Tükenmez, “A review of vehicle routing with
simultaneous pickup and delivery,” Computers & Operations Research,
work aims to explore the adaptability and performance of the vol. 122, p. 104987, 2020.
algorithm in a broader range of complex routing scenarios, [21] M. M. Solomon, “Algorithms for the vehicle routing and scheduling
thus contributing to the advancement of the field. problems with time window constraints,” Operations Research, vol. 35,
no. 2, pp. 254–265, 1987.
[22] P. Moscato and C. Cotta, “A modern introduction to memetic algo-
rithms,” Handbook of Metaheuristics, pp. 141–183, 2010.
R EFERENCES [23] F. Neri and C. Cotta, “Memetic algorithms and memetic computing op-
[1] E. Angelelli and R. Mansini, “The vehicle routing problem with timization: A literature review,” Swarm and Evolutionary Computation,
time windows and simultaneous pick-up and delivery,” in Quantitative vol. 2, pp. 1–14, 2012.
Approaches to Distribution Logistics and Supply Chain Management. [24] F. Glover and J.-K. Hao, “The case for strategic oscillation,” Annals of
Springer, 2002, pp. 249–267. Operations Research, vol. 183(1), pp. 163–173, 2011.
[2] K. Braekers, K. Ramaekers, and I. Van Nieuwenhuyse, “The vehicle [25] M. Li, J.-K. Hao, and Q. Wu, “Learning-driven feasible and infeasible
routing problem: State of the art classification and review,” Computers tabu search for airport gate assignment,” European Journal of Opera-
& Industrial Engineering, vol. 99, pp. 300–313, 2016. tional Research, vol. 302, no. 1, pp. 172–186, 2022.
[3] H. Min, “The multiple vehicle routing problem with simultaneous [26] W. Sun, J.-K. Hao, X. Lai, and Q. Wu, “Adaptive feasible and infeasible
delivery and pick-up points,” Transportation Research Part A: General, tabu search for weighted vertex coloring,” Information Sciences, vol.
vol. 23, no. 5, pp. 377–386, 1989. 466, pp. 203–219, 2018.
[4] J. Dethloff, “Vehicle routing and reverse logistics: The vehicle routing [27] Y. Zou, J.-K. Hao, and Q. Wu, “A two-individual evolutionary algorithm
problem with simultaneous delivery and pick-up: Fahrzeugeinsatzpla- for cumulative capacitated vehicle routing with single and multiple
nung und redistribution: Tourenplanung mit simultaner auslieferung und depots,” IEEE Transactions on Evolutionary Computation, 2024.
rückholung,” OR-Spektrum, vol. 23, pp. 79–96, 2001. [28] Y. Nagata, O. Bräysy, and W. Dullaert, “A penalty-based edge assembly
[5] J. Hof and M. Schneider, “An adaptive large neighborhood search with memetic algorithm for the vehicle routing problem with time windows,”
path relinking for a class of vehicle-routing problems with simultaneous Computers & Operations Research, vol. 37, no. 4, pp. 724–737, 2010.
pickup and delivery,” Networks, vol. 74, no. 3, pp. 207–250, 2019. [29] J. Vermorel and M. Mohri, “Multi-armed bandit algorithms and em-
[6] S. Liu, K. Tang, and X. Yao, “Memetic search for vehicle routing pirical evaluation,” in European Conference on Machine Learning.
with simultaneous pickup-delivery and time windows,” Swarm and Springer, 2005, pp. 437–448.
Evolutionary Computation, vol. 66, p. 100927, 2021. [30] V. Kuleshov and D. Precup, “Algorithms for multi-armed bandit prob-
[7] H. Wu and Y. Gao, “An ant colony optimization based on local search lems,” arXiv preprint arXiv:1402.6028, 2014.
for the vehicle routing problem with simultaneous pickup–delivery and [31] K. Sörensen and M. Sevaux, “MA|PM: memetic algorithms with popu-
time window,” Applied Soft Computing, vol. 139, p. 110203, 2023. lation management,” Computers & Operations Research, vol. 33, no. 5,
[8] T. Vidal, T. G. Crainic, M. Gendreau, and C. Prins, “A hybrid genetic pp. 1214–1225, 2006.
algorithm with adaptive diversity management for a large class of [32] D. C. Porumbel, J.-K. Hao, and P. Kuntz, “An evolutionary approach
vehicle routing problems with time-windows,” Computers & Operations with diversity guarantee and well-informed grouping recombination for
Research, vol. 40, no. 1, pp. 475–489, 2013. graph coloring,” Computers & Operations Research, vol. 37, no. 10, pp.
[9] P. He and J.-K. Hao, “General edge assembly crossover-driven memetic 1822–1832, 2010.
search for split delivery vehicle routing,” Transportation Science, vol. 57, [33] F. Wilcoxon, “Individual comparisons by ranking methods,” in Break-
no. 2, pp. 482–511, 2023. throughs in statistics: Methodology and distribution. Springer, 1992,
[10] D. Cattaruzza, N. Absi, D. Feillet, and T. Vidal, “A memetic algorithm pp. 196–202.
for the multi trip vehicle routing problem,” European Journal of Oper- [34] R. M. Aiex, M. G. Resende, and C. C. Ribeiro, “Ttt plots: a perl program
ational Research, vol. 236, no. 3, pp. 833–848, 2014. to create time-to-target plots,” Optimization Letters, vol. 1, pp. 355–366,
[11] H.-F. Wang and Y.-Y. Chen, “A genetic algorithm for the simultaneous 2007.
delivery and pickup problems with time window,” Computers & Indus- [35] C. C. Ribeiro, I. Rosseti, and R. Vallejos, “Exploiting run time dis-
trial Engineering, vol. 62, no. 1, pp. 84–95, 2012. tributions to compare sequential and parallel stochastic local search
[12] C.-H. Liang, H. Zhou, and J. Zhao, “Vehicle routing problem with algorithms,” Journal of Global Optimization, vol. 54, pp. 405–429, 2012.
time windows and simultaneous pickups and deliveries,” in 2009 16th
International Conference on Industrial Engineering and Engineering
Management. IEEE, 2009, pp. 685–689.
[13] M. Lai and E. Cao, “An improved differential evolution algorithm
for vehicle routing problem with simultaneous pickups and deliveries
and time windows,” Engineering Applications of Artificial Intelligence,
vol. 23, no. 2, pp. 188–195, 2010.
[14] S. Kassem and M. Chen, “Solving reverse logistics vehicle routing
problems with time windows,” The International Journal of Advanced
Manufacturing Technology, vol. 68, pp. 57–68, 2013.
[15] C. Wang, F. Zhao, D. Mu, and J. W. Sutherland, “Simulated annealing for
a vehicle routing problem with simultaneous pickup-delivery and time
windows,” in Advances in Production Management Systems. Sustainable
Production and Service Supply Chains: IFIP WG 5.7 International
Conference, APMS 2013, State College, PA, USA, September 9-12, 2013,
Proceedings, Part II. Springer, 2013, pp. 170–177.
[16] C. Wang, D. Mu, F. Zhao, and J. W. Sutherland, “A parallel simulated
annealing method for the vehicle routing problem with simultaneous
 1

Supplementary materials of paper “A Memetic

Algorithm for Vehicle Routing with Simultaneous
Pickup and Delivery and Time Windows”
Zhenyu Lei and Jin-Kao Hao (Corresponding author)

I. C ALCULATIONS OF CONCATENATION OPERATION route R0 can be computed using the values of D, Tv , and Lmax
We provide detailed calculations of the concatenation oper- by Equation (22) in constant time. It’s important to note that
ation, denoted as ⊕, for constant-time move evaluation of the multiple matrices must be maintained for each route to store
local search operators presented in Section III-C3. the values of these metrics. These matrices can be updated
Since the move of the local search operators can be promptly whenever a move alters the corresponding route.
represented as a rearrangement of subsequences of routes,
the evaluation of the move can be computed efficiently by ∆feval (R0 , R) = D(R0 ) − D(R) (22)
P
maintaining a set of metrics for each route, including travel ck 0
+ λ1 P (Tv (R ) − Tv (R))
distance (D), duration time (Td ), earliest and latest arrival tk
times (Te and Tl ), wait time (Tw ), violated time window value 2 · max ck − min ck
+ λ2
Q
(Tv ), load values before and after the sequence (Lin and Lout ),
(max(Lmax (R0 ) − Q, 0) − max(Lmax (R) − Q, 0))
and the maximum load (Lmax ).
For a subsequence containing a single node, denoted as
σ 0 = {vi }, these metrics can be defined:
II. S ENSITIVITY ANALYSIS OF THE PARAMETERS
D(σ 0 ) = 0 (1)
Td (σ 0 ) = si (2)
This section presents a sensitivity analysis of the parameters
of the algorithm to evaluate their impact on performance
Te (σ 0 ) = ei (3)
and calibrate them for best results. The analysis focuses on
Tl (σ 0 ) = li (4)
three key parameters: the coefficient of the fitness function
Tw (σ 0 ) = 0 (5)
ξ, the coefficient adjustment factor κ, and the initial reten-
Tv (σ 0 ) = 0 (6) tion ratio γr0 . The parameters are varied within the ranges:
Lin (σ 0 ) = di (7) ξ ∈ {0.1, 0.2, . . . , 0.9}, κ ∈ {0.1, 0.2, . . . , 0.9}, and γr0 ∈
0
Lout (σ ) = pi (8) {0.1, 0.2, . . . , 0.9}.
Lmax (σ 0 ) = max(di , pi ) (9) Using the sensitivity analysis tool SALib [1], we randomly
sampled 8000 parameter combinations from the parameter
For two subsequences of routes, σ1 = {vi , . . . , vj } and space and evaluated them on a training set comprising 10
σ2 = {vk , . . . , vl }, the corresponding metrics of the combined representative and challenging benchmark instances. The av-
sequence of routes σ1 ⊕ σ2 can be calculated by: erage gap value against the best-known solutions (BKS) on
the training set served as the performance metric, and Sobol’s
method [2] was used to evaluate parameter sensitivity. Detailed
D(σ1 ⊕ σ2 ) = D(σ1 ) + cj,k + D(σ2 ) (10)
1 results are presented in Table A.I and Figure 1.
∆t = Td (σ1 ) + Tw (σ ) + tj,k − Tv (σ1 ) (11)
The analysis reveals that the coefficient adjustment factor
∆w = max(Te (σ2 ) − ∆t − Tl (σ1 ), 0) (12)
κ significantly influences the algorithm’s performance, with a
∆v = max(Te (σ1 ) + ∆t − Tl (σ2 ), 0) (13)
high total-order (ST) value of 0.974937 and a first-order (S1)
Td (σ1 ⊕ σ2 ) = Td (σ1 ) + tj,k + Td (σ2 ) (14)
index of 0.976788, accompanied by low confidence intervals.
Te (σ1 ⊕ σ2 ) = max(Te (σ1 ), Te (σ2 ) − ∆t ) − ∆w (15)
Conversely, the coefficient of the fitness function ξ exhibits
Tl (σ1 ⊕ σ2 ) = min(Tl (σ1 ), Tl (σ2 ) − ∆t ) + ∆v (16)
relatively low sensitivity, suggesting a minor impact on per-
Tw (σ1 ⊕ σ2 ) = Tw (σ1 ) + ∆w + Tw (σ2 ) (17)
formance. Similarly, the initial retention ratio γr0 has the lowest
Tv (σ1 ⊕ σ2 ) = Tv (σ1 ) + ∆v + Tv (σ2 ) (18)
sensitivity index, implying minimal impact. The consistency
Lin (σ1 ⊕ σ2 ) = Lin (σ1 ) + Lin (σ2 ) (19)
between the total-order (ST) and first-order (S1) indices,
Lout (σ1 ⊕ σ2 ) = Lout (σ1 ) + Lout (σ2 ) (20)
along with negligible second-order indices (S2), indicates low
Lmax (σ1 ⊕ σ2 ) = max(Lmax (σ1 ) + Lin (σ2 ), Lout (σ1 ) + Lmax (σ2 ))
(21) interaction effects among the parameters.
Given the low interaction effects, we conducted a one-at-a-
The difference of D(·), V1 (·), and V2 (·) in evaluation time sensitivity analysis to determine suitable values for each
function values between the original route R and modified parameter. Each parameter was tested independently while
 2

TABLE A.I: Sensitivity analysis results

and the reference algorithms.
Parameter Sensitivity Index Confidence Interval
TABLE A.II: CPU information and time conversion ratio γ
Total-order indices (ST)
Algorithm Processor Base frequency CPU mark γ
ξ 0.024685 0.002019 MA-FIRD AMD EPYC 7282 2.80 GHz 1829 1.00
κ 0.974937 0.057852 ALNS-PR [3] Intel Core i5-6600 3.30 GHz 2260 1.24
MATE [4] Intel Xeon E5-2699A 2.40 GHz 1945 1.06
γr0 0.000060 0.000007
First-order indices (S1)
To account for different experimental environments in the
ξ 0.025848 0.014347 proposed algorithm and reference algorithms, we apply a time
κ 0.976788 0.071350
γr0 0.000146 0.000637 conversion ratio γ to standardize the computation time. This
ratio is defined as the ratio of CPU scores obtained from
Second-order indices (S2)
PassMark1 , a recognized CPU benchmarking platform. Table
(ξ, κ) -0.002324 0.028654
A.II presents the detailed information about the CPU used in
(ξ, γr0 ) -0.001171 0.020575
(κ, γr0 ) -0.002530 0.076821 our experiments and those used by the reference algorithms,
and the calculated time conversion ratio γ.

R EFERENCES
[1] J. Herman and W. Usher, “Salib: An open-source python
library for sensitivity analysis,” Journal of Open Source
Software, vol. 2, no. 9, p. 97, 2017.
[2] I. M. Sobol, “Global sensitivity indices for nonlinear
mathematical models and their monte carlo estimates,”
Mathematics and Computers in Simulation, vol. 55, no.
1-3, pp. 271–280, 2001.
[3] J. Hof and M. Schneider, “An adaptive large neighbor-
hood search with path relinking for a class of vehicle-
routing problems with simultaneous pickup and delivery,”
Networks, vol. 74, no. 3, pp. 207–250, 2019.
[4] S. Liu, K. Tang, and X. Yao, “Memetic search for vehicle
Fig. 1: Visualization of the sensitivity analysis results. routing with simultaneous pickup-delivery and time win-
dows,” Swarm and Evolutionary Computation, vol. 66, p.
100927, 2021.
keeping the others fixed at default values: ξ = 0.5, κ = 0.5,
and γr0 = 0.5. Each parameter setting was evaluated on the
training set over 30 independent runs. Figure 2 illustrates the
average gap values with different parameter settings. The best
parameter settings were found to be ξ = 0.8, κ = 0.5, and
γr0 = 0.9. These values are used in the experiments.

Fig. 2: Visualization of the sensitivity analysis results.

III. CPU INFORMATION AND TIME CONVERSION RATIO

This section provides detailed CPU information and the
calculated time conversion ratio γ for the proposed algorithm 1 https://www.passmark.com/