1. Introduction
In recent years, cities around the world have been transforming into more sustainable living spaces, one of the main drivers of which is the effects of climate change. According to a 2020 report by the OECD, urban areas account for 70% of global greenhouse gas emissions, with a third of major cities’ greenhouse gas emissions coming from transportation. This makes it crucial to provide more sustainable transportation options in urban areas.
Free-floating bike-sharing systems (FFBSSs) are an alternative in a sustainable transport system, which has become increasingly important globally in recent years. The most important reason is that FFBSSs offer the potential for cities to achieve the sustainable development goals. FFBSSs have numerous benefits, providing a low-carbon mobility model for short-distance trips and a supplementary means for other types of transport, e.g., buses or subways, reducing car usage and carbon emissions, and contributing to healthy living and physical activity benefits. Furthermore, FFBSSs have contributed to many sustainable development goals. Consequently, it is of great significance to design a low-cost and efficient bike rebalancing system and to provide high-quality service and attract more customers to join low-carbon travel mode.
Related research can be mainly distinguished into two categories in terms of the operation type: dynamic rebalancing and static rebalancing. The former considers the active scenarios where the rebalancing operation would react to the real-time system state. We refer the interested reader to papers by Nair and Miller-Hooks [
1], Caggiani and Ottomanelli [
2], Li et al. [
3], Zhang et al. [
4], Brinkmann et al. [
5], Caggiani et al. [
6], Legros [
7], Warrington and Ruchti [
8], You [
9] and Tian et al. [
10].
The static rebalancing assumes that the demand in the system is low or the demand is forecasted in advance. Some studies focus on single-vehicle scenarios, at which time the static bike rebalancing problem (SBRP) can be seen as a variant of the One-Commodity Pickup and Delivery TSP. Exact methods, branch-and-cut, for the single-vehicle SBRP were suggested by Chemla et al. [
11], Erdogan et al. [
12], Kadri et al. [
13], and Bruck et al. [
14]. For the heuristics, Ho and Szeto [
15] proposed an iterated tabu search; Szeto et al. [
16] used the enhanced chemical reaction optimization; Li et al. [
17] designed a combined hybrid genetic algorithm; and Cruz et al. [
18] developed an iterated local-search-based heuristic. Particularly, some works [
11,
14,
18,
19] permitted the split load.
More researchers focus on multi-vehicle SBRP, which is more realistic and is a variant of One-Commodity Pickup and Delivery VRP, an essential family in the VRP. The branch-and-cut framework was used to design the exact solution methods for multi-vehicle SBRP [
20,
21,
22]. Raviv et al. [
23] proposed a 9.5 approximation method. Several heuristics were used, such as the PILOT, GRASP, and VNS approaches [
24], the three-step approach [
25], a two-stage heuristic [
26], a destroy and repair metaheuristic [
27], a cluster-first route-second heuristic [
28], a hybrid large neighbourhood search [
29,
30,
31], and an artificial bee colony algorithm [
32,
33]. The objectives considered in the multi-vehicle SBRP vary. Most only consider a single objective, such as total travel cost, total travel time, or total CO
emission [
20,
22,
27,
29,
34], while the others consider the weighted sum objectives, e.g., total duration and total loading actions [
24], penalty cost and total travel time [
25,
30], deviation from the tolerance of total demand dissatisfaction and service time [
32], or unmet demand and the total cost or time [
33,
35].
This paper focuses on the Heterogeneous Fleet and Multi-Depot Static Bike Rebalancing Problem with Split Load (HFMDSBRP-SD), a variant of the heterogeneous fleet SBRP. The repositioning vehicles of different types are available at several depots. The limitation that each station can only be visited by one repositioning vehicle is relaxed, making the problem more challenging to solve than classical SBRP, because each station’s pickup or delivery quantity emerged as a decision variable. Therefore, the HFMDSBRP-SD is used to determine a set of feasible routes compatible with their pickup/delivery quantity at each visited station for a fleet of heterogeneous repositioning vehicles at several depots to rebalance all the stations in an FFBSS at the minimum total cost.
To address and solve this issue, this study develops a branch-and-price-and-cut (BPC) algorithm. BPC algorithms have been used to solve many variants of VRP, such as electric VRP with time windows [
36], the cable-routing problem in solar power plants [
37], and synchronized VRP with split delivery, proportional service time and multiple time windows [
38]. Those showed that a combination of column generation and cutting planes was much more effective than each of those techniques taken alone.
The main contributions of this paper are summarized as follows.
We introduce a more realistic SBRP that considers the split load, heterogeneous fleet and multiple depots.
We use Dantzig–Wolfe Decomposition to reformulate the arc-flow formulation of the HFMDSBRP-SD into a set partitioning master problem and several pricing subproblems. As mentioned before, a route’s pickup/delivery quantity at each visited station is the decision variable. We consider the decision variables in the pricing subproblem to avoid an exponential number of constraints in the master problem.
A Tabu Search Column Generator and a heuristic label-setting algorithm are proposed to accelerate the column generation procedure.
Three types of non-robustness valid inequalities are used to increase the quality of the lower bound and accelerate the entire running process.
The rest of this paper is structured as follows.
Section 2 provides the problem description and an arc-flow formulation for the HFMDSBRP-SD.
Section 3 uses Dantzig–Wolfe Decomposition to reformulate the arc-flow formulation as a set partitioning model and several pricing subproblems.
Section 4 proposes a BPC algorithm and mainly focuses on the heuristic column generators and valid inequalities.
Section 5 presents the computational results and
Section 6 draws the conclusion.
2. Problem Description and Arc-Flow Formulation
The HFMDSBRP-SD is defined on a directed graph , where is the vertex set and A is the arc set. is the set of stations and is the set of depots. can be divided into two subsets , where is the set of pickup stations with each station’s demand , and is the set of delivery stations with demand . A service time , where is the time taken to move a bike. In the arc set , each arc has travel cost and travel time , assumed to the triangle inequality.
There are types of repositioning vehicles. represents the set of repositioning vehicle types. For the type repositioning vehicle, the capacity is , the maximum route duration is , and the fix cost is . There are number of type h repositioning vehicles available at depot . Let denote the set of available type h repositioning vehicles at depot . Let .
The problem aims to minimize the total cost, including the fixed cost for each repositioning vehicle in operation and the total traveling cost. We assume that
The demand of each station can be satisfied by multiple vehicles;
Vehicles can only visit a station once, and preemption (the station is temporarily used as storage) is not permitted;
There is no capacity limit for the depot.
The decision variables are given as follows.
: a binary variable equal to one if arc is traversed by vehicle , and 0 otherwise;
: a variable indicating the pickup/delivery quantity of station i serviced by vehicle , where the pickup quantity , and the delivery quantity ;
: a binary variable equal to one if station is visited by vehicle , and zero otherwise;
: a binary variable equal to one if station is visited by vehicle , and zero otherwise;
: a non-negative variable indicating the load weight of vehicle after leaving depot or station .
The objective function (
1) aims to minimize the total cost, including the fixed cost per repositioning vehicle in operation and the total traveling cost. Constraints (
2) ensure that each station’s demand is satisfied. Constraints (
4) and (
5) ensure that variables
induce a route in
A, indicating that vehicle
departs from depot
, visits a subset of stations, and returns to the same depot. Constraints (
6) and (
7) make sure that each pickup/delivery occurs when the station is visited. Constraints (
8) and (
9) ensure the pickup/delivery quantity is within the allowable limitation. Constraints (
10) and (
11) ensure that the vehicle capacity limitations are respected. Constraints (
12) ensures that the total travel time of vehicle
cannot exceed the maximum route duration
. Finally, constraints (
13) and (
14) ensure the binary requirements of variables
.
4. Branch-and-Price-and-Cut
Branch-and-price-and-cut (BPC) is a generic framework used to solve the problem in the VRP family [
40]. Based on the branch-and-bound approach, the
restricted linear relaxation of the master problem (RLMP) is solved iteratively by column generation procedure in each node of the search tree, and the valid inequalities are added to strengthen the RLMP to accelerate the solution process. Moreover, the best-first strategy is used in the branch-and-bound search tree.
4.1. Column Generation
Column generation is an iterative procedure. In each iteration, the pricing subproblem obtains its objective function parameters from the dual prices of constraints in RLMP. The solution to the subproblem with negative reduced cost could be added to the RLMP as a column.
This paper proposes two heuristics column generators, tabu search (TS) and heuristic label-setting (HLS), and applies one exact label setting (ELS) proposed by Ding et al. [
41] to solve the pricing subproblem. These column generators are used in an ordered sequence from the fastest to slowest, namely, TS, HLS, and ELS. When the previous one fails to find a negative-reduced-cost column, the following column generator is applied until the ELS cannot find a negative-reduced-cost column, i.e., the pricing subproblem is solved optimally.
4.1.1. Exact Method for Optimal Pickup and Delivery Pattern
In Ding et al. [
41], an exact method was proposed for optimal pickup and delivery pattern, viz., the optimal solution of the LP-MBKP (
35)–(
38). Several definitions and properties were given and proven, and are generalized as follows.
Subroute. A route could be partitioned into several vertex-disjoint subroutes. For each subroute, the vehicle load before reaching the first vertex of the subroute and the vehicle load when leaving the last vertex of the subroute are either 0 or Q.
Split pickup/delivery. There is, at most, one vertex in the route, for which a fraction of its pickup/delivery demand is satisfied. Such a pickup/delivery is called a split pickup/delivery. The pickup/delivery quantities of the other vertices in the subroute are either their lower or upper bounds.
End state. Define the end state of a subroute as full when the vehicle is a full load at the last vertex of the subroute. Define the end state of a subroute as empty when the vehicle load is empty at the last vertex of the subroute.
Subroute generation. During the process that each station’s pickup/delivery quantity is considered case by case, once a violated constraint emerges, a subroute arises. If the end state of the subroute is different from the previous one, the pattern adjustment only happened in the current subroute. If they are the same, the two adjacent subroutes will combine into one subroute with the same end state.
The explanation of the exact method requires the following additional notations.
: the kth subroute in route r, , is the length of subroute .
: the feasible pickup and delivery pattern compatible with the kth subroute in route r, .
: the vehicle load when the vehicle leaves the last station of subroute , . Define .
s: the route segment, i.e., the last few stations in the partial route that have not been attached to a subroute. , where l is the length of the route segment and l could be 0 when all the stations in the partial route have been attached to a subroute.
: the feasible pickup and delivery pattern compatible with the route segment s; the pickup/delivery quantities of the stations in s are either their upper or lower bounds.
: the vehicle load when the vehicle leaves the last station of the route segment, i.e., the last station of the partial route, .
k: the current number of subroutes, .
As shown in
Figure 1, given a route
, the exact method
Optimal Pattern considers each station’s pickup/delivery quantity case by case from
to
. Each time the paritial route
r with pattern
is extended to a new station
, the sub-procedure
Extend Paritial Route is executed to obtain the optimal pickup and delivery pattern of the new partial route, achieving
.
Figure 2 is the flow diagram of the sub-procedure
Extend Paritial Route . In each execution, firstly input the partial route and its compatible pickup and delivery pattern with the subroute partition information
,
as well as the newly reached station
. Then, initialize the pickup/delivery quantity of the newly reached station
to its upper or lower bound (
) according to its dual price
. Add
to the end of the partial route, namely, the end of the route segment, and check whether the vehicle capacity constraint (37) at station
is satisfied. Once violated, we will adjust the pickup and delivery pattern of the route segment; otherwise, terminate the sub-procedure.
In the case that the violation is an excess of
Q, a new subroute with
full end state would arise after the execution of the sub-procedure
Pattern Adjustment with Full End State shown in
Figure 3. We should first check whether the
end state of the route segment
s is the same as the last subroute
. If so, the last subroute
would combine with the route segment
s, becoming a new route segment
, and the number of subroutes would drop by one,
. Then, the new route segment
s and the last subroute
with their pickup and delivery patterns are inputted into the the sub-procedure
Pattern Adjustment with Full End State . Otherwise, the last subroute
and the route segment
s with their patterns will be inputted into the sub-procedure directly.
In the case that the violation is a vehicle load less than zero, a new subroute with
empty end state would arise after the execution of the sub-procedure
Pattern Adjustment with Empty End State shown in
Figure 4. Similar to the previous case, we check the
end states first to confirm whether to combine the last subroute and the route segment and then input the last subroute and the route segment into the sub-procedure
Pattern Adjustment with Empty End State .
In the sub-procedure
Pattern Adjustment with Full End State in
Figure 3, calculate the excess quantity
W. Firstly, select and sort the vertices in the route segment, and then decrease the pickup/delivery quantity of the vertex to its lower bound case by case according to this sequence until
Each time a vertex’s quantity is reduced, check whether the vehicle capacity constraints from this vertex to the end of the segment 0 ≤
ωrk +
qvi − Δ ≤
Q,
μ = {
g,…,
l} are satisfied. If not, the route segment is partitioned into two parts, the previous one has become a subroute, and the latter, still with excess quantity, should be handled by the
Pattern Adjustment with Full End State (
rk,
prk,
s,
ps) again. The sub-procedure
Pattern Adjustment with Full End State (
rk,
prk,
s,
ps) in
Figure 4 is the same way.
4.1.2. Exact and Heuristic Label-Setting
Ding et al. [
41] developed a Label-Setting algorithm embedded with the exact method for optimal pickup and delivery pattern, which is capable of exactly solving the pricing subproblem (
23)–(
34).
The label is defined as , corresponding to the elementary (partial) route ending up with vertex , and the optimal pickup and delivery pattern , obtained by the exact method described above. Moreover,
is the total time consumed along (partial) route r, including the travel time and the service time, ,
is the reduced cost of (partial) route r with the compatible pickup and delivery pattern , including the travel cost and the dual variable cost with respect to , , and
is the vehicle load at vertex v.
Hence, the initial label at vertex 0 is defined as = . A label can be extended to vertex along arc , generating a new label . The extension functions are
Extend Partial Route in
Figure 2, where
,
,
,
,
and
,
,
,
.
The proposed dominance rule is shown as follows. If E dominates , then the following conditions hold:
,
,
,
,
.
The objective function of the pricing subproblem (
23) shows that the solution of the pricing subproblem is seriously affected by the term
. Therefore, to speed up the exact label-setting algorithm, we use a heuristic strategies that eliminate some stations with the minimal
. Each time the heuristic label-setting (HLS) is used, all the stations are sorted in descending order according to the value of
. Then, a percentage of stations at the top of the list will remain, and the rest are rejected from the graph
. The rejection procedure is executed at the beginning of the HLS. The HLS can be executed multiple times with different parameters of the remaining percentage from small to large. The suitable times and parameters of the remaining percentage will be explored in
Section 5.1.
4.1.3. Tabu Search Column Generator
It is unnecessary to solve the pricing subproblem optimally in each column generation iteration. It is more efficient to find a negative reduced cost column heuristically. Thus, a tabu search (TS) column generator is proposed in this section to accelerate the column generation process. The TS method chooses the stations and determines their order, while the exact method in
Figure 1 determines the compatible pickup and delivery pattern.
Consider the following notations.
S: the set of routes with their compatible pickup and delivery patterns with negative reduced cost found by the TS method.
: the maximum number of the routes in S.
: the set of routes with their compatible pickup and delivery patterns associated with the basic columns of the RLMP in current column generation iteration.
: the maximum number of TS iterations for each basic column in .
: the maximum size of the tabu list.
TS begins with the set of basic columns ; namely, the columns with zero reduced cost. Then, it is independently executed on each column in . In each execution, two types of neighborhoods of the column are explored by applying REMOVE and ADD times.
REMOVE: select a station
, remove it from
r, and calculate the compatible optimal pickup and delivery pattern by the exact method
in
Figure 2.
ADD: select a station , add it to r, and calculate the compatible optimal pickup and delivery pattern by the exact method , finding the cheapest ADD.
The procedure terminates until the size of the set S reaches . Algorithm 1 shows the detailed procedure of the Tabu Search Column Generator.
4.2. Cutting
This section extends three types of non-robust valid inequalities for the HFMDSBRP-SD, namely, the subset-row inequality, the strong minimum number of vehicles inequality, and the elementary inequality. These inequalities are all defined on the variables in the MP. Thus, they are non-robust valid inequalities that will change the structure of the pricing subproblems and increase the problem complexity. However, the non-robust valid inequalities present a better performance of tightening the lower bound of the MP.
4.2.1. Subset-Row Inequalities
The subset-row (SR) inequalities were proposed by Jepsen et al. [
42], corresponding to a subset of the rank-1 Chvátal-Gomory inequalities. Given a subset of stations
, the SR inequalities can be expressed as
where the binary parameter
if the pickup/delivery demand of station
i is fully satisfied in pattern
p for route
r, namely,
, is the subset of the station set, and
.
Varying the cardinality of C and the value of yields four types of SR inequalities, as follows.
3-SR: the SR inequalities with .
4-SR: the SR inequalities with .
5,1-SR: the SR inequalities with .
5,2-SR: the SR inequalities with .
Preliminary test results showed that the SR inequalities for subset
C with
are rarely violated in the HFMDSBRP-SD. Therefore, only 3-SR inequalities are considered in this paper. The 3-SR inequalities are separated by complete enumeration.
Algorithm 1: Tabu Search Column Generator. |
|
Define
as the dual variables of the 3-SR inequalities. The reduced cost of variable
becomes
In the three column generators used in this paper, the pickup and delivery pattern will change with the extension of the partial route. Therefore, the term will be calculated when the route extends to the depot.
4.2.2. Strong Minimum Number of Vehicles Inequalities
Let
, and let
be the set of stations, such that the station
and
. The strong minimum number of vehicles (SMV) inequalities can be expressed as
where the binary parameter
if the pickup/delivery demand of station
i is fully met in pattern
p for route
r, and the binary parameter
if a split pickup/delivery or a zero pickup/delivery is performed at station
i in pattern
p for route
r.
Define
as the dual variables of the SMV inequalities. The reduced cost of variable
becomes
The same is true of the SR inequalities; the dual prices of SMV inequalities is calculated when the route extends to the depot.
4.2.3. Elementary Inequalities
The elementary inequalities were first introduced by Balas [
43], inspired by the fact that if route
r is used, for a node
i not visited by route
r, the route visiting
i cannot visit any node visited by route
r; otherwise, at least one node visited by route
r will be visited twice.
Pecin et al. [
44] developed enhanced elementary inequalities proven to dominate the elementary inequalities proposed by Balas [
43], which is also a subset of the rank 1 Chvátal–Gomory inequalities. Given a station subset
, and a station
, considering two multipliers
and
, the enhanced elementary inequalities are expressed as
A local search separation algorithm is used to find the violated elementary inequalities. The local search separation algorithm requires the following notations.
R: the set of all feasible routes, .
: the set of all routes with their compatible pickup and delivery patterns, .
: given a station , is the subset of such that a full pickup/delivery is performed at station i in pattern p for route r for each .
.
: the set of stations visited by r with pattern p and receiving a full pickup/delivery.
.
The local search is executed for each pair of station
and
, as shown in Algorithm 2.
Algorithm 2: Local Search Separation Algorithm for Elementary Inequalities. |
|
Define
as the dual variables of the elementary inequalities. The reduced cost of variable
becomes
The same is true for the SR inequalities; the dual prices of elementary inequalities are calculated when the route extends to the depot.
4.2.4. Cutting Strategies
In our proposed BPC algorithm, the three types of valid inequalities are separated and added to the RLMP only in the branch-and-bound nodes with depths less than three. The preliminary test results show that this is better than the strategies with depths less than 5 and 10. There are two types of cutting strategies.
All valid inequalities are separated and added to the RLMP in each iteration.
The valid inequalities are ordered based on the separating time. Then, only one type of valid inequality is sought and added to the RLMP in an iteration. It turns to the following type of valid inequalities until this type of valid inequality cannot be found anymore. It terminates until the last type of valid inequality cannot be found.
The two cutting strategies will be tested and compared in
Section 5.2.
4.3. Branching Strategies
To retain the structures of three column generators, branching on the decision variables is not feasible. Hence, we use four types of branching strategies, namely, branching on (1) the number of each type of vehicle used, (2) the number of times each station is visited, (3) the number of times each arc in A is visited, and (4) the number of times two consecutive arcs are visited. Note that the last branching strategy will change the structure of the column generator. However, it is rarely used in the solution process and is obligatory to guarantee integrality.
5. Computational Experiments
The instances used to test the proposed BPC algorithm are generated based on the instance generation scheme in Hernandez-Perez and Salazar-Gonzalez [
45] and Chemla et al. [
11]. The stations are randomly located in the square
. If one only considers one depot, it is located at
. If one considers two depots, they locate at
and
. The demand of each customer is randomly selected from
. We generate three instance classes with
, named as
, respectively. Each class has 10 instances. Define four vehicle capacities
. Three cases are considered individually for
instance classes. Case 1: one depot and one repositioning vehicle with
, Case 2: one depot and two vehicles with the capacity combination
, and Case 3: two depots and one vehicle with
. Thus, the total number of instances is 360.
All tests are coded in C++ and performed on a Windows PC equipped with 224 GB of RAM and an Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz. All the MPs are solved by the CPLEX solver. A time limit of two hours is set for each test. The computational experiments mainly focus on comparing the column generator results, the valid inequalities results, and the integer solution results.
5.1. Column Generator Results
Firstly, we conduct the experiments for the HLS parameter selection. The HLS parameter is the percentage of the stations that remain in the label-setting algorithm. The computational experiments are conducted on the RLMP at the root node of the branch-and-bound tree. Twenty instances are selected. The first 10 instances belong to the
instance class with one depot and one vehicle with
, expressed as
in
Table 1. The last 10 instances belong to the
instance class with one depot and one vehicle with
, expressed as
in
Table 1. Then, we consider four column-generator-application strategies, namely one-phrase HLS + ELS, two-phrase HLS + ELS, TS + one-phrase HLS + ELS, and TS + two-phrase HLS + ELS. For the one-phrase HLS, five parameters are considered, i.e., 0.70, 0.75, 0.80, 0.85, and 0.9. For the two-phrase HLS, nine parameter combinations are considered, i.e., 0.60 + 0.80, 0.65 + 0.80, 0.70 + 0.80, 0.60 + 0.85, 0.65 + 0.85, 0.70 + 0.85, 0.60 + 0.90, 0.65 + 0.90, and 0.70 + 0.90.
Table 1 reports the average time taken to solve the RLMP at the root branch-and-bound node on
instance classes in four column-generator-application strategies. The results indicate that the column-generator-application strategy with TS is much better than that without TS, and the efficiency and robustness of the average time for the TS + two-phrase HLS + ELS strategy are the best among the four strategies. Therefore, the TS + 2-phrase HLS + ELS is applied in the following computational experiments, and the parameter combination 0.65 + 0.80 is selected.
Table 2 displays the number of column generation iterations (nIt) the objective value (Obj) and the total computational time in seconds to solve the RLMP in the root branch-and-bound node (Time) on the
instance classes with one depot and one vehicle with
. As shown, in the
instance class, HLS + ELS provides the minimal average computational time because label setting can solve the small-scale problem more efficiently than TS. However, the results on
prove the effectiveness of TS. Although the HLS and TS increase the number of column generation iterations, the total computational time is reduced.
5.2. Valid Inequality Results
To evaluate the performance of various valid inequalities, we conduct the computational experiments on
instance classes with one depot and one vehicle with
to get (1) The lower bounds obtained at root branch-and-bound node without any valid inequality
; (2) The lower bounds at root node achieved by adding 3-SR inequalities
, SMV inequalities
, and elementary inequalities
individually; (3) The lower bounds at root node achieved by adding two or three types of valid inequalities (SVM + E, SVM + 3-SR, 3-SR + EC, and SMV + 3-SR + E) in cutting strategy 1 and 2, as described in
Section 4.2.4; and (4) the upper bound
.
Table 3 reports the average gap improved by various inequalities; that is, (
)/(
), where
x refers to any valid inequalities or combinations.
As shown, the SMV inequalities are the most effective valid inequalities among the three types. Moreover, for the most valid inequality combinations, the average gaps in cutting strategy 1 are better than those in cutting strategy 2. The average gap improved by SMV + 3-SR + E in Strategy 1 is the best in both instance classes. Therefore, all three types of valid inequalities and cutting strategy 1 are used in the following computational experiments.
5.3. Integer Solution Results
Finally, the computational experiments are conducted on all instances.
Table 4,
Table 5 and
Table 6 report the summary of the results, including the number of depots (NB dpt), the number of repositioning vehicles (NB veh), the capacity of each type of repositioning vehicle (Q), the number of instances in the instance class (NB inst), the number of solved instances (NB solved), and the following averages over the solved instances: the lower bound achieved by column generation at root branch-and-bound node (LB), the lower bound obtained by adding valid inequalities at root branch-and-bound node (LB_cut), the integer programming solution (IP), the time in seconds for LB (LB time), the time in seconds for solving the root branch-and-bound node (Root time), the time in seconds for IP (IP time), the time in seconds for separating valid inequalities (cut time), the number of nodes in the branch-and-bound tree (Tree size), the total number of SMV inequalites (nSMV), the total number of 3-SR inequalities (n3SR), the total number of elementary inequalities (nE), and the number of split pickup/delivery (nSplits). The detailed computational results are reported in the
Table A1,
Table A2,
Table A3,
Table A4,
Table A5,
Table A6,
Table A7 and
Table A8 in the
Appendix A.
As shown in
Table 4 and the detailed results in
Table A1,
Table A2 and
Table A3, among the 120 instances, the BPC algorithm obtains the optimal integer solutions for 118 instances within 2 s, and the integer feasible solutions for
and
with two depots, one vehicle, and
are not obtained due to insufficient memory. Among the solved 118 instances, there are 59 instances whose cut time value is zero, and the Tree size value is one, which indicates that these problems are solved optimally only using the column generation algorithm to solve the root node linear relaxation problem of these instances. There are eight instances in which the Tree size value is one, and the cut time value is non-zero, indicating that these instances can obtain the optimal integer solution after using the cutting plane method at the root node. Moreover, the larger the values of nSMV and n3SR, the more LB cut is improved relative to LP. It can be seen from the nSplits value of each instance that the smaller the vehicle capacity, the more often the split pickup or delivery occurs in the optimal solution. In addition, it can be seen from the IP values that in the instances with one depot and one vehicle, the larger the vehicle capacity, the lower the IP value; in the instances with one depot and two vehicles, the IP value of each vehicle type combination is not much different; in the instances with two depots and one vehicle, the greater the capacity used, the greater the IP value. Among the three cases, the IP of the instances with one depot and two vehicles is the smallest.
As shown in
Table 5 and the detailed results in
Table A4,
Table A5 and
Table A6, among the 120 instances, the BPC algorithm obtains the optimal integer solutions of 119 algorithms. The remaining instance
n20_2 with two depots, one vehicle and
only obtains integer feasible solutions, and the optimality has not been verified. The gap between the upper and lower bounds is 2.77%. Among all the examples, there are 15 instances. The
cut time value is zero, and the
Tree size value is one, which indicates that these problems are solved optimally only using the column generation algorithm to solve the root node linear relaxation problem of these instances. There are 19 instances whose
Tree size value is one, and the value of
cut time is non-zero. In these cases, the optimal integer solution to the problem is obtained after using the cutting plane method at the root node. It can be seen that the smaller the vehicle capacity, the longer the
IP time, the more valid inequalities obtained by the separation algorithm, and the split pickup or delivery occurs more frequently in the optimal solution. The difficulty of solving the instances gradually increases, and it takes longer to obtain the optimal solution. These results are also in line with reality.
As shown in
Table 5 and the detailed results in
Table A7 and
Table A8, for the 120 instances of
n30 instances, the BPC algorithm obtains the optimal integer solutions for 64 instances and obtains the integer feasible solutions for four instances. For the rest of the instances, the algorithm fails to obtain its integer feasible solution within the specified time.
It can be observed from
Table 4,
Table 5 and
Table 6 that with the expansion of the problem scale, the running time of the separation algorithm becomes longer, and the effective inequalities separated by the three separation algorithms increase. Valid inequalities play a more significant role in improving the lower bound of the problem. In addition, the expansion of the problem scale leads to a surge in the number of nodes in the branch-and-bound search tree; that is, it is more difficult for the algorithm to find an integer solution to the problem.
6. Conclusions
This paper studies a more realistic problem considering the split load, heterogeneous fleet and multiple depots. We propose an arc-flow formulation for the problem and use Dantzig–Wolfe Decomposition to reformulate it into a set partitioning master problem and several pricing subproblems. The pricing subproblem is a combination of RCESPP and LP-MBKP. This paper develops a BPC algorithm to solve the HFMDSBRP-SD. We design a Tabu Search column Generator and a heuristic label-setting algorithm to accelerate the accelerate the column generation procedure. Three types of valid inequalities, SR, SMV, and elementary inequalities, are extended and applied to reduce the solution space and accelerate the solution process. The computational experiments demonstrate that TS and HLS significantly improve the performance of the column generation algorithm, and the extensions of the three types of valid inequalities are particularly effective for enhancing the BPC algorithm, where the SMV inequalities are the most effective and the elementary inequalities are the weakest. We conduct the computational experiments in 360 instances, where 298 instances are capable of achieving optimality within a two-hour time limit.
Since our branch-and-price-and-cut algorithm only optimally solves 298 instances among 360 instances, there is much space to improve our solution procedure, such as designing heuristic algorithms for the problem to obtain a better initial solution so as to shorten the computational time.