1. Introduction
The operational parameters of a scheduling problem cannot be fixed or predetermined. For example, the processing time might be influenced by machine breakdowns or altered by the number of ordered items (products), the release date might be delayed by unexpected factors, and the due date might have to be adjusted earlier/later for customers. Therefore, these parameters in a scheduling problem can be treated as uncertain. For example, the travel time of a caregiver might vary and the service time of an elder could be prolonged in the home health industry. In home healthcare situations, assigning caregivers and routes of services is a practical and important issue. For example, [
1] supplied a framework to make a robust scheduling problem for an institution. Other examples were presented in the facility location problem by [
2], in the discrete time/cost trade-off problem by [
3], and in random quality deteriorating hybrid manufacturing by [
4].
To address these situations, researchers attempt to search job sequences or job schedules to reduce the risk or to minimize one or several aspects of loss measures. This led to the robust approach to solve a scheduling model with scenario-dependent effects [
5,
6]. There were two deterministic methods (in contrast to stochastic methods) to model the uncertain parameters of a scheduling problem discussed in the relevant literature. The uncertain parameters were either bounded within an interval (continuous case) or described by a finite number of scenarios (discrete case) [
5,
6]. The discrete scenario robust scheduling problem was first studied by [
6]. Taking a finite number of scenarios into consideration, they discussed three robust measures in a single-machine setting, i.e., the absolute robust, the robust deviation, and the relative robust deviation [
7]. The main objective function is to seek an optimal job schedule among all possible permutations over all possible scenarios. Other studies pertinent to machine scheduling in the face of uncertainty for discrete cases include [
8,
9]. Additionally, some recent distributionally robust optimization scheduling problems with uncertain processing times or due dates include [
10,
11].
The total tardiness measure is very important in practice. Tardiness relates to backlog (of order) issues, which may cause customers to demand compensation for delays and loss of credit. In addition to the single-machine operation environment, the total tardiness minimization has been investigated in other work environments, e.g., flow-shop, job shop, parallel, and order scheduling problems (refer to [
12,
13], respectively). Readers can refer to the review papers by [
14,
15] on the total (weighted) tardiness criterion in the scheduling area. More recently, [
16] introduced the scenario-dependent idea into a parallel-machine order scheduling problem where the tardiness criterion is minimized.
However, the performance measure total tardiness together with the uncertainty of parameters is rarely discussed in the literature for deterministic models of scheduling problems. Moreover, [
17] noted that “single-machine models often have properties that have neither machines in parallel nor machines in series. The results that can be obtained for single machine models not only provide insights into a single machine environment, they also provide a basis for heuristics that are applicable to more complicated machine environments…” (see page 35 in [
17]). Therefore, we introduce scenario-dependent due dates and scenario-dependent processing times to a single-machine setting, in which the criterion is the sum of tardiness of all given jobs.
This work provides several contributions: We introduce a scenario-dependent processing time and due date concept to a single-machine scheduling problem to minimize the sum of job tardiness in the worst case. We derive a lower bound and several properties to increase the power capability of a branch-and-bound (B&B) method for an exact solution. In addition, we propose three different values for a parameter in one local search heuristic method and a population-based iterative greedy algorithm for near-optimal solutions. The organization of this article is as follows:
Section 2 states the formulation of the investigated problem.
Section 3 derives one lower bound and eight dominances used in the B&B method, introduces three different values for a parameter in one local search heuristic method and a population-based iterative greedy (PBIG) method, and proposes details of the branch-and-bound method.
Section 4 presents tuning parameters in the iterative greedy algorithm.
Section 5 reports and analyses the related simulation results. The conclusions and suggestions are summarized in
Section 6.
2. Problem Statement
In this section, the study under consideration is formally described. There is a set
J of
n jobs {
J1,
J2, …,
Jn} to be processed on a single-machine environment. The machine can execute at most one job at a time, preemption is not allowed, and their ready times are zero. Another assumption is that there are two scenarios of job parameters (due dates and processing times) to capture the uncertainty of the two parameters discretely. Assume that
represents the processing time while
represents the due date of job
for scenario
v = 1, 2. Furthermore, for scenario
v, let
represent the completion time of job
, where
is a job sequence (schedule). The tardiness of
is defined to be
, and the total tardiness for all
n jobs is
for scenario
v. Without considering the uncertainty of the parameters, the classic tardiness minimization scheduling problem in a single-machine environment, denoted by
, has been shown to be an NP-hard problem. Accordingly, the considered problem can be denoted by
and as an NP-hard problem ([
18,
19]) That is, under the assumptions that
and
are uncertain and can be captured discretely by scenario
v = 1, 2, we are required to find an optimal robust schedule
such that
. In other words, the aim of this study is to seek a robust single-machine schedule incorporating scenario-dependent due dates and scenario-dependent processing times in which the tardiness measurement for the worst scenario can be minimized (optimal).
5. Computational Experiments and Analysis of Results
This section performs several problem instance tests to check the computational behaviors of the proposed heuristics and metaheuristics. The processing times, integers
and
, were generated independently from two different uniform distributions (i.e., U [1, 100] and U [1, 200], respectively, see [
16]), while the due dates,
, are integers generated from a uniform distribution
, where
,
represent the range of the due dates and the tardiness factor (see [
28]), respectively. The values of
were designed as 0.25 and 0.5, and the values of
were 0.25, 0.5, and 0.75. For each combination of
and
, 100 instances were generated as the test bank. In addition, as the number of explored nodes exceeds 10
8, the branch-and-bound method is terminated and advances to the next set of instances. To determine the behavior of the B&B method, three local heuristics, and the PBIG algorithm, the experiments were examined for job sizes
n = 8, 10, and 12 for the small-size problem and
n = 60, 80, and 100 for the large-size problem. In total, 1800 problem instances were generated to solve the proposed problem. The four proposed algorithms were coded in FORTRAN 90. They were executed on a 16 GB RAM, 3.60 GHz, Intel(R) Core™ i7-4790 personal computer (64 bits) with Windows 7.
The represent the optimal values received by running the B&B method, and the record the solution obtained by each heuristic (or algorithm) for the test instances for the small-size case. To evaluate the performances of the three heuristics and the PBIG algorithm, the average error percentage (AEP =) is used.
Table 1.
The behavior of the B&B.
Table 1.
The behavior of the B&B.
| | | Node | CPU_Time |
---|
n | | | Mean | Max | Mean | Max |
---|
8 | 0.25 | 0.25 | 1364.95 | 8041 | 0.00 | 0.03 |
| | 0.5 | 924.09 | 5872 | 0.00 | 0.03 |
| | 0.75 | 649.00 | 3344 | 0.00 | 0.02 |
| 0.5 | 0.25 | 2169.63 | 12,975 | 0.01 | 0.03 |
| | 0.5 | 1441.85 | 11,199 | 0.00 | 0.03 |
| | 0.75 | 850.82 | 5707 | 0.00 | 0.02 |
10 | 0.25 | 0.25 | 18,383.94 | 155,332 | 0.08 | 0.58 |
| | 0.5 | 11,381.24 | 89,998 | 0.05 | 0.37 |
| | 0.75 | 6970.97 | 44,421 | 0.03 | 0.19 |
| 0.5 | 0.25 | 47,294.00 | 634,898 | 0.18 | 2.25 |
| | 0.5 | 29,192.77 | 673,842 | 0.12 | 2.42 |
| | 0.75 | 25,154.69 | 717,700 | 0.10 | 2.59 |
12 | 0.25 | 0.25 | 319,906.05 | 4,022,178 | 1.91 | 22.11 |
| | 0.5 | 181,384.58 | 1,733,253 | 1.35 | 13.63 |
| | 0.75 | 77,810.74 | 603,931 | 0.63 | 4.62 |
| 0.5 | 0.25 | 2,155,579.66 | 44,957,346 | 13.91 | 265.44 |
| | 0.5 | 517,837.47 | 19,204,302 | 3.74 | 125.30 |
| | 0.75 | 108,610.00 | 2,626,869 | 1.00 | 19.44 |
| | mean | 194,828.14 | 4,195,067 | 1.28 | 25.51 |
Table 1 presents the capability of the B&B method. All tested problem instances could be solved before 10
8 nodes. The computation CPU times, including the average execution times and maximum execution times (in seconds), increased dramatically as
n increased (Columns 6 and 7,
Table 1). As
n increased, the mean and maximum nodes also increased (Columns 4 and 5,
Table 1).
Table 2 reports the results of CPU times and node numbers for the small-size
n,
, and
.
Table 2.
Summary of the results of the B&B method.
Table 2.
Summary of the results of the B&B method.
| | Node | CPU_Time |
---|
| | Mean | Max | Mean | Max |
---|
n | 8 | 1233.390 | 7856.333 | 0.002 | 0.027 |
| 10 | 23,062.935 | 386,031.833 | 0.093 | 1.400 |
| 12 | 560,188.083 | 12,191,313.167 | 3.757 | 75.090 |
| 0.25 | 68,752.840 | 740,707.778 | 0.450 | 4.620 |
| 0.5 | 320,903.432 | 7,649,426.444 | 2.118 | 46.391 |
| 0.25 | 424,116.372 | 8,298,461.667 | 2.682 | 48.407 |
| 0.5 | 123,693.667 | 3,619,744.333 | 0.877 | 23.630 |
| 0.75 | 36,674.370 | 666,995.333 | 0.293 | 4.480 |
Regarding the behaviors of the three heuristics and the PBIG, their AEPs are displayed in
Table 3. All AEPs of the three heuristics and the PBIG algorithm increased slightly as
n increased. Overall, the PBIG algorithm, with a mean AEP of less than 0.14% for
n = 8, 10, and 12, performed the best.
Figure 9 indicates the AEPs (output results) of the three heuristics and the PBIG algorithm. Since the computer execution times are all less than 0.1 s, they are not reported here.
Table 3.
The AEPs of the heuristics and the PBIG algorithm.
Table 3.
The AEPs of the heuristics and the PBIG algorithm.
| | HA_025 | HA_050 | HA_075 | PBIG |
---|
n | 8 | 10.301 | 8.121 | 9.948 | 0.004 |
| 10 | 14.283 | 9.408 | 12.235 | 0.027 |
| 12 | 17.045 | 16.276 | 21.333 | 0.130 |
| 0.25 | 18.925 | 13.966 | 19.432 | 0.011 |
| 0.5 | 8.827 | 8.570 | 9.578 | 0.096 |
| 0.25 | 7.850 | 7.623 | 7.642 | 0.020 |
| 0.5 | 10.806 | 8.890 | 12.193 | 0.046 |
| 0.75 | 22.972 | 17.291 | 23.681 | 0.094 |
For the behaviors of the proposed four algorithms, we performed an analysis of variance (ANOVA) on the AEPs. As shown in
Table 4 (Columns 2 and 3), the Kolmogorov–Smirnov test was significant, with a
p value
0.01. This implies that the samples of AEPs do not follow the normal distribution. Therefore, based on the ranks of AEPs, the Kruskal–Wallis test was utilized to determine if the populations of AEPs came from the same population or not. Column 2 of
Table 5 confirmed that the proposed three heuristics and the PBIG algorithm were indeed significantly different, with a
p value < 0.001.
Table 4.
Normality Tests for small n and large n.
Table 4.
Normality Tests for small n and large n.
| Small n | Large n |
---|
Method of Normality Test | Statistic | p Value | Statistic | p Value |
---|
Shapiro–Wilk normality test | 0.7957 | <0.0001 | 0.6225 | <0.0001 |
Kolmogorov–Smirnov test | 0.1753 | <0.0100 | 0.1805 | <0.0100 |
Cramer–von Mises normality test | 0.5897 | <0.0050 | 0.6418 | <0.0050 |
Anderson–Darling normality test | 3.5444 | <0.0050 | 4.2936 | <0.0050 |
Table 5.
Kruskal–Wallis Test.
Table 5.
Kruskal–Wallis Test.
Kruskal–Wallis Test |
---|
| Small n | Large n |
---|
Chi-square | 40.8017 | 29.6735 |
DF | 3 | 3 |
Pr > Chi-square | <0.0001 | <0.0001 |
In addition, heuristics HA_025, HA_050, HA_075, and the PBIG were further used to conduct pairwise differences. The Dwass–Steel–Critchlow–Fligner (DSCF) procedure was applied (see [
29]).
Table 6 confirms that the mean ranks of AEPs were grouped into two subsets under the level of significance of 0.05. From Columns 3 and 4 of
Table 6, the PBIG (with the AEP of 0.005) was placed in a better behavior group, while HA_025 (with the AEP of 0.014), HA_050 (with the AEP of 0.011), and HA_075 (with the AEP of 0.015) were placed in a worse performance set.
Table 6.
DSCF pairwise comparison.
Table 6.
DSCF pairwise comparison.
Pairwise Comparison | DSCF |
---|
Small n | Large n |
---|
Between Algorithms | Statistic | p Value | Statistic | p Value |
---|
HA_025 vs. HA_050 | 1.2976 | 0.7955 | 0.1790 | 0.9993 |
HA_025 vs. HA_075 | 0.4922 | 0.9855 | 2.3714 | 0.3359 |
HA_025 vs. PBIG | 7.2880 | <0.0001 | 5.2350 | 0.0012 |
HA_050 vs. HA_075 | 1.5213 | 0.7044 | 3.0426 | 0.1371 |
HA_050 vs. PBIG | 7.2880 | <0.0001 | 5.5035 | 0.0006 |
HA_075 vs. PBIG | 7.2880 | <0.0001 | 6.7563 | <0.0001 |
Regarding the test performance on the large-sized jobs, we tested the number of jobs at
n = 60, 80, and 100. For each combination of
, and
n, we generated one hundred problem instances to evaluate the performances of the proposed methods. Overall, we examined and tested 1800 random problem instances. The measurement is the relative percent deviation (RPD), where RPD is defined as
). It was noted that
was the value of the objective function found in each algorithm, and
was the smallest objective function between the four methods. All of the average RPDs of the four algorithms are recorded in
Table 7. As shown in
Table 7, PBIG provided the lowest value of RPDs, no matter the value of
n.
Figure 10 displays the boxplots of the RPDs for the three heuristics and the PBIG algorithm.
Furthermore, using ANOVA to determine whether the RPDs follow a normal distribution or not,
Table 4 (Columns 4 and 5) indicates that the normality assumption is not met, since the
p value < 0.01. Therefore, Column 3 of
Table 5 indicates that a Kruskal–Wallis test, which is based on ranks of RPDs, clearly states that “the RPD samples belong to different distributions” when the
p value < 0.001. Thus, the DSCF procedure was adopted to compare the pairwise differences between the four methods. Columns 4 and 5 of
Table 6 report that the PBIG algorithm was placed in a better set; meanwhile, the other three heuristics belong to another set for a large number of job cases. Furthermore, the boxplots in
Figure 10 show that the RPDs of the PBIG had a smaller range than those of the three heuristics. This implies that the PBIG could find a stable and accurate solution when compared to the other three heuristic methods in the large-size problem cases. For the computational time or CPU times (in seconds),
Figure 11 displays the boxplots of the times for the heuristics and PBIG algorithm. Three heuristics take less than one second, while PBIG takes less than 15 s.
Table 7.
The RPD values of the four algorithms.
Table 7.
The RPD values of the four algorithms.
| | HA_025 | HA_050 | HA_075 | PBIG |
---|
n | 60 | 214.519 | 72.929 | 47.249 | 2.521 |
| 80 | 69.519 | 33.830 | 50.049 | 1.773 |
| 100 | 71.282 | 35.274 | 52.304 | 2.665 |
τ | 0.25 | 232.728 | 90.140 | 90.587 | 3.387 |
| 0.5 | 4.152 | 4.549 | 9.148 | 1.253 |
ρ | 0.25 | 4.678 | 4.578 | 6.358 | 1.855 |
| 0.5 | 30.137 | 11.730 | 27.567 | 2.163 |
| 0.75 | 320.506 | 125.724 | 115.677 | 2.941 |