You are on page 1of 6

2018 5th International Conference on Control, Decision Thessaloniki, Greece

and Information Technologies (CoDIT’18) April 10-13, 2018

Dynamic and flexible scheduling for a single machine*


Z. Bahroun., A. Shamayleh, and R. S. Zakaria

Abstract—Manufacturing and service enterprises rely heavily dynamic arrival of urgent orders, and material availability. The
on Planning and Scheduling in decision making because they uncertainty can be in tasks duration, transportation times, and
suffer from high cost resources and high workload situation machines set-up times [1,2,3,4]. Both unexpected events and
making it vital to dynamically adjust and schedule resources. uncertainty necessitate the need for flexibility in scheduling
Different approaches were developed and used to tackle the which consists of developing explicitly or implicitly different
scheduling problems. Most of these approaches are based on the scheduling solutions to adapt to a new situation.
use of combinatorial optimization to build deterministic
schedules; however, these schedules are not efficient due to A wide range of methods were developed to solve the
uncertainties such as processing times and machine failure rates scheduling problems where combinatorial optimization was
which can create a volatile environment; thus, necessitating the the basis for most of these methods. Some of the developed
need for the use of dynamic scheduling. The objective of this methods were exact methods based on mathematical
work is to propose a dynamic and flexible scheduling tool for the programming and many of them were heuristics leading to
single machine case based on the overlapping load adjustment good or near optimal scheduling solutions. Most of the work in
approach. We developed for this purpose a tool that generates all the literature deal with building static schedules that can be
the possible schedules starting from a valid loading solution. We easily and completely disturbed by any unexpected event,
conducted an experimental study on a set of problems. We then therefore, dynamic scheduling is recommended since it is more
studied the quality of the generated schedules in terms of suited to better take into consideration different types of
schedule ending time of the last task (Cmax), number of tasks disturbances [2, 5]. There are three main types of approaches
delayed and total processing time delay. The results showed that to tackle dynamic scheduling problems [3, 6] which are:
a very large number of schedules can be generated from the
reactive scheduling approaches [7, 8], Predictive-Reactive
loading solution which shows the flexibility of the approach. In
addition, the study showed the quality of the generated schedules
approaches [9, 10, 11], and pro-active scheduling [12, 13, 14,
with an average Cmax gap of 1.7% to the best generated schedule. 15].
There is an increasing trend in the literature stressing the
Keywords-Dynamic scheduling; Flexible scheduling; Single need for real-time data for scheduling and studying its impact
machine; Load adjustment; on quality of schedules developed. Luo et al. [16] discussed the
I. INTRODUCTION AND LITERATURE REVIEW implementation of RFID technologies to enable the shop floor
visibility and reduce uncertainties in the real-time scheduling
Scheduling is essential in many manufacturing and service for hybrid flow-shop (HFS) production. They created a
fields such as in computer systems (scheduling tasks on the ubiquitous manufacturing environment by deploying advanced
different processors), transportation (routing problems), wireless devices into value-adding points for the collection and
healthcare (operating rooms scheduling, appointment systems) synchronization of real-time data. Also, they developed a
and manufacturing systems [1,2]. It is defined as the planning multi-period hierarchical scheduling mechanism to divide the
and allocating of different resources to execute a set of related planning time horizon into multiple shorter periods. Al-
tasks. The scheduling problems generally include a set of tasks, Behadili et al. [17] proposed an optimization model for robust
precedence constraints, processing times and a set of resources; scheduling that considers utility, stability and robustness
the objective is to assign tasks to the resources to optimize one measures to generate robust schedules that minimize the effect
or more criteria such as completion time of all jobs, lateness or of different real-time events on the planned schedule. Wang et
tardiness of the jobs, flow time or the weighted number of late al. [18] proposed a reinforcement learning approach for Real-
tasks [1]. There are two main types of scheduling approaches Time Decision of the dynamic flow-shop Scheduling problem;
which are deterministic and on-line. The deterministic models they discussed settings for orders, performance measurements,
prepare a full schedule of the operations to be run on the and learning methods in detail to construct a controlled
different resources assuming the data is well known with no environment.
randomness. The on-line models consider tasks arrival one by
one “on-line”; combinatorial optimization and simulation In this work, we will investigate and develop an original
approaches can sometimes be used to solve such scheduling predictive-reactive approach for dynamic and flexible
problems. manufacturing scheduling. The goal is exploring the use of the
overlapping load approach presented in section II as a
Online scheduling is becoming the center of focus since in scheduling tool as shown in section III. An experimental study
real systems, unexpected events and uncertainty are of great of the work is presented in section IV and a conclusion of our
importance because they can disturb the prepared schedule. work will be presented in section V.
Examples of unexpected events are: machines failures,

*Research supported by the American University of Sharjah FRG Grant A. Shamayleh is Assistant Professor at the American University of
N° FRG16-R-19 Sharjah, PO Box 26666, Sharjah, UAE (e-mail: ashamayleh@aus.edu).
Z. Bahroun is Associate Professor at the American University of Sharjah, R.S. Zakaria, Engineering System Management Master at the American
PO Box 26666, Sharjah, UAE (corresponding author: +971 6 515 2981; fax: University of Sharjah, PO Box 26666, Sharjah, UAE (e-mail:
+ 971 6 515 2979; e-mail: zbahroun@aus.edu). g00024154@alumni.aus.edu).

978-1-5386-5065-3/18/$31.00 ©2018 IEEE -616-


II. THE OVERLAPPING LOAD ADJUSTMENT APPROACH be found by subtracting the total load of this lapse (22 hours)
from the total capacity of three periods or days (8*3= 24 hours)
The overlapping load adjustment approach was initially which is equal to 2 hours. We can note that Job A has a
developed by Dillenseger [19] for the single machine case. It scheduling segment that is overlapping in period 3 but not
consists in calculating for each lapse of time composed of a completely included in the lapse [3,5] so its load will not count
set of consecutive periods: in the cumulative load of this lapse.
- A cumulative capacity which represents the sum of the
capacities of all the periods that compose the lapse. We can also check in Fig. 2 that all the lapses have a
remaining capacity greater or equal to 0 which means that there
- A direct load representing the sum for the loads of all tasks exists a possible loading solution according to preemptive
with a scheduling interval (from the ready date to the due possibility. If in some lapses the remaining capacity is
date) that exactly matches the lapse of time. negative, then we delay the minimum number of jobs with a
scheduling interval finishing exactly at the end of the
- A cumulative load that sums all the load for all tasks with considered lapse.
a scheduling interval completely included in the laps of
time. TABLE I. PROCESSING TIME, READY AND DUE DATE OF EXAMPLE 1
Dillenseger [19] proved that if for any lapse of time, the Job Process Ready Date Due Date Total Margins
cumulative capacity is greater than the cumulative load, we ID time (day) (day) in days
(hours)
always can find a loading solution of all the tasks within their A 3 2 3 1
scheduling interval and according to pre-emptive possibility.
B 4 1 2 1
Bahroun et al. [20-21] then applied the overlapping load C 2 1 2 1
approach for flow shops and job shops with and without D 3 2 2 0
interruption. E 5 3 4 1
F 3 3 4 1
Table 1. shows an example that considers a set of 12 jobs
G 2 3 4 1
to be processed on a single machine. The processing time (in
hours), ready date (in days) and the due date (in days) in H 3 1 1 0
columns 2 to 4 respectively. It is assumed a week of 5 days I 5 4 5 1
(periods) and 8 hours are available each day (period). The total J 2 4 5 1
margin in periods is calculated and reported in the same table. K 2 3 5 2
For Job K, for example, the scheduling interval is from M 3 4 5 1
beginning of period (day) 3 till end of period (day) 5. Job K
needs 2 hours on the single machine. The total margin for each
job is the total available time when you deduct from the
scheduling interval the number of periods (rounded up) needed
to process the job. For Job K, one period is needed to process
the job so 2 periods are left as margin.
In Fig. 1, the Capacity Requirement Planning (CRP) is
developed which draws the scheduling interval for each job.
For example, Job K needs 2 hours as processing time and can
be scheduled anytime between day 3 to day 5. Job B has a
processing time of 4 hours and can be scheduled anytime in
day 1 and 2.
The next step is to build the Feasibility Control Graph
(FCG), shown in Fig. 2, to check the feasibility of the loading Figure 1. Capacity Requirement Planning (CRP) of Example 1
solution. We consider 5 days, so we will have 5 lapses
composed of one day ([11], [22], [33], [44], [55]), 4 lapses
composed of two consecutive days ([12], [23], [34], [45]), 3
lapses composed of 3 consecutive days ([12], [24], [35]), 2
lapses of 4 consecutive days ([14], [25]) and finally only one
lapse composed of 5 consecutive days ([15]). For each lapse of
time (a set of consecutive days), the following is calculated: the
cumulative capacity, the direct and cumulative load and the
remaining capacity. For example, the lapse [3,5] that is
composed of days 3, 4 and 5, there is only one job K with a
scheduling interval starting exactly at period 3 and finishing
exactly in period 5 (Fig. 1) so the direct load for this lapse is
equal to 2 hours. In addition, we can see in Fig. 1 that there are
7 jobs with a scheduling segment completely included in this
lapse [3,5] (E, F, G, I, J, K, M). The cumulative load for this Figure 2. Feasibility Control Graph (FCG) of Example 1
lapse is equal hence to 22 hours. The remaining capacity can

CoDIT'18 / Thessaloniki, Greece - April 10-13, 2018 -617-


III. SCHEDULING USING THE OVERLAPPING LOAD
ADJUSTMENT APPROACH
(ending time of last task scheduled on the machine), the
The objective of this work is to investigate the use of the number of tasks delayed and the total delay in hours with the
overlapping load approach as a scheduling tool. The best generated solution and the Earliest Due Date (EDD)
overlapping approach ensures the existence of schedules with heuristic. For the example of Fig. 1, using Algorithm 1, we
preemption (at most one task may be interrupted by period). generated 2520 different schedules based on the validated
The idea is to build dynamically and progressively the schedule CRP; the solution with the minimum Cmax and then the
where at the end of each task the user will have a choice to minimum number of tasks delayed is considered as the best
launch one task between a set of tasks. generated solution.
Let us consider the CRP of Fig. 1. At period 1, we can see In our example, the best obtained Cmax is equal to 38 hours.
that three tasks can be loaded, so we can choose to start with From all the schedules with a Cmax = 38 the minimum number
task B, C or H. When we start period 2, in addition to the tasks of delayed tasks is equal to 0. The best generated solution is
that has not been started in period 1, new tasks can be thus with a Cmax = 38 hours with no delayed tasks. Out of 2520
scheduled (A and D). Tasks E, F, G and K can be scheduled generated schedules, there were 1512 schedules having the best
starting from period 3. The decision on whether to start a task Cmax equal to 38 and with no delayed tasks. Fig. 3 represents a
or not can be made at the starting of the period; if there is a sample of some generated schedules.
problem with one particular task, the user has a set of choices Over the 2520 schedules, we obtained the statistics reported
to replace this task without changing the Capacity Requirement in Table II. The minimum, maximum, average Cmax obtained
Planning. In some cases, and depending on previous choices, a are 38, 42, and 38.57 respectively. the average is close to the
task must be interrupted to replace it with more urgent task. minimum value meaning that a large number of schedules have
This is happening in particular cases and at most once each a Cmax close to 38. Also, many schedules have a Cmax equal to
period and only when we have a task that has a scheduling
38 but with 1 or 2 delayed tasks. The total number of tasks
interval greater or equal than 3 periods. If we start this task in
delayed varies between 0 and 2 for the 2520 generated
its first period and if in the second period, there are tasks than schedules with an average equal to 0.46. Finally, the total
can only been scheduled in that period, we may be obliged to processing delay is varying between 0 and 3 hours with an
interrupt this task at the end of period 1, process the tasks that average of 0.57 hours. We also compared our result with the
only can be scheduled in the second period and finish the most famous heuristic for the single machine problem, the
considered task in period 3. Earliest Due Date (EDD) heuristic approach. For this example,
We developed a tool (coded in Microsoft Visual Basic) to the EDD also gave the best solution with a Cmax of 38 hours
generate all the possible schedules from a validated Capacity and without delayed tasks. As we can see, in this example,
Requirement Planning and Feasibility Control Graph (all the many good scheduling solutions can be available starting from
lapses have a remaining capacity greater or equal than 0). We a validated CRP. The advantage is that we do not prepare a
compared then the results in terms of the makespan or Cmax fixed schedule that will become obsolete very quickly, we

Figure 3. A sample of generated schedules for Example 1

CoDIT'18 / Thessaloniki, Greece - April 10-13, 2018 -618-


rather build a schedule progressively and dynamically taking TABLE II. EXAMPLE 1 RESULTS
into account all the events that could happen in the shop floor. Min Max Average Standard
Algorithm 1: Feasible Schedules Generation based on either the Jobs’ Deviation
Ready Dates or Due Dates
Cmax 38 42 38.57 1.316
Input: Total tasks delay 0 2 0.46 0.691
Validated Single Machine with JOBS:= {j1, j2,…,jn} Total processing delay (hours) 0 3 0.57 0.965

Process:
For each ji ϵ JOBS IV. EXPERIMENTAL STUDY
get ready_date (RDi) An experimental study will be presented in this section to
get due_date (DDi) assess the value of the overlapping approach as a dynamic
if RDi ∉ RDn scheduling tool. We generated 25 instances inspired by Sels
RDn = RDi ∪ RDn and Vanhoucke [22] and according to the following
if DDi ∉ DDn assumptions:
DDn = DDi ∪ DDn - All instances are for 12 jobs.
- The period is a day of 8 hours (8 sub-periods).
//Sort elements of RDn in ascending order - The processing time for each job is generated
//Sort elements of DDn in ascending order
according to a uniform distribution between 1 and 5
hours.
For each RDi ϵ RDn
- The ready date follows a uniform distribution between
get jobs where ready_date = RDi
Store jobs in RDJi 1 and (total processing time of all tasks * a). The
RDJn = RDJi ∪ RDJn coefficient “a” has been fixed to 0.1 after many trials.
- The due date is generated according to a uniform
For each RDJi ϵ RDJn distribution between the ready date of the considered
permute jobs in RDJi task and (total processing time of all tasks * b). As for
store permutations for every RDJi the coefficient “a”, the coefficient “b” also has been
fixed to 0.125 after many trials.
//Starting from RDJ1, combine every permutation
under RDJ1 with every permutation in RDJ2. Then For each instance, we first build the CRP and the FCG using
combine every result under RDJ1+RDJ2 with RDJ3 to the developed tool. Lapses with negative remaining capacity
get RDJ1+RDJ2+ RDJ3 and so on. means that the instance does not have any feasible schedule.
//Eventually, a set of feasible schedules (FS1) based We delay in that case a minimum number of tasks to validate
on Ready Date will be formed under RDJ1+RDJ2+ the Feasibility Control Graph. The next step is to generate all
RDJ3+…+ RDJn. feasible schedules according to the validated CRP and using
Algorithm 1; we capped the number of generated solutions to
For each DDi ϵ DDn 100,000 schedules. After that, we compared the generated
get jobs where due_date = DDi schedules with the result obtained by the EDD heuristic and
Store jobs in DDJi also by the best generated solution in terms of Cmax, number
DDJn = DDJi ∪ DDJn
of tasks delayed and total processing delay.
For each DDJi ϵ DDJn
permute jobs in DDJi
As discussed earlier, the best generated solution is the one
store permutations for every DDJi
with the minimum Cmax and then the minimum number of
//Starting from DDJ1, combine every permutation delayed tasks. To determine the best generated solution, we
under DDJ1 with every permutation in DDJ2. Then first determine the minimum Cmax from all the generated
combine every result under DDJ1+DDJ2 with DDJ3 to solutions. Then, in the set of solutions having this minimum
get DDJ1+DDJ2+ DDJ3 and so on. Cmax, we select the solution having the minimum number of
//Eventually, a list of feasible schedules (FS2) based delayed tasks. In other words, we can have a solution having
on Due Date will be formed under DDJ1+DDJ2+ a higher Cmax than the defined best solution but with lower
DDJ3+…+ DDJn. number of delayed tasks.

Output: In Table III, we report our results of the best generated


FEASIBLE_SCHEDULES = FS1 ∪ FS2 solution for each instance as well as the EDD solution. We can
//The result FEASIBLE_SCHEDULES has all the obviously see that the results obtained from our approach are
feasible schedules for the given Validated Single much better than the EDD heuristic even that it is important to
Machine; no duplicates. mention that our objective is to dynamically built a flexible
schedule from many possibilities and not determining a

CoDIT'18 / Thessaloniki, Greece - April 10-13, 2018 -619-


unique solution like the EDD approach. TABLE IV. NUMBER OF GENERATED SCHEDULES, NUMBER OF BEST
SCHEDULES, AND STATISTICS FOR CMAX

Ins. Num. of Num. % of Cmax Cmax Aver. SD %


Table IV reports first for each instance, the number of sched. sche. best min max Cmax Cmax Gap-
generated schedules (capped to 100,000 schedules), the equal sol. Cmax
best sol. min
number of schedules having the same Cmax and the same 1 38880 960 2.5 33 44 33.69 2.15 2.1
number of delayed tasks than the best solution defined in 2 18720 480 2.6 33 36 33.15 0.66 0.5
3 34992 384 1.1 37 41 37.04 0.37 0.1
Table III. We also reported the percentage of best solutions 4 100000 1440 1.4 29 33 29.15 0.73 0.5
obtained compared to all the generated solutions. We finally 5 18720 1280 6.8 33 41 33.35 1.45 1.1
mention some statistics about the Cmax over all the generated 6 60912 48 0.1 33 37 33.02 0.23 0.1
7 30528 192 0.6 39 47 39.27 1.29 0.7
solutions as well as the Gap in percentage between the average 8 27072 1152 4.3 33 38 33.17 0.88 0.5
Cmax and the minimum Cmax. It can be noticed that there are 9 2016 288 14.3 44 56 45.36 3.48 3.1
many generated solutions that are equal to the best solution. In 10 3744 384 10.3 41 58 45.83 6.57 11.8
11 31680 1872 5.9 35 41 35.34 1.14 1.0
addition, the gap between the best Cmax and the average one is 12 19008 3600 18.9 39 47 39.61 1.93 1.6
very small (1.7 % on average). This shows that the 13 35136 288 0.8 35 38 35.04 0.34 0.1
overlapping approach, despite its important flexibility, 14 100000 2078 2.1 46 50 46.01 0.22 0.0
15 17712 1248 7.0 37 44 37.15 0.95 0.4
suggests in many cases a good solution. 16 26208 48 0.2 34 36 34.02 0.17 0.1
17 86784 192 0.2 39 42 39.01 0.14 0.0
Table V reports statistics about the number of delayed tasks 18 3744 288 7.7 42 42 42 0 0.0
19 43200 2880 6.7 34 46 37.22 4.77 9.5
and the total processing delay. We can see that for some 20 2304 56 2.4 42 51 43.82 3.25 4.3
instances, the minimum number of delayed tasks is sometimes 21 61344 1168 1.9 36 37 36.01 0.09 0.0
22 31104 12000 38.6 35 43 35.41 1.59 1.2
lower than the one of the best solution. This is because we
23 15264 480 3.1 39 42 39.19 0.72 0.5
gave the priority first to the Cmax when we defined the best 24 17558 64 0.4 36 38 36.03 0.22 0.1
solution. 25 10368 864 8.3 46 59 47.37 3.34 3.0
Aver. 5.9 Aver. 1.7
TABLE III. BEST GENERATED SOLUTION AND EDD RESULTS
Instance Best generated solution EDD TABLE V. STATITICS FOR NUMBER OF DELAYED TASKS AND TOTAL
PROCESSING DELAY
Cmax Num. of tasks Cmax Num. of tasks
delayed delayed Instance Number of tasks delayed Total processing delay
1 33 2 44 8 (hours)
2 33 0 33 0 Min Max Avg SD Min Max Avg SD
3 37 0 38 0 1 2 8 4 1.03 5 55 20.3 6.85
4 29 0 29 0 2 0 5 3.5 0.99 0 26 16.3 5.19
5 33 1 39 1 3 0 5 2 0.72 0 23 12.5 5.35
6 33 0 37 1 4 0 6 4.1 0.87 0 48 32.7 8.4
7 39 0 44 5 5 0 6 2.6 0.98 0. 25 10.6 2.98
8 33 3 38 5 6 0 7 3.9 1.01 0 35 20.3 5.86
9 44 2 49 2 7 0 9 3.5 1.1 0 44 30.6 6.38
10 41 1 50 5 8 0 8 4.9 1.09 0 43 25.8 7.19
11 35 3 38 1 9 2 8 3.6 1.32 5 46 21.3 6.76
12 39 2 45 5 10 0 7 3.5 1.98 0 69 21.3 19.8
13 35 1 37 1 11 1 7 4.3 1.22 1 35 21.6 7.22
14 46 1 47 1 12 2 8 3.3 1.17 9 37 22.1 5.5
15 37 2 44 5 13 1 5 3.5 0.91 2 40 20.0 8.92
16 34 0 36 4 14 1 6 3.1 0.85 1 48 31.9 6.96
17 39 0 39 0 15 2 6 3.5 0.8 5 38 25.3 4.93
18 42 0 42 0 16 0 8 5.5 1.01 0 50 35.93 6.19
19 34 1 45 8 17 0 7 3.6 1.01 0 38 21.9 6.67
20 42 3 47 2 18 0 6 3.6 1.38 0 27 14.5 6.07
21 36 0 36 0 19 1 9 3.7 2.55 5 63 19.4 15.49
22 35 2 43 8 20 2 8 4.2 1.35 9 38 23.8 6.02
23 39 0 42 4 21 0 4 2.1 0.82 0 14 7.81 4.09
24 36 0 36 0 22 2 10 2.8 1.35 4 40 14.97 5.24
25 46 2 59 10 23 0 5 4.1 1.00 0 36 22.9 7.43
24 0 6 3.7 0.92 0 31 16.1 5.93
25 0 10 3.7 1.7 0 70 22.7 10.79

CoDIT'18 / Thessaloniki, Greece - April 10-13, 2018 -620-


V. CONCLUSION [7] S. Nguyen, M. Zhang, M. Johnston, K. Tan, “Automatic design of
scheduling policies for dynamic multi-objective job shop scheduling
We investigated in this paper the development of a dynamic via cooperative coevolution genetic programming, IEEE Transactions
and flexible scheduling tool based on the overlapping on Evolutionary Computation,” vol. 18, no. 2, pp. 193-208, 2013.
[8] L. Nie, L. Gao, P. Li, X. Li, “A GEP-based reactive scheduling
adjustment approach. The main advantage of the overlapping policies constructing approach for dynamic flexible job shop
load adjustment approach is that it ensures the existence of scheduling problem with job release dates,” Journal of Intelligent
feasible schedules without explicitly generate them. After Manufacturing, vol. 24, no. 4, pp. 763-774, 2013.
[9] M.A. Adibi, M. Zandieh, M. Amiri, “Multi-objective scheduling of
generating and validating the Capacity Requirement Planning
dynamic job shop using variable neighborhood search,” Expert
and the Feasibility Control Graph, the user can choose at the Systems Application, vol. 37, no. 1, pp. 282-287, 2010.
beginning of each period, dynamically and in real time one of [10] G. Chryssolouris, E. Subramanian, “Dynamic scheduling of
the tasks that can be loaded at that period. manufacturing job shops using genetic algorithm,” Journal of
Intelligent Manufacturing, vol. 12, pp. 281-293, 2001.
[11] C. Raquel, X. Yao, “Dynamic multi-objective optimization: a survey
We developed in this paper a tool that, starting from a of the state-of-the-art,” in Evolutionary Computation for Dynamic
validated Capacity Requirement Planning, generates all the Optimisation Problems, Springer-Verlag, Berlin, pp. 58-106, 2013.
[12] D. Lei, “Simplified multi-objective genetic algorithms for stochastic
possible schedules. We conducted an experimental study on a job shop scheduling,” Applied Soft Computing, vol. 11, no. 8, pp.
set of 25 problems for a weekly schedule of 5 days and 40 4991-4996, 2011.
hours. We then generated for each problem and from the [13] L. Liu, H. Gu, Y. Xi, “Robust and stable scheduling of a single
machine with random machine breakdowns,” International Journal of
validated Capacity Requirement Planning, all the possible Advanced Manufacturing Technology, vol. 31, pp. 645-654, 2007.
schedules. After that, we assessed the results in terms of [14] M.T. Jensen, “Generating robust and flexible job shop schedules using
Cmax, number of delayed tasks and total processing time genetic algorithms,” IEEE Transactions on Evolutionary Computation,
vol. 7, no. 3, pp. 275-288, 2003.
delay. In all instances, we were able to generate a large [15] N. Al-Hinai, T.Y. ElMekkawy, “Robust and stable flexible job shop
number of schedules which shows the flexibility of our scheduling with random machine breakdowns using a hybrid genetic
approach. 5.9% of the generated schedules correspond to the algorithm,” International Journal of Production Economics, vol. 132,
no. 2, pp. 279-291, 2011.
best generated solution. Moreover, the gap between the Cmax
[16] H. Luo, J. Fang, G. Q. Huang, “Real-time scheduling for hybrid
of any generated solution and the best Cmax is only equal to flowshop in ubiquitous manufacturing Environment,” Computers &
1.7% on average. Industrial Engineering, vol. 84, pp. 12-23, 2015.
[17] M. Al-Behadili, D. Ouelhadj, D. Jones, “Multi-objective Particle
Swarm Optimisation for Robust Dynamic Scheduling in a Permutation
In the future, we will conduct a larger experimental study Flow Shop,” in International Conference on Intelligent Systems
to assess the quality and the number of the generated solutions Design and Applications, pp. 498-507, 2016.
under different conditions. [18] J. Wang, S. Qu, J. Wang, J. O. Leckie, R. Xu, “Real-Time Decision
Support with Reinforcement Learning for Dynamic Flowshop
Scheduling,” Smart SysTech, 2017.
We also would like to extend this tool to the flow-shop case [19] F. Dillenseger, “Conception d’un système de planification à moyen
composed of a set of machines on a production line. We terme pour fabrications à la commande,” Ph.D. thesis, INSA Lyon,
France, 1993.
should in that case dynamically manage the link between the [20] Z. Bahroun, M. Fendouli, J-P. Campagne, “A new load adjustment
Capacity Requirement Planning of each machine and make approach for job-shops,” in Advances in Robotics Automation and
sure that they stay all time validated for all the machines. Control, edited by I-Tech books Croatia, pp. 151-170, 2008.
[21] Z. Bahroun, An advanced load adjustment approach without tasks
Indeed, each scheduling decision on any stage will impact the interruption for flow shops, International Journal of Industrial and
considered stage but also the next stages. Systems Engineering, vol. 7, no. 3, 381-413, 2011.
[22] V. Sels, M. Vanhoucke, “A hybrid dual-population genetic algorithm
for the single machine maximum lateness problem,” in EvoCOP 2011,
European Conference on Evolutionary Computation in Combinatorial
REFERENCES Optimization, pp. 14-25, 2011.
[1] J-C. Billaut, A. Moukrim, E. Sanlaville, Flexibility and Robustness in
Scheduling, Edited by Wiley, 2008.
[2] P. Fattahi, A. Fallahi, “Dynamic scheduling in flexible job shop
systems by considering simultaneously efficiency and stability, CIRP
Journal of Manufacturing Science and Technology,” vol. 2, pp. 114-
123, 2010.
[3] X-N. Shena, X. Yaob, “Mathematical modeling and multi-objective
evolutionary algorithms applied to dynamic flexible job shop
scheduling problems,” Information Sciences, vol. 298, pp. 98-224,
2015.
[4] B. J. Joo, Y. C. Choi, P. Xirouchakis, “Dispatching Rule-based
Algorithms for a Dynamic Flexible Flow Shop Scheduling Problem
with Time-dependent Process Defect Rate and Quality Feedback,”
Procedia CIRP, vol. 7, pp. 163-168, 2013.
[5] A. V. Barenji, R. V. Barenji, M. Hashemipour, “A frameworks for
structural modelling of an RFID-enabled intelligent distributed
manufacturing control system,” South African Journal of Industrial
Engineering, vol. 25, pp. 48-66, 2014.
[6] D. Ouelhadj, S. Petrovic, “A survey of dynamic scheduling in
manufacturing systems, Journal of Scheduling,” vol. 12, no. 4, pp.
417-431, 2009.

CoDIT'18 / Thessaloniki, Greece - April 10-13, 2018 -621-

You might also like