A Two-machine Jobshop Scheduling Problem with
Convex Processing Times
1st Kailiang Xu
School of Automation and Information Engineering
Xi’an University of Technology
Xi’an, Shaanxi, P.R.China
Email: klxu@xaut.edu.cn
3rd Yuchuan Sun
School of Automation and Information Engineering
Xi’an University of Technology
Xi’an, Shaanxi, P.R.China
Email: 18991159381@126.com
Abstract—A two-machine job-shop scheduling problem is con-
sidered in this paper, where job-processing times are controllable,
and modeled by a non-linear convex resource consumption func-
tion. The objective is to minimize the resource consumption under
the condition that makespan should not exceed the given deadline.
The problem is strongly 𝒩 𝒫-hard. A polynomial time resource
allocation algorithm is presented first, which optimally allocate
resource to job operations with arbitrarily given processing order.
After that, a heuristic algorithm is presented, which decides the
near-optimal processing sequence of the operations of the jobs.
Index Terms—Two-machine jobshop, Convex resource con-
sumption function, Heuristic
I. I NTRODUCTION
The job-shop problem is well-studied in literature. In most
research work, it is assumed that processing time of job
operations is constant and known in advance. However, in
many practical environments, processing time can be con-
trolled by the consumption of an extra resource, such as gas,
fuel, electricity power, cash and labors. For example, Janiak[1]
described an application in steel mills, where batches of ingots
have to be preheated before being hot-rolled in a blooming
mill, and both the preheating time and the rolling time are
inversely proportional to the gas flow intensity. Therefore,
the classical job-shop scheduling problem need be generalized
with controllable processing time of the job operations.
In this paper we study a two-machine job-shop scheduling
problem where processing time and resource consumption
is modeled by a non-linear convex function introduced by
Monma et al.[10] and Shabtay and Steiner[8], stated formally
as follows:
( )𝑘
𝜔𝑗,𝑖
𝑝𝑗,𝑖 (𝑢𝑗,𝑖 ) = 𝑗 = 1, 2, . . . , 𝑛 𝑖 = 1, 2, . . . , 𝑚
𝑢𝑗,𝑖
(1)
where 𝜔𝑗,𝑖 > 0 represents the workload of operation 𝑂𝑗,𝑖 that
belongs to job 𝑗 and is processed on machine 𝑖, 𝑢𝑗,𝑖 > 0
is the amount of the resource allocated to the operation, and
𝑘 is a constant positive parameter. The problem is formally
978-1-5386-3758-6/18/$31.00 2018
c IEEE
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 19,2024 at 17:19:51 UTC from IEEE Xplore. Restrictions apply.
2nd Jing Luo
School of Computer Science Engineering
Xi’an University of Technology
Xi’an, Shaanxi, P.R.China
Email: luojing@xaut.edu.cn
stated as follows: A set of jobs 𝒥 = {1, 2, . . . , 𝑛} are to
be processed in a two-machine job-shop without recircula-
tion. Each job has at most two operations, 𝑂𝑗,1 and 𝑂𝑗,2 .
The processing time of each operation, 𝑝𝑗,1 or 𝑝𝑗,2 , is a
decreasing non-linear convex function with respect to the
resource consumption 𝑢𝑗,1 or 𝑢𝑗,2 described by Eq.1. The
objective is to determine the processing sequence as well as
the resource allocation of the operations on both machines,
such that the makespan is limited within deadline 𝐾, while
the total resource consumption is minimized. Using the three
field problem classification introduced by Graham et al.[5]
and extended by Shabtay and Steiner[8],
∑2 ∑this problem can be
𝑛
denoted as 𝐽2 ∣𝑐𝑜𝑛𝑣, 𝐶max ≤ 𝐾 ∣ 𝑚=1 𝑗=1 𝑢𝑗,𝑚 .
In this paper a polynomial time resource allocating algo-
rithm is presented in Section II, which optimally allocates
resource to operations under the condition that their processing
sequence is already determined on both machines. After that,
several properties of the optimal solutions are discussed in
Section III, which helps reducing the problem into a job-
sequencing problem with controllable processing times. Based
on this, a heuristic algorithm is presented in Section IV. The
conclusion is drawn in Section V.
II. O PTIMAL RESOURCE ALLOCATION
Consider in this section the problem of optimally allocating
resource to operations under an arbitrary processing sequence.
The problem is first decomposed into a set of sub-problems
solved by the equivalent workload method. After that, a
dynamic program method is applied to search for the optimal
solution among them.
According to the equivalent workload method defined by
Monma et al.[10], when 𝑛 operations are to be processed in
serial with makespan no larger than deadline 𝐾, the optimal
2769
solution is equivalent workload
⎧ 𝑘 ΩΛ𝑙 = ΩΛ𝑙 ,1 + ΩΛ𝑙 ,2

 𝜔𝑗𝑘+1 ⎛ ⎞ 𝑘+1 ⎛ ⎞ 𝑘+1

 𝑝𝑗 = 𝐾 𝑗 = 1, 2, . . . , 𝑛

 ∑𝑛 𝑘
∑
𝑘
∑
𝑘

 𝜔𝑖𝑘+1
⎝
𝑘
𝑘+1 ⎠ ⎝
𝑘
𝑘+1 ⎠
⎨ 𝑖=1
( ) 𝑘1 = 𝜔ℎ,1 + 𝜔ℎ,2 (4)
𝑘 ∑𝑛 𝑘
1

 𝑢𝑗 = 𝜔𝑗𝑘+1 𝑖=1 𝜔𝑖
𝑘+1
𝐾− 𝑘 𝑗 = 1, 2, . . . , 𝑛 (2) 𝑂ℎ,1 ∈Λ𝑙 𝑂ℎ,2 ∈Λ𝑙



 ( ) 𝑘+1 while the equivalent workload of all the operations is

 ∑𝑛 𝑘 𝑘

⎩ 𝑈 = 𝜔𝑗𝑘+1 𝐾− 𝑘
1
( 𝑘 ) 𝑘+1
𝑗=1 𝑘 𝑘
ΩΛ = ΩΛ𝑘+11
+ ⋅ ⋅ ⋅ + Ω 𝑘+1
Λ𝑚 (5)
and operations can be regarded as a single one with equivalent
( ) 𝑘+1 Given deadline 𝐾, the processing time of section Λ𝑙 can be
∑𝑛 𝑘 𝑘
calculated as
workload Ω𝑠 = 𝜔
𝑖=1 𝑖
𝑘+1
. When 𝑛 operations are to ( ) 𝑘
ΩΛℎ 𝑘+1
be processed in parallel, the optimal solution is 𝑝Λ𝑙 = 𝐾 (6)
Ω𝜎
⎧ and the processing time of the operations within the section is
⎨ 𝑝𝑗 = 𝐾 𝑗 = 1, 2, . . . , 𝑛
𝑢𝑗
1
= 𝜔𝑗 𝐾 − 𝑘 𝑗 = 1, 2, . . . , 𝑛 (3) ( ) 𝑘
⎩ ∑𝑛 𝜔1,𝑙 ΩΛℎ 𝑘+1
𝑈 =
1
−𝑘 𝑝1,𝑙 = 𝐾 𝑖+1≤𝑙 ≤𝑗 (7)
𝑖=1 𝜔𝑖 𝐾 Ω1,Λℎ Ω𝜎
( ) 𝑘
and operations can also be regarded as a single operation with 𝜔2,𝑙 ΩΛℎ 𝑘+1
∑𝑛 𝑝2,𝑙 = 𝐾 𝑖 ≤ 𝑙 ≤ 𝑗 − 1 (8)
equivalent workload Ω𝑝 = 𝑖=1 𝜔𝑖 . Ω2,Λℎ Ω𝜎
A job can at most have two operations, and they must be Therefore, the problem left is to determine the critical prece-
processed in sequence. Suppose resource has been optimally dence constraints that optimally divide operations into sec-
allocated to the operations. If the second operation of a job tions.
starts immediately after the first completes, the precedence In order to solve this problem, operations are transformed
constraint between them is named as a critical precedence into an acyclic directed graph, as Fig.1 shows. On the graph,
constraint. Otherwise, the constraint is named as slack prece- each node represents a precedence constraint between two
dence constraint, because the relaxation of the constraint has operations. Draw arcs between nodes. Each arc represents
no influence to the resource allocation. a section that is bounded by the precedence constraints
In the schedule, operations are divided into sections by represented by the emanating and sinking node of the arc.
critical precedence constraints. Within each section, there For example, let 𝐴𝑖,𝑗 be the arc between node 𝑖 and node
only exist slack constraints. Since slack constraints have no 𝑗. Assume both constraints are critical, and there exists no
influence to the resource allocation, the equivalent workload other critical constraint between them. Calculate the equivalent
method can be applied. For example, suppose operations are workload as well as the relative processing times of the
divided into sections Λ = {Λ1 , Λ2 , . . . , Λ𝑚 }, where Λ𝑙 is operations within section 𝐴𝑖,𝑗 according to Eq.4 and Eq.7 and
bounded by the precedence constraints of job 𝑖 and job 𝑗. Eq.8.
Because the calculation assumes the relaxation of the prece-
1) 𝑂𝑖,1 ≺ 𝑂𝑖,2 and 𝑂𝑗,1 ≺ 𝑂𝑗,2 : Λ𝑙 contains operations dence constraints, it is necessary to check the feasibility of
𝑂𝑖,1 ≺ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,1 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂𝑗,1 on machine 1, and the result. For every operation that has a predecessor also
operations 𝑂𝑖,2 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,2 ⪯ ⋅ ⋅ ⋅ ≺ 𝑂𝑗,2 on machine belonging to 𝐴𝑖,𝑗 , check their start time and completion time.
2; If none of these operations starts earlier before its predecessor
2) 𝑂𝑖,2 ≺ 𝑂𝑖,1 and 𝑂𝑗,1 ≺ 𝑂𝑗,2 : Λ𝑙 contains operations completes, the calculation is feasible, and let the equivalent
𝑂𝑖,1 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,1 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂𝑗,1 on machine 1, and workload be the length of the arc. Otherwise, the calculation
operations 𝑂𝑖,2 ≺ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,2 ⪯ ⋅ ⋅ ⋅ ≺ 𝑂𝑗,2 on machine is infeasible, and the arc is removed from the graph.
2; After all the feasible arcs are enumerated, there exist a
3) 𝑂𝑖,1 ≺ 𝑂𝑖,2 and 𝑂𝑗,2 ≺ 𝑂𝑗,1 : Λ𝑙 contains operations variety of paths from node 𝑈 to node 𝑆. Each arc represents
𝑂𝑖,1 ≺ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,1 ⪯ ⋅ ⋅ ⋅ ≺ 𝑂𝑗,1 on machine 1, and a feasible solution to the resource allocation problem, and the
operations 𝑂𝑖,2 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,2 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂𝑗,2 on machine equivalent workload of the solution is
2; ⎛ ⎞ 𝑘+1
4) 𝑂𝑖,2 ≺ 𝑂𝑖,1 and 𝑂𝑗,2 ≺ 𝑂𝑗,1 : Λ𝑙 contains operations ∑ 𝑘
𝑘

𝑘+1 ⎠
𝑂𝑖,1 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,1 ⪯ ⋅ ⋅ ⋅ ≺ 𝑂𝑗,1 on machine 1, and Ω=⎝ Ω𝑖,𝑗 (9)
operations 𝑂𝑖,2 ≺ ⋅ ⋅ ⋅ ⪯ 𝑂ℎ,2 ⪯ ⋅ ⋅ ⋅ ⪯ 𝑂𝑗,2 on machine 𝐴𝑖,𝑗 ∈𝑃
2.
where 𝑃 is a path from node 𝑈 to node 𝑆. Obviously, there
According to the equivalent workload method, operations exists at least one path that has the minimum equivalent
within section Λ𝑙 can be regarded as a single operation with workload. Since the directed graph is acyclic, this path can

2770 2018 13th IEEE Conference on Industrial Electronics and Applications (ICIEA)

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 19,2024 at 17:19:51 UTC from IEEE Xplore. Restrictions apply.
 A1,4
A1,3
1 A1,2 2 A2,3
Fig. 1. The directed graph that contains all the feasible sections of the operations.
be easily found using a dynamic programming method.
Based on the above discussion, the optimal resource allo-
cation algorithm is presented in the follows:
Algorithm 1. Optimal resource allocation
1) Enumerate precedence constraints in the schedule. Sup-
pose there are 𝑚 precedence constraints, let nodes
1, 2, . . . , 𝑚 represent these constraints. Let node 0 and
𝑚 + 1 represent the start and
2) For each 𝑗 = 1, 2, . . . , 𝑚 + 1, draw arcs 𝐴𝑖,𝑗 for
𝑖 = 0, 1, . . . , 𝑗 − 1. Calculate the relative processing time
of the operations in the arcs according to Eq.7 and Eq.8.
Check the feasibility of each arc by examining if no
operation starts earlier than its predecessor. If the arc is
feasible, calculate the equivalent workload according to
Eq.4, and let the value be the length of the arc. If the arc
is infeasible, remove the arc from the graph;
3) For each 𝑗 = 1, 2, . . . , 𝑛, calculate Ω𝑗 using the above
recursive equations.
4) Search for the shortest path from node 0 to node 𝑛, where
each arc on the path represents a section of the operations
in the optimal solution. Calculate processing time of the
operations using Eq.7 and Eq.8.
III. P ROPERTIES OF THE OPTIMAL SOLUTION
Some of the properties of the problem will be discussed
in this section. For convenience, operations are classified into
three subsets:
1) 𝒪1 : Operations that have a successor;
2) 𝒪2 : Operations that have a predecessor;
3) 𝒪3 : Operations that have no successor or predecessor.
Property 1. There exists an optimal schedule, where all the
operations in 𝒪1 are scheduled first on machine 1 and machine
2.
Proof. Suppose there is a schedule 𝒮1 where the processing
time of the operations are already optimized and known. It can
be observed easily, for any an operation in 𝒪1 , postponing its
successor will not expand the makespan as long as the machine
for the successor is kept busy. Therefore, a new schedule
𝒮2 can always be generated by having all the operations in
𝒪2 delayed until operations in 𝒪1 are finished. Since the
makespan of the schedule 𝒮2 is no larger than that of 𝒮1 ,
2018 13th IEEE Conference on Industrial Electronics and Applications (ICIEA)
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 19,2024 at 17:19:51 UTC from IEEE Xplore. Restrictions apply.
A3,n
3 A3,4 4 ... ... ... ... n
A4,n
A2,4
when resource allocation is optimized, the total consumption
will not be increased. The same conclusion also fits operations
in 𝒪3 , since they can be assumed to be successors of dummy
operations with processing time equal to 0. Therefore, the
property is proved.
Property 2. There exists an optimal schedule, where all the
operations in 𝒪3 are scheduled before operations in 𝒪2 on
machine 1 and machine 2.
Proof. The property can be proved by assuming every opera-
tion in 𝒪3 to be the predecessor of a dummy operation with
processing time equal to 0.
Property 3. There exists an optimal schedule, where op-
erations in 𝒪2 are scheduled in the same order as their
predecessors.
Proof. Suppose there are two operations, 𝑂𝑖,1 and 𝑂𝑗,1 pro-
cessed in sequence in schedule 𝒮1 , while their successors are
processed with 𝑂𝑗,2 ≺ 𝑂𝑖,2 . Generate a new schedule 𝒮2
by swapping 𝑂𝑗,2 and 𝑂𝑖,2 with processing time of all the
operations the same as 𝒮1 . It can be easily observed that the
makespan in schedule 𝒮2 is no larger than 𝒮2 . Therefore, when
the resource allocation of 𝒮2 is optimized, the total resource
consumption will be no larger than that of 𝒮1 . The property
is proved.
According to the above properties, there always exists an
optimal solution, where operations in 𝒪1 are scheduled first,
then operations in 𝒪3 , and finally operations in 𝒪2 in the same
order as their predecessors. Since operations in 𝒪3 have no
predecessor or successor, they can be processed in arbitrary
order. Therefore, the problem of scheduling all the operations
is simplified to scheduling operations with a successor.
Classify jobs with two operations into subsets 𝒥1,2 and 𝒥2,1 ,
where 𝒥1,2 contains jobs processed on machine 1 first, while
𝒥2,1 contains jobs processed on machine 2 first. Following
property shows only jobs in either one subset need be se-
quenced.
Property 4. In any a solution generated according to property
1-3, if not all the precedence constraints are slack, then jobs
either in 𝒥1,2 or 𝒥2,1 have no critical precedence constraint
(unless the job is the only member of the set).
2771
Proof. Consider a solution generated according to property
1-3. Suppose some jobs in 𝒥1,2 have critical precedence
constraints. Let job 𝑖 be the first processed job among them,
and let job 𝑗 be the last processed job in 𝒥2,1 . According to
the above properties, 𝑂𝑗,2 and other first operations of the jobs
in 𝒥2,1 are processed before 𝑂𝑖,2 , while their successors are
processed after 𝑂𝑖,1 . It can be easily observed, unless job 𝑗 is
the only member in 𝒥2,1 , and job 𝑖 is the last processed job
in 𝒥1,2 , then the completion time of the first operation of the
jobs in 𝒥2,1 must be strictly earlier than the start time of their
successors. Obviously, the same conclusion can also be drawn
when some jobs in 𝒥2,1 have critical precedence constraints.
The property is proved.
According to property 4, in any a solution generated by
above principles, precedence constraints have no influence
to the resource allocation of the jobs either in 𝒥1,2 or 𝒥2,1 .
Therefore, they can always be scheduled in arbitrary order.
As a result, the scheduling problem is further simplified
to sequencing the first operations of the jobs in 𝒥2,1 or
𝒥1,2 , respectively. Furthermore, since the successors of
these operations have the same processing sequence as their
predecessors, the problem is therefore equivalent to a job
sequencing problem.
IV. T HE HEURISTIC METHOD
In the classical two-machine job-shop scheduling problem
where processing time is constant, a so-called SPT(1)-LPT(2)
heuristic method is normally used to minimize the makespan.
For our problem, although the processing time is unknown, it
has a close relation to the workload of the operations. There-
fore, a similarly heuristic method named SWL(1)-LWL(2) is
proposed in this section, which can be stated as follows:
Algorithm 2. SWL(1)-LWL(2) method
1) Let 𝐽1,2 denote the set of the jobs that are processed on
machine 1 first, while 𝐽2,1 the set of the jobs processed
on machine 2 first. Jobs with only one operation are not
considered;
2) Sort jobs in 𝐽1,2 as follows: Jobs whose first operation has
smaller workload than the second operation are ordered
first, while others are ordered second. For the first group,
job whose first operation has smaller workload is order
first. For the second group, job whose second operation
has larger workload is order first;
3) Sort operations in 𝐽2,1 in the same manner;
4) Schedule the first operations of the jobs in 𝐽1,2 on ma-
chine 1 in their sorted order. Schedule the first operations
of the jobs in 𝐽2,1 on machine 2 in their sorted order;
5) Schedule jobs with only one operation on machine 1 and
machine 2 in arbitrary order;
6) Schedule all the left operations on machine 1 and machine
2 in the same order as their predecessors;
7) Calculate resource allocation of the operations.
Although the problem is in theory strongly 𝒩 𝒫-hard, the
SWL(1)-LWL(2) method is very likely to generate the optimal
2772 2018 13th IEEE Conference on Industrial Electronics and Applications (ICIEA)
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 19,2024 at 17:19:51 UTC from IEEE Xplore. Restrictions apply.
solution. In the schedule generated by this method, predeces-
sors and successors are normally separated with a ”distance”
between them, so the precedence constraints in most cases
have no influence to the resource allocation. If this is not
the case, the solution is normally close to the optimum, and
a branch and bound or a tabu-search algorithm can further
improves it.
V. C ONCLUSION
A two-machine jobshop scheduling problem with control-
lable processing times is studied in this paper. The objective
is to make sure all the jobs can be completed no later
than the deadline, while the necessary resource consumption
can be minimized. A polynomial time resource allocation
algorithm is presented first, which minimizes the resource
consumption under the condition that the processing sequence
of the operations is determined. After that, several properties of
the optimal schedule are discussed, and a heuristic algorithm
for optimal or near-optimal processing sequence is presented.
ACKNOWLEDGMENT
This research is supported by National Natural Science
Foundation of China (No.61203183).
R EFERENCES
[1] A. Janiak. ”Minimization of the blooming mill
standstills-mathematical model, suboptimal algorithms”,
Mechanika, vol 8, no. 2, pp. 37-49, 1989.
[2] J. Grabowski and A. Janiak, ”Job-shop scheduling with
resource-time models of operations”, European Journal
of Operations Research, vol. 28, no. 1, pp. 58-73, 1987.
[3] Janiak A and Szkodny T, ”Job-shop scheduling with con-
vex models of operations”, Methematical and Computer
Modelling, vol. 20, no. 2, pp. 59-68, 1994.
[4] Jansen K., Mastrolilli M., Solis-Oba R, ”Approximation
schemes for job shop scheduling problem with control-
lable processing times”, European Journal of Operations
Research, vol. 167, no. 2, pp. 297-319, 2005.
[5] R.L. Graham, E.L Lawler, J.K. Lenstra, A.H.G Rinnooy
Kan, ”Optimization and approximation in deterministic
sequencing and scheduling: a survey”, Annals of Discrete
Mathematics, vol. 5, no. 1, pp. 287-326, 1979. 287-326.
[6] Vickson RG. ”Choosing the job sequence and processing
times to minimize total processing plus flow cost on a
single machine”, Operations research, vol. 28, no. 5, pp.
1155-1167, 1980.
[7] Biskup D, Cheng T.C.E. ”Single-machine scheduling
with controllable processing times and earliness, tardi-
ness and completion time penalties”, Engineering Opti-
mization, vol. 31, no. 3, pp. 329-336, 1999.
[8] Shabtay D, Steiner G. ”A survey of scheduling with
controllable processing times”, Discrete Applied Mathe-
matics, vol. 155, no. 13, pp. 1643-1666, 2007.
[9] Kaspi M, Shabtay D. ”Convex resource allocation for
minimizing the makespan in a single machine with job
 release dates”, Computers and Operations Research, vol.
31, no. 9, pp. 1481-1489, 2004.
[10] Monma CL, Schrijver A, Todd, MJ, Wei VK. ”Convex
resource allocation problems on directedacyclic graphs:
duality, complexity, special cases and extensions”, Math-
ematics of Operations Research, vol. 15, no. 4, pp. 736-
748. 1990.
[11] Shabtay D, Kaspi M. ”Minimizing the total weighted
flow time in a single machine with controllable process-
ing times”, Computers and Operations Research, vol. 31,
no. 13, pp. 2279-2289, 2004.
[12] Shabtay D, Kaspi M, Steiner G. ”The no-wait two-
machine flow-shop scheduling problem with convex
resource-dependent processing times”, IIE Transaction,
vol. 39, no. 5, pp. 539-557, 2007.
[13] Xu K, Feng Z, Jun K. ”A tabu-search algorithm for
scheduling jobs with controllable processing times on a
single machine to meet due dates”, Elsevier Science Ltd.,
vol. 181, no. 1, pp. 303-324, 2010.
[14] Xu K, Feng Z, Ke L. ”A branch and bound algorithm for
scheduling jobs with controllable processing times on a
single machine to meet due dates”, Annals of Operations
Research, vol. 181, no. 1, pp. 303-324, 2010.
[15] Xu K, Feng Z, Ke L. ”Single machine scheduling with
total tardiness criterion and convex controllable process-
ing times”, Annals of Operations Research, vol. 186, no.
1, pp. 383-391, 2011.

2018 13th IEEE Conference on Industrial Electronics and Applications (ICIEA) 2773

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 19,2024 at 17:19:51 UTC from IEEE Xplore. Restrictions apply.