Sie sind auf Seite 1von 24

OR Spectrum (2013) 35:711–733

DOI 10.1007/s00291-012-0291-6

REGULAR ARTICLE

Reducing ergonomic risks by job rotation scheduling

Alena Otto · Armin Scholl

Published online: 4 April 2012


© Springer-Verlag 2012

Abstract The proverb “a change is as good as a rest” expresses one of the main
advantages of job rotation. In this article, we examine ways to set-up effective job
rotation schedules that balance ergonomic risks among workers. The practical rele-
vance of the problem is comprehensively discussed on examples from the automobile
industry. We present the ergonomic job rotation scheduling problem and show that
it is NP-hard in the strong sense. Therefore, the development of specialized solution
methods is important. Exploiting the problem structure, we propose a fast and effective
smoothing heuristic which can be integrated into solution methods for computing ini-
tial solutions and/or as a local re-optimization procedure. We find that integrating the
smoothing heuristic into a suited tabu search approach is particularly recommendable.
In computational experiments, this combination of approaches is able to solve almost
all instances of a practical data set in very short computation times of some seconds,
whereas the standard solver FICO Xpress and the best-known heuristic from literature
perform considerably worse concerning computation time and solution quality.

Keywords Scheduling · Mixed-model assembly lines · Manufacturing ·


Ergonomic risk assessment · Job rotation scheduling

1 Introduction

In the last years, the question of human well-being at the working place rocketed
to one of the key issues in production planning. On the one hand, firms got obliged
by law to keep workplace ergonomic risks, or risks for the health of workers due to

A. Otto · A. Scholl (B)


Chair of Management Science, Friedrich-Schiller-University of Jena,
Carl-Zeiß-Straße, 07743 Jena, Germany
e-mail: armin.scholl@uni-jena.de

123
712 A. Otto, A. Scholl

working environment, below the critical level (EU Machinery directive 2006/42/EC;
89/391/EEC; Occupational Safety and Health act of 1970 among others). On the other
hand, management of manufacturing firms realized that improvement of working con-
ditions reflects in higher productivity, less failures and in overall higher satisfaction
of workers, higher company loyalty and less absenteeism (e.g., Eklund 1995; Moreau
2003; Falck et al. 2010).
The important factor here is that ergonomic risks depend not only on the design and
conditions of the workplace, but on the content of work, repetitivity and duration of
certain tasks, in particular of those requiring awkward postures or application of forces.
Therefore, assembly line balancing (Otto and Scholl 2011), mixed-model sequencing
and job rotation scheduling can lead to a significant improvement of worker well-being
and, thus, to prevention of work-related diseases, enhancement of job satisfaction and
improvement in productivity of work (Neumann et al. 2006; Lodree et al. 2009). Field
studies on job rotation widely confirm these results. Thus, in their study of job rotation
in a refuse collecting department, Kuijer et al. (1999) found a positive effect of job rota-
tion on the perceived, energetic and postural load. The follow-up study (Kuijer et al.
2005) confirmed that job rotation reduces the need for recovery, but it was inconclusive
about the risk of low back complaints due to possible selection effects in assigning
workers to the groups. In the field study of Rissén et al. (2002), supermarket cash-
iers positively evaluated the effect of job rotation in self-ratings. Furthermore, careful
laboratory experiments (Raina and Dickerson 2009; Keir et al. 2011) confirmed that
job rotation balances the exposure to risks among workers, whereby the degree of the
observed effects depends on the muscle groups involved.
Job rotation has become one of the most widespread organizational routines. This
happened, not to the least extent, due to its ability to balance ergonomic risks swiftly at
low or even no costs. The National Survey of Establishments 1997 targeting American
for-profit organizations found that job rotation was adopted by 55.5 % of enterprises
and this number is even growing (Osterman 2000). In particular, it has been recog-
nized in different industries that smoothing ergonomic risks among workers is one of
the major impacts of job rotation. Tharmmaphornphilas and Norman (2004) under-
line an utmost importance of an ergonomic objective in a saw mill. Jorgensen et al.
(2005), in a survey of job rotation in the Midwest US manufacturing firms, found the
factor “decrease of incidence/symptoms of work-related injuries” (ergonomic risks)
among the two most important perceived benefits from implementing the job rotation
principle.
The ergonomic job rotation scheduling problem (EJRSP), which aims at smoothing
ergonomic risks between workers by minimizing the ergonomic load for the worker
most exposed to ergonomic risks, is introduced by Carnahan et al. (2000). These ideas
are further extended by Tharmmaphornphilas and Norman (2007) who persuasively
show the need for specialized effective algorithms for EJRSP, since standard software is
not able to solve instances of real-world sizes (e.g., 12 workers and 8 rotation periods).
Carnahan et al. (2000) propose a genetic algorithm to solve EJRSP and use a neuronal
network algorithm to find promising rules for job rotation. Tharmmaphornphilas and
Norman (2007) develop a greedy heuristic for EJRSP. Further, Diego-Mas et al. (2009)
describe the software DPI-ASEPEYO for calculating job rotation schedules based on
a genetic algorithm. Costa and Miralles (2009) integrate job rotation scheduling into

123
Reducing ergonomic risks by job rotation scheduling 713

the assembly line balancing problem. Seçkiner and Kurt (2007, 2008) implement ideas
of EJRSP to formulate a shift assignment problem with an ergonomic objective, which
they solve with ant colony optimization and simulated annealing algorithms.
However, so far, no reflections on the complexity of EJRSP and on the problem
structure are published. In our article, we address these questions in detail and propose
effective solution procedures that exploit the problem structure.
To narrow the existing gap between theory and practice, we also provide detailed
descriptions of the relevant problem settings that arise in enterprises, especially in
automobile industry where job rotation is highly relevant. It is necessary to exam-
ine state-of-the-art methods for measuring ergonomic risks as those are required for
defining the ergonomic parameters and the objective function of the EJRSP. In the lit-
erature, so far, the ergonomic aspects of EJRSP have not been considered sufficiently
such that practitioners do not consider the theoretical achievements as applicable in
real-world settings.
The remainder of the paper is organized as follows. We describe job rotation sched-
uling routines, common in the automobile and related industries, in Sect. 2. In Sect. 3,
we present the ergonomic job rotation scheduling problem and reflect on its com-
putational complexity. Section 4 provides solution methods that exploit the problem
structure, smoothing heuristic and tabu search. The design and the results of a com-
prehensive computational experiment comparing new and former solution approaches
are presented in Sect. 5. We conclude with an outlook on future research directions in
Sect. 6.

2 Practices of job rotation scheduling

In this section, we combine a literature overview with insights from our cooperation
projects and interviews with our partners from automobile industry.

2.1 Job rotation in the automobile industry

Job rotation, or rotation of workers between different jobs, is usually performed within
a team. A team is the smallest and rather stable organizational unit and it usually con-
sists of 6–20 members. Each team is managed by a team supervisor and the members
have, as a rule, the same shift schedule and meet each other regularly to discuss quality
issues. Because of high flexibility requirements in automobile industry, especially at
assembly lines, each team member has qualifications to work at any workplace of the
team’s line segment (see Boysen et al. 2009a, for a detailed description of the assembly
line production system and related decision problems).
For the reasons discussed in Sect. 1, the total workload of the team is distributed
among the team members by systematically applying the principle of job rotation, i.e.,
workers rotate between the different workplaces. The terms “workplace” and “job”
are synonyms for the job rotation scheduling; a job is defined by the work content
of a workplace. All the operations that have to be performed by workers, including
their times and constituent movements of workers, are exactly analyzed, e.g., based
on the methods-time measurement (MTM) system (Longo and Mirabelli 2009), and

123
714 A. Otto, A. Scholl

Table 1 Illustrative example of


Rotation period Worker 1 Worker 2 Worker 3
job rotation schedule
Period 1: 8:00 to 10:00 a1 a2 a3
Period 2: 10:20 to 13:40 a3 a1 a2
Period 3: 14:20 to 17:00 a2 a3 a1

documented in the firm’s information systems. Therefore, it is common practice and


justified to assume deterministic workloads and deterministic ergonomic risks for each
worker at each workplace. Usually, rotation schedules of a team are manually com-
posed by the team supervisor for the next shift, week or even month, who assigns
workers to jobs in each rotation period.
Table 1 shows an illustrative example problem which represents a team of three
workers who have to operate three workplaces a1 , a2 and a3 , i.e., to perform three
different jobs, within an 8-h shift. To simplify presentation, the shift is subdivided
into three rotation periods only, here defined as time spans between two main breaks.
In each of the three periods, each worker is assigned to one of the three jobs. For
example, worker 1 does job a1 in period 1, job a3 in period 2, and job a2 in period 3.
In each period, every job is performed by one of the workers.
Often, a team has a certain degree of freedom in choosing its frequency of rotation.
Typically, rotation is performed during the major or minor pauses, once in 1 or 2 h or
even every 30 min, especially, when the work load is very demanding.

2.2 Estimation of ergonomic risks

Ergonomic risk estimation methods evaluate risk factors present in tasks that are to be
performed by a worker, such as awkward postures, application of forces or repetitive
actions (Otto and Scholl 2011). As a rule, a higher ergonomic risk estimate, measured
as a coefficient or as ergonomic points, means higher risks present. Higher ergonomic
risks imply a higher probability of getting occupational diseases, as well as a more
severe form of its’ expression. Therefore, higher ergonomic risks imply an exponential
increase of costs (Tharmmaphornphilas and Norman 2007).
Ergonomic risks estimation methods calculate ergonomic risks on a shift-basis.
This is motivated by the reasonable assumption that workers fully regenerate after a
sufficiently long break (e.g., night), so that the ergonomic workloads of 2 days are inde-
pendent of each other. As the exact way of measuring ergonomic risks is not essential
for job rotation scheduling, we assume, w.l.o.g., that ergonomic risks are measured by
a widespread method in the European automobile industry called European assembly
worksheet (EAWS; cf. Schaub et al. 2010). For a survey of this and other methods,
which can alternatively be applied, see Otto and Scholl (2011).
As already mentioned, (the whole-body section of) EAWS calculates ergonomic
points (EP) for each workplace (or job) assuming that the same worker performs this
job all over the shift. The workplaces are classified according to a traffic light system
based on the measured shift-related EP values, e.g., as follows: green = low ergonomic
risk (2–25 EP), yellow = potential risk (26–50 EP), red = high risk (more than 50 EP).

123
Reducing ergonomic risks by job rotation scheduling 715

Table 2 Estimations of
Rotation period Period-specific EP: et (a j )
ergonomic risks for the example
problem Workplace a1 Workplace a2 Workplace a3

Period 1: 120 min 4 16 20


Period 2: 200 min 8 24 20
Period 3: 160 min 4 16 20

In case of job rotation, however, each worker may perform several different jobs
within one shift. A common approach in practice is to estimate ergonomic risks as a
time-weighted average of the respective ergonomic points for the different jobs (see
the presentation of MTMErgonomics tool, which is based on EAWS method on www.
dmtm.com). Tharmmaphornphilas and Norman (2004) describe this type of aggrega-
tion based on the so-called Job severity index, another risk measurement method for
lifting operations. Following this logic, we define period-specific ergonomic points
et (a j ) of any job a j in a period t, which are already weighted by the duration of
period t relative to the entire shift. For example, if job a1 has a measured EP value of
16 for an 8-h shift, it gets the period-specific value e1 (a1 ) = 4 if period 1 lasts 2 h
(see Table 2).
Besides the different durations of rotation periods, there is often a variation of ergo-
nomic risks from one period to the other. Moreover, the composition of operations to be
completed on the same workplace may be different from period to period. An example
from the automobile industry would be a workplace at a mixed-model assembly line
that combines mounting tail lamps (usually a task with lower risks) with mounting a
sun roof (usually a task with high ergonomic risks as it requires work over-shoulder or
even overhead). The sun roof is an optional feature of the car. So, the portion of cars
with sun roofs might change from one rotation period to another depending on the
(customer-driven) sequence of car models on the assembly line (Boysen et al. 2009b).
Reflecting this, the extent of ergonomic risks at the considered workplace may change
significantly in consecutive rotation periods.
In practice, such changes in the composition of operations between rotation peri-
ods are, due to a lack of suited methodology, usually not taken into consideration in
rotation plans. Instead, the estimation of ergonomic risks on a workplace is done on
basis of an average composition of tasks, e.g., it is assumed that an average car is
manufactured at an assembly line during the entire shift. Nevertheless, as this seems
to be a deficit and to define a more general and precise model, we allow ergonomic
points to be dynamic, i.e., the period-specific values vary from period to period not
only due to different period lengths but also due to changing model-mix.
Let us reconsider the example of Table 1 with three workplaces and three rotation
periods which is complemented by the dynamic, period-specific EP values in Table 2.
Assuming workplace 1 being the one mounting tail lamps and sun roofs, the portion
of customer orders with sun roofs in the model-mix is obviously larger in period 2
(smaller in period 3) compared to period 1. If the model-mix was identical to that
of period 1 in the other periods, we would get 4 · 200 120 = 6.67 EP in period 2 and
4 · 120 = 5.33 EP in period 3.
160

123
716 A. Otto, A. Scholl

Table 3 Different rotation schedules and their ergonomic risks

Rotation schedule S0 : Rotation schedule S1 : Rotation schedule S2 :


workplace / period-spec. EP workplace / period-spec. EP workplace / period-spec. EP

Worker 1 Worker 2 Worker 1 Worker 1 Worker 2 Worker 3 Worker 1 Worker 2 Worker 3

Period 1 a1 / 4 a2 / 16 a3 / 20 a1 / 4 a2 / 16 a3 / 20 a1 / 4 a3 / 20 a2 / 16
Period 2 a1 / 8 a2 / 24 a3 / 20 a2 / 24 a3 / 20 a1 / 8 a2 / 24 a3 / 20 a1 / 8
Period 3 a1 / 4 a2 / 16 a3 / 20 a2 / 16 a1 / 4 a3 / 20 a2 / 16 a1 / 4 a3 / 20
PEP 16 56 60 44 40 48 44 44 44
green red red yellow yellow yellow yellow yellow yellow

Table 3 shows and evaluates three feasible rotation schedules. We calculate the
personal ergonomic points (PEP), which measure the individual ergonomic risks a
worker is exposed to during executing his/her jobs during a shift, as described above.
For example, according to rotation schedule S2 , worker 1 spends period 1 at work-
place a1 with e1 (a1 ) = 4 EP and period 2 and 3 at workplace a2 with e2 (a2 ) = 24 and
e3 (a2 ) = 16 EP. Then, we get 4 + 24 + 16 = 44 PEP for worker 1 as given in the last
row of Table 3. This value corresponds to the yellow zone, i.e., possible ergonomic
risks are present but moderate. Also, as we see in this example, the rotation schedule
may significantly influence the distribution of ergonomic risks. Rotation schedule S0
results in “green” ergonomic conditions for worker 1 and in “red” conditions for work-
ers 2 and 3. To the contrary, rotation schedule S2 evenly distributes ergonomic risks
and leads to “yellow” conditions for all three workers. Note that regardless of the
rotation schedule, the sum of ergonomic points for a team remains the same—in this
example it equals to 132 points.
Though workers have different physical abilities and, thus, could cope with dif-
ferent levels of ergonomic risk, it is common practice in the automobile industry not
to consider individual ergonomic risks or individual “traffic light” categories. This
is mainly due to legal requirements which claim to avoid discrimination of workers,
especially at assembly lines (e.g., EN 614-1). So, we assume worker-invariant ergo-
nomic risk estimation in our basic model described in Sect. 3. However, as some risk
estimation methods utilize scales that differ for different sexes (e.g., EAWS in man-
ual handling for men and women, see Schaub et al. 2010), for state of health (e.g.,
Ergo-FWS in Sinn-Behrendt et al. 2004) and for fitness (e.g., Siemens method, see
Bokranz and Landau 2006) and to keep our approach general, we additionally show
how to integrate individual ergonomic risks into the model.

3 Ergonomic job rotation scheduling problem (EJRSP)

In the following, we first describe a basic job rotation scheduling problem which
assigns jobs to workers such that the ergonomic risks measured in EP are reduced and
balanced between the workers. Then, we show that this problem is NP-hard in the
strong sense.

123
Reducing ergonomic risks by job rotation scheduling 717

3.1 Formulation of EJRSP and its peculiarities

The basic ergonomic job rotation scheduling problem (EJRSP) can be formulated
as follows (cf. Carnahan et al. 2000). A set J of n workplaces (defining n jobs)
a j ∈ J, j = 1, . . . , n, and a number of periods T ∈ Z+ are given. Jobs are assigned
to workers, which we express by a permutation function πt (i) meaning that in period
t = 1, . . . , T worker i = 1, . . . , n has to perform job aπt (i) . Each job a j is character-
ized by period-specific ergonomic risk points ε jt = et (a j ) > 0 as defined in Sect. 2.2.
Without loss of generality, we assume ergonomic points to be integral. 
Theobjective is to find a set ofT permutations of jobs aπ1 (1) , aπ1 (2) , . . . , aπ1 (n)
, . . . , aπT (1) , aπT (2) , . . . , aπT (n) such that the maximal (shift-related)
 risk exposure
T
E of any worker, i.e., E = max t=1 et (aπt (i) ) |i = 1, . . . , n , is minimized.
A binary linear program representing EJRSP can be formulated in the following
way:

Minimize E (1)

subject to


n
xi jt = 1 ∀ j ∈ {1, . . . , n} , ∀t ∈ {1, . . . , T } (2)
i=1
n
xi jt = 1 ∀ j ∈ {1, . . . , n} , ∀t ∈ {1, . . . , T } (3)
j=1


T 
n
ε jt · xi jt ≤ E ∀i ∈ {1, . . . , n} (4)
t=1 j=1
xi jt = {0, 1} ∀i ∈ {1, . . . , n} , ∀i ∈ {1, . . . , n} , ∀t ∈ {1, . . . , T } (5)

Job rotation schedules are defined by the binary decision variables xi jt (see constraint
5), which show whether worker i is assigned (xi jt = 1) to workplace j in period t or
not (xi jt = 0). Constraints (2) and (3) ensure that each worker is assigned to exactly
one workplace (job) in each period. The left-hand side of each constraint (4) defines
the ergonomic points of a worker i (PEP) as the sum of (time-weighted) period-specific
EP of the assigned jobs. In combination with (4), the objective function (1) defines E
as the highest risk any worker has to face with and minimizes this value. As the sum
of all ergonomic points over jobs and periods is given by the production program and,
thus, is constant, the highest risk value E can only be lowered if ergonomic points
of other workers will be increased. So, the objective function contributes to balanc-
ing the risks of all workers in the schedule. This is very important in practice, since
higher ergonomic risks imply exponentially increasing costs. Furthermore, a balanced
schedule is more equitable within a team. The model defined above is seen as a basic
model because it describes the situation found in most industries most adequately (cf.

123
718 A. Otto, A. Scholl

the discussion in Sect. 2.2). This model will be denoted as EJRSP-VT (with VT for
varying tasks). Two other versions of the model can also be defined:
• EJRSP-Ind Ergonomic points are dynamic and individual (Ind), i.e., we need to
define period-specific ergonomic points εi jt which measure the ergonomic risk of
worker i for performing job j in period t. As mentioned in Sect. 2.2, individual
risk estimations are not desired in practice, but can be simply modeled by replacing
ε jt with εi jt in constraints (4).
• EJRSP-S Simple or static EJRSP with equidistant periods and static (time-invariant)
ergonomic points, modeled by ε j = ε jt ∀t.
Although EJRSP is similar to other well-known problems, it has some differing char-
acteristics such that solution procedures for these related problems cannot be applied
directly. EJRSP is different from the job rotation problem of Butković and Lewis
(2007), since it cannot be expressed in an assignment matrix with constant transition
costs between jobs. Furthermore, EJRSP is similar to the parallel machine problem
(see, e.g., Cheng and Sin 1990) with workers defining machines that have to perform
successive jobs and the objective of minimizing the maximal ergonomic load of any
worker (makespan). However, EJRSP shows special (block-wise) “precedence” rela-
tions between the jobs. Namely, all the jobs of the first period have to be performed
by exactly one worker each, before any job of the next period becomes available.

3.2 On the complexity of EJRSP

Here we show that EJRSP-VT is NP-hard in the strong sense. That means that even in
case of small or moderately differing ergonomic points, there exist instances requiring
computational time growing exponentially in the number of workplaces and periods.
This statement has far-reaching consequences. For example, even small instances from
practice with four periods and 15 workplaces cannot be solved by standard MIP soft-
ware within 1-h time limit (as shown in our computational experiments in Sect. 5).
Nevertheless, as we show in our experiments, the application of suited heuristic solu-
tion procedures allows finding optimal or near-optimal solutions within a few seconds
(see Sect. 5.2).

Theorem 1 EJRSP-VT is NP-hard in the strong sense.


We prove the statement by showing that the decision version of EJRSP-VT is NP
complete in the strong sense. Given a maximal  risk level E, this decision problem  asks,
whether there exist T sequences of jobs aπ1 (1) , aπ1 (2) , . . . , aπ1 (n) , . . . , aπT (1) ,
 T
aπT (2) , . . . , aπT (n) such that maxi t=1 et (aπt (i) ) does not exceed E. This problem
is shown to be NP complete in the strong sense, because it is, obviously, in NP and
the 3-partition problem (3PP) can be polynomially transformed into it.

3-Partition problem (3PP) B is a finite set of 3m elements bt (t = 1, . . . , 3m) A ∈ Z+


is a bound and s(bt ) ∈Z+ are the sizes of the elements bt such that each s(bt ) satisfies
3m
4 < s(bt ) < 2 and t=1 s(bt ) = m
· A. Find whether B be partitioned into m dis-
A A

joint sets B1 , B2 , . . . , Bm such that bt ∈B j s(bt ) = A ∀ j = 1, . . . , m. This problem


is known to be NP complete in the strong sense (Garey and Johnson 1979).

123
Reducing ergonomic risks by job rotation scheduling 719

Table 4 Transformation of
Rotation schedule
3PP-solution to EJRSP-VT
Job / weighted erg. points

Worker 1 Worker 2

Period 1 b1 / 5 f1 / 1
Period 2 b2 / 25 f2 / 1
Period 3 f3 / 1 b3 / 5
Period 4 f4 / 1 b4 / 21
Period 5 b5 / 17 f5 / 1
Period 6 f6 / 1 b6 / 5
Total EP 50 50

Polynomial transformation of 3PP into EJRSP-VT We construct the following


transformation from any instance of 3PP to an instance of (the decision version of)
EJRSP-VT, recalling that ergonomic risk estimates are dynamic but not individual,
given as ergonomic points et (a j ) ∈ Z+ ( j = 1, . . . , n, t = 1, . . . , T ). Let there be
T = 3m periods and n = m jobs. The first job has ergonomic points et (a1 ) = s(bt )+1
in all periods t = 1, . . . , 3m. The other (m − 1) jobs are “technical fillers” with iden-
tical ergonomic points in each period: et (a j ) = 1 ∀ j = 2, . . . , m, ∀t = 1, . . . , T.
Finally, let E = A + 3m.
Below, we show that each solution of a 3PP-instance corresponds to a solution of
the constructed EJRSP-VT and vice versa.
Let m disjoints sets B1 , B2 , . . . , Bm be a solution of 3PP. We build the correspon-
dent solution of the constructed EJRSP-VT in the following way. Assign job a1 to the
first worker in the three periods t for which bt ∈ B1 holds. In those periods, (arbi-
trarily) assign the remaining jobs (or “technical fillers”) j = 2, . . . , m to the remaining
workers. We repeat this step for each further set Bs (s = 2, ..., m) by assigning job
a1 to the worker s in the three periods t for which bt ∈ Bs and filling the remaining
assignments with “technical fillers”. In the constructed solution of EJRSP-VT, the
ergonomic points for each worker exactly equal to E = A + 3m.
Consider an example of a 3PP instance with 6 elements, bound 44 and sizes 4, 24,
20, 20, 16, 4. Let a known solution of this instance be {{b1 , b2 , b5 } , {b3 , b4 , b6 }} . The
correspondent EJRSP-VT has T = 6 periods, m = 2 jobs, ergonomic points of the
first job equal to 5, 25, 21, 21, 17, 5 in periods 1 to 6 and ergonomic points of the
second job equal to 1 in each period, and E = 44 + 6 = 50. The corresponding
solution is shown in Table 4 with bt representing job 1 and f t denoting the filler job 2
in period t.
In its turn, from each solution of the constructed EJRSP-VT instance a solution for
the 3PP-instance can be constructed.
Further, the decision problem is NP complete in the strong sense, because the results
hold even if the constructed EJRSP-VT instance is not allowed to have very large val-
ues for the input data, i.e., ergonomic points. More precisely, the  expression
for the
maximum number of the constructed EJRSP-VT instance Max I ejr sp is polynomi-

ally bounded by its length Length I ejr sp , because the same is true for 3PP (Garey
and Johnson 1979). 


123
720 A. Otto, A. Scholl

4 Heuristic solution procedures

EJRSP has a very specific structure, exploiting of which may significantly enhance
performance of exact and heuristic solution methods.

4.1 Naïve construction procedure

Assume that worker-job assignments are given for periods 1, . . . , t − 1. Then, a very
simple and intuitive procedure for assigning the jobs to the workers in period t, called
one-period assignment (OPA), works as follows: if a worker in the previous periods
has had the highest exposure to ergonomic risks, then he/she will be assigned to the
job with the lowest ergonomic risks in the current period (ties are broken arbitrarily).
Indeed, this logic is often used by team supervisors to construct the job rotation
schedules for their team period-by-period. This construction procedure shall be called
naïve construction procedure and can be formalized as follows.
Naïve construction procedure (NCP) In period 1, arbitrarily assign the jobs to work-
ers. W.l.o.g., worker i receives job i, or π1 (i) = i. Then, iteratively apply OPA for
the periods t = 2, . . . , T, i.e., for each t consider workers in non-increasing order of
 t−1
τ =1 eτ (aπτ (i) ) and assign the jobs j in ascending order of et (a j ) to these workers
(cf. the similar approach of the RLPT rule for the parallel machine scheduling problem
in Walter 2011).
For our example of EJRSP-VT with T = 3 periods and n = 3 jobs shown in rota-
tion schedule S1 of Table 3, we assign worker 1 to job 1 (4 EP), worker 2 to job 2 (16
EP) and worker 3 to job 3 (20 EP) in the first period. In period t = 2, we assign the
“easiest” job 1 (8 EP) to worker 3 with the highest ergonomic load, job 3 (20 EP) to
worker 2 and job 2 to worker 1 (24 EP). Now, the (weighted) ergonomic loads of the
workers are 4 + 24 = 28, 36 and 28, respectively. Therefore, in period t = 3, job 1 (4
EP) is assigned to worker 2 with the highest cumulated ergonomic load up to now, job
2 (16 EP) to worker 1 and job 3 (20 EP) to worker 3. The job assignment of workers
is {a1 , a2 , a2 } , {a2 , a3 , a1 }and {a3 , a1 , a3 } and the shift-related ergonomic loads are
44, 40 and 48, respectively. Thus, the objective function value equals to 48, which is
the maximal ergonomic load of a worker. This schedule S1 is, obviously, much better
than the schedule without job rotation S0 . Still, S1 is a suboptimal solution and further
improvement is possible.

4.2 Smoothing heuristic

The procedure starts with applying NCP to construct an initial feasible solution (phase
1), which is then improved by iteratively applying the greedy procedure OPA (phase
2). In opposite to its application as a part of the construction procedure NCP, we now
have a complete feasible solution, i.e., worker-job assignments exist for all periods
t = 1, . . . , T. Nevertheless, we might release the assignments for some period t and
apply OPA to re-assign workers to jobs for this very period. Provided that the assign-
ments of the other periods are already fixed, OPA leads to the best possible solution.
To be more specific and formal, the smoothing heuristic (SH) works as follows.

123
Reducing ergonomic risks by job rotation scheduling 721

Phase 1: initial solution Apply NCP to compute an initial feasible solution which
is stored as incumbent (currently best) solution.
Phase 2: local improvement procedure Starting with the incumbent solution, apply
OPA successively to the periods t = 1, . . . , T and call this an improvement step. If
the objective value has been improved, then the current schedule is stored as the new
incumbent solution. Repeat performing improvement steps until a stopping criterion
is fulfilled. The incumbent solution defines the resulting schedule.
Stopping criteria

• If the schedule has not changed at all, then terminate the procedure as further steps
will not change the solution anymore.
• To avoid being caught into the permanent re-visiting of the same solutions (cycling
of the search), we allow maximally ten consecutive improvement steps without
improving the value of the objective function. Note that changing assignments need
not change the objective function value.

Example Consider our example EJRSP-instance in Tables 2 and 3. The initial solu-
tion received by NCP in the first phase is given by schedule S1 in Table 3. The partial
ergonomic loads of the workers without the assignment of period t = 1 are e2 (a2 ) +
e3 (a2 ) = 24 + 16 = 40 for worker 1, as well as 24 and 28 for workers 2 and 3, respec-
tively. OPA is applied to re-assign the jobs of the first period. Worker 1 receives job 1,
worker 3, job 2 and worker 2, job 3. We get the schedule S2 of Table 3 with worker-job
assignment {{a1 , a2 , a2 } , {a3 , a3 , a1 } , {a2 , a1 , a3 }} and maximal ergonomic load of
44. As all workers have this maximal ergonomic load, the solution is optimal.

Extensions The described smoothing heuristic can be used without modifications to


solve the simple EJRSP with period-invariant ergonomic points (EJRSP-S).
In case of EJRSP-Ind, we replace the simple OPA procedure by solving a linear
bottleneck assignment problem (LBAP) which is known to be polynomially solvable
2.5
in O( √ n ) (Burkard et al. 2009). The objective function of LBAP minimizes the
log(n)
maximal “costs” among the worker-job assignments. We construct LBAP for a period
t by setting the number of jobs and workers to n and the costs ci j of the assignment
of worker i to job j to the sum of the given current ergonomic workload of worker
i and the ergonomic points of job j in the period t, i.e., ci j = τ =t eiτ (aπτ (i) ) +
eit (a j ).
For example, let us have an instance of EJRSP-Ind with T = 3 periods, n = 2
jobs, individual ergonomic points, and the current schedule S1 as shown in Table 5.
The cumulated workload of workers, when skipping the assignment in period 1, is 12
and 28, respectively. Hence, for the LBAP, the costs of the first worker to be assigned
to the first job are c11 = 12 + e11 (a1 ) = 16, costs of the second worker to be
assigned to the first job are c12 = 28 + e21 (a1 ) = 40 and so on. The cost matrix
of the LBAP for the first period is given as matrix C in Table 5. To minimize the
maximal ergonomic load, we have to assign job a 1 to worker 2 and job a 2 to worker
1 in period 1. The received (here: optimal) schedule has a maximal ergonomic load of
40 EP.

123
722 A. Otto, A. Scholl

Table 5 Illustration of the LBAP for period t = 1

Period-specific ergonomic points of Period-specific ergonomic points of


worker 1 worker 2

Workplace a1 Workplace a2 Workplace a1 Workplace a2

Period 1 4 16 Period 1 12 16
Period 2 8 24 Period 2 12 12
Period 3 4 16 Period 3 12 16
Assignment S1 : Re-assignment of period 1
workplace/ergonomic points (cost matrix C for the LBAP)
Worker 1 Worker 2 Worker 1 Worker 2
Period 1 a1 / 4 a2 / 16 Workplace a1 16 40
Period 2 a1 / 8 a2 / 12 Workplace a2 28 44
Period 3 a1 / 4 a2 / 16
Total EP 16 44

4.3 Tabu search heuristic

The smoothing heuristic described above results in a locally optimal solution. To over-
come local optimality, sophisticated meta-heuristics are available. We propose a rather
straight-forward tabu search (TS) procedure (cf. De Werra and Hertz 1989; Glover
1990) well-suited for permutation-based problems like EJRSP and easy to implement
in decision support software of car manufacturers. It performs simple swap moves
and utilizes a static tabu list. The smoothing heuristic is applied to provide the initial
solution and as a local re-optimization procedure within TS (see below).
Neighborhood definition The neighborhood consists of all the feasible job rota-
tion schedules that can be reached from the current solution by a swap of two jobs
between two workers within the same period, involving the worker having currently
the highest risks. The total number of pairs in the neighborhood that has to be exam-
ined for possible swaps is (n − 1)T if there is one worker with the highest workload
and never exceeds n(n−1)2 T, since we can maximally have (n − 1) workers with the
highest workload in a suboptimal solution. All swaps lead to feasible solutions. For
example, in case of the current assignment shown in the matrix S1 of Table 3, the
neighborhood solutions can be achieved by swapping the jobs of worker 3 with one
of the other two workers within three periods, such that there are six swaps defining
the neighborhood.
Each time, the best solution from the neighborhood, i.e., the one with the low-
est maximal ergonomic load of workers, becomes the current solution; if this low-
ers the objective value, the improved solution becomes the new incumbent solution.
First-level ties are broken by selecting a solution with the minimal standard devia-
tion of ergonomic loads. By this means we prefer solutions, where the ergonomic
loads of the remaining workers are distributed more equally. Second-level ties are

123
Reducing ergonomic risks by job rotation scheduling 723

broken randomly. Due to the permutation character of the problem, this situation is
not atypical, so that the designed TS has a stochastic component.
Note that another meaningful kind of tie-breaker would be to prefer solutions with
the maximal standard deviation of ergonomic loads as the first-level tie-breaker. The
rationale behind this is that the algorithm may receive more degrees of freedom on fur-
ther steps to shift points away from the bottleneck worker. However, our preliminary
tests of these two kinds of tie-breakers did not reveal any significant changes in the
results, both in terms of speed and solution quality. Therefore, we (arbitrarily) restrict
to the former one in our experiments of Sect. 5.
Tabu list After performing a swap, we save the period and the jobs that have been
swapped in the tabu list. We prohibit swaps between these jobs in the given period for
TL number of iterations, where TL is the length of the tabu list.
Such an approach is very restrictive and prohibits not only cycling to the already
visited solutions, but it also restricts the ability to examine not yet visited neighbor-
hood solutions. However, experiments with less restrictive neighborhoods (including
storing complete solutions and, thus, prohibiting the return to the visited solution only)
showed worse results. It can be explained, that this approach supports a diversification,
i.e., considerably differing parts of the solution space are explored.
Local re-optimization procedure Periodically, after some swaps, we applied the
local improvement procedure from step 2 of the smoothing heuristic to the current
solution to perform a local re-optimization. If an improved solution is achieved, it is
saved as new incumbent. This re-optimization is a means of intensification, i.e., to
consider the promising parts of the solution space more intensively.
In some preliminary computational experiments, we tested two strategies: the “inva-
sive” local improvement, where the locally improved solution updates the current
solution for the next iteration of TS, and the “non-invasive” local improvement which
considers locally improved solutions only with respect to improving the incumbent.
The tests showed that there was no dominance of one strategy over the other. So, we
include the “non-invasive” strategy as it seems to provide a more diversified search.
Termination criterion We stop executing our tabu search procedure as soon as either
the objective value of the best solution found equals to the lower bound or if a maximal
number of iterations is achieved. We employ the simple LP-relaxation lower bound of
the binary linear program described in Sect. 3.1.

5 Computational experiments

We conduct three experiments. In the first one, we examine the performance of the
proposed naïve construction procedure (NCP), smoothing heuristic (SH) and the tabu
search procedure (TS) for the EJRSP-VT, the problem setting that is common in
automobile industry. We also compare it to standard software FICO Xpress and to
the best reported and tested solution method proposed for EJRSP until now, which
is the greedy heuristic with random diversification procedure (GRD) proposed by
Tharmmaphornphilas and Norman (2007). After that, in our second experiment, we
investigate the main drivers of the performance of our tabu search procedure and per-
form a sensitivity analysis. In experiment 3, we look at the performance of the proposed

123
724 A. Otto, A. Scholl

solution methods in further settings of EJRSP, namely EJRSP-S and EJRSP-Ind. As


discussed in Sect. 2, EJRSP-S refers to a situation with homogenous products assem-
bled at the line and equal rotation periods. EJRSP-Ind allows for worker-specific
ergonomic points. Although due to possible discrimination issues, it is not allowed to
use EJRSP-Ind for assembly lines in Europe, this model may be appropriate for special
industries in some countries (e.g., for a saw mill in US, cf. Tharmmaphornphilas and
Norman 2004).
GRD starts with a randomly generated feasible initial solution. It employs the same
definition of the neighborhood as TS and moves to a solution in the neighborhood
that brings the largest improvement to the current solution. If no improving solution
in the neighborhood is available, then a random diversification procedure is applied,
which swaps workers in each period with probability 30 %. The diversified solution
becomes the current solution, irrespectively of its quality, and the improvement process
is restarted.

5.1 Data generation procedure

Since no benchmark data set exists for EJRSP so far, new data sets are required. For
each problem type, we generate both a structured data set, which is created closely
referring to the situation we observed at real-world assembly lines, and a random data
set, which we use to control our findings. We also diversify the size of the problem
instances. The shift length is set to 8 h. The number of periods is set to 4 (rotation is
performed once in 2 h), 8 (once an hour) or 16 (once half an hour). The number of
jobs is chosen to be 10, 15 or 20, which corresponds to typical team sizes in practice.
As basis for our data generation, we take the whole-body section of EAWS risk
estimation tool, whose modifications are widely used by car manufacturers in Europe
(Volkswagen, FIAT, Opel among others). Note that all procedures introduced in Sect. 4
are also suitable when any other ergonomic risk estimation tool is utilized. Measured
in EAWS, ergonomic points for a worker’s assignment per shift start from 2 and rarely
exceed 100 in practice. Therefore, we never allow ergonomic points of a period be
lower than T2 or larger than 100T . Further, we assume T rotation periods of equal length,
so that for a single rotation period the shift-related ergonomic points are to be weighted
by 1/T. Overall, we aligned the distribution of ergonomic risks to the situation that we
observed at assembly lines in automobile industry as this is one of the most important
fields of application for job rotation.
EJRSP-S data sets For the structured data set, the chosen distribution of the period-
specific ergonomic points of jobs (identical in each period) is constructed close to
real-world data. It is a normal distribution with expectation at 30 T points, standard
deviation of 20 points, truncated from below at 2
points and from above at 100 . For
T T  2 100 T
the random data set, we take the uniform distribution in the interval T , T .
EJRSP-VT data sets In line with our observations of real-world assembly lines, we
allow that a third of the jobs (variable jobs, i.e., workplaces that mount optional parts)
in the structured data set change their points in the range of ±25 %. First, “basic”
ergonomic points are generated as described in the previous paragraph for each job.
Then, to receive the ergonomic points for this job in a specific period, the correspondent

123
Reducing ergonomic risks by job rotation scheduling 725

add-ups or downs to the “basic” points are chosen from a uniform distribution between
−25 and +25 %. For the random data set, we allow the ergonomic points of all the
workplaces in each period to be independently uniformly distributed in the interval
[ T2 , 100
T ].
EJRSP-Ind data sets In the structured data set, we let one third of workers have
lower fitness or some restrictions due to their health state. We model it as add-ups
to the ergonomic points of the job, uniformly distributed from 0 to 50 %. Of course,
only variable jobs are allowed to have variable individual add-ups between the peri-
ods. In the random data set, we generate ergonomic points independently for each
worker-period-job constellation.
We perform a full factorial design with ten instances in each cell, i.e., 3 numbers of
periods ×3 numbers of jobs ×10 replications = 90 instances in each of six data sets
(3 problem types, each in a structured and a random version).

5.2 Experiment 1: comparing the performance of heuristics for EJRSP-VT

We run the exact optimization software FICO Xpress on the binary linear program
(1)–(5) described in Sect. 3.1 with a time limit of 1 h. For TS, we set the maximal
number of iterations to 20,000, the length of the tabu list to TL = 0.5 · (n − 1) · T (as
discussed in our sensitivity analysis in experiment 2) and allow the local improvement
to be performed in each iteration. Since an iteration of the greedy heuristic with random
diversification procedure GRD is computationally less expensive than that of TS, to
make the computational times comparable, we let GRD run with 300,000 iterations
for each problem instance. For GRD and TS, which have stochastic components, we
perform ten repetitions for each problem instance and report average values.
The computational experiments are performed on a personal computer with 2.67
GHz Intel QuadCore processor and 3 GB RAM. Only 1 core out of four is used in
each case.
Out of 180 EJRSP-VT problem instances (90 structured, 90 random), optimal solu-
tions for 176 are found either by Xpress or by one of the heuristics, in cases where
the heuristically determined solution value (upper bound) equals the LP-relaxation
lower bound (LB). When we speak about deviation from optimality, we either mean
the deviation of the found upper bound from the optimal solution value, if known, or
from LB, otherwise.
Table 6 summarizes the results reporting the number of found optima and best
solutions among the applied heuristics, deviations from optimality measured in abso-
lute ergonomic points EP and as a percentage, and the average computation time per
instance.
NPC, which often serves as a rule of thumb for team supervisors in the real world,
performs very poorly. For none of the instances, it is able to find the best known solu-
tion, the average relative deviation from optimality is about 7.5 % and, even worse,
the maximal relative deviation amounts to almost 27.9 %. In other words, the obtained
solution was observed to be up to 27.9 % worse than the optimum. The measured max-
imal absolute deviation from the optimum reached about 10 EP which is very critical
as it typically leads to critical (“red”) workplaces where workplaces with medium risks

123
726 A. Otto, A. Scholl

Table 6 Comparison of solution methods for EJRSP-VT

EJRSP-VT, structured (90 inst.) EJRSP-VT, random (90 inst.)

NCP SH GRD Xpress TS NCP SH GRD Xpress TS

# optimal solutions found 0 32 11.2 59 86.6 0 13 4.6 27 83.5


# best solutions found 0 32 12.8 61 88.6 0 13 6.2 29 85.5
Average deviation from optimality
EP 2.4 0.3 0.5 0.04 0.01 3.1 0.5 0.68 0.10 0.02
% 7.5 0.8 1.5 0.1 0.03 6.1 1.0 1.3 0.2 0.04
Maximal deviation from optimality
EP 10.00 2.25 1.30 0.25 0.25 10.50 4.75 1.63 0.50 0.25
% 27.9 6.7 3.8 0.9 0.7 23.7 9.9 3.4 0.9 0.5
Avg. time per instance (s) <0.05 <0.05 147 1505 2 <0.05 <0.05 144 2,779 3

(“yellow”) are possible, a planning result completely inacceptable in practice. Note


that the relative and the absolute deviations may achieve their maxima for different
instances.
SH proves itself as an effective and fast heuristic. In very short time, for no instance
it required more than 0.05 s, it is able to solve 36 % (14 %) of the instances from the
structured (random) data set to optimality. However, the maximum deviation consti-
tuted rather substantial 2.25 ergonomic points or 6.7 % in the structured data set.
GRD takes much more time than SH and provides less optimal solution as well as
larger average deviations. However, as it is an improvement procedure with explicit
diversification strategy it shows lower maximal deviations.
Xpress achieves a very good solution quality but misses optimal solutions (due to
time outs) quite often and requires large computation times not acceptable in practice.
Tabu search spends on average just about 2–3 s and maximally 1.4 min per EJRSP-
VT instance. On average in ten runs, it found optimal solution in 96 and 93 %
of the structured and random instances, respectively. With the best result over ten
runs, all 176 known optimal solutions of both data sets were found. The maximal
(worst case) deviation from optimality never exceeds negligible 0.25 points (less than
1 %).
In Fig.1, we examine the results of the first experiment for the practical
structured data set in more detail based on the percentage of optimal solutions
found. Results for the random data set reveal the same pattern. In addition to
the above findings, it becomes obvious that GRD and Xpress show a good per-
formance only for small problem instances. While Xpress requires large compu-
tation times, GRD fails to find optimal solutions for instances with more than
four periods at all. SH performs especially well on larger problem instances,
whereas TS reveals the best performance for all problem sizes. Remarkably,
no deterioration in performance with growing problem size can be detected for
TS.

123
Reducing ergonomic risks by job rotation scheduling 727

100%

NCP

80%

SH

60%

GRD*

40%

Xpress

20%

TS*

0%
T= 4 8 16 4 8 16 4 8 16

n = 10 n = 15 n = 20

Fig. 1 Comparison of solution methods for the structured data set: % of times the optimum was found.
Asterisk average of ten iterations

5.3 Experiment 2: main drivers of performance for tabu search

In this experiment, we examine the main drivers of performance of TS for EJRSP-VT:


introduction of the tabu list, employment of the local improvement procedure and
usage of a good initial solution by SH. Since GRD employs the same definition of
the neighborhood, we use its results after 20,000 iterations as the basis of comparison
(Fig. 2). Then, we compare it with a version of TS, which employs a random initial
solution and uses no local re-optimization procedure. Afterwards, we examine TS
with a random initial solution and with local re-optimization in each step. The overall
bar shows performance of TS with the settings as described in Sect. 4.2, which solves
170.1 instances on average over ten runs.
Introducing a tabu list has the largest impact, it is “responsible” for 81 % of optimal
solutions. This result is not surprising, because due to combinatorial character of the

Total: 170.1 instances solved


100% 0% with smoothing
10%
heuristic as
80%
initial solution
60% with local
improvement
81%
40%
with tabu list
20%

0% 9% GRD
% from # of optimal
solutions found
Fig. 2 Drivers of performance of TS

123
728 A. Otto, A. Scholl

# of optimal solutions
found for EJRSP-VT
175 varied length
170 of TL
165
160 varied
155 frequency of
loc re-opt
150
145 varied #
TL: 0.2(n-1)T TL: 0.3(n-1)T TL: 0.4(n-1)T TL: 0.5(n-1)T iterations
Freq: 3T Freq: 2T Freq: T Freq: 0
# iter: 5K # iter: 10K # iter: 15K # iter: 20K
Fig. 3 Sensitivity analysis of TS to parameter changes, average over ten runs. Total number of instances is
180. TL axis for changes in length of the tabu list. Freq axis for changes in frequency of local re-optimization
(re-opt). # iter axis for changes in number of iterations

problem, avoiding getting stuck in local optima and jumping away far enough from
the current solution is important. The application of the local improvement procedure
is responsible for further 10 % of optimal solutions found by TS. Employment of a
good initial solution has no remarkable effect as already known for TS applications to
other problems.
The conclusions made above were confirmed by the (local) sensitivity analysis
(Fig. 3). In these tests, we investigate the performance of TS at different number
of iterations, different lengths of the tabu list and different frequencies of applying
the local re-optimization procedure. We change one parameter at a time. The default
values, as described at the beginning of Sect. 5, are set to 20,000 iterations, local
re-optimization in each step and tabu list length to T L = 0.5 · (n − 1) · T.
Thus, a 60 % reduction of the length of the tabu list, from 0.5 · (n − 1) · T to
0.2 · (n − 1) · T, leads to a substantial 10 % reduction in number of optimal solutions.
Note that the impact of the length of the tabu lists flattens and achieves its maximum at
about 0.5·(n − 1)·T up to 0.8·(n − 1)·T . Longer tabu lists (given the current concept
of the tabu moves) become very restrictive. For example, T L = (n − 1)· T sometimes
results in situations, where all the moves from the current solution are prohibited. On
the other hand, it is possible to quarter the number of iterations without a significant
impact on the performance (number of optimal solutions found on average over ten
runs reduces from 170.1 to 165.7).

5.4 Experiment 3: performance of the procedures for other problem versions

To examine whether our new procedures also perform well for the two variations of
EJRSP, we repeat the experiments for the EJRSP-S and EJRSP-Ind data sets. The
results are documented in Tables 7 and 8, respectively.
For the simple problem version with equidistant periods and static ergonomic risks
(EJRSP-S), the procedures altogether are able to find and prove optima for 159 (84 +
75) instances. Concerning the relative performance of the procedures, we get very sim-
ilar results as for the dynamic version EJRSP-VT. However, the solution quality of all

123
Reducing ergonomic risks by job rotation scheduling 729

Table 7 Comparison of solution methods for EJRSP-S

EJRSP-S structured (90 inst.) EJRSP-S random (90 inst.)

NCP SH GRD Xpress TS NCP SH GRD Xpress TS

# optimal solutions found 0 27 8.3 44 79.7 0 13 2.9 20 70.3


# best solutions found 0 28 12.3 50 85.7 0 13 8.1 33 84.3
Average deviation from optimality
EP 1.93 0.34 0.46 0.08 0.03 3.16 0.51 0.73 0.19 0.08
% 5.9 1.0 1.5 0.3 0.1 6.4 1.0 1.4 0.4 0.2
Maximal deviation from optimality
EP 8.50 3.00 1.28 0.50 0.50 10.00 2.25 2.85 1.50 1.00
% 26.9 10.7 4.7 1.5 1.5 27.2 4.7 4.4 2.3 2.3
Avg. time per instance (s) <0.05 <0.05 147 2,007 4 <0.05 <0.05 143 2,951 8

Table 8 Comparison of solution methods for EJRSP-Ind

EJRSP-Ind structured (90 inst.) EJRSP-Ind random (90 inst.)

NCP SH GRD Xpress TS NCP SH GRD Xpress TS

# optimal solutions found 0 0 0.0 11 4.2 0 12 3.0 63 44.2


# best solutions found 0 0 0 49 34 0 12 3.0 90 46.8
Average deviation from optimality
EP 2.72 0.76 2.31 0.21 0.24 3.52 1.29 13.42 0.11 0.23
% 8.0 2.2 6.9 0.6 0.7 26.8 10.1 113.1 1.0 2.1
Maximal deviation from optimality
EP 9.00 2.50 3.93 0.75 0.70 7.50 3.50 22.93 0.88 1.20
% 20.5 7.2 10.4 2.3 2.3 59.0 31.3 241.2 8.8 12.0
Avg. time per instance (s) <0.05 <0.05 135 3,248 301 <0.05 <0.05 98 1,260 302

procedures (except for NCP) slightly declines while the computation times of Xpress
and TS increase. Probably, the increased level of symmetry in the problem complicates
fathoming by Xpress (as many partial solutions with identical bound values exist) and
finding promising search trajectories by TS and GRD.
For the problem with dynamic and individual ergonomic points (EJRSP-Ind), the
procedures altogether are able to optimally solve only 16 instances in the structured
data set, which seems to be the hardest one, and 63 instances in the random data set.
In the following, we concentrate on differences to former findings.
It has to be recalled that OPA, which is an ingredient of NCP, SH, and TS, now
consists in solving a respective LBAP instance as formulated in Sect. 4.2. As a conse-
quence, GRD which only performs a local search and escapes from local optimality
by a random diversification shows very poor results with extremely large deviations,
especially in the random data set where ergonomic points are unsystematically (and
unrealistically) distributed among workers. In this case, finding promising regions in

123
730 A. Otto, A. Scholl

Table 9 Two-tailed Wilcoxon test

EJRSP-VT EJRSP-VT EJRSP-S EJRSP-S EJRSP-Ind EJRSP-Ind


struc random struc random struc random

Wilcoxon test:
SH versus GRD
p value <0.0025 <0.0025 <0.0025 <0.005 <0.0025 <0.0025
# non-zero differences 87 90 84 86 89 90
in ranks
W + / W− 1,014 / 2814 1,227 / 2,868 1,053 / 2,517 1,212 / 2,529 75 / 3,930 41 / 4,054
Wilcoxon test:
TS versus XPress
p value <0.0025 <0.0025 <0.0025 <0.0025 <0.2 <0.0025a
# non-zero differences 30 64 39 54 76 46
in ranks
W + / W− 2 / 464 33 / 2,047 0 / 780 0 / 1,485 1,176 / 1,750 1,081 / 0
a second method shows a better performance

the search space is more complicated and requires better guidance of the search. But
even for the structured data set, the performance of GRD is disappointing as it is only
slightly better than that of NCP despite of much larger computation times.
Due to solving an LBAP instance in each iteration to re-optimize the current solu-
tion, the computation times of TS drastically increase (about 5 min on average, and up
to 32.5 min). Despite of this increased effort, the deviations of TS, as well as those of
the other procedures, seem to be much larger than for the other problem versions. One
reason is the rather low number of known optimal solutions such that the deviations
for many instances refer to the weak LP-relaxation lower bounds.
In opposite to the former results, Xpress now performs better than TS in terms of
number of optimal and best solutions, especially for the random data set where the
lack of structure in the data complicates finding the best search direction. However, the
average and maximal deviations reveal only small differences between the two meth-
ods, indicating that TS achieves competitive results in much smaller computational
times than Xpress.

5.5 Statistical foundation of the results

To check the statistical significance of differences in performance between the exam-


ined solution procedures found in Sects. 5.2 and 5.4, we perform a Wilcoxon signed-
rank test. As performance measure for different algorithms in each instance, we take
the average (absolute) deviation from optimality. The Wilcoxon test compares the sum
of ranks for positive (W+ ) and negative (W− ) differences. The larger the difference
between W+ and W− is, the higher is the probability that one method is better than
another. We neglect moderate interaction effects with the problem size.

123
Reducing ergonomic risks by job rotation scheduling 731

For no instance, SH finds a worse solution than NCP as well as for no instance the
solution of TS is inferior to that of NCP, SH or GRD. Thus, there remain two pairs of
procedures to be examined by the two-tailed Wilcoxon test (see Table 9).
SH shows a better performance than GRD in all six data sets at the significance
level of α < 0.005. TS exhibits a statistically significantly better performance than
Xpress for instances of EJRSP without individual effects at α < 0.0025. In case of
structured instances of EJRSP-Ind, the Wilcoxon test could not reject hypothesis on
the equal performance of TS and XPress with 1 h limit even at α = 0.1. Only for
randomly generated EJRSP-Ind instances, TS shows a slightly inferior performance
compared with XPress.

6 Discussion

In this article, we show that the dynamic ergonomic job rotation scheduling problem
with variable tasks is NP-hard in the strong sense. However, for our experimental data
set generated closely to the situation at automobile assembly lines, it was possible
to achieve rotation schedules with negligible deviation from optimality in reasonable
computation time.
Further, we introduce a smoothing heuristic, which can be incorporated in various
solution methods for constructing initial solutions and/or as a local re-optimization
procedure. The smoothing heuristic proves to be a fast heuristic, providing an imme-
diate and in many cases a good approximation of the optimal solution. In situations,
where the solutions are required “at the push of a button”, it should replace straight-
forward construction procedures often used in practice.
If several seconds of computational time are available, then we recommend integrat-
ing the smoothing heuristic into our tabu search procedure. The proposed tabu search
algorithm is the best and the most robust from the examined solution methodologies
with very small deviations from optimality. The found solutions differed maximally
by 0.5 ergonomic points, or 1.5 %, from the optimum in our computational experi-
ments for instances generated closely to practical settings. To achieve such results,
our tabu search methodology requires just a couple of seconds in relevant for practice
settings.
In case of individual ergonomic risks estimation, which is not common for assembly
lines in Europe, our tabu search still provides robust and competitive results. However,
computational times increase substantially and equal to 5 min on average. We see a
potential for improvement here.
Several questions remain open for further research. First of all, the significant impact
of the tabu list, discovered in our sensitivity analysis, and the speed of tabu search
indicate that there is potential to develop an exact algorithm that will be much more
effective than standard MIP software for the dynamic ergonomic job rotation sched-
uling problem. Second, the question of the complexity of the simple ergonomic job
rotation problem (with time-invariant ergonomic risks) still remains open. However,
our experiments, especially difficulties of the standard software Xpress to solve even
small problem instances within an hour time limit, point on a possible NP-hardness in
the strong sense even of this restricted problem. Third, improved solution procedures

123
732 A. Otto, A. Scholl

are required for the problem version with individual ergonomic risk evaluations which
are able to better utilize the problem structure of this problem.

Acknowledgments This article was supported by the Federal Program “ProExzellenz” of the Free State
of Thuringia. We would like to thank Dr. Malte Fliedner and two anonymous reviewers for valuable hints
that helped to improve the paper considerably.

References

Bokranz R, Landau K (2006) Produktivitätsmanagement von Arbeitssystemen. Schäffer-Poeschel, Stuttgart


Boysen N, Fliedner M, Scholl A (2009a) Production planning of mixed-model assembly lines: overview
and extensions. Prod Plan Control 20:455–471
Boysen N, Fliedner M, Scholl A (2009) Sequencing mixed-model assembly lines: survey, classification and
model critique. Eur J Oper Res 192:349–373
Burkhard R, Dell’Amico M, Martello S (2009) Assignment problems. Society for Industrial and Applied
Mathematics, Philadelphia
Butković P, Lewis S (2007) On the job rotation problem. Discrete Optim 4:163–174
Carnahan BJ, Redfern MS, Norman B (2000) Designing safe job rotation schedules using optimization and
heuristic search. Ergon 43:543–560
Cheng TCE, Sin CCS (1990) A state-of-the-art review of parallel-machine scheduling research. Eur J Oper
Res 47:271–292
Costa AM, Miralles C (2009) Job rotation in assembly line employing disabled workers. Int J Prod E
120:625–632
De Werra D, Hertz A (1989) Tabu search techniques: a tutorial and an application to neural networks. OR
Spectr 11:131–141
Diego-Mas JA, Asensio-Cuesta S, Sanchez-Romero MA et al (2009) A multi-criteria genetic algorithm
for the generation of job rotation schedules. Int J Ind Ergon 39:23–33
EN 614-1 (2006) Safety of machinery—ergonomic design principles—Part 1: Terminology and general
principles. European Committee for Standardization
Eklund JA (1995) Relationships between ergonomics and quality in assembly work. Appl Ergon 26:15–20
Falck AC, Örtengren R, Högberg D (2010) The impact of poor assembly ergonomics on product quality:
A cost-benefit analysis in car manufacturing. Hum Factors Ergon Manuf 20:24–41
Garey MR, Johnson DS (1979) Computers and intractability: a guide to the theory of NP-completeness. W.
Freeman and Company, New York
Glover F (1990) Tabu search: a tutorial. Interfaces 20:74–94
Jorgensen M, Davis K, Kotowski S, Aedla P, Dunning K (2005) Characteristics of job rotation in the
Midwest US manufacturing sector. Ergon 28:1721–1733
Keir PJ, Sanei K, Holmes MWR (2011) Task rotation effects on upper extremity and back muscle activity.
Appl Ergon 42:814–819
Kuijer PPFM, Visser B, Kemper HCG (1999) Job rotation as a factor in reducing physical workload at a
refuse collecting department. Ergon 42:1167–1178
Kuijer PPFM, van der Beek AJ, van Dieën JH, Visser B, Frings-Dresen MHW (2005) Effect of job rotation
on need for recovery, musculoskeletal complaints, and sick leave due to musculoskeletal complaints:
a prospective study among refuse collectors. Am J Ind Med 47:394–402
Lodree EJ, Geiger CD, Jiang X (2009) Taxonomy for integrating scheduling theory and human factors:
Review and research opportunities. Int J Ind Ergon 39:39–51
Longo F, Mirabelli G (2009) Effective design of an assembly line using modelling and simulation. J Simul
3:50–60
Moreau M (2003) Corporate ergonomics programme at automobiles Peugeot-Sochaux. Appl Ergon 34:
29–34
Neumann WP, Winkel J, Medbo L, Magneberg R, Mathiassen SE (2006) Production system design
elements influencing productivity and ergonomics. Int J Oper Prod Manag 26:904–923
Occupational Safety and Health Administration (2000) Final Ergonomics Program Standard. Federal
register #64, pp 65768–66078
Osterman P (2000) Work reorganization in an era of restructuring: trends in diffusion and effects on
employee welfare. Ind Labor Relat Rev 53:176–196

123
Reducing ergonomic risks by job rotation scheduling 733

Otto A, Scholl A (2011) Incorporating ergonomic risks into assembly line balancing. Eur J Oper Res
212:277–286
Raina SM, Dickerson CR (2009) The influence of job rotation and task order on muscle fatigue: a deltoid
example. Work 34:205–213
Rissén D, Melin B, Sandsjö L, Dohns I, Lundberg U (2002) Psychophysiological stress reactions, trapezius
muscle activity, and neck and shoulder pain among female cashiers before and after introduction of
job rotation. Work Stress 16:127–137
Schaub K, Caragnano G, Britzke B, Bruder R (2010) The European Assembly Worksheet. In: Mandel-
o P, Karwowski W, Saarela K, Swuste P, Occhipinti E (eds) Proceedings of the VIII international
conference on occupational risk prevention, Valencia, 5–7 May 2010
Sinn-Behrendt A, Schaub K, Winter G, Landau K (2004) Ergonomisches Frühwarnsystem „Ergo-FWS“. In:
Landau K (ed) Montageprozesse gestalten: Fallbeispiele aus Ergonomie und Organisation, Ergonomia,
Stuttgart
Seçkiner SU, Kurt M (2007) A simulated annealing approach to the solution of job rotation scheduling
problems. Appl Math Comput 188:31–45
Seçkiner SU, Kurt M (2008) Ant colony optimization for the job rotation scheduling problem. Appl Math
Comput 201:149–160
Tharmmaphornphilas W, Norman BA (2004) A quantitative method for determining proper job rotation
intervals. Ann Oper Res 128:251–266
Tharmmaphornphilas W, Norman BA (2007) A methodology to create robust job rotation schedules. Ann
Oper Res 155:339–360
Walter R (2011) Comparing the minimum completion times of two longest-first scheduling-heuristics.
Central Eur J Oper Res. doi:10.1007/s10100-011-0217-4

123
Copyright of OR Spectrum is the property of Springer Science & Business Media B.V. and its
content may not be copied or emailed to multiple sites or posted to a listserv without the
copyright holder's express written permission. However, users may print, download, or email
articles for individual use.

Das könnte Ihnen auch gefallen