This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS 1

Simulation-Based Optimization Methodology for a

Manual Material Handling Task Design That
Maximizes Productivity While Considering
Ergonomic Constraints
Yaar Harari , Avital Bechar, Member, IEEE, and Raziel Riemer

Abstract—Design of workplaces that include human–machine measure-based methodology for the effective ergonomic design
systems and manual material handling should consider both the of workstations. The authors applied their methodology to a
productivity of workers and the risk of injury. In this study, a hose pressure test workstation and reduced the work process
simulation-based optimization methodology for a manual material
handling task design was developed. A new formulation of the op- time by 38% while reducing the oxygen consumption of the
timization problem is presented, whose objective is to maximize worker by 20%. Longo and Mirabelli [9] used DHM to develop
worker productivity and at the same time not to exceed ergonomic an effective assembly line design for heater production that
thresholds (which represent injury-risk measures). The workplace takes into consideration work measurements, line balancing,
and work process were simulated using digital human modeling and ergonomic factors. By manipulating various design config-
software (Jack), and the best design was found using a genetic
algorithm. The results show that the new formulation of the opti- urations (i.e., changing the height of a workstation) better line
mization problem improved the predicted productivity by 105%, balancing was achieved, which increased productivity by 47%
compared to the formulation used in previous studies that used while improving the working postures of the workers. Battini
a multi-objective function. Meanwhile, the risk of injury did not et al. [10] developed a methodological framework to improve
exceed ergonomic thresholds. productivity and ergonomics in assembly system design, while
Index Terms—Computational human modeling, ergonomics, hu- taking into account technological variables (e.g., work times),
man performance, manual material handling task design, opti- environmental variables (e.g., workforce motivation), and er-
mization.
gonomics evaluations. The authors applied their methodology
I. INTRODUCTION by redesigning a shower enclosure workstation, and increased
productivity by 15% while lowering the risk of injury. Shewchuk
ORK-RELATED musculoskeletal disorders (MSDs)
W are responsible for 30% of days lost to injury and result
in annual costs of $45–54 billion in the U.S. alone [1], [2].
et al. [11] developed a methodology for simulation of workers
during physical tasks, considering both the workers’ motion
and ergonomic assessments. This approach was applied to pan-
Therefore, design of workplaces that include human–machine elized residential construction and resulted in a software that
systems (HMS) should consider not only the workers’ produc- enables to simulate physical tasks while considering the work-
tivity, but also their risk of injury [3]–[7]. Based on that, the best ers’ ergonomics. However, in all of the above studies, the new
HMS design should yield maximum productivity while main- and improved HMS workplace designs were selected out of a
taining the injury risk below physiological and biomechanical limited number of manually designed configurations with no
thresholds. optimization process. This means that it is very likely that a
Several studies have used digital human modeling (DHM) for better solution exists.
HMS workplace design, while considering both worker produc- To overcome this limitation, several studies have offered
tivity and the risk of injury. Cimino et al. [8] proposed a multiple frameworks for solving HMS workplace design as an optimiza-
tion problem. Ben-Gal and Bukchin [12] presented a method-
Manuscript received April 15, 2018; revised October 31, 2018, December ology for workstation design consisting of multi-objective opti-
5, 2018, and January 22, 2019; accepted February 6, 2019. This work was
supported in part by the Helmsley Charitable Trust through the Agricultural, mization that considered both production and ergonomics fea-
Biological and Cognitive Robotics Initiative of Ben-Gurion University of the tures and was applied using the response surface optimization
Negev. This paper was recommended by Associate Editor L. Rothrock. (Corre- method. The authors demonstrated their methodology by im-
sponding author: Yaar Harari.)
Y. Harari and R. Riemer are with the Department of Industrial Engineering proving a fruit-packaging work process and reduced the pro-
and Management, Ben-Gurion University of the Negev, Beer Sheva 8410501, cess cycle time by 17.5% while improving ergonomic measures
Israel (e-mail:,[email protected]; [email protected]). by up to 33%. del Rio Vilas et al. [13] proposed a general
A. Bechar is with the Institute of Agricultural Engineering, Agricultural
Research Organization, Bet Dagan 7505101, Israel (e-mail:, avital@volcani. framework for manufacturing workstation design, combining
agri.gov.il). ergonomic and operational considerations. The authors simu-
Color versions of one or more of the figures in this paper are available online lated the workplace and process in DHM software and used a
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/THMS.2019.2900294 multi-objective function which combined both ergonomic and
2168-2291 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS

operational performance measures. Implementing their method-

ology resulted in improved work process cycle time and reduced
both the workers’ risk of injury and their energy expenditure. Ore
et al. [14] presented an application of HMS design optimization
that included human–robot collaboration. The application was
demonstrated for a handover task. The tradeoff between the cy-
cle time and the rapid upper limb assessment (RULA) ergonomic
measure [15] was investigated. The authors of the current study
recently solved the workplace design optimization problem by
using DHM and applying a two-step grid search, which com-
prised a coarse search of the entire solution span and a fine Fig. 1. Box-conveying work process at the pepper packing house. (a) Lifting a
search around the best solution from the coarse search [16]. box from a conveyor. (b) Carrying the box. (c) Lowering the box onto a platform.
However, although all of the above studies that address work-
place design optimization are innovative and useful, they all
combined productivity and risk-of-injury measures into one
multi-objective function. Yet, in industry, the common practice
is to consider the risk of injury as a constraint (e.g., the compres-
sion force on the lower back should not exceed 3400 N [17]).
Also, in all of the above studies except one [16], execution of
the optimization process and work with the DHM software was
conducted manually. This is a very time-consuming process, and
therefore the workplaces could not be optimized in a reasonable
time [12], [18], [19]. Even in the one study that did offer an
automated optimization process [16], the two-step grid search
optimization method was computationally inefficient, with the
possibility of missing a global optimum.
This study offers a new automated optimization methodology Fig. 2. Box-conveying simulation in Jack.
for HMS workplace design, based on a genetic algorithm (GA)
using a DHM simulation (Jack) that is a commonly used tool
for workplace design [20], [21]. This methodology could be industries. An example of such an HMS is a packing house
used as a decision-making tool to enhance the capabilities of for peppers. Here, boxes of peppers continuously arrive on a
practitioners (e.g., ergonomists and industrial engineers). conveyor belt. The workers perform the following continuous
The methodology presented in this study includes a new for- sequential work process (see Fig. 1):
mulation of the optimization problem, in which the objective of 1) lifting a box from the conveyor;
the optimization is to maximize productivity under ergonomic 2) carrying the box in front of the body for three meters;
constraints. For demonstration purposes, we solved a case study 3) lowering the box onto a shipping platform.
of a box-conveying task. The ergonomic constraints were thresh- After lowering the box, the workers return to the conveyor
olds on the compression forces acting on the lower back verte- to lift the next box. The distance between the conveyor and
brae (Lower Back Analysis, LBA; [22]), the worker’s oxygen platform was set at three meters.
consumption rate (VO2 ) for multiple-task work process [23]
and the RULA ergonomic measure [15]. The work process cy- B. Digital Human Modeling
cle time was calculated using time-prediction models from [24], The work process was simulated using the task simulation
and the productivity was calculated by multiplying the box mass builder module in Jack (see Fig. 2). The worker was repre-
by 1/cycle time. sented using a virtual male mannequin with a height of 1.75
m and weight of 79 kg. This height and weight represents the
II. METHOD anthropometrics of a median male, according to the ANSUR
The aim of the proposed methodology is to determine the op- database [25]. The simulation inputs were the mass of the box,
timal workplace design by controlling the properties of different the conveyor height, and the platform height. The outputs of
entities in the HMS (i.e., shelf and conveyor heights, handled the simulation were the joint angles of various body parts of
object mass). To demonstrate our methodology, we chose the the virtual mannequin, and the compression forces acting on the
case study of a box-conveying task, modeled it in Jack, formu- L5/S1 vertebra joints during the simulation at 30 Hz.
lated the optimization problem with constraints and solved for
the optimal design using a GA. C. Overview of the HMS Workplace Design Methodology
The HMS workplace design optimization methodology con-
A. Case Study—Box Conveying Work Process sisted of
To demonstrate our optimization methodology, we chose the 1) the DHM environment for designing the workplace and
case study of a box-conveying task, which is common in various simulating the work process (Jack software by Siemens);
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HARARI et al.: SIMULATION-BASED OPTIMIZATION METHODOLOGY FOR A MANUAL MATERIAL HANDLING TASK DESIGN 3

D. Development of the Optimization Methodology

Our objective was to maximize productivity while remaining
within the ergonomic constraints. The following sections de-
scribe the problem formulation and the optimization algorithm
code.
1) Objective Function: The productivity measure selected
for the objective function was the production rate (PR). The PR
is defined as the total mass of boxes transferred per unit time
PR = m/CT (1)
where m is the mass handled per work cycle and CT is the
cycle time—the time required to complete the task. The CT
was calculated as the total time (min) for completing the lifting,
carrying, lowering, and returning tasks
CT = 60/ (tlift + tcarry + tlower + treturn ) (2)
where tlift , tcarry , tlower , and treturn are the times required to lift
the mass from the conveyor, to carry the mass for three meters,
to lower the mass onto the platform, and to return to the initial
lifting point without carrying the mass, respectively.
tlift , tcarry , tlower , and treturn were calculated using the time-
prediction models from [24] (3)–(8). These models were se-
lected because they consider the influence of the mass of the box
and the lifting and lowering heights on the task duration. Fur-
thermore, these models were found to be more accurate for this
type of work process than methods-time measurement, which is
the time-prediction model in Jack [24]

tlift = 2.099 + 0.0418 ∗ m − 2.211 ∗ LFH + 1.1658

Fig. 3. Overview of the optimization methodology, which is comprised of
(a) manual setting of the initial workplace, GA parameters and ergonomic ∗LFH2 − 0.117 ∗ LWH + 0.0752 ∗ LWH2 (3)
thresholds; and (b) execution of the automated optimization process.
vcarry = 1.2521 + 0.0073 ∗ m−0.0004 ∗ m2 − 0.1815 ∗ LFH
2) an objective function for the worker’s productivity; + 0.1077 ∗ LFH2 −0.0966 ∗ LWH+0.1389 ∗ LWH2
3) a set of ergonomic constraints which consider physiolog- (4)
ical and biomechanical injury-risk thresholds;
tcarry = dist/vcarry (5)
4) an optimization procedure using the GA method;
5) a main program (in the Python language) that managed tlower = 2.1684 + 0.036 ∗ m − 0.1903 ∗ LFH + 0.0963
and integrated the objective function, the ergonomic con-
straints, the GA procedure and the Jack software simula- ∗LFH2 − 1.8512 ∗ LWH + 0.9359 ∗ LWH2 (6)
tion. vwalk = 1.1015 + 0.0088 ∗ m − 0.0006 ∗ m −0.0289 ∗ LFH 2
The main program received as input the work process to be
simulated, the design parameters (e.g., the box mass), the virtual + 0.0127 ∗ LFH2 −0.1487 ∗ LWH+0.0661 ∗ LWH2
mannequin’s anthropometrics (gender, height, and weight), and (7)
parameter values for the GA.
twalk = dist/vwalk . (8)
For each solution that was generated by the GA procedure,
the main program redesigned the workplace and simulated the Above, m is the box mass (kg), LFH is the lifting initial height
work process. Then, data were extracted from Jack and the (m), LWH is the lowering final height (m), dist is the distance
objective function score was calculated. In addition, using between the lifting and lowering stations (m), and vcarry and
the data from Jack, the program checked whether the solution vwalk are the carrying and walking velocities, respectively (m/s).
met the ergonomic constraints. In this study, we included a 6 s allowance time between cycles
A flowchart of the proposed methodology for optimizing the in order for the workers to rest.
HMS workplace design is presented in Fig. 3. The optimization 2) Ergonomics Measures: The lower back compression
process was performed on a Toshiba Satellite P50 PC with an force (LBCF), RULA score, and VO2 were chosen as the er-
Intel i7-4710MQ processor (6 MB cache, 2.5 GHz, 1600 MHz gonomic measures. These measures were selected because each
front-side bus). evaluates a different injury-risk factor (RULA evaluates pos-
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS

tures, LBCF evaluates forces on the spine, and VO2 evaluates ergonomic measures are constrained (16)–(20)
continuous effort or metabolic rate).
MAX PR, (16)
The LBCF was calculated at 30 Hz during the simulation
using the LBA tool in Jack [22], and indicated the peak com- s.t.
pression force (N) acting on the L5/S1 vertebra joints during
LBCF < LBCFcr , (17)
the work process simulation. The RULA score was also calcu-
lated at 30 Hz during the simulation using our own customized RULA < RULAcr , (18)
Python code. During the simulation, this code extracted the an-
VO2 < VO2 cr , (19)
gles for different body joints and parts (e.g., shoulder, trunk) of
the virtual mannequin in Jack. Using the joint angles, the code BM < BMcr . (20)
followed the RULA protocol [15] and determined the RULA
Here, LBCFcr , RULAcr , and VO2cr are the ergonomic thresh-
score. The RULA measure indicated the highest RULA score
olds for lower back compression forces, RULA score, and
during the work process simulation. The VO2 (ml/min) was
oxygen-consumption rate, respectively. BMcr is the maximum
calculated using the prediction equations from [23] for a lift–
box mass that may be handled. In this study, the threshold values
carry–lower process (9)–(14), and indicated the worker’s oxygen
were set as follows (21)–(24):
consumption rate during the work process simulation
LBCFcr = 3400N [17] (21)

VO2 = −899.1 + 9 ∗ bw + 184.7 ∗ freq RULAcr = 5 [15] (22)

L −L

+ 35 ∗ freq ∗ dist + 36.7 ∗ m (9) V O2cr = 1000 ml/min [23] (23)

VO2 = −775 + 9.5 ∗ bw + 53.8 ∗ freq BMcr = 23 kg NIOSH for lifting tasks [17] . (24)
L −M

+ 48.1 ∗ freq ∗ dist + 31.3 ∗ m (10) The LBCFcr was set at 3400 N since this represents the cutoff
value for lower back injury risk [17], based on cadaver studies
VO2 L −H = −771.6 + 8.7 ∗ bw + 122.9 ∗ freq (e.g., [26], [27]) and biomechanical models (e.g., [28], [29]).
+ 32.6 ∗ freq ∗ dist + 40.4 ∗ m (11) The RULAcr was set to 5 since lower values represent a low
risk of MSDs [15]. The VO2cr was set at 1000 ml/min (approx-
VO2 H −L = −606 + 7.7 ∗ bw + 77.2 ∗ freq imately 5 kcal/min) since this represents the cutoff value for
+ 39 ∗ freq ∗ dist + 26.3 ∗ m (12) prevention of aerobic and muscle fatigue, which is also related
to risk of MSDs [30]–[32]. The BMcr was set to 23 kg, which
VO2 H −M = −680.4 + 9.7 ∗ bw + 22.1 ∗ freq was stated by NIOSH to be the maximal acceptable weight for
+ 35.3 ∗ freq ∗ dist + 20.6 ∗ m (13) lifting, regardless of task design [17].
4) Genetic Algorithm: A GA is a biologically inspired opti-
VO2 H −H = −733.9 + 7.9 ∗ bw + 70.7 ∗ freq mization method that first examines a generation of solutions.
+ 40.1 ∗ freq ∗ dist + 26.5 ∗ m. (14) The descendants of the solutions are then examined in the next
generation by combining pairs of solutions and by creating ran-
dom changes in other solutions. We chose a GA since it is par-
In VO2 X −Y , X is the initial lifting height (L for heights ticularly suitable for problems with characteristics similar to the
below 90 cm and H for heights above 90 cm) and the Y is the problem presented in this study (non-differentiable, discontin-
final lowering height (L for heights below 80 cm, M for heights uous objective functions, and multiple local minima [16], [33],
between 80 and 120 cm, and H for heights above 120 cm). bw is [34]). In this study, we developed a GA in the Python language.
the worker’s body weight (kg), dist is the distance between the Each solution (a GA chromosome) represented a different HMS
lifting and lowering platform (m), m is the mass of the box (kg), workplace design and was comprised of three parameters:
and freq is the number of times the work process is conducted 1) the mass of the box to be handled;
per minute, which was calculated as follows (15): 2) the height of the lifting platform;
3) the height of the lowering platform.
The box masses ranged between 2 and 23 kg, in increments
freq = 1/CT (15) of 0.5 kg. The lifting and lowering heights ranged between 20
and 160 cm above the floor level, in increments of 2 cm. The
following GA parameters values were used:
where CT is the time required to complete the work process in 1) population sizes of 25, 50, 75, and 100 chromosomes per
minutes. generation;
3) Optimization Problem Formulation: The aim of the op- 2) mutation rates of 1%, 5%, 10%, and 15%;
timization methodology is to find the HMS workplace design 3) total number of generations between 1 and 10, in steps of
in which the productivity is maximal, while the workers do not 1 generation.
exceed the ergonomic thresholds. Therefore, we developed an Reproduction was implemented using a one-point crossover,
objective function that maximizes the PR, and in which the and the selection operator was chosen to be the roulette wheel
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HARARI et al.: SIMULATION-BASED OPTIMIZATION METHODOLOGY FOR A MANUAL MATERIAL HANDLING TASK DESIGN 5

TABLE I
FOUR CONFIGURATIONS THAT WERE APPLIED FOR SOLVING THE BOX-CONVEYING CASE STUDY

Each design configuration represents a different combination of the formulation of the optimization problem, optimization method, and time-prediction model.

technique. Elitism was implemented by passing the top 10% of were fixed. The LBCFcr ranged between 1400 and 3400 N in
solutions in each generation to the next one. steps of 800 N, the RULAcr ranged between 3 and 7 in steps
of 1, the VO2cr ranged between 800 and 1200 ml/min in steps
E. Analyses of 100 ml/min, and the BMcr ranged between 4 and 23 kg in
steps of 4 kg.
To evaluate the performance of our new methodology for Finally, we investigated the effect of the GA parameters (the
workplace design, the following analyses were performed. number of generations, the number of solutions examined, the
1) Our new formulation of the optimization problem (max- mutation rate and the population size) on the PR value of the best
imum productivity with ergonomic constraints; config- solution the GA obtained. This investigation explores whether
uration #1 in Table I) was compared to that of previ- GAs with different parameter values converge to the same op-
ous studies (multi-objective approach; configuration #2 in timal solution. It might also clarify which parameter values are
Table I). Both were solved using the GA algorithm, where preferable for solving this problem.
the multi-objective approach (configuration #2 in Table I)
was formulated as follows (25):

U = PR−1 ∗ LBCF1 ∗ RULA1 ∗ VO2 1 . (25) III. RESULTS

A. Optimal Solution Obtained Using Different Optimization
2) Our GA procedure (configuration #1 in Table I) was com- Configurations
pared to the optimization method from a previous study Optimal HMS workplace designs were obtained using three
(the two-step grid search; configuration #3 in Table I), different configurations of the optimization method (see Ta-
with both solving the optimal design for maximum pro- ble II). First, we compared the new formulation of the op-
ductivity with ergonomic constraints. timization problem, which maximizes productivity under er-
3) It is possible that users will choose to use the time- gonomic constraints (configuration #1), with the multi-objective
prediction models currently implemented in Jack (i.e., approach (configuration #2). With the new formulation of the
MTM-1 [35]). Therefore, we will run the optimization optimization problem the PR was higher by 105%. The optimal
methodology using the time prediction models in Jack solution of the new formulation of the optimization problem
(configuration #4 in Table I) and analyzes the feasibil- resulted in higher RULA, LBCF, and VO2 in comparison to the
ity of using the methodology with Jack time prediction multi-objective approach. Yet, these measures did not exceed
models. injury-risk thresholds, and are therefore likely to be acceptable
To perform these analyses, we solved the optimization prob- for safe workplace design.
lem in four configurations. The parameters in each configuration Comparing the GA (configuration #1) to the two-step grid
were the formulation of the optimization problem, the optimiza- search (configuration #3) revealed that the optimal design ob-
tion method, and the time-prediction model (see Table I). tained by the GA yielded a PR that was higher by 69% than the
In this study, we solved the optimization problem using guide- two-step grid search.
line ergonomic thresholds as constraints in the optimization. Yet, Since users might wish to apply the methodology using the
it is possible that, due to the preference (of an ergonomist or current time-prediction model implemented in Jack, we ran the
production engineer) or an updated guideline, the threshold will methodology using Jack’s time instead of the models of [24].
need to be changed. Therefore, to test the effect of changes in The optimal solution resulted in a box mass of 19 kg, lifting
the ergonomic thresholds, a sensitivity analysis was performed. height of 92 cm, and lowering height of 104 cm. This solution
In this analysis, we ran the optimization multiple times (using resulted in a PR of 77.6 kg/min (using the time models of Jack)
configuration #1 in Table I) where, in each run, the value of or 74 kg/min (using the models of [24]). The RULA score was
one constraint changed while the values of the other constraints 4, the LBCF 2514 N, and the VO2 999 ml/min.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS

TABLE II
OPTIMAL SOLUTIONS OBTAINED USING THREE DIFFERENT CONFIGURATIONS, WHICH INCLUDE DIFFERENT FORMULATIONS OF THE OPTIMIZATION PROBLEM,
OPTIMIZATION METHODS, AND TIME-PREDICTION MODELS

The GA’s shortest time for finding the best solution was 10 min and 24 s. On average, the GA found the best solution in 25 min and 24 s.

B. Relation Between Productivity and Ergonomics Thresholds population sizes of 25 and 50 chromosomes evaluated fewer
The LBCF threshold (constraint) was the limiting factor for solutions before reaching the best solution [see Fig. 6(c) and (d)].
values between 1400 and 2200 N, the RULA constraint was Analysis of the effect of the number of chromosomes revealed
the limiting factor for values between 3 and 4, the VO2 con- that, for all cases but one, the GA found the optimal solution in
straint was the limiting factor for values between 0.8 and 1.2 less than ten generations [see Fig. 5(b), 25 chromosomes).
l/min, and the box mass constraint was the limiting factor for
values between 4 and 20 kg. Each constraint threshold had a IV. DISCUSSION
different effect on the PR. The maximum PRs that were ob-
tained from changing each ergonomic constraint threshold, and A. Comparison of the Proposed Design Methodology to
the constraint values of the optimal solution (configuration #1, Different Design Configurations
Table II) are presented in Fig. 4. The results show that the limit- In this study, a new formulation of the optimization problem
ing constraint for our case study was the VO2 constraint, since was presented. This new formulation uses an objective func-
it is the only constraint for which increasing its value resulted tion of productivity, and solves the optimization problem with
in an increased PR [see Fig. 4(b)]. ergonomic thresholds as constraints. This new formulation re-
flects common practice in workspace design, in which the design
objective is to maximize productivity as long as the injury-risk
C. Investigation of the Genetic Algorithm Configuration
measures’ values are below the acceptable thresholds. The for-
We conducted multiple runs of the optimization program with mulation in this study is an innovation, compared to previous
different GA parameter values (i.e., number of examined solu- studies that used multi-objective function formulations to find a
tions, population sizes, mutation rates) and investigated the ef- design that maximizes productivity and minimizes the risk for
fects of these GA parameters on the optimal solution (see Figs. 5 injury at the same time [12], [13], [16]. Using our new formula-
and 6). Our investigation revealed the following. The smallest tion increased the PR by 105% while not exceeding injury-risk
number of solutions evaluated by the GA before finding the op- thresholds.
timal solution was 125 [see Fig. 5(d), 25 chromosomes]. The Another advantage of our new formulation of the optimiza-
average number of solutions evaluated by the GA before find- tion problem is that it offers a more objective design, since only
ing the best solution was 305, which is only 5.6% of the 5445 the PR is considered in the objective function and all the er-
solutions examined by the two-step grid search [24]. gonomic constraints must remain under commonly used guide-
Using the hardware detailed in Section II-B, it took approxi- lines. Therefore, it is less susceptible to subjective interpretation.
mately five seconds for the automated optimization program to In comparison to the relatively objective design of the current
examine one solution. We compared the run-time of the opti- study, two subjective decisions influenced the optimal design in
mization program to the run-time of only the Jack simulation the previous studies that used the multi-objective function. The
component. The results show that 90% of the optimization run- first is the formulation of the objective function, which can be
time was attributed to the simulation in Jack. selected from a vast number of formulations and may result in
The GA’s shortest time for finding the best solution was 10 different optimal design solutions [36]. Furthermore, even after
min and 24 s. On average, the GA found the best solution in 25 choosing the function formulation, the weights that are given to
min and 24 s. the production and injury-risk measures in the objective func-
Comparison of the different mutation rates shows that the tion represent the user’s preference [12]. Thus, changing one’s
average number of solutions examined before finding the best preference for the relative importance of each of the measures
solution was 250, 275, 438, and 300 for mutation rates of 1%, (the weights) will result in very different optimal designs [16].
5%, 10%, and 15%, respectively. For the low mutation rates It is possible that users might choose to use the proposed
of 1% and 5%, population sizes of 75 and 100 chromosomes methodology with the time models already implemented in Jack.
evaluated fewer solutions before reaching the best solution [see However, the time-prediction models from [24] and Jack yielded
Fig. 6(a) and (b)]. For the high mutation rates of 10% and 15%, two different design solutions. Jack software time predictions
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HARARI et al.: SIMULATION-BASED OPTIMIZATION METHODOLOGY FOR A MANUAL MATERIAL HANDLING TASK DESIGN 7

Fig. 5. Best solution achieved versus the number of solutions (in hundreds)
examined by the GA, for different population sizes and mutation rates of (a)
1%, (b) 5%, (c) 10%, and (d) 15%.

Fig. 6. Best solution achieved for each generation of the GA, for different
population sizes and mutation rates of (a) 1%, (b) 5%, (c) 10%, and (d) 15%.
Fig. 4. Production rate as a function of the constraint threshold values. =
values for the optimal solution of the proposed methodology (see configuration
#1 in Table II). (e.g., [10], [37]), while others showed that reducing the risk
of injury will reduce productivity (e.g., [14], [16]). The results
are based on the MTM-1 method [28]. Yet, in the past, we found of the current study support the second group, since increasing
that the time-prediction models from [24] were more accurate the ergonomic thresholds (and, as a result, increasing the work-
than MTM-1 for predicting the task times in the current case ers’ risk of injury) did increase the productivity (see Fig. 4).
study. Thus, we compared the two workspace design solutions However, since several ergonomic constraints were considered,
using the time models from [24]. The workplace design obtained increasing the threshold increased the productivity only in the
with the models from [24] resulted in a 5% higher PR than range of values for which the constraint was the limiting factor.
the solution using the Jack time models. The solution using Therefore, while these findings hold for the optimal solution in
Jack time models resulted in the same RULA and similar VO2 , which the constraints are the limiting factor, it is possible that
compared to the one using the time models from [24], but it improving a poor design will improve both the productivity and
resulted in 11.7% more LBCF. Thus, the results show that using the ergonomic measures.
the Jack time models might result in a sub-optimal solution, The results of the sensitivity analysis emphasized the impor-
which could decrease the workers’ productivity or increase the tance of accurate selection of the constraint thresholds. From
values of the ergonomic measures. the physiological standpoint, while VO2cr of 1000 ml/min is
considered an acceptable threshold [23], other studies offered a
more conservative threshold of 800 ml/min [17]. The results of
B. Relationship Between Productivity and Ergonomics the current study show that, if the VO2cr were reduced to 800
In previous studies, two opposite perspectives have been pre- ml/min as suggested by [17], the PR would be lower by 20%.
sented regarding the relationship between worker productivity On the other hand, allowing 10% more oxygen consumption
and ergonomics. One group of researchers suggested that reduc- (i.e., VO2cr = 1100 ml min) would increase the productivity
ing the workers’ risk of injury will increase their productivity by 12%.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS

C. Investigation of the Genetic Algorithm Configuration VI. CONCLUSION

Common practice in the industry, and in previous studies for This paper presents an innovative framework for formulat-
HMS workplace design using DHM software, requires an expe- ing workplace design as an optimization problem that max-
rienced ergonomist or industrial engineer to manually perform imizes productivity while maintaining ergonomic assessment
the design and simulation run, and calculate the measure values values below commonly used thresholds. Furthermore, we have
and the objective function score [8], [10], [12], [13], [38]. This demonstrated an automated solution of this workplace design
is a highly time-consuming task and even for a skilled engineer methodology using a GA algorithm. This methodology offers
it takes about 15 min to examine each design [12]. Therefore, the potential for future development of tools which could be used
there is a need for a method that could solve this type of problem by practitioners (e.g., industrial engineers and ergonomists). The
faster. automation of DHM software could enable a user to evaluate
The optimization problem in this paper is classified as a dis- many possible configurations in a relatively short time (5 s to an-
crete event simulation-optimization problem [39]. A previous alyze each configuration), in comparison to manual evaluation
study presented an optimization framework for DHM work- using the software (approximately 15 min per configuration;
place design using the response surface methodology as the [12]).
optimization method [12]. In general, this method is suitable Using the new formulation of the optimization problem,
for solving this type of problem [26] and is faster than a GA. which maximizes productivity while not exceeding injury-risk
However, the HMS workplace design problem using DHM may thresholds, resulted in a design with higher productivity by
result in several local-optimum solutions in some cases. There- 105% than the previously used formulation in which the produc-
fore, methods such as the response surface methodology could tivity and ergonomics measures are combined into one objective
converge to a local optimum and “miss” the global optimum, function.
whereas using a GA increases the probability of avoiding these Applying a genetic algorithm for solving the DHM work-
local optimum solutions and finding a solution near the global place design optimization enables avoiding local optima. Fur-
optimum within a reasonable time and computational cost [26], thermore, the method found the best design within the specified
[27]. constraint conditions in a relatively short time.
Our results show that, on average, the solution using the GA
was obtained after 25 min and 24 s, which is much faster than A. Future Study
the previous two-step method (a reduction of 94.6% in com-
putation time). Yet, there is a need for more research into the The productivity and injury-risk measures used in this study
behavior of this type of objective function and the best optimiza- were selected out of a large number of possible measures. A
tion algorithm for solving this type of problem. About 90% of future study should test the influence of other measures on the
the solution time was attributed to the run-time of the simula- optimal design. In addition, in many cases, industrial settings are
tion in Jack. Therefore, adding more ergonomic constraints to more complex than the case study presented in this study (e.g.,
the formulation would probably not have a considerable effect there may be multiple workers collaborating; larger number of
on the solution time. However, changes in the granularity of workstations; larger number of design variables etc.). There-
the workplace design parameters (e.g., changing the platform fore, future work should apply this optimization methodology
height in steps of 4 cm instead of 2 cm), or the number of design to more complex workplaces and work processes. Furthermore,
parameters, will affect the solution time since it will change in this study, we have used the genetic algorithm successfully
the number of feasible solutions, and as a result influence the for solving the optimization problem. Yet, there are many other
number of simulations in Jack that will be required in order to optimization algorithms (e.g., simulated annealing, gradient de-
find the optimal solution. scent) that might be even better for these type of problems and
should be considered.

V. LIMITATIONS ACKNOWLEDGMENT
The optimization methodology presented in this study is The authors declare the following interest: the authors have a
general. Yet, the productivity and injury-risk measures were provisional patent relating to material pertinent to the submitted
selected from a vast number of possible measures. Future users article #62/648,963.
of such a methodology could choose to use different measures
such as the comprehensive lifting model [40], the comprehen- REFERENCES
sive manual handling limits for lowering, pushing, pulling, and [1] Bureau of Labor Statistics, Nonfatal Occupational Injuries and Illnesses
carrying activities [41], the NIOSH lifting index [17], or the Requiring Days Away From Work. Washington, DC, USA: U.S. Dept.
maximal acceptable weight [42]. Obviously, using other mea- Labor, 2015.
[2] National Academy of Sciences, Musculoskeletal Disorders and the Work-
sures might result in a different optimal workplace design. The place: Low Back and Upper Extremities. Washington, DC, USA: Nat.
time-prediction models used in the optimization methodology Acad. Press, 2001.
[24] were developed based on an experiment in which the range [3] R. Riemer and A. Bechar, “Investigation of productivity enhancement
and biomechanical risks in greenhouse crops,” Biosyst. Eng., vol. 147,
of box masses was 2–14 kg. Thus, using it with box masses up pp. 39–50, 2016.
to 23 kg is an extrapolation of the model and might result in an [4] P. L. Jensen and L. Alting, “Human factors in the management of produc-
inaccurate time prediction. tion,” CIRP Ann.- Manuf. Technol., vol. 55, no. 1, pp. 457–460, 2006.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HARARI et al.: SIMULATION-BASED OPTIMIZATION METHODOLOGY FOR A MANUAL MATERIAL HANDLING TASK DESIGN 9

[5] W. Zhao et al., “A human-centered activity tracking system: Toward [30] Work Practices Guide for Manual Lifting (No. 81-122), U.S. Dept. Health
a healthier workplace,” IEEE Trans. Hum.-Mach. Syst., vol. 47, no. 3, Hum. Services, Public Health Service, Centers Disease Control and Pre-
pp. 343–355, Jun. 2017. vention, Nat. Inst. Occupational Safety Health, Division of Biomedical
[6] I. Gilad and M. Elnekave, “Inserting cost effectiveness to the ergonomic and Behavioral Science, Atlanta, GA, USA, 1981.
equation when considering practical solutions: (Part II of two part paper),” [31] A. Mital, A. S. Nicholson, and M. Ayoub, A Guide to Manual Materials
Int. J. Prod. Res., vol. 44, no. 24, pp. 5415–5441, 2006. Handling. London, U.K.: Taylor & Francis, 1993.
[7] A. Jevtić, G. Doisy, Y. Parmet, and Y. Edan, “Comparison of interaction [32] D. B. Chaffin, “Some effects of physical exertion,” Research Monograph,
modalities for mobile indoor robot guidance: Direct physical interaction, Dept. Ind. Oper. Eng., Univ. Michigan, Ann Arbor, MI, USA, 1972.
person following, and pointing control,” IEEE Trans. Hum.-Mach. Syst., [33] Y. Carson and A. Maria, “Simulation optimization: Methods and applica-
vol. 45, no. 6, pp. 653–663, Dec. 2015. tions,” in Proc. 29th Conf. Winter Simul., 1997, pp. 118–126.
[8] A. Cimino, F. Longo, and G. Mirabelli, “A multimeasure-based method- [34] G. Renner and A. Ekárt, “Genetic algorithms in computer aided design,”
ology for the ergonomic effective design of manufacturing system work- Comput. Aided Des., vol. 35, no. 8, pp. 709–726, 2003.
stations,” Int. J. Ind. Ergonom., vol. 39, no. 2, pp. 447–455, 2009. [35] H. B. Maynard, G. J. Stegemerten, and J. L. Schwab, Methods-Time Mea-
[9] F. Longo and G. Mirabelli, “Effective design of an assembly line using surement. New York, NY, USA: McGraw-Hill, 1948.
modelling and simulation,” J. Simul., vol. 3, no. 1, pp. 50–60, 2009. [36] R. T. Marler and J. S. Arora, “Survey of multi-objective optimization
[10] D. Battini, M. Faccio, A. Persona, and F. Sgarbossa, “New methodological methods for engineering,” Struct. Multidisciplinary Optim., vol. 26, no. 6,
framework to improve productivity and ergonomics in assembly system pp. 369–395, 2004.
design,” Int. J. Ind. Ergonom., vol. 41, no. 1, pp. 30–42, 2011. [37] B. Das and A. K. Sengupta, “Industrial workstation design: A systematic
[11] J. P. Shewchuk, M. A. Nussbaum, S. Kim, and S. Sarkar, “Simulation ergonomics approach,” Appl. Ergonom., vol. 27, no. 3, pp. 157–163, 1996.
modeling and ergonomic assessment of complex multiworker physical [38] V. G. Duffy, Handbook of Digital Human Modeling: Research for Applied
processes,” IEEE Trans. Hum.-Mach. Syst., vol. 47, no. 6, pp. 777–788, Ergonomics and Human Factors Engineering. Boca Raton, FL, USA: CRC
Dec. 2017. Press, 2016.
[12] I. Ben-Gal and J. Bukchin, “The ergonomic design of workstations using [39] J. R. Swisher, P. D. Hyden, S. H. Jacobson, and L. W. Schruben, “A survey
virtual manufacturing and response surface methodology,” IIE Trans., of recent advances in discrete input parameter discrete-event simulation
vol. 34, no. 4, pp. 375–391, 2002. optimization,” IIE Trans., vol. 36, no. 6, pp. 591–600, 2004.
[13] D. del Rio Vilas, F. Longo, and N. R. Monteil, “A general framework [40] J. Hidalgo, A. Genaidy, W. Karwowski, D. Christensen, R. Huston,, and J.
for the manufacturing workstation design optimization: A combined er- Stambough, “A comprehensive lifting model: Beyond the NIOSH lifting
gonomic and operational approach,” Simulation, vol. 89, no. 3, pp. 306– equation,” Ergonomics, vol. 40, no. 9, pp. 916–927, 1997.
329, 2013. [41] C. Shoaf, A. Genaidy, W. Karwowski, T. Waters, and D. Christensen,
[14] F. Ore, B. R. Vemula, L. Hanson, and M. Wiktorsson, “Human–industrial “Comprehensive manual handling limits for lowering, pushing pulling
robot collaboration: Application of simulation software for workstation and carrying activities,” Ergonomics, vol. 40, pp. 1183–1200, 1997.
optimisation,” Procedia CIRP, vol. 44, pp. 181–186, 2016. [42] S. H. Snook and V. M. Ciriello, “The design of manual handling tasks:
[15] L. McAtamney and E. N. Corlett, “RULA: A survey method for the inves- Revised tables of maximum acceptable weights and forces,” Ergonomics,
tigation of work-related upper limb disorders,” Appl. Ergonom., vol. 24, vol. 34, no. 9, pp. 1197–1213, 1991.
no. 2, pp. 91–99, 1993.
[16] Y. Harari, A. Bechar, U. Raschke, and R. Riemer, “Automated simulation- Yaar Harari received the B.Sc. and M.Sc. degrees
based workplace design that considers ergonomics and productivity,” Int. in industrial engineering and management from the
J. Simul. Model., vol. 16, no. 1, pp. 5–18, 2017. Ben-Gurion University of the Negev, Beer Sheva,
[17] T. R. Waters, V. Putz-Anderson, A. Garg, and L. J. Fine, “Revised Israel, in 2014 and 2015, respectively.
NIOSH equation for the design and evaluation of manual lifting tasks,” Since 2015, he has been working on his Ph.D. the-
Ergonomics, vol. 36, no. 7, pp. 749–776, 1993. sis on investigation of worker’s biomechanics and op-
[18] J. Krüger, “Automated vision-based live ergonomics analysis in assembly timization of workplace design with the Ben-Gurion
operations,” CIRP Ann., vol. 64, no. 1, pp. 9–12, 2015. University of the Negev. His research interests in-
[19] A. Enomoto, N. Yamamoto, and T. Suzuki, “Automatic estimation of the clude occupational ergonomics, biomechanics, work-
ergonomics parameters of assembly operations,” CIRP Ann., vol. 62, no. 1, place design, and artificial intelligence.
pp. 13–16, 2013.
[20] D. B. Chaffin, “Digital human modeling for workspace design,” Rev. Hum. Avital Bechar (M’04) received the B.Sc. degree in
Factors Ergonom., vol. 4, no. 1, pp. 41–74, 2008. aerospace engineering and the M.Sc. degree in agri-
[21] M. C. Leu et al., “CAD model based virtual assembly simulation, planning cultural engineering from the Technion—Israel Insti-
and training,” CIRP Ann., vol. 62, no. 2, pp. 799–822, 2013. tute of Technology, Haifa, Israel, and the Ph.D. degree
[22] U. Raschke, B. J. Martin, and D. B. Chaffin, “Distributed moment his- in industrial engineering from Ben-Gurion University
togram: A neurophysiology based method of agonist and antagonist trunk of the Negev, Beer Sheva, Israel.
muscle activity prediction,” J. Biomech., vol. 29, no. 12, pp. 1587–1596, He is currently a Senior Research Scientist and the
1996. Head of the Department of Production, Growing and
[23] P. G. Dempsey, V. M. Ciriello, R. V. Maikala, and N. V. O’Brien, “Oxygen Environmental Engineering, Institute of Agriculture
consumption prediction models for individual and combination materials Engineering (IAE), Agriculture Research Organiza-
handling tasks,” Ergonomics, vol. 51, no. 11, pp. 1776–1789, 2008. tion (ARO),Bet Dagan, Israel. His research interests
[24] Y. Harari, R. Riemer, and A. Bechar, “Factors determining workers’ pace include robotics and automation for agriculture, human–robot collaborative
while conducting continuous sequential lifting, carrying, and lowering systems, and production techniques and methodologies for agricultural work
tasks,” Appl. Ergonom., vol. 67, pp. 61–70, 2018. processes.
[25] C. C. Gordon, T. Churchill, C. E. Clauser, B. Bradtmiller, J. T. McConville,
I. Tebbetts and R. A. Walker, “Anthropometric survey of US army per-
sonnel: Summary statistics 1988,” Anthropology Research Project Inc., Raziel Riemer received the B.Sc. degree in mechan-
Yellow Springs, OH, USA, 1989. ical engineering and the M.Sc. degree in industrial
[26] M. Jäger and A. Luttmann, “Biomechanical analysis and assessment of engineering and management from Ben-Gurion Uni-
lumbar stress during load lifting using a dynamic 19-segment human versity of the Negev, Beer-Sheva, Israel, in 1993 and
model,” Ergonomics, vol. 32, no. 1, pp. 93–112, 1989. 1997, respectively, and the Ph.D. degree in mechan-
[27] P. Brinckmann, M. Biggemann, and D. Hilweg, “Fatigue fracture of human ical engineering from the University of Illinois at
lumbar vertebrae,” Clin. Biomech., vol. 3, pp. S1–S23, 1988. Urbana-Champaign, Champaign, IL, USA, in 2007.
[28] D. Chaffin, “A computerized biomechanical model—development of and He is currently a Senior Lecturer and the Direc-
use in studying gross body actions,” J. Biomech., vol. 2, no. 4, pp. 429–441, tor of the Biomechanics and Robotics Laboratory,
1969. Department of Industrial Engineering and Manage-
[29] S. M. McGill and R. W. Norman, “Partitioning of the L4-L5 dynamic ment, Ben-Gurion University of the Negev. His main
moment into disc, ligamentous, and muscular components during lifting,” research interests include the science of human motion, and interaction with the
Spine, vol. 11, no. 7, pp. 666–678, 1986. physical environments.