International Journal of Production Research

Vol. 50, No. 4, 15 February 2012, 921937

Comparing dynamic risk-based scheduling methods with MRP via simulation

Li Suna, Sunderesh S. Heragua*, Lijian Chena and Mark L. Spearmanb
a

Industrial Engineering Department, University of Louisville, Louisville, KY 40292, USA; bFactory Physics,
Inc., 3600 East 29th Street, Bryan, TX 77802
(Received 25 May 2010; final version received 13 October 2010)

It is well known that material requirements planning (MRP) is a deterministic planning tool which heavily
relies on accurate demand forecast, and that its performance under perturbed, constantly changing
environment becomes questionable. Although many companies still use basic MRP as a planning tool, there
are other alternative tools available. For example, Factory Physics Inc. has developed a tool called dynamic
risk-based scheduling (DRS), which creates a set of policy parameters (e.g. work in process (WIP) level, lot
sizes, re-order point, and re-order quantity) that work for a range of situations. The main objective of this
paper is to compare DRS and MRP scheduling systems for single-machine and multi-machine systems via
simulation. Their performance measures are compared for various systems assuming three levels of demand
uncertainty. The simulation results suggest that DRS outperforms MRP in terms of robustness, fill rate and
inventory level.
Keywords: production planning; MRP

1. Introduction
Material requirements planning (MRP) is one of the earliest computerised production scheduling approaches, and
is still regarded as one of the most widely used systems for production scheduling (Mohan and Ritzman 1998, Hopp
and Spearman 2008). MRP got an extensive boost in 1972 because the American Production and Inventory Control
Society (APICS) launched its MRP crusade to promote its use. MRP system is designed to deliver an optimal
production plan under deterministic parameters which are best guesses or expected values. Because MRP is
essentially a deterministic decision support tool, and has some fundamental problems (i.e. the assumption of infinite
capacity and fixed lead time), its practical performance is questionable because most real-world systems experience
uncertainty at multiple levels. To improve MRPs performance, there was a trend toward the development of more
and more detailed finite-capacity scheduling systems (i.e. Manufacturing Resource Planning (MRP II), Enterprise
Resource Planning (ERP), and Advanced Planning and Optimisation (APO)). When there are demand forecast
errors, MRPs ability as well as those of its variants to handle these perturbations and therefore its usefulness is
a concern. In order to investigate the performance of MRP under perturbation, many papers have employed
simulation to assess the effects of variability and randomness in an MRP system (Grasso and Taylor 1984, Lee and
Adam 1986, Wemmerlov 1986, Ho and Ireland 1998, Yeung et al. 1998, Enns 2001).
Instead of relying on deterministic models, Factory Physics Inc. has developed the dynamic risk-based
scheduling (DRS) system, which uses analytical models to create a set of policy parameters (e.g. work in process
(WIP) level, lot sizes, re-order point and re-order quantity) that work for a range of demand scenarios to generate
the production schedule in real time. The policy depends on the actual demand and actual production, which, of
course, are constantly changing. The policy is risk-based in that it considers random events. It also assumes there is
inaccuracy in the forecast. The optimal execution obtained from DRS policy results in a manufacturing system that
is robust enough to accommodate moderate changes in demand and/or capacity without the need to reschedule.
It can also detect when the assumption regarding demand and capacity used to determine the dynamic policy is no
longer valid and indicates the need for either more or less capacity.
In many ways, DRS is an extension of the Toyota Production System (TPS) to accommodate a varied and
changing product mix. Figure 1 shows how scheduling is done in TPS using a takt time. The takt time is,
essentially, the reciprocal of the production frequency based on estimated demand. The advantage of using a takt

*Corresponding author. Email: s.heragu@louisville.edu

ISSN 00207543 print/ISSN 1366588X online
2012 Taylor & Francis
http://dx.doi.org/10.1080/00207543.2011.556152
http://www.tandfonline.com

922

L. Sun et al.

Figure 1. Toyota production system scheduling.

time is that it provides a smooth flow to the main assembly line which then propagates to any feeder lines. In other
words, the takt time prevents the intrinsic variability of demand from entering the factory thereby reducing queuing
which then reduces overall cycle time and WIP. Of course, the inherent variability in demand does not simply
disappear. Because of this variability, the demand will never (i.e. with probability one) equal the throughput of the
factory set by the takt time. If demand is higher, it must be backlogged. If demand is lower, maintaining the takt
time will create products that have no demand. There is always an implicit time buffer between the demand and the
line running according to the takt time. We call this buffer a virtual queue virtual because it does not contain any
material WIP, only information regarding new jobs to be released. By monitoring the virtual queue, Toyota would
adjust the takt time as demand changes. An increasing queue indicates falling demand and so the takt time would be
increased to reduce production. A decreasing queue would indicate the opposite with a response of decreasing the
takt time. By monitoring the virtual queue, Toyota is able to track demand and maintain production without
excessive inventory.
Interestingly, it now appears that during the recent downturn Toyota either forgot how the TPS worked or
simply ignored an increasing virtual queue and made no changes to the takt time of its factories. Consequently,
Toyota was caught with an excess of ageing inventory and ended its fiscal year in March 2009 with a $4.3 billion
loss, its first since 1950.
A second critical component of TPS is a tight coupling of the production line by limiting and controlling WIP
(e.g. kanban). If a problem arises on the assembly line, the worker is authorised to stop the line and fix the problem.
This also happens on any feeder lines. Such a practice provides the ability to track quality and other production
issues back to their source and to eliminate them (eliminating defects appears to be something else that Toyota
forgot). Of course, such problems also inevitably cause the plant to miss its takt time and, thereby, its daily
production quota. Thus, Toyota adds a make-up period to each shift to ensure that the production quota (as
established by the takt time) is not missed. The classic instance of this was two 12 hour shifts per day with only 10
hours of work scheduled for each thereby providing an extra two hours to accommodate any problems arising
during the 10 hour period. In scheduling literature, such an arrangement is known as having recourse to additional
capacity. Of course, if Toyota had stopped after adding four non-productive hours to each work day, no one would
have heard of the Toyota Production System it would have collapsed under reduced productivity. Instead, Toyota
used reduced WIP to identify root causes of problems, eliminate them, and, over time, increase the effective capacity
of the line. Improved capacity allowed shorter make-up periods and improved productivity. Thus, using a tightly
coupled process with a takt time/daily quota and the virtual queue, Toyota was able to produce with very little
inventory, high quality, and, ultimately, high productivity (low costs).
DRS takes the basic ideas of TPS and extends them to situations that are more complex than an automotive
plant driven by a main assembly line and its associated feeder lines (see Figure 2). Indeed, the motivation for DRS
was a medical device manufacturer with a low volume/high mix environment. The plant had neither a main
assembly line to pace the entire plant nor an established takt time since there was no main unit of demand (e.g.
a car) but, instead, many small complex devices.
DRS establishes its takt time and WIP control using the same mechanismCONWIPCONstant WIP which
provides an easily implemented WIP control for an environment with a rich and changing product mix. The takt
comes from the fact that CONWIP naturally accumulates WIP in front of bottleneck processes thereby resulting in a
stable flow. Work orders (jobs) arrive from both make-to-order demands and make-to-stock requirements. These
are accumulated in the virtual queue and are released when the WIP in the line falls below the CONWIP level.

International Journal of Production Research

923

Figure 2. Dynamic risk-based scheduling.

Figure 3. Setting WIP levels.

The final component of DRS is the optimal setting of the dynamic control parameters. These are the CONWIP
level, capacity trigger, along with re-order points (ROP) and re-order quantities (ROQ) for each item. Typically the
ROPs are converted into a planned lead time and safety stock which are then used by standard MRP logic.
Likewise the ROQ may be converted into a periodic order quantity in order to provide a more flexible order
quantity. However, unlike MRP, the new work orders are not released on the computed release date. Instead, the
release date is the date that the order goes into the virtual queue. This is accomplished by having a planning lead
time that includes some virtual queue time while the offset for release in the virtual queue does not. If the WIP in the
production flow is below the CONWIP level, the new order is released; otherwise it remains in the virtual queue. The
optimal WIP level is set by considering the characteristic curves for the production flow (see Figure 3). The WIP
should be high enough that the potential throughput of the line exceeds demand but not so high that cycle times are
unduly increased. The optimal planned lead times allow for some time in the virtual queue. If many orders exceed
this time, more capacity is needed. If the virtual queue becomes low, less capacity is required.
The optimisation uses a non-linear optimisation algorithm that seeks to minimise total inventory carrying cost
(WIP and stocks) along with any out of pocket costs (such as lost material during a changeover). This optimisation

924

L. Sun et al.

uses a queuing network model integrated into a stochastic inventory model to evaluate different scenarios before
choosing an optimal set of ROPs and ROQs. Thus, DRS explicitly makes use of the randomness and variability in
demand and when setting optimal dynamic parameters. DRS also provides the essential elements of TPS (WIP
control, takt time, capacity recourse) that are optimised for a rich product mix environment with no natural pacing
mechanism.
A number of papers study the comparison between MRP system and other production strategies (Kamarkar
1991, Wight 1995, Herer and Masin 1997, Benton and Shin 1998, Chase et al. 1998, Huang et al. 1998, Damodaran
and Melouk 2002, Cheraghi and Dadashzadeh 2008). This paper is designed to compare MRP with DRS strategy.
In this paper, we investigate the impact of demand uncertainty, process variability and updating frequency on the
system performance. Demand uncertainty is an important factor that affects the performance of a MRP system, and
is investigated by a number of researchers (Lee and Adam 1986, Wemmerlov 1986, Ho and Ireland 1998, Enns
2001). MRPs deterministic calculation is based on the forecast demand, and it is well known that the actual demand
cannot be predicted exactly. MRP determines the start times of the jobs by offsetting the due dates of the jobs by
fixed planned lead times, which is rarely achieved in practice due to the uncertainty in the system, e.g. process
variability. Updating frequency is another key determinant of the effectiveness of an MRP system (Yeung et al.
1998). If we update too frequently, the shop is busy constantly changing planned order releases. If we update too
infrequently, we can end up with old plans that are often out of date. Hence, in this paper, we chose forecast errors,
process variability and updating frequency as the factors to be investigated to compare MRP system with DRS
system under different demand and operating conditions. In order to improve the system performance, we also
introduce several constraints into the models, such as introducing capacity and recourse constraints into MRP. We
then compare the MRP and DRS under these constraints. The objective is to demonstrate via simulation the
conditions under which static and dynamic models are preferred. By comparing the model performance under
different scenarios, we not only predict which policy will be better under a specific situation, but we also quantify the
impact of uncertainty and variability in the systems for different policies.
It should be noted that there are numerous additional papers that address uncertainty. Some of these are
referenced here. The primary source of uncertainty is the stochastic demand which has been extensively studied
(Louveaux and Thisse 1985, Iida 2002, Panda et al. 2006, Inderfurth 2009, Feng et al. 2010). In particular, detailed
reviews on demand uncertainty and buffer method are summarised by Srivastava (2000) and Koh and Saad (2003).
Inderfurth (2009) discussed the techniques to control risks in demand uncertainty, and Koh and Saad (2003) and
Koh et al. (2002) presented solution techniques for problems with stochastic demand. Gupta and Brennan (1995)
consider stochastic supply as well as stochastic demand. A simulation based optimisation technique for MRP is
proposed by Grasso and Taylor (1984).
This paper compares the key performance measures, e.g. fill rate, inventory, and backorder, between DRS and
MRP systems via simulation. By comparing the key system performance measures, we are able to evaluate the
policys performances when there is uncertainty and variability in the system. The rest of the paper is structured as
follows. We present the simulation modelling of MRP and DRS system in Section 2, and show results from our
simulation models in Section 3. We present a case study of a DRS system in Section 4. We conclude in Section 5 by
demonstrating the conditions under which DRS could outperform the MRP.

2. Simulation modelling of MRP and DRS systems

2.1 MRP simulation modelling
The MRP system procedure essentially consists of three main steps. The first step is to determine the net
requirements by deducting the on-hand inventory and any scheduled receipts from the demand. The next step is to
divide the net requirements into appropriate lot sizes to form jobs. The last step is to determine start times of the
jobs by offsetting the due dates of the jobs by planned lead times. Correspondingly, three components are included
in an MRP model. In the first component, we update the inventory position based on forecast demand for a
planning time horizon. Then we calculate the net requirement for each product deterministically based on the
updated inventory. We then calculate the planned order receipts and determine the planned order release by taking
into consideration the production lead times (see Figure 4). Table 1 lists the notation used in our calculations.
Consider a set of forecast demand, denoted as Di(t), t 0, . . . , n, where Di t equals the forecast demand for
product i in period t. Note that Di(0) is the sum of the demands before the first period. The period in our model is
one day and the default planning horizon is four weeks.

925

International Journal of Production Research

Forecast
demand

Update
inventory
position

Update net
requirement

Determine
planned order
receipts

Determine
planned order
releases

Figure 4. Planning procedure in an MRP model.

Table 1. Notation.
Notations
Di t
IPi t
Di 0
wi t
NRi t
POi t
PWi t

Forecast demand for product i in period t

Projected inventory position for product i in period t
Demands due before the first period for product i
Work orders for product i in period t
Net requirements for product i in period t
Planned order receipts for product i in period t
Planned work orders for product i in period t

The projected inventory position in period t, IPi(t) is computed as:

IPi t IPi t  1  Di t

IPi 0 wi 0  Di 0

wi(0) is the current on-hand inventory. The next step is to get the net requirement for each product in each
period, which is the demand beyond what the on-hand inventory and the scheduled receipts can cover. NRi(t) is
calculated as



3
NRi t min Di t, max 0,  IPi t
Equation (3) makes the net requirement equal to the magnitude of the first negative projected inventory or the
demand for the period, whichever is smaller. Then we compute the planned order receipts POi(t). In our model we
use re-order quantity Q as the lot size. Thus the planned order release is an integer multiple of Q.
We determine the planned work orders by taking into consideration the production lead times. We calculate the
planned lead time l by dividing re-order point by average daily demand. Then, the planned work orders, PWi(t) are
given by
PWi t POi t l

The second part of the MRP model triggers production based on the order release plan schedule, which is
illustrated in Figure 5. First, we check the production plan each day to see if there are any planned order releases for
the products. If so, we trigger a production and update the inventory level.
The third part of the MRP system is about examining the actual order demand to update the inventory level,
which is illustrated in Figure 6. When a demand is realised, we update the inventory level and compare it to the
re-order point. If the inventory level is greater than the reorder point, the order is filled directly from stock.
Otherwise, it becomes a backorder.
To improve the MRP performance, several constraints are introduced into the MRP model, such as capacity
constraints, and recourse policy. In the basic MRP model, the production scheduling plan is calculated based on
demand forecast without the consideration of capacity. It is clear that the incorporation of capacity constraints is
more realistic. However, it also makes the problem a constrained optimisation problem which is usually
computationally demanding. Turning to the original MRP model, in addition to calculating the net requirements
NRi(t) in each period, we can also assume known production capacity Ci(t) in each period. Therefore, we need to
find a feasible solution for planned work orders PWi(t) subject to the constraints,
PWi t  Ci t

926

L. Sun et al.
Check daily
production plan

Yes

Planned order
release = 0?

Dispose

No
Update
inventory
level

Trigger
production

Update OH
inventory

Figure 5. Production triggers in an MRP model.

Order
arrives

Update
inventory
level

Inventory >
ROP?

No
Dispose

Yes
Fill order

Figure 6. Order generation in an MRP model.

To obtain a production schedule with consideration of capacity constraints, we must first obtain planned work
orders PWi(t) based on the original MRP procedure. Then we compare PWi(t) with capacity Ci(t). If it is greater
than the available capacity, we release the excess production plan PWi(t)  Ci(t) into subsequent days which have
excess capacity; hence we get new planned work orders PWi(t), which does not violate the capacity constraints.
Recourse is another useful way to improve the performance. After we calculate the planned order release based
on the MRP procedure, we compare planned work orders PWi(t) with capacity Ci(t). If the former is greater than
the latter, recourse is introduced into the model, which means that a second-shift is introduced into the system. It is
obviously a way to decrease the cycle time and improve the customer service.

2.2 DRS simulation modelling

In DRS model, the production is triggered based on the actual demand. When a demand occurs, the inventory level
is updated. If the updated inventory level is greater than the re-order point, the order is filled. If not, a production is
triggered. The number of batches of production is calculated as:


AvailableInventory  ROP


6


ROQ
de is the smallest integer greater or equal to . The model is illustrated in Figure 7.
For the DRS model, CONWIP and recourse constraints are also introduced to improve its performance.
CONWIP constraint is illustrated in Figure 8. CONWIP constraint is introduced to control the production process.
After we trigger an order production, we check the WIP to see if it is less than the WIP cap which is set as a constant
according to the system capacity. If the current WIP is less than the WIP cap, the order can be released into the
production line directly. Otherwise, the order is held outside the production line until the WIP drops below the WIP
cap. In this case, by limiting order releases into the system, the CONWIP controls WIP and hence results in a lower
average WIP level. This also makes cycle time shorter due to Littles Law.
Introducing CONWIP into the system also brings another benefit for monitoring the status of the system.
We monitor the items waiting outside the production line which is like a virtual queue. When we apply recourse
to improve the system performance, a second shift can be easily scheduled based on the size of the virtual queue.
We check the queue outside the production line at the end of each day (a period). If the number in the virtual queue

927

International Journal of Production Research

Order
Arrives

Update
Inventory
Level

Yes

Inventory Level
> ROP ?

Fill Order

Dispose

No
Trigger Production
and Update
Inventory Level

Production
Process

Update
On-hand
Inventory

Figure 7. Production process in a DRS system.

Trigger Order
Production

Check WIP

yes

WIP<WIP

Cap?

no

Production
Process

Dispose

Hold Order
Outside until
WIP < WIP Cap

Figure 8. CONWIP constraint in a DRS system.

is greater than the limit cap which is set up based on the system capacity, a second shift would be scheduled.
Otherwise, the process line produces items in one shift.

3. Simulation experiments and results

We begin with a single-machine problem and then extend to a multiple machine problem later. In the single-machine
case, the system contains only one machine to produce 23 parts. In the multiple-machine problem, the system has
multiple work centres and each of the work centres consist of multiple machines. Multiple products need to be
produced in the system as well and each product needs to go through multiple work centres. For all the MRP and
DRS models in this section, we run 100 replications of the simulation for 10 years with a one-year warm-up period.
We use Rockwell Arena 10.0 as the simulation software on an x-86 workstation.

3.1 Single machine problem

In the single-machine problem, we consider three levels of demand scenarios and two levels of process variability in
the system to compare the MRP model with the DRS model. In the first demand scenario, the actual order demands
follow a Poisson distribution with the mean demand equal to the predicted order demands; which is called the base
scenario. The second (over-estimated demand) scenario assumes that the forecast demand is over-estimated by 20%
for each product, which means that the actual mean demand for each product is 20% lower than the forecast
demand and is called the over-estimated scenario. For each demand scenario, we consider two levels of process
variability in the system. One is with moderate variability and the other is without variability. With moderate
variability, the number of actual orders per period follows a Poisson distribution, the order size is normally
distributed with truncation, and the process time is exponentially distributed. In the no-variability scenario, all the
parameters, including the number of orders per period, order size, and process time, are constant. The specific data
for the problems is from real industries provided by Factory Physics Inc., as shown in Table 2.
To compare the performance between MRP and DRS strategies, we mainly focus on the following
measurements:
.
.
.
.
.

Fill rate based on time: the fraction of time the system does not have backorders.
Fill rate based on units: the fraction of demand (based on units) that will be filled from stock.
Average backorder.
Average inventory position.
Average on-hand inventory.

Fill rate based on time and fill rate based on units both represent a reasonable definition of customer service.
The higher the fill rate, the greater is the expected customer satisfaction. Backorder is another measure of customer

928

L. Sun et al.

Table 2. Input data.

ItemID

NrOrders

VarOrders

AvgOrdSize

StDevOrdSize

OnHandQty

ROQ

ROP

Part1
Part2
Part3
Part4
Part5
Part6
Part7
Part8
Part9
Part10
Part11
Part12
Part13
Part14
Part15
Part16
Part17
Part18
Part19
Part20
Part21
Part22
Part23

16.5826
9.8444
97.4500
203.7260
112.5884
31.9135
46.2692
51.9272
6.3083
1.9786
8.2923
125.3411
36.0400
157.1427
183.6258
3.3139
41.3034
54.3913
14.6444
18.6798
14.0424
7.7521
8.7343

16.5826
9.8444
97.4500
203.7260
112.5884
31.9135
46.2692
51.9272
6.3083
1.9786
8.2923
125.3411
36.0400
157.1427
183.6258
3.3139
41.3034
54.3913
14.6444
18.6798
14.0424
7.7521
8.7343

230
90
260
500
260
370
130
660
120
750
130
170
50
1800
310
790
880
230
270
1090
1250
230
320

46
18
52
100
52
74
26
132
24
150
26
34
10
360
62
158
176
46
54
218
250
46
64

0
1399
4094
37622
4782
1892
2184
5087
630
0
2023
5951
1505
70660
11732
4796
11336
3446
2493
2958
3319
3488
1614

9890
9990
10370
25920
19620
4930
7260
7480
7510
4930
4930
9000
4760
80640
10080
9780
18140
7510
7510
7510
9890
4930
4930

519
145
2142
7483
2396
1315
604
3358
145
455
188
1712
199
21751
4249
649
3781
1204
560
2681
2538
317
479

*NrOrders, average number of orders per period (one month); VarOrders, variance of the number of orders; AvgOrdSize,
average order size for each part; StDevOrdSize, standard deviation of order size; OnHandQty, onhand inventory; ROQ, re-order
quantity; ROP, re-order point.

satisfaction, which is better when it is lower. Inventory incurs a holding cost, so a lower inventory is preferred. In
an ideal situation, we want the fill rate to be as high as possible to meet customer demand, while maintaining as low
inventory as possible. We compare the overall average performance measurements (i.e. fill rate based on time,
backorder, fill rate based on units and inventory) for the six scenarios. Figure 9 shows a summary of the
comparison.
We find that the performance measures show consistent results for the DRS model among the three scenarios for
the system without variability (i.e. base scenario, over-estimated scenario, under-estimated scenario), which means
that DRS model is very robust even with a forecast bias. While comparing the results for MRP model among the
three scenarios, we find that the performance measures change significantly according to the forecast bias: overestimated demand causes a big increase in inventory position level (i.e. from 9525 in base scenario to 10,504 in overestimated scenario), and under-estimated demand results in a dramatic decrease in fill rate (i.e. from 93.65% fill rate
based on time in base scenario to 76.77% in under-estimated scenario). We can therefore conclude that MRP model
is not robust for the forecast bias.
Comparing the performance between DRS and MRP models, we find that the DRS system outperforms MRP
in the base scenario in terms of higher fill rate and lower inventory position level (except in the system without
variability where the DRS provided slightly lower fill rate based on units than an MRP system). For the overestimated forecast scenario, DRS obtained slightly lower fill rate based on units, but a much lower inventory
position, which still reflects a relatively better performance. For the under-estimated scenario, DRS appears to
outperform MRP with much higher fill rates and relatively small increase in inventory.

3.2 Multiple-machine problem

3.2.1 DRS results
We extend our study to multiple-machine problems. Several constraints are introduced into the models capacity
and recourse constraints into the MRP and CONWIP models and recourse constraints into a basic DRS model.
MRP and DRS models are built with four work stations and three products. The first work station has four

929

International Journal of Production Research

Fill rate based on units

Fill rate based on me

100.00%

100.00%

95.00%

95.00%

90.00%

90.00%

85.00%

85.00%

80.00%

80.00%

75.00%

75.00%

70.00%

70.00%
under-esmated

base scenario

over-esmated

Backorder

5000
4500
4000
3500
3000
2500
2000
1500
1000
500
0

under-esmated

11000

base scenario

over-esmated

Inventory Posion

10500
10000
9500
9000
8500
8000
under-esmated

base scenario

over-esmated

under-esmated

base scenario

over-esmated

On-hand inventory
9000
8500
8000
7500
7000
6500
6000
5500
5000
4500
4000
under-esmated

base scenario

over-esmated

Figure 9. Comparison summary for single machine problem.

machines and the others all have three machines each. The order inter-arrival time is exponentially distributed.
The order size is constant. All the three parts need to be processed by all the four work stations. The process times
are exponentially distributed. We develop the MRP and DRS models and compare their performance. As in the
single machine problem, three demand scenarios are considered (i.e. base scenario, over-estimated and underestimated scenarios).
We find that the bottleneck resource (Workcenter 2) in the base scenario is utilised 89.04%. Because the
utilisation is relatively high, if the forecast error increases, the bottleneck resource utilisation also increases
significantly and the production line build-up begins to increase even more. This occurs even if we only increase the
forecast error by 10%. Figure 10 shows that the comparison summary for the performance of the DRS model
among base scenario, 10% under-estimated forecast scenario and 10% over-estimated scenario. We find that
forecast error does not affect inventory significantly when compared to the base scenario. For the over-estimated
scenario, cycle time decreases compared to the base scenario, resulting in lower backorder and higher fill rate.
However, for the under-estimated scenario, cycle time increases dramatically compared to base scenario. Because
the bottleneck resource utilisation increases to 97.38% in the under-estimated scenario, which is rather high, we can
reasonably conjecture that high bottleneck utilisation causes a long waiting line in the system, making the cycle time
long, backorder high, and fill rate low.
We add a CONWIP constraint to the basic DRS model where we use the average WIP level from the base
scenario as a cap on the inventory. Figure 11 compares the results obtained in the under-estimated, over-estimated
and base scenarios. Cycle time is the same as that in the basic model and it represents the time between when a job is
released until it reaches the end point of the production line. The other labelled as CT O represents the cycle time
plus the time the job waits outside the production line in the virtual queue. Note that it waits in the virtual queue
when the current WIP level is greater than the WIP cap.
From the results, we find that the inventory level is similar to that in the basic CONWIP model. For the underestimated scenario (where the actual demand is 10% greater than the forecast demand), we find that cycle time plus
the waiting time outside the production line increases dramatically which explains why the backorder level increases
and fill rate decreases significantly. This is much worse when compared with the basic DRS model. From the

930

L. Sun et al.
Inventory posion

Backorder

260

120

250

100

240

80

230

60

220

40

210

20

200

0
under-esmated

base scenario

over-esmated

base scenario

under-esmated

Cycle me

over-esmated

Fill rate based on me

500

400

0.8

300

0.6

200

0.4

100

0.2
0

0
under-esmated

base scenario

base scenario

under-esmated

over-esmated

over-esmated

On-hand inventory

160
140
120
100
80
60
40
20
0

under-esmated

base scenario

over-esmated

Figure 10. Comparison summary for DRS.

Backorder

Inventory posion

260

5000

258

4000

256

3000

254

2000

252

1000

250

0
under-esmated

base scenario

DRS

over-esmated

under-esmated

DRS with CONWIP

DRS

Cycle me
8000

6000

0.8

over-esmated

DRS with CONWIP

Fill rate based on me

0.6

4000

0.4

2000

0.2
0

0
under-esmated
DRS

150

base scenario

base scenario

DRS with CONWIP

over-esmated

CT O for DRS with CONWIP

under-esmated
DRS

base scenario

over-esmated

DRS with CONWIP

On-hand inventory

100
50
0
under-esmated
DRS

base scenario

over-esmated

DRS with CONWIP

Figure 11. Comparison summary DRS versus CONWIP.

difference between cycle time and CT O, we can conclude that the production batches spend a significant time
waiting outside the production line in the virtual queue which results in a deterioration of the overall performance.
The long external waiting times result from the CONWIP constraints, where we use the same WIP cap as in the
basic scenario while the larger actual demand increases the number of production batches waiting outside the

931

International Journal of Production Research

Inventory posion
260
250
240
230
220
210
200

100
50
0
under-esmated

base scenario

DRS with recourse

over-esmated

under-esmated

DRS

base scenario

DRS with recourse

over-esmated
DRS

Fill rate based on me

Cycle me
1

500
400
300
200
100
0

0.8
0.6
0.4
0.2
under-esmated

base scenario

over-esmated

Cycle me for DRS with recourse

CT O for DRS with recourse
DRS

150

Backorder

150

under-esmated

base scenario

DRS with Recourse

over-esmated
DRS

On-hand inventory

100
50
0
under-esmated

base scenario

DRS with Recourse

over-esmated
DRS

Figure 12. Comparison summary DRS versus DRS with recourse.

production line. We conclude that the CONWIP cap must be chosen carefully. An inappropriate number can
worsen the performance. We add the CONWIP constraint in the basic DRS model for the over-estimated scenario
(where the actual demand is 10% lower than forecast demand). We find that the results are similar to the basic
model, which can be explained by the closeness of the cycle time and CT O values.
From the results of CONWIP constraint, we find that CONWIP by itself is not effective in improving the
performance. However, when it is used along with a recourse constraint, wherein additional capacity, for example a
second shift, is added to the production line dynamically whenever there is an excessive number of jobs waiting
outside the network, its performance improves. Because the CONWIP cap represents a reasonable capacity of the
production line, monitoring the number of waiting items outside the production line in the virtual queue could be an
acceptable way to decide whether or not a second shift is needed. In our model, we set up a cap for the waiting items
outside the production line, which could be a signal to schedule the second shift. When the number of the waiting
items outside the production line is greater than the cap, a second shift is scheduled.
Figure 12 shows the results for the base scenario, under-estimated scenario and over-estimated scenario for the
recourse constraint model. We find that for the base scenario and over-estimated scenario, the recourse constraint
does not result in an obvious improvement when compared to the basic DRS model. This can be interpreted as
follows. Without scheduling the second shift, the basic DRS model already had a good performance. This means
that the original capacity was already able to meet the demand and recourse is not necessary. In the under-estimated
scenario, the backorder decreases and fill rate increases dramatically compared to the basic model. So we can
conclude that recourse based on CONWIP constraint is an effective way to check when the production line capacity
cannot meet the production requirements. The average inventory in Figures 10 to 15 refers to the average inventory
position. The average inventory position is about the same for all three demand scenarios in all the DRS models
because the increase (decrease) in on-order amounts for the under-estimated (over-estimated) scenario is offset by a
corresponding increase (decrease) in backorder quantities, making the average inventory position about the same in
steady state.
3.2.2 MRP results
We now consider the MRP model and test the effect of updating frequency on the MRP model. We consider three
different updating frequencies (i.e. update once each week, every two weeks, and every four weeks). We also
introduce capacity, and recourse constraints into the MRP model.

932

L. Sun et al.

Inventory posion

Backorder

300

200

280

150

260

100

240

50

220
200

0
under-esmated

base scenario

MRP 1WK

MRP 2WK

over-esmated

under-esmated

MRP 4WK

MRP 1WK

Cycle me

base scenario

over-esmated

MRP 2WK

MRP 4WK

Fill rate based on me

550

450

0.8

350

0.6

250

0.4

150

0.2
under-esmated

base scenario

MRP 1WK

MRP 2WK

over-esmated

under-esmated

MRP 4WK

MRP 1WK

base scenario
MRP 2WK

over-esmated
MRP 4WK

On-hand inventory
200
150
100
50
0
under-esmated

base scenario

MRP 1WK

MRP 2WK

over-esmated
MRP 4WK

Figure 13. Comparison results for MRP models with 1 week, 2 week and 4 week updating frequency.

Inventory posion

Backorder

290
280
270
260
250
240
230
220
210
200

200
150
100
50

under-esmated

base scenario

over-esmated

0
under-esmated

Cycle me

base scenario

over-esmated

Fill rate based on me

550

500

0.9

450

0.8

400

0.7

350

0.6

300

0.5

250

0.4

200

0.3

150
under-esmated

base scenario

over-esmated

0.2
under-esmated

base scenario

over-esmated

On-hand inventory
200
180
160
140
120
100
80
60
40
20
0
under-esmated

base scenario

over-esmated

Figure 14. Comparison summary for MRP models with an upper bound on capacity and different updating frequencies.

Figure 13 shows the comparison summary for MRP when the update is done each week, every two weeks, and
every four weeks, for the base, under-estimated and over-estimated scenarios. We compare the three scenarios and
conclude that forecast error affects the performance of the MRP models. When the actual demand is higher than the
forecast demand (i.e. under-estimated scenario), the cycle time and backorder increase dramatically, hence the fill

933

International Journal of Production Research

Inventory posion

Backorder

290

200

280
270

150

260
250

100

240
230

50

220
210
200

0
under-esmated

base scenario

over-esmated

under-esmated

Cycle me

base scenario

over-esmated

Fill rate based on me

550

500

0.9

450

0.8

400

0.7

350

0.6

300

0.5

250

0.4

200

0.3

150

0.2
under-esmated

base scenario

over-esmated

under-esmated

base scenario

over-esmated

On-hand inventory
200
180
160
140
120
100
80
60
40
20
0
under-esmated

base scenario

over-esmated

Figure 15. Comparison summary for MRP with recourse.

rate decreases dramatically. When the actual demand is lower than the forecast demand, the cycle time and
backorder decrease, inventory increases, and hence fill rate improves.
We also find that the updating frequency affects the MRP performance. In the two-week update model,
the inventory is lower than in the one-week update model in the base scenario. However, the fill rate is also lower
than in the one-week update model. So it is hard to conclude which one is better. In the under-estimated scenario,
the inventory is lower in the two-week update model than in the one-week scenario which indicates better
performance. However, the fill rate is also lower. In the over-estimated scenario, the inventory is higher than in the
one-week update model and the fill rate is similar. We can say that in the over-estimated scenario, the one-week
update model is better than the two-week update model. A comparison between the four-week update model and
the two-week update model yields similar observation as in the comparison between the two-week update and the
one-week update model.
In the basic MRP model, capacity constraints are not considered. The implicit assumption is that the production
line has infinite capacity. Because this assumption is not reasonable, we introduce capacity constraints into the basic
MRP model, which is what we did with the CONWIP constraint in the DRS model. According to the WIP cap in
the CONWIP DRS model, we set up an upper bound on capacity for each part type in the MRP model, which is
calculated based on some specific proportion of CONWIP cap. For example, part 1 capacity is calculated by:
1
 WIP Cap
1 2 3

where 1, 2 and 3 represent the arrival rates of parts 1, 2 and 3.

We add the MRP capacity constraint to the basic MRP model, considering all three scenarios (base, underestimated and over-estimated scenarios) with three different updating frequencies.
Figure 14 summarises the results. The comparison among the three scenarios for the MRP with capacity
constraint is similar as in the basic MRP model. In the base scenario, the capacity constraint results in a little shorter

934

L. Sun et al.
Table 3. Results for the base scenario under different policies.

DRS
DRS CONWIP
DRS CONWIP recourse
MRP
MRP capacity
MRP recourse

Average fill rate

Average inventory position

89.72%
80.34%
91.83%
88.49%
89.00%
89.45%

218.77
218.79
218.79
222.85
223.05
225.22

cycle time, hence a little higher inventory level, lower back order and a little higher fill rate compared with the basic
model, which shows that the chosen capacity upper bound is effective. Since the capacity upper bound is calculated
based on right forecast (base scenario), when the demand is under-estimated, apparently the capacity upper bound
is not effective. It may even worsen the performance resulting in a much higher cycle time, hence higher backorder.
In the over-estimated scenario, there is no obvious difference between the model with capacity constraint and the
basic model. The capacity constraint does not have much of an effect.
As in the DRS model, we introduce the recourse constraint into the basic MRP model. As discussed in the
previous section, we use the capacity upper bound of each part type as the signal to schedule a second shift. Each
day, we check the daily production release plan based on the basic MRP model. If the release plan is greater than the
capacity upper bound, a second shift is scheduled. The performance should improve. In this case, we also build the
model for all three scenarios with three different update frequencies.
Figure 15 summarises the comparison. The comparison among the three scenarios for the MRP with recourse
constraint is similar as in the basic MRP model. Since the production planning is based on the right forecast (base
scenario), the recourse does not have much effect for the base scenario, and only results in slightly better
performance in yielding a slightly lower cycle time, hence slightly lower backorder, and higher fill rate. While in the
under-estimated scenario, when the production requirements are more than the forecast, the resource model is
effective in reducing the cycle time, which results in a much lower backorder and higher fill rate. In the overestimated scenario, there is no obvious difference between the basic model and the model with the recourse
constraint. Because there is no need for excess production when the actual demand is lower than the forecast, the
system seldom schedules the second shift.

3.2.3 Summary
We summarise the comparison results for the multiple-machines case. We built the DRS and MRP models under
different scenarios (i.e. base scenario, under-estimated scenario and over-estimated scenario). We focused on two
performance measures for the comparison, inventory position representing the cost, and fill rate representing
customer service. Table 3 shows the aggregate results for the three parts for the base scenario under different
policies. Here the MRP models are all based on one-week updating frequency.
From the example, we can say that CONWIP is not necessarily an effective way to improve the performance,
which even worsens the performance compared to the basic DRS model resulting in lower average fill rate and
higher average inventory, but it is an effective way to provide recourse constraints. DRS with CONWIP and
recourse constraints obtains the highest fill rate and relatively low inventory which is only higher than the basic
DRS model among all the policies. Even the basic DRS model outperforms all MRP model in terms of higher fill
rate and lower inventory.
Table 4 compares the results among basic DRS model and basic MRP models with different updating frequency.
Overall, the DRS system obtains better performance than MRP systems resulting in higher fill rate and lower
inventory (except slightly higher inventory in under-estimated scenario compared to MRP with one-week update).
Comparing MRP performance among different update frequency, we find that the update frequency affects the
MRP performance along with the forecast error. The more frequent the update, the higher the fill rate and the
inventory. Overall the DRS system outperforms the MRP system.

935

International Journal of Production Research

Table 4. Results for the basic DRS model and the basic MRP models with different updating frequency.
Base scenario

DRS
MRP 1wk
MRP 2wks
MRP 4wks

Under-estimated scenario

Over-estimated scenario

Average
fill rate

Average
inventory

Average
fill rate

Average
inventory

Average
fill rate

Average
inventory

89.72%
88.49%
87.21%
84.50%

218.77
222.85
219.05
216.20

37.23%
26.86%
21.51%
20.55%

218.80
216.48
204.93
185.62

99.19%
99.12%
99.01%
98.94%

218.79
230.46
234.05
243.26

Figure 16. Product flows and CONWIP.

Case study Medical spinal devices manufacturer

A client of Factory Physics Inc. called the latter to help improve the current low on-time delivery rates and to
reduce the existing long lead times. The process involved the fabrication of numerous components, purchase of some
common components, and the assembly and test of the final device. The current situation was to start production of
components using MRP (MAPICS). Because most jobs were late, the manufacturing crib became a pass through
point instead of a controlled stock. After analysing the situation, FPI recommended increasing and controlling the
crib inventory in order to assemble to order the high volume devices. This strategy would greatly reduce lead
times and improve on-time delivery. To facilitate these changes, the DRS production control system would be
employed.
The plant was divided into product flows defined by the machining cell at the beginning of the line
(see Figure 16). Each flow was controlled using DRS in the following way:
(1) Once per month, new CONWIP levels, planned lead times, safety stocks, and ROQ parameters would be
generated using two modules, flow optimiser (see Figure 3) to establish WIP levels and the cash flow
optimiser to optimise the ROP and ROQ for each part of a software package now known as CSUITE.
(2) The optimal parameters would be loaded into MAPICS to generate new work orders.
(3) The work orders would go into the virtual queue of each flow and be released when the WIP in the flow fell
below the CONWIP level.
(4) If the virtual queue exceeded a set level, additional capacity would be scheduled on the weekend.
The results of the implementation were a decrease in WIP inventory and a corresponding increase in crib
inventory. However the service levels rose from around 50% to more than 98%. Moreover during this period,

936

L. Sun et al.

Figure 17. Results (service level).

demand increased and the lead time to the customer was significantly reduced. However, the implementation
was not without glitches. After repeated assurances by the client that the bottleneck would always be the machining
cell, the FPI team became suspicious when cycle times did not decrease as predicted. It was discovered that the
actual bottleneck was the QC (quality control) operation, a bad place for a bottleneck. When this bottleneck
was broken by increasing the staff, the system began to respond as predicted and the improvements followed
(see Figure 17).

4. Conclusion
In this paper, we compared a new production strategy, dynamic risk-based scheduling (DRS) developed by Factory
Physics Inc., with the traditional MRP-based strategies via simulation. We began with a single-machine problem.
We compared the performance between MRP and DRS for varying levels of uncertainty in forecast demand (i.e.
base scenario, under-estimated scenario and over-estimated scenario) and different levels of variability in the system
(i.e. moderate variability and without variability). We found that in the system with moderate variability, the DRS
model had better performance than the MRP model resulting in:
(1) Higher fill rate and lower inventory in the base scenario.
(2) Higher fill rate with slightly higher inventory in under-estimated scenario.
(3) Much lower inventory with slightly lower fill rate in over-estimated scenario.
In the system without variability, the DRS model had better performance than the MRP model in terms of:
(1) Higher fill rate and much lower inventory in the base scenario.
(2) Much higher fill rate and slightly higher inventory in under-estimated scenario.
(3) Higher fill rate and much lower inventory in over-estimated scenario.
In addition, the DRS model obtained similar performance under the three scenarios which told us that the DRS
model is more robust when there is forecast error.
When the experiment was extended to the multiple-machine cases, we introduced more constraints into both the
DRS and MRP models to improve their performance. The constraints were CONWIP, capacity, and recourse.
We also tested the performance of the MRP models under different updating frequencies. We found that
CONWIP, without a virtual queue and capacity recourse, is not an effective way to improve the system
performance, but an effective way to obtain recourse constraints. Introducing capacity or recourse constraints
into the basic MRP model improved its performance, but the DRS model with CONWIP and recourse constraints
yielded better performance.

International Journal of Production Research

937

References
Benton, W. and Shin, H., 1998. Manufacturing planning and control: The evolution of MRP and JIT integration. European
Journal of Operational Research, 110 (3), 411440.
Chase, R.B., Aquilano, N.J., and Jacobs, F.R., 1998. Production and operations management: Manufacturing and services.
Columbus, OH: Irwin/McGraw-Hill.
Cheraghi, S.H. and Dadashzadeh, M., 2008. Comparative analysis of production control systems through simulation. Journal
of Business & Economics Research, 6 (5), 87103.
Damodaran, P. and Melouk, S., 2002. Comparison of push and pull systems with transporters: a metamodelling approach.
International Journal of Production Research, 40 (12), 29232936.
Enns, S.T., 2001. MRP performance effects due to lot size and planned lead time settings. International Journal of Production
Research, 39 (3), 461480.
Feng, K., Rao, U., and Raturi, A., 2010. Setting planned orders in master production scheduling under demand uncertainty.
International Journal of Production Research, 48 (1), 119.
Grasso, E. and Taylor, B., 1984. A simulation-based experimental investigation of supply/timing uncertainty in MRP systems.
International Journal of Production Research, 22 (3), 485497.
Gupta, S. and Brennan, L., 1995. MRP systems under supply and process uncertainty in an integrated shop floor control
environment. International Journal of Production Research, 33 (1), 205220.
Herer, Y.T. and Masin, A.M., 1997. Mathematical programming formulation of CONWIP based production lines and
relationships to MRP. International Journal of Production Research, 35 (4), 10671076.
Ho, C.J. and Ireland, T.C., 1998. Correlating MRP system nervousness with forecast errors. International Journal of Production
Research, 36 (8), 22852299.
Hopp, W.J. and Spearman, M.L., 2008. Factory physics. 3rd ed. New York, NY: McGraw-Hill.
Huang, M., Wang, D., and Ip, W.H., 1998. A simulation and comparative study of the CONWIP, Kanban and MRP production
control systems in a cold rolling plant. Production Planning & Control, 9 (8), 803812.
Iida, T., 2002. A non-stationary periodic review production-inventory model with uncertain production capacity and uncertain
demand. European Journal of Operations Research, 140 (3), 670683.
Inderfurth, K., 2009. How to protect against demand and yield risks in MRP systems. International Journal of Production
Economics, 121 (2), 474481.
Karmarkar, U., 1991. Push, pull and hybrid control schemes. Tijdschrift voor Economie en Management, 26 (3), 345363.
Koh, S., Saad, S., and Jones, M., 2002. Uncertainty under MRP-planned manufacture: review and categorisation. International
Journal of Production Research, 40 (10), 23992421.
Koh, S. and Saad, S., 2003. MRP-controlled manufacturing environment disturbed by uncertainty. Robotics and ComputerIntegrated Manufacturing, 19 (12), 157171.
Lee, T.S. and Adam, E.E., 1986. Forecasting error evaluation in material requirements planning (MRP) production inventory
systems. Management Science, 32 (9), 11861205.
Louveaux, F. and Thisse, J.-F., 1985. Production and location on a network under demand uncertainty. Operations Research
Letters, 4 (4), 145149.
Mohan, R.P. and Ritzman, L.P., 1998. Planned lead times in multistage systems. Decision Sciences, 29 (1), 163191.
Netessine, S. and Rudi, N., 2003. Centralised and competitive inventory models with demand substitution. Operations Research,
51 (2), 329335.
Panda, G., Khan, D. and Ray, U., (2006). A jels stochastic inventory model with random demand [online]. Stochastic
Programming E-Print Series. Available from: http://www.speps.org [Accessed 15 December 2009].
Srivastava, R., 2000. A review of techniques for buffering against uncertainty with MRP systems. Production Planning & Control,
11 (3), 223233.
Wemmerlov, U., 1986. A time-phased order-point system in environments with and without demand uncertainty: a comparative
analysis of non-monetary performance variables. International Journal of Production Research, 24 (2), 343358.
Wight, O.W., 1995. Manufacturing resource planning: MRP II: Unlocking Americas productivity potential. Hoboken, NJ:
John Wiley and Sons.
Yeung, J.H.Y., Wong, W.C.K., and Ma, L., 1998. Parameters affecting the effectiveness of MRP systems: a review. International
Journal of Production Research, 36 (2), 313332.

Copyright of International Journal of Production Research is the property of Taylor & Francis Ltd 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.