Card based production planning and control:

(1)

Card based production planning and control 1

Card based production planning and control:

Investigating workload balancing capability of load-based pull

systems

Fiodor Bodnar S2136120

Master thesis, MSc Technology & Operations Management University of Groningen, Faculty of Economics and Business

June 20, 2016

(2)

Abstract

Current literature is lacking when it comes to addressing workload balancing effectiveness of load-based pull system. In order to test to what degree load-load-based cards contribute to the effective workload balancing of pull systems discrete event simulation is used. This research compares performance of unit- and load-based m-CONWIP against the performance of the CONWIP under different experimental factors. Previous research has found CONWIP not to be effective in terms of reducing total throughput times, thus having no workload balancing capability. Therefore, CONWIP’s performance can be used as a performance benchmark. This paper demonstrates that addition of load-based cards proves to be valuable for workload balancing effectiveness. However, only under specific long/ short order ratio.

(3)

Thus far, much of the research on production planning and control was dedicated to effective improvement of pull production control systems and development of techniques for selection of pull systems in different environments. For instance, Gaury, Kleijnen & Pierreval (2001) propose a novel approach to design customized pull control systems for single product flow production lines. González-R & Framinan (2009) build on the idea of ‘customized pull systems’ which was introduced by Gaury et al. (2001). Their paper formalizes a new way of material flow control by customized token based systems which is found to perform better than traditional CONWIP. As for the pull system selection strategies, for example Gaury, Pierreval & Kleijnen (2000) present an evolutionary approach to choose among different pull systems. Whereas Khojasteh & Sato (2015) introduce a framework for a pull production control system selection in multi-stage production processes. When it comes to performance, lower congestion levels, easier control and work in process capping are the three main underlying reasons for improved performance of pull production control systems (Spearman, Woodruff, & Hopp, 1990).

(5)

that ideally demand has to be met from the stock. The proportion of demand that is fulfilled from the stock is commonly referred to as service level. Unfortunately, neither of these performance measures are useful in a high variety and customization make to order (MTO) environment which we are interested in. Germs & Riezebos (2010) state that in such environments throughput time performance is of great importance as firms focus on fast order fulfillment as means of gaining a competitive advantage.

In order to realize shorter throughput times in an MTO setting, a pull system has to possess an effective workload balancing capability. Land & Gaalman (1998) define the workload balancing capability as an ability of the system to evenly distribute the work among the workstations on the shop floor. Several studies have addressed the workload balancing capability of pull production control systems i.e. CONWIP, m-CONWIP and POLCA (Germs & Riezebos, 2010; Ziengs, Riezebos, & Germs, 2012). In both instances, the pull systems used were unit-based, meaning that the workload was limited by controlling the amount of orders on the shop floor. Alternatively, the shop floor workload could be limited basing on the processing time requirements of orders and such systems are referred to as load-based pull systems.

(6)

workload balancing. To put it differently, such systems are able to reduce the time between order release and completion by means of reducing queue lengths, which results from better control of order arrival time at the workstations. This in turn might also reduce the time between order arrival and its completion (Land & Gaalman, 1996). Therefore, this makes for an interesting case to investigate how a load-based mechanism adds to the effective workload balancing capability of these systems in a high variety and customization setting (Ziengs et al., 2012).

From the practical standpoint the load-based pull systems might be more complicated in implementation and use, which makes them more expensive. Implementation requires an additional system, which assesses processing time requirements of orders and subsequently checks whether required capacity is available. An additional personnel training, which would provide an explanation on how the system works, is also needed. Accordingly, from the theoretical standpoint this paper will aid in understanding of how a load-based order release mechanism contributes to the effective workload balancing capability of pull systems. It is important to realize that in case load-based pull systems are able to balance the workload effectively better than their unit-based counterparts, an improved throughput time performance will be realized. As for the managerial relevance, if a significant throughput time reduction can be achieved by implementing the load-based pull system, it will be worthwhile to incur extra investments over the unit-load-based system. This brings us to the following research question.

(7)

The paper is structured as follows: section 2 reviews the background literature, section 3 discusses the methodology used, section 4 presents the results followed by a discussion and, finally, section 5 summarizes the main results, as well as comments on limitations of the paper and provides suggestions for further research.

2. Background 2.1 Production control systems

In the late seventies lean practices were being employed and popularized by Toyota while American firms were predominantly using material resource planning (MRP) systems which were promoted by the American Production and Inventory Management Association (Hopp & Spearman, 2004). Since lean approach helped Toyota to achieve an impressive competitive advantage over competitors and significantly increase their market share, many firms became interested in lean manufacturing. Pull production control system being one of the tools employed by the lean philosophy advocates received most attention from both academia and practitioners (Kabadurmus, 2009).

(8)

worth noting that Kanban cards became associated with pull systems, while MRP became synonymous with push. The clear distinction between the two is provided by Hopp & Spearman (2004):

A pull production system is one that explicitly limits the amount of work in process that can be in the system. By default, this implies that a push production system is one that has no explicit limit on the amount of work in process that can be in the system. (p. 142)

During the late 20th_{century many different extensions of pull systems have been developed} (Liberopoulos & Dallery, 2000). The most notable include Constant WIP (CONWIP; Spearman et al., 1990) and paired-cell overlapping loops of cards with authorization (POLCA; Suri, 1998). Even though the idea of restricting WIP in pull control systems is the same, there are two ways to implement control mechanisms. They will be discussed further once two different aspects of pull systems that are distinguished by Gaury (2000), namely, structure and configuration, are introduced.

(9)

Route-specific control uses cards to assign products to specific routings, e.g. POLCA and m-CONWIP structures. At the same time, non-route-specific control, such as m-CONWIP structure, assigns cards without taking into account product type and routing.

Pull system configuration refers to the number of cards which is placed in each of the control loops (Gaury, 2000). Gaury (2000) also provides an overview of available techniques for selection of the optimal configuration which include (1) empirical formulas, (2) optimization based on analytical models and (3) optimization based on simulation models. An in-depth discussion of selection techniques is beyond the scope of this paper. It is worth pointing out that if the number of cards in each loop is considerably large, the system does not limit the workload and can be used to represent a pure push system which releases orders into production the moment they arrive. Riezebos (2010) also makes a distinction between two configuration aspects, that is, between product-specific and product-anonymous control. Product-specific type of control assigns cards to a specific product category. However, due to the requirement of a buffer for each product type such control is not well suited to environments with a vast variety of products and/or routings. In contrast with product-specific control, an order in product-anonymous control proceeds downstream regardless of the product type or routing as soon as another order leaves the system and card becomes available. Both CONWIP and m-CONWIP are examples of product-anonymous control, with the only difference concerning their routing specificity.

(10)

research. We have already established that unit-based systems are more common due to their simplicity in use (Land & Gaalman, 1998). In such systems workload is limited basing on the number of orders regardless of processing time requirements of this order. Whereas load-based equivalents limit the workload by taking into account processing time requirements of orders, i.e. load-based POLCA (Vandaele, Van Nieuwenhuyse, Claerhout, & Cremmery, 2008). Workload control can be implemented in a variety of different ways. Land & Gaalman (1996) distinguishes between three different concepts that are addressed in literature: Becthte’s WLC concept (Bechte, 1994), Bertrand’s WLC concept (Bertrand & Wortman, 1981) and Tatsiopoulos WLC concept (Tatsiopoulos, 1993). We are not going to discuss these concepts in detail as this is beyond the scope of this research. Instead, we will address main benefits of workload control systems. It all boils down to the control of average queue length by limiting the number of orders on the shop floor, which results in shorter and more predictable time between order release and completion. Consequently, time between order arrival and its completion might also be reduced. Therefore, implementation of a load-based mechanism will show us whether effectiveness of pull systems’ workload balancing can be improved in an MTO environment.

2.2 Workload balancing capability

(11)

This in turn reduces the required average queues in front the workstations to obtain a predetermined utilization level. Under these circumstances the time between order release and departure, which is also referred to as shop floor throughput time (STT), is decreased. Furthermore, this might also result in a reduction of the time between order arrival and departure from the production system, which is called total throughput time (TTT).

There is little research about the workload balancing when card-based pull systems in the context of the MTO production setting are concerned. Even when the balancing capability of pull systems is addressed in literature (Germs & Riezebos, 2010; Ziengs, Riezebos & Germs, 2012) only unit-based pull systems are used. It seems counter intuitive that unit-based systems have been researched to a larger extent than load-based ones, especially when the following fact is considered. Even though unit-based systems reduce the workload imbalance, once processing time variability is introduced the workload balancing capability of unit-based systems drastically decreases (Germs & Riezebos, 2010).

(12)

STT along with the average order pool time (OPT) constituted TTT (Land, Oosterman. & Gaalman, 2000). Figure 1 presents a graphical representation of the order flow through a production line. Bearing this in mind, the system was only deemed capable of effective workload balancing when a decrease in STT was accompanied by a reduction in TTT when compared to a non-restricted system. To put it differently, an increase in OPT did not have to offset a drop in STT for a system to possess effective workload balancing.

Figure 1. Throughput time components of a production system with controlled release (adopted from Land & Gaalman, 1996).

(13)

workload balancing capability. Hence, the magnitude of the effective workload balancing is also dependent on pull structure and configuration. Comparatively to POLCA, m-CONWIP obtained a larger STT reduction. It is a direct result of POLCA’s overlapping loops nature that are not able to perfectly detect and send information about imbalance in the workload distribution upstream (Germs & Riezebos, 2010). In case of m-CONWIP, its loops do cover each individual routing from start to an end, therefore information is communicated upstream without any obstacles along the way.

(14)

Figure 2. Performance of a pull system in terms of throughput times (adopted from Ziengs et al.,

2012)

The following section discusses additional questions that have to be addressed in order to answer the main research question.

2.3 Additional sub-questions

(15)

based on the outcomes of Germs & Riezebos (2010) and it shows that non-route-specific unit-based CONWIP has no workload balancing capability whereas route-specific unit-unit-based m-CONWIP possesses the aforementioned capability. Next, question marks represent yet to be uncovered effects of load-based control systems on the workload balancing capability.

Table 1

Workload balancing capabilities of unit- and load-based pull systems

Pull systems

Product-anonymous type of control Non-route-specific Route-specific

Card type CONWIP m-CONWIP

Unit-based - +

Load-based ? ?

Note. Divergent production setting. + and - signs represent presence or lack of workload balancing capability, respectively.

(16)

influence effectiveness of the aforementioned capability. Last, but not least, once the load-based m-CONWIP is simulated, the combined effect of load-based cards and the route-specific structure on workload balancing will be checked. If it turns out to be greater than the one of unit-based m-CONWIP, it would lead us to believe that load-based cards contribute to the effect the structure has on effectiveness of the workload balancing capability. However, if not, we would have to conclude that the effective workload balancing capability of m-CONWIP can be solely attributed to its route-specific structure. Based on the discussion above, additional sub-questions to be addressed in order to answer the main research question are as follow:

Do load-based cards have an effect on the workload balancing capability of CONWIP?

Does route-specific structure of unit-based m-CONWIP have an effect on the workload balancing capability?

What is the effect of load-based cards on the workload balancing capability of route-specific m-CONWIP?

(17)

3. Methodology 3.1 Research method

There is a variety of modelling techniques available, i.e. simulations, experiments with the real systems and other modelling approaches, which range from simple paper calculations to extremely complex mathematical programming and heuristic methods. Since the nature of operations systems is subject to variability, complexity and interconnectedness a modelling technique, which is to be used in this paper, needs to be able to cope with these factors. Unlike other methods, simulation models are able to deal with variability, complexity and interconnectedness of systems extremely well (Robinson, 2004, pp. 4-6). This allows simulation models to forecast performance of the system, compare alternative system designs and determine how system performance is affected by alternative policies. Furthermore, there are also four additional advantages of simulation models when compared to experimentations with real systems. These include 1) cost, 2) time, 3) experimental condition control and 4) existence of real systems (Robinson, 2004, p. 8).

Due to the above discussed reasons, effects of pull systems’ structure and configuration based characteristics on effectiveness of the workload balancing capability will be measured using a discrete event simulation. It is also worth pointing out that discrete event simulation is the most commonly used method for operations systems modelling (Robinson, 2004, p. 33). Next, tools which will aid in constructing the discrete event simulation are presented.

(18)

There are three types of software for simulations development: spreadsheets, programming languages and specialist simulation software. In table 2 a comparison of these approaches is presented.

Table 2. A comparison of spreadsheets, programming languages and specialist simulation software for simulation modelling

Feature Spreadsheet Programming language Specialist simulation software

Range of application Low High Medium

Modelling flexibility Low High Medium

Duration of model build Medium Long Short

Ease of use Medium Low High

Ease of model validation Medium Low High

Run-speed Low High Medium

Time to obtain software skills

Short (medium for macro use)

Long Medium

Price Low Low High

Note. Adopted from Robinson, 2004, p. 42.

(19)

process-Card based production planning and control 19

based discrete event simulation library, it can simulate highly complex operation systems. The following section presents the simulation model design.

3.2 Model design

Figure 3 presents a divergent topology of interest and pull structures of CONWIP and m-CONWIP systems.

Figure 3. Divergent production topology (adapted from Germs & Riezebos, 2010)

(20)

seven stations, ranging from 1 to 7. At each consecutive stage of the divergent production line the number of workstations doubles. The number of routings is determined by looking at the number of workstations at the last stage of the topology. As can be seen from figure 3, the 3rd stage of the production system consists of four workstations, namely, 4, 5, 6 and 7. This results in four different routings, the occurrence of which is equally likely. This in turn means there are four m-CONWIP loops (1:4; 1:5; 1:6; 1:7). Each of these pairs of numbers comprises one m-CONWIP loop where the first number represents a starting station and the second one stands for an end station, i.e. 1:4 loop means the starting station is 1 and the end station is 4. As for the CONWIP loop, it covers all stages and is referred to as 1:[4,5,6,7]. This loop differs from m-CONWIP loops in a sense that there is one starting station, which is represented by the first number, 1, but an end station can be any of stations 4, 5, 6, or 7. Furthermore, for both systems the routing becomes known immediately once the order arrives. The decision making within the simulation model is governed by the flowchart which is presented in appendix B.

(21)

the structures of control systems used, workload can only be balanced before it enters production. Since unit- and load-based card types are used, it is worth addressing the card allocation process as it differs between the two. First, the unit-based card allocation process is discussed which is followed by explanation of the load-based card allocation.

In the unit-based system once the order arrives, a corresponding routing is already known. Processing time requirements are also known, however, the system of this type does not make use of them. Let’s assume the order has to proceed through control loop 1:4. Before it can enter production the order waits for card 1:4, where 1 stands for the first workstation and 4 for the last one in that route, to become available. Once card 1:4 is available, it gets assigned to the order which is waiting to be processed at the specific 1:4 routing. After the order is processed at workstation 4 and leaves the system, card 1:4 gets detached and is available for the following order.

(22)

processed at all workstations within a route under constant processing times, that is, one unit of processing time at the first station in the route and two, four at stations two and three, respectively. The second card needs to be attached once the order’s processing time requirement exceeds this threshold. The cards are detached right after the order’s processing at all workstations of the route is completed and become available for the following orders.

(23)

tend to prioritize orders that have shorter processing time requirements and have to be processed at less congested routes, thereby balancing the workload among different routings.

Order release mechanisms can also differ in terms of its complexity. In this case, a 3-card load-based mechanism will be used in order to test whether additional performance improvement can be realized. The underlying idea for testing such a release mechanism is based on the fact that once up till three cards can be attached instead of two, the production system might become more flexible in terms of being able to balance the workload more efficiently among the routings. From figure 4 we can see the Erlang-2 distribution of processing times and respective percentage probabilities of orders falling within one of three ‘buckets’ of processing times. The order receives the first card if its processing time does not exceed 4.2. Two cards are attached if the processing time requirements fall within 8.1 units of time. Finally, three cards get attached to the order if the processing time requirements exceed 8.1 units of time.

(24)

Orders with two cards will be prioritized over the orders with three cards for less congested routes. As in the simple two-card release mechanism, orders requiring one card are going to be prioritized over orders that need two. When the order leaves the production system all the cards are detached and can be used again. In the next section experimental factors of the production system are introduced.

3.2 Experimental design

(25)

queue lengths in front of workstations, thereby improving choice of orders in front of the control loops. This in turn improves the workload balancing capability of pull systems (Germs & Riezebos, 2010).

Table 3

Experimental factors

Factor Experimental levels

Structure

Type of structure Conwip, m-Conwip

Configuration characteristics Type of cards

Number of cards Order arrival pattern

Unit-based, Load-based 1-20, ∞

Inter-arrival time Constant

Utilisation 90%

Batch size

Processing time variability Processing time

1

Constant, random (Erlang-2)

(26)

This leaves us with two sets of experiments, namely CONWIP and m-CONWIP, for which we use a full factorial design for the combination of all above mentioned experimental factors. This leaves us with eight individual experiment where each contains 20 independent experiments with a run length of 10,000 time units and consists out of 100 runs. The warm up period of 2,500 time units is selected in order to reduce the initial transient thus allowing the system to reach its steady state. In order to produce sufficient amount of accurate data, 100 runs are going to be performed. Both the number of runs and the warm up length are conservatively selected. Moreover, Welch’s (Heidelberger & Welch, 1983) procedure proves that sufficient warm-up time is chosen. The graphical method confirms that number of runs selected is sufficient. More details on both procedures can be found in appendix C. Next, STT and TTT of each experiment are going to be recorded. Subsequently, the optimal total throughput times of simulated systems will be used to answer the research questions of this paper.

Table 4

Simulation model of two control systems: CONWIP and m-CONWIP

Parameters (1:[4,5,6,7]) (1:4) (1:5) (1:6) (1:7)

Conwip n ∞ ∞ ∞ ∞

m-Conwip ∞ n n n n

Note. One parameter represents one loop which can include multiple workstations i.e. (1:4 comprises W1; W2; W4). Parameter (1:[4,5,6,7]) stands for CONWIP loop. Value n represents the number of cards in the corresponding loop.

(27)

4. Results and Discussion

This section presents outcomes of the simulation experiments. It analyses effectiveness of the workload balancing capabilities of CONWIP and m-CONWIP production control systems.

Table 5

Optimal throughput time performance of CONWIP and m-CONWIP.

Constant inter-arrival time Constant inter-arrival time

Unit-based Load-based

CONWIP m-CONWIP CONWIP m-CONWIP

Batch size

Utilization %TTT %STT %TTT %STT %TTT %STT %TTT %STT

Constant processing time

1 90% 0.00 0.00 5.44 52.06 0.00 0.00 5.44 52.00

Random processing time

1 90% 2.60 2.95 1.40 93.12 2.60 2.78 1.30 92.45

Note. Values in the table represent percentage decreases in TTT and STT under different experimental factors.

(28)

on effectiveness of the workload balancing capability. Since we are also interested in how a more complicated release mechanism performs in terms of TTT, the performance of a mechanism which can attach to the order up to three cards instead of two will also be evaluated.

(29)

(30)

Figure 5. Unit- and load-based M-CONWIP performances under constant processing times compared against the benchmark.

In contrast with Germs & Riezebos (2010) who observe no performance improvement when the unit-based m-CONWIP is simulated under random processing time, we do observe 1.4% reduction in TTT, which is caused by the route-specific m-CONWIP structure. Further, this effect needs to be corrected by 2.6% due to above presented reasons. This results in an increase in TTT by 1.2%. With this in mind, we can conclude that the unit-based m-CONWIP balances workload, as when corrected, STT is reduced by 90.2%. However, the workload balancing capability is not effective as TTT is not reduced when compared to CONWIP performance. Next, with the processing time variability the load-based m-CONWIP reduced TTT by 1.30%, which was by 0.1% lower than TTT performance of the unit-based counterpart. However, once corrected TTT

21 21,2 21,4 21,6 21,8 22 22,2 22,4 22,6 10 12 14 16 18 20 22 24 TTT STT

(31)

was actually increased by 1.3%. Similar to the case of the unit-based m-CONWIP, load-based one possesses workload balancing capability, as when corrected STT was reduced by 89.7%, but it was not effective because TTT was not decreased. For illustrative purposes, both unit- and load-based m-CONWIP performances under random processing times are shown in figure 6.

Figure 6. Unit- and load-based M-CONWIP performances under random processing times compared against the benchmark.

Before we delve into discussion of a possible reason for such an outcome we first assess whether more complex release mechanism yields better results.

In order to test whether a more complex release mechanism yields a better performance, we compare the performance of a simple release mechanism with two cards against the one which can attach up to three cards. As it can be seen from the table 6, the three-card release mechanism

3400 3500 3600 3700 3800 3900 4000 4100 4200 4300 0 500 1000 1500 2000 2500 3000 3500 TTT STT

(32)

does not yield any additional performance over the one with two cards. One of the possible reasons for such an outcome might be the fact that there are still orders that could cause blocking of routings for shorter orders, thus limiting the system’s ability to improve performance. Also, an arbitrary selection of processing time ‘buckets’ as shown in figure 4 could have contributed to order blocking.

Table 6

Throughput time performance of unit- and load-based m-CONWIP

Unit-based m-CONWIP Card count 20 17 15 13 11 9 7 5 TTT 3472.09 3460.01 3462.83 3475.81 3473.74 3489.42 3499.88 3544.62 STT 269.45 247.01 222.45 193.23 175.77 150.94 128.22 92.35 Load-based m-CONWIP with 2 card release mechanism

Card count

25 22 21 20 19 18 15 10

TTT 3464.83 3464.36 3464.31 3464.07 3464.39 3464.21 3465.44 3467.98 STT 309.24 282.89 273.95 265.01 255.80 246.39 217.58 162.12 Load-based m-CONWIP with 3 card release mechanism

Card count

25 22 21 20 19 18 15 10

TTT 3464.83 3464.36 3464.31 3464.07 3464.39 3464.21 3465.44 3467.98 STT 309.24 282.89 273.95 265.01 255.80 246.39 217.58 162.12

(33)

Next, order release mechanisms will be addressed, which will also help to understand how blocking occurs.

(34)

Figure 7. Decision making in the current implementation of the load-based order release mechanism.

(35)

replicate the benefit of the current implementation, that is, to balance workload among routes, but would also provide an additional benefit of balancing workload within the routes (figure 8).

Figure 8. Decision making in the potential implementation of the load-based order release mechanism.

(36)

how current implementation works in section 3.2. If, for example, there is only one 1:4 card available and there is an order somewhere in the queue which requires only one 1:4 card, it will be prioritized and released into production earlier than the order which needs two cards. Perhaps, using this variation of a load-based order release mechanism would allow m-CONWIP system to yield improved TTT performance as compared to the current implementation of the release mechanism.

4.1 Post hoc analysis

In order to better understand the nature of the load-based m-CONWIP performance under random processing times and constant inter-arrival times, additional experiments will be performed.

(37)

As shown in figure 9 the initial load-based order release mechanism attached the second card to an order once processing time requirements exceeded seven units of time. In that case 40% of total orders required attachment of the second card. Figure 10 presents TTT performances of the load-based m-CONWIP system under two different long to short order ratios. With the initial ratio the load-based m-CONWIP was not capable to balance workload effectively, which lead to 1.3% increase in TTT. This potentially happened due to long blocking orders. However, when the order ratio is changed so that only 20% of total orders are orders that require two cards to be attached, performance improvement of 7.78% in TTT is observed, which, once corrected, equals to 5.18%. Therefore, we can conclude that the load-based m-CONWIP is capable of balancing workload better than its unit-based equivalent, however, only under a certain ratio of long and short orders.

Figure 10. Load-based m-CONWIP performance under different two different order mixes

3000 3200 3400 3600 3800 4000 4200 4400 0 500 1000 1500 2000 2500 3000 3500 TTT STT

(38)

5. Conclusion

Companies in the MTO industry rely on short throughput times as means of gaining an advantage over their competitors. Pull production control systems are designed to achieve shorter throughput times by limiting the amount of workload on the shop floor. Even though work in process (WIP) restriction results in shorter shop floor throughput times (STT), the time orders spend in the order pool (OPT) increases as a consequence of order blocking. In order to reduce the total throughput times (TTT) pull systems have not only to restrict the WIP, but also to improve the balance of workload on the shop floor. Thus far, academic literature has only paid attention to effectiveness of workload balancing capability of unit-based variants of pull systems. In addition to that, when load-based pull systems were addressed, the workload balancing capability was not among the discussed topics.

(39)

TTT performance was not improved. In the load-based case, the most likely reason for not being able to improve TTT was order blocking which resulted from implementation of the current order release mechanism. Next, TTT performance of the load-based m-CONWIP was assessed using the smaller ratio of long to short orders. In this case the load-based order release mechanism improved TTT performance by a sizeable margin. Therefore, we can conclude that even though load-based prioritization has a downside of blocking orders with implementation of the current order release mechanism, it can be counteracted by specifying a different ratio of long to short orders.

As for limitations of this paper, increasing the number of stages and subsequently finding an optimal configuration would pose a challenge. Currently, an exhaustive search strategy is employed and it would require extremely big computational expenses to continue using such a strategy for optimum selection. Further, it is important to realize that the basic version of the workload balancing mechanism was implemented and that more complex workload concepts could potentially be more effective in terms of workload balancing.

(40)

References

1. Bechte, W. (1994), “Load-oriented manufacturing control just-in-time production for job shops”, Production Planning & Control, Vol. 5(3), pp. 292–307.

http://doi.org/10.1080/09537289408919499

2. Bertrand, J. W. M., & Wortmann, J. C. (1981), Production control and information systems for component manufacturing shops. Elsevier Science Inc.

3. Gaury, E. (2000), Designing pull production control systems: Customization and

robustness.Tilburg University, School of Economics and Management.

4. Gaury, E. G. a, Kleijnen, J. P. C., & Pierreval, H. (2001), “A methodology to customize pull control systems”, Journal of the Operational Research Society, Vol. 52(7), pp. 789– 799. http://doi.org/10.1057/palgrave.jors.2601153

5. Gaury, E. G. A., Pierreval, H., & Kleijnen, J. P. C. (2000), “Evolutionary approach to select a pull system among Kanban, Conwip and Hybrid”, Journal of Intelligent Manufacturing, Vol. 11(2), pp. 157–167. http://doi.org/10.1023/A:1008938816257

6. Germs, R., & Riezebos, J. (2010), “Workload balancing capability of pull systems in MTO production”, International Journal of Production Research, Vol. 48(8), pp. 2345–2360.

http://doi.org/10.1080/00207540902814314

7. Goldrat, E. M., & Fox, R. E. (1986), The race. New York: North River Press.

(41)

9. Heidelberger, P., & Welch, P. D. (1983), “Simulation run length control in the presence of an initial transient” Operations Research, Vol. 31(6), pp. 1109–1144. http://doi.org/10.1287/opre.31.6.1109

10. Hopp, W. J., & Spearman, M. L. (2004), “To pull or not to pull: what is the question?”

Manufacturing & Service Operations Management, Vol. 6(2), pp. 133–148.

http://doi.org/10.1287/msom.1030.0028

11. Kabadurmus, O. (2009), “A comparative study of POLCA and generic CONWIP production control systems in erratic demand conditions”, IIE Annual Conference.

Proceedings, p. 1197.

http://proquest.umi.com/pqdweb?did=1848728371&Fmt=7&clientId=47883&RQT=309 &VName=PQD

12. Khojasteh, Y., & Sato, R. (2015), “Selection of a pull production control system in multi-stage production processes”, International Journal of Production Research, Vol. 53(14), pp. 4363–4379. http://doi.org/10.1080/00207543.2014.1001530

13. Land, M., & Gaalman, G. (1996), “Workload control concepts in job shops a critical assessment”, International Journal of Production Economics, Vol. 46-47, pp. 535–548. http://doi.org/10.1016/S0925-5273(96)00088-6

14. Land, M. J., & Gaalman, G. J. C. (1998), “The performance of workload control concepts in job shops: Improving the release method”, International Journal of Production

Economics, Vol. 56-57, pp. 347–364. http://doi.org/10.1016/S0925-5273(98)00052-8

(42)

on workload balancing”, International Journal of Production Economics, Vol. 68(1), pp. 107–119.

16. Liberopoulos, G., & Dallery, Y. (2000), “A unified framework for pull control mechanisms in multi‐stage manufacturing systems”, Annals of Operations Research, Vol. 93(1-4), pp. 325–355. http://doi.org/10.1023/A:1018980024795

17. Paris, J.-L., & Pierreval, H. (2001), “A distributed evolutionary simulation optimization approach for the configuration of multiproduct Kanban systems”, International Journal of

Computer Integrated Manufacturing, Vol. 14(5), pp. 421–430.

http://doi.org/10.1080/09511920010021784

18. Python. (2016). Retrieved from http://www.python.org

19. Riezebos, J. (2010), “Design of POLCA material control systems”, International Journal

of Production Research, Vol. 48(5), pp. 1455–1477.

http://doi.org/10.1080/00207540802570677

20. Robinson, S. (2004), Simulation: the practice of model development and use. Palgrave Macmillan. http://doi.org/10.1057/palgrave.jos.4250031

21. Spearman, M. L., Woodruff, D. L., & Hopp, W. J. (1990), “CONWIP - A pull alternative to Kanban”, International Journal of Production Research, Vol. 28(5), pp. 879–894. 22. Sugimori, Y., Kusunoki, K., Cho, F., & Uchikawa, S. (1977) “Toyota production system

and Kanban system materialization of just-in-time and respect-for-human system”,

International Journal of Production Research, Vol. 15(6), pp. 553–564.

(43)

23. Suri, R. (1998), “Don't push or pull-POLCA”, APICS, The Performance Advantage, Vol. 8, pp. 32-39.

24. Tatsiopoulos, I. P. (1993), “Simplified production management software for the small manufacturing firm”, Production planning & control, Vol. 4(1), pp. 17-26.

25. Vandaele, N., Van Nieuwenhuyse, I., Claerhout, D., & Cremmery, R. (2008), “Load-based POLCA: an integrated material control system for multiproduct, multimachine job shops”,

Manufacturing & Service Operations Management, Vol. 10, pp. 181–197.

http://doi.org/10.1287/msom.1070.0174

26. Welcome to SimPy. (2016). Retrieved

from https://simpy.readthedocs.io/en/latest/index.html

27. Ziengs, N., Riezebos, J., & Germs, R. (2012), “Placement of effective work-in-progress limits in route-specific unit-based pull systems”, International Journal of Production

(44)

Appendix A

Literature search is performed by doing the backward and forward search based on the González-r, P. L., Framinan, J. M., & Pierreval, H. (2012). Token-based pull production control systems: an introductory overview. Journal of Intelligent Manufacturing, 23(1), 5-22. Forward search yielded 14 papers, while backward search produced 76 papers. Our inclusion criteria were as follows: (1) use of discrete event simulation to conduct the research and (2) papers that discussed pull production control systems. Based on the inclusion criteria 4 out of 14 and 23 out of 76 papers were selected from forward and backward search, respectively. Next, selected papers were coded as follows:

Literature coding

Topic Labels Code

Production control systems

- Production control system description - Production planning

1A 1B

Pull systems - CONWIP

- KANBAN - HYBRID

- CONWIP comparison with other systems - POLCA 2A 2B 2C 2D 2E

Push system - Push system 3

Generic system - Generic model - Simulation - Evolutionary algorithm 4A 4B 4C Workload balancing capability - Workload balancing

- Processing and inter-arrival variability

5 5A

(45)

Appendix B

The following figure describes the decision making during a production run.

Appendix C I. Run length selection

(46)

Figure 1. Run length selection

II. Warm-up time selection

Welch procedure calculates and plots the moving averages. As we can see from the graph the line levels out at around 800 units of time. Given the fact that 2500 is selected as the warm up time we can safely conclude that it is sufficient enough to remove the initial transient.

Card based production planning and control: