Investment in Production Resource Flexibility:

Transcription

1 Investment in Production Resource Flexibility: An emirical investigation of methods for lanning under uncertainty Elena Katok MS&IS Deartment Penn State University University Park, PA William Tarantino Center for Army Analysis Fort Belvoir, VA Terry P. Harrison MS&IS Deartment Penn State University University Park, PA August 2001 Corresonding author

2 Investment in Production Resource Flexibility: An emirical investigation of methods for lanning under uncertainty Elena Katok, William Tarantino and Terry P. Harrison We examine several methods for evaluating resource acquisition decisions under uncertainty. Traditional methods may underestimate equiment benefit when art of this benefit comes from decision flexibility. We develo a new, ractical method for resource lanning under uncertainty, and show that this aroach is more accurate than several commonly used methods. We successfully alied our aroach to an investment roblem faced by a major firm in the aviation information industry. Our recommendations were acceted and resulted in estimated annual savings in excess of $1 million (US). Keywords: manufacturing flexibility, stochastic rogramming, and samling In recent years, many firms have found it increasingly imortant to invest substantially in technology to maintain a cometitive edge. Technological imrovements often require suerior roduction methods, and some firms find themselves constantly evaluating oortunities for investments in new roduction resources. These decisions can easily become crucial to survival in a cometitive market lace. While essential to the well being of firms, roduction investment decisions are extremely difficult because they involve lanning under uncertainty. For examle, when a new roduction resource rovides manufacturing flexibility, the benefit of this flexibility can be easily underestimated. As Jordan and Graves [16] oint out: in caacity and flexibility lanning, investment costs for flexible oerations are tyically quantified; however, it is less common to quantify the benefits because demand uncertainty is not exlicitly considered by the lanners. Since flexibility is exensive, this tyically results in decisions not to invest in it. The benefits of a new roduction resource can emerge in three ways: 1. Lower cost due to suerior erformance 2. Increased caacity 3. Increased decision flexibility. -1-

3 The first two sources of benefit are fairly intuitive: cost savings may result if a new resource rovides a more efficient roduction rocess or introduces a new dedicated rocess. If a new resource is added to the current roduction system at a articular stage, caacity at that stage may increase. If that stage reviously formed a bottleneck, the throughut of the entire system increases (Goldratt and Cox [12]), otentially yielding cost savings. The third source of benefit comes from increased decision flexibility (Benjaafar et al. [2]). Decision flexibility is the ability to ostone decisions until more information is obtained. When a new roduction resource is added to the current system, it can increase decision flexibility by either roviding additional caacity where it is needed, or by roviding an additional routing for a art. To correctly estimate the imact of a flexible resource, a model must include all three sources of benefit. With this study we contribute to manufacturing flexibility lanning research in three ways: We describe a new and better method that accurately accounts for all three sources of flexibility benefit, and that is ractical enough to be used for large and comlex real-world roblems. We imlement and use the new method to hel with a real flexibility-lanning decision faced by Jeesen Sanderson, Inc. (Jeesen), a manufacturing comany, and generate annual savings in excess of $1 million. We clearly demonstrate, using the Jeesen investment roblem as a case study, how other commonly used methods can consistently under-estimate the benefit of flexibility. This aer is organized as follows. First, we describe the roblem and summarize relevant literature in Section 1. In Section 2, we develo a formal mathematical model for flexibility lanning and show analytically that some commonly used methods generally under-value flexibility. We also describe our new samling-based otimization algorithm that assesses the benefit of manufacturing flexibility more accurately than existing methods. In Section 3, we -2-

4 describe the flexibility-lanning roblem faced by Jeesen and the alication of our method to this roblem. In Section 4 we demonstrate that alternative methods significantly under-estimate the benefit of flexibility at Jeesen. We discuss the imact of our work at Jeesen in Section 5 and summarize this work s contributions in Section Problem Descrition Flexibility lanning has been studied extensively during the last decade, and here we do not attemt to rovide an exhaustive survey. For a summary of flexibility categories and measures see Sethi and Sethi [21] or Guta and Goyal [14]. Recent frameworks for flexibility lanning san the sectrum from qualitative and descritive (Gerwin [11]), to urely theoretical (degroote [8], van Mieghem [28]), to emirical (Suarez et al. [25]), to managerial (Uton [27]). Our view is that the roblem of evaluating an investment in a new roduction resource in general, and in a flexible resource in articular, consists of two arts. The first art is how to accurately estimate the future benefits the new resource will generate (for examle an uncertain stream of cash flows). The second question is how to roerly determine the value of these benefits. In this aer, we focus on the first art: roerly estimating the future benefits attributed to an investment. A large decision analysis and real otions ricing 1 literature already addresses the second question. Smith and Nau [24] show the circumstances under which the real otions and the decision analysis aroaches are consistent. Consider a roblem setting where a manufacturing firm has an oortunity to imrove its roduction rocess by urchasing a new iece of equiment. While the cost of the new resource is 1 The idea behind the real otion ricing aroach is to aly finance methods for valuing ut and call otions to real rojects. If one can construct a ortfolio of financial instruments that exactly relicates the real roject s cash flows in every ossible state of the world, the market value of this ortfolio is the same as the value of the real roject. See Smith and McCardle [23] for a detailed summary of real otion ricing aroach. -3-

5 known, its real benefit is not. To find the true benefit of the new resource, we need to be able to comare the erformance of the resulting roduction system with and without the new resource. We have two ways to do this comarison. We can try to model the systems either as it is actually used or as it should be used. The first method can be achieved with a simulation model, and the second with an otimization model (for examle, stochastic rogramming). Examles of the ure simulation aroach include Azzone and Bertele [1], Suresh [26] and Das and Nagendra [7]. Simulation models are concetually easy to understand and imlement, but they can lead to sub-otimal results. Ramasesh and Jayakumar [20] take simulation one ste further, by using it to generate realizations of uncertain arameters that are then used as data in an otimization model. We will say more about the Ramasesh and Jayakumar aroach in Section 2.2. Alternatively, otimization methods such as stochastic rogramming in theory yield otimal solutions, but real-sized, multi-stage stochastic rogramming models with recourse are often intractable. Examles of aroaches based on stochastic rogramming include Sinha and Wei [22], Guta, Gerchak and Buzacott [13] and Fine and Freund [10]. Recently much work has been done on develoing aroximation methods for certain classes of stochastic rogramming roblems (see Birge and Louveaux [3] and Infanger [15]). Successful alications of stochastic rogramming include the work of Een, Martin and Schrage [9] who develoed a model for General Motors that uses a scenario aroach to select the tye and level of roduction caacity. Mulvey, Gould and Morgan [19] describe an asset-andliability management system develoed for Towers Perrin-Tillinghast that uses stochastic rogramming to hel its clients make major business decisions. Carino et al. [6] describe another asset and liability management system develoed for Frank Russell Comany and The Yasuda Fire and Marine Insurance Co., Ltd., to determine the otimal investment strategy. -4-

6 The new method we describe here is an extension of the aroach first described by Katok [17]. It combines otimization with samling to aroximate system erformance under uncertainty. The dynamics of the algorithm are consistent with decision-making ractices shown to be suerior by Benjaafar et al. [2]. The new method is more intuitive and is easier to imlement than stochastic rogramming, and is more robust and general than ure simulation. 2. Model Develoment To determine the benefit of investing in a new roduction resource, we wish to estimate the additional cash flows that the new resource will generate. To accomlish this we model the current system (without the new resource) to determine the base cash flows. We then model the system with the new resource to determine the cash flows from the enhanced system. The difference between the two sets of cash flows can be attributed to the new resource. If the value of these additional cash flows (determined via decision analysis or real otion ricing) exceeds the cost of the new resource, the new resource is worth obtaining. Theoretically, the roer way to determine the erformance of a system under uncertainty is with a multi-stage stochastic rogramming. The objective function value of this model reresents the system s erformance. Since such large roblems are notoriously difficult to solve to otimality, we develo aroximate solutions. In the following sections we develo the stochastic rogramming formulation (also called the recourse roblem) of the resource acquisition decision. 2.1 Problem Formulation We use the stochastic rogramming notation of Birge and Louveaux [3], where random variables are denoted in bold. Consider a manufacturing firm that roduces a set of roducts { 1, P} P = using a general assembly rocess. Each P reresents either a finished roduct or a sub-assembly. We can secify any tye of bill of materials (BOM) structure by defining a set -5-

7 S (successors of ) for each roduct to include immediate successors of in the BOM. We also let k, j be the number of units of required to make one unit of j when j S. If is an end-item, S =. Let Ρ be the set of roduction resources. Since each P reresents a roduct at a articular roduction stage, we assume without loss of generality, that it needs to be rocessed only by one resource at each stage, although there may be alternative ways to rocess a roduct at a stage. Finally, let us assume that the model can be naturally decomosed into convenient time blocks = { = 1, J} J l, in such a way that there are not many interactions among the time blocks (ideally no interactions at all). Secifically, we assume that inventory cannot be carried across time blocks. We also assume that backorders across time blocks are allowed, but the interretation of a backorder during the last eriod of a time block changes to unmet demand, so there is never any backorder that has to be met in the first eriod of a time block. Each time block J, in turn consists of time eriods t T. Therefore, the model can be decomosed into searate multistage stochastic rograms for each time block. We identify each eriod in the model by a air of indexes, ( t, ), reresenting the time eriod t of the time block. We introduce the time block notation for convenience, and without loss of generality. If time horizon cannot be broken into time blocks, we simly have a single time block in the roblem. Since the Ramasesh and Jayakumar [20] model requires the use of time blocks, we introduce them here, to ensure consistency among models. Let d t, be the demand for the end items only (in eriod t of time block for roduct ) and a random variable. When is an intermediate item, d t, = 0. If demand in eriod t is not filled, a t, unit backorder cost λ is charged for the eriod. The rocessing time for roduct on resource r -6-

8 at time t of time block is a, and we assume that rocessing times do not san multile eriods. t,, r If a roduct does not need to be rocessed on a articular resource then t, ar, = 0 t,. Different resources involve different oerating requirements, so let resource r. Finally, each resource has Let w r be the cost of one unit of time on t, c r units of caacity available at time t of time block. t, ξ denote the vector of random arameters at time t of time block. The elements forming t, t,, ξ are demands (,, t d1 dp ) l. The decision variables are x t,, r, reresenting the number of units of roduct rocessed on resource r at time t of time block. These roduction decisions x are made at the beginning of t,, r time eriod t of time block, before the demand d t, for that time eriod is known. After the roduction decisions are made the demand ( d inventory decisions are made for the next eriod ( h ( b t, t, ) is revealed. At the end of the eriod the t+ 1, -7- ) along with the backordering decisions ). So the inventory and the backorder variables are recourse variables that absorb the uncertainty in each eriod. Assume for convenience and without loss of generality that the beginning and ending inventory levels are 0. Also, let ρ t, be the comounded discount rate from the beginning of the lanning horizon until eriod ( t, ); (1) - (5) is a mathematical rogramming formulation of the stochastic roduction-lanning recourse roblem (SPP): 1 t, t, t, t, min zspp = E wrar, r, t ( 1 t, ) λ ξ + x b (1) + ρ r Subject to

9 h + x k x d + b b h = 0, t 1, (2) t, t, t, t, t, t 1, t+ 1,, r, j j, r r r, j S h + x k x d + b h = 0, (3) 1, 1, 1, 1, 1, 2,, r, j j, r r r, j S t, t, t, a r, x r, c r r,, t (4) t, t, t, h, x, r, b 0, rt,, (5) Equation (1) is the objective function that minimizes the total exected discounted roduction and backorder cost, with exectation taken with resect to the random vector ξ. If the lanning horizon is sufficiently long, we should include the inventory holding costs as well. Equations (2) and (3) are the set of material balance constraints that ensures that no roduct is rocessed until all its redecessors are available. Note that equation (3) is for the first eriod of a time block, where backorder from the revious time block does not have to be met. Equation (4) is the set of caacity constraints. In ractice, more simlistic rocedures than stochastic rogramming are used to determine the value of flexibility, and we review two such rocedures in the next section. Sometimes simle simulation-based methods do an adequate job, correctly aroximating the benefit of flexible resources; nevertheless, at times, as we will demonstrate, simlistic methods may systematically underestimate the benefit of flexible equiment. 2.2 Alternative Methods The Wait and See Model If uncertainty can be aroximated by a set of scenarios, then one way to determine the value of flexibility is to solve the so-called wait and see roblem (WS). If we let the individual -8-

10 scenarios corresond to realizations of the random variable ξ, then equations (6) - (10) can define the otimization roblem associated with one articular scenario ξ. 1 t, t, t, t, min z ( ξ ) = wa r rxr b t ( 1 t, ) + λ (6) + ρ r Subject to (7) h + x, k, x, d + b b h = 0, t 1, t, t, t, t, t, t 1, t+ 1, r j j r r r,j S (8) h + x, k, x, d + b h = 0, 1, 1, 1, 1, 1, 2, r j j r r r,j S t, t, t, a r, x r, c r r,, t (9) t, t, t, h,x, r,b 0, r, t, (10) Here all variables and arameters indexed by t and reresent quantities in eriod t of time-block. * Denote an otimal solution to (6) - (10) as x ( ) ξ (since the x variables uniquely determine the h and the b variables), and the corresonding objective function value as z ( ξ ). We can then comute z E z( ) WS = ξ ξ, as the exected value of objective function values of deterministic subroblems corresonding to realizations of the random variables in all scenarios. This solution is known in the literature as the wait and see solution (Birge and Louveaux [3]). Comuting z E z( ) WS = ξ ξ exactly is unlikely to be ractical because the number of scenarios can be extremely large. If this is the case, we must aroximate z E z( ) WS = ξ ξ with a samle-mean estimate of z WS. This is the aroach we take for the emirical comarisons discussed in Section

11 2.2.2 The Aggregate Model A natural method to simlify comutations of the otimal value of the objective function for the deterministic roduction lanning roblem and establish a base line on the benefit of new equiment is to consider an aggregate formulation (APP). This method is esecially convenient when random variables can be naturally searated into several time blocks, with not many interactions among the time blocks. This is the Ramasesh and Jayakumar [20] aroach. Een, Martin and Schrage [9] use a similar aroach, aggregating their caacity lanning model develoed for General Motors into five yearly time blocks. In the aggregate formulation, the lanning horizon consists of time blocks, { 1, J} J = =. We aggregate the roducts into end-items. In this case the set Π of roducts includes end-items only, and the decision variables x, r reresent the number of units of the enditem rocessed on resource r during the time-block. We can measure the er unit requirement of resource r by roduct in time block, A = a t,, r q, r t T q where q was in the BOM for in the SPP model. The demand for roduct is now the aggregate demand for the time block, D = d t T t. The caacity of resource r is the aggregate caacity for the time block, C r = c t T t r. If caacity is insufficient to fill current time-block demand, the roduct is backordered, and -10- B is the total backorder of roduct for time block. Note that the nature of backordering can be different in APP than in SPP, since in APP backordering reresents the unmet demand, while in SPP backorders can be filled in subsequent eriods. If we allow some demand at the end of a time block to remain unmet in SPP, that unmet demand has the same meaning as cost λ = t, B in APP. The backorder λ when t is the last eriod in a time block. Since no inventory is carried across time

12 blocks, we do not need the inventory variables. The discount factor ρ in the aggregate model is the single eriod discount factor t, = t T ρ, comounded over the time block, ρ ( ρt, ) The aggregate mathematical formulation is similar to the roblem described by Ramasesh and Jayakumar [20], and we make every attemt to use notation consistent with theirs. Subject to 1 min zapp = E wr A,r,r ( 1 ) λ ξ x + B (11) + ρ r xr, + B = D, (12) r A r, x r, C r r, (13) xr,, B 0, r, (14) If several of the resources (r ) are interchangeable, constraint (13) becomes A x C. r, ' r, ' r' r, ' r' Even though the otimization model described by (11) through (14) is smaller than SPP, and searates into one roblem for each time block, just as SPP does, solving it directly may not be comutationally feasible. However, Ramasesh and Jayakumar [20] develo and test an efficient method for finding aroximate solutions. Following the aroach of Ramasesh and Jayakumar [20], we assume the demand is known at the beginning of each time block and is different for subsequent time blocks. Again following the aroach of Ramasesh and Jayakumar [20], we can estimate the system erformance over time by drawing realizations of uncertain arameters from their distributions, and solving the aggregate roblem several times. Ramasesh and Jayakumar [20] -11-

13 show that this aroach gives solutions very close to otimal solutions to the aggregate roblem. However, APP is a relaxation of SPP, and therefore z APP is a lower bound on z SPP. Now let us analyze APP s estimates for the benefit of a flexible resource. First, let us say that we have the base-line system consisting of a set of resources Ρ, and a new system, consisting of a set of resources R ', where ' = { r } R R. Let V ( ') = z ( ') z ( ) new R R R reresent the APP APP APP APP estimate of the benefit of the new set of roduction resources{ r new}, and also let ( ') = ( ') ( ) V R z R z R reresent the WS estimate of the benefit of{ r }. Recall that we WS WS WS ostulated that there are three sources of benefit of a resource: (1) lower roduction cost, (2) caacity, and (3) decision flexibility. The roblem APP considers the cost of oerating a resource (unlike the Ramasesh and Jayakumar [20] formulation that looks at the time rather than the cost), so the ortion of the new resource benefit due to any roductivity imrovement that results in lower roduction cost is addressed by APP. APP only artially accounts for benefit due to caacity. Problem APP has a caacity constraint that reserves the aggregate caacity for the time block. It is ossible, however, to observe the aggregate caacity constraint while violating caacity constraints for single eriods. For examle, if each day has eight hours of caacity, and the time block has two days, the aggregate caacity constraint tells us that we cannot exceed the 16 hours caacity in a two-day eriod. But a roduction lan requiring 10 hours on day 1 and 6 hours on day 2 is still aggregate-feasible, although the lan exceeds day 1 caacity and allows an unrealistic shift of available hours. A stronger caacity constraint would force the model to allocate hours roerly and highlight the benefits from having the additional caacity on days when it is required. Since APP has a weaker new -12-

14 caacity constraint than SPP, the benefit of the new resource due to caacity can be underestimated because the model will not identify benefits on caacitated days. When decisions are made in APP, all relevant time-block information is known. Benjaafar et al. [2] show that decision flexibility rovides no benefit if no relevant future information is exected. This result imlies that APP does not account for any benefit of the new resource due to an increase in decision flexibility, but it does rovide an aroximation for the benefit from efficiency gain and artial benefit from caacity gain. Similarly, V ( ') V ( ') estimate for the gains from caacity not catured in ( ') V SPP ( ') V ( ') WS APP WS R R rovides an APP V R. And most imortantly, R R rovides an estimate for the gains from decision flexibility. 2.3 The New Method Both, APP and WS make a art of the SPP recourse roblem into a deterministic roblem and then solve a sequence of deterministic roblems with arameters reresenting realizations of stochastic arameters. The solution to a roblem where stochastic arameters are relaced with their realizations is called a wait-and-see solution. Birge and Louveaux [3] (. 140) rove that the wait-and-see solution is a lower bound on the recourse roblem solution (in our terminology, z WS z SPP ). Birge and Louveaux [3] also describe the notion of the exected result of using the exected value solution (EEV) (. 139). This is the exected system erformance that results if, at the beginning of the lanning horizon, we solve a roblem relacing all stochastic arameters with their exected values and imlement the solution. Birge and Louveaux [3] show that EEV is an uer bound on the recourse roblem solution z SPP, because, by construction, EEV is always a feasible solution to the recourse roblem. It turns out that tyically, EEV is a weak uer bound. Our goal is to develo an algorithm with a stronger uer bound on z SPP. -13-

15 We begin by looking at the decision-making rocess under uncertainty. Benjaafar et al. [2] ostulate that there are two general aroaches to flow control decisions in manufacturing. The lanning-based aroach alies when a roduction lan is determined rior to the beginning of roduction (at t = 0) and is rigidly adhered to. It is similar in flavor to the EEV solution. The real time based, or oortunistic aroach allows decision making to be contemoraneous with action imlementations. Decisions are made based on the state of the system, and no decision is imlemented until it has to be. Benjaafar et al. [2] show that under conditions of uncertainty, oortunism is suerior to lanning. To generate an uer bound on z SPP we aly the oortunistic decision making rocess and use the rolling horizon strategy (see Bitran and Sarkar [4] and Bitran and Yanasse [5]). Under this strategy, we solve a multi-eriod roblem each eriod, but only imlement first-eriod decisions and kee track of the first eriod erformance measures. Algorithm 1 formally outlines the method, and the resulting rocess for estimating the erformance of a roduction system under uncertainty. We wish to estimate the difference in the total exected cost of oeration over the T-eriod lanning horizon of n different systems with different sets of roduction resources. Let Ρ i denote the set of resources in the i th t system under consideration. Let z ( ) R be the cost during eriod t of a i system with resources Ρ i. We estimate the objective function value of the recourse roblem z SPP with an estimate zc SPP by reeating the oortunistic decision-making task m times. -14-

16 j 0; z SPP 0 foreach Ρ i (i = 1,, n) do while (j m) do foreach t 1,...,T do fix decision variables for eriods t < t randomly generate realizations of uncertain arameters 2 solve the deterministic flexibility-lanning roblem endfor endfor endwhile zc SPP ( R ) i m imlement decisions for the current eriod t R current eriod cost z j= 1 ( ) i j z SPP z m j SPP t z + z ( R ) j SPP i Algorithm 1. Samling Algorithm for finding a bound on the objective function for a multi-stage stochastic rogram. By construction, the Algorithm 1 solution is feasible in the m instances of random arameter realizations. As m grows large P{ z z } SPP SPP goes to one, in other words, as m grows large, it becomes more likely that zc SPP is a valid uer bound to the objective function value of the recourse roblem 3. When it is an uer bound, it is likely to be a stronger uer bound than EEV 4 in exectation, because, again by construction, EEV is a solution where m =1, and all random variables are simly relaced by their exected values, while zc SPP is determined using a larger value of m. The major benefit of our modeling technique is that it accounts for the oortunistic decisionmaking rocess, exlicitly modeling decision flexibility. Therefore, unlike the APP and WS 2 We actually use the same set of scenarios for all the systems we comare. 3 For examle, if we actually solve the roblem for every ossible scenario, ~ z SPP becomes a feasible solution to the recourse roblem by construction. 4 Although it is ossible to construct examles where it is not a stronger uer bound. -15-

17 models, our new model will be less likely to under-estimate the benefit of a resource as much as APP and WS when that resource rovides decision flexibility. 3. Jeesen Sanderson, Inc. The alied ortion of this work focuses on Jeesen s roduction system for flight manual revision service. For a more comlete descrition of Jeesen see Katok, Tarantino and Tiedeman [18]. Airway safety considerations dictate that all ilots on all flights must have a set of airort mas, enroute charts, and other flight information for the area within a 200-mile radius of the lanned route. Flight information changes constantly, so this material must be udated regularly. For examle, about 75% of all charts are revised at least once annually, and many charts are amended much more often. Enroute charts that cover large areas change on average four times a year. A tyical Jeesen chart is shown in Figure 1. Jeesen usually configures its flight manuals by geograhical area. Many ilots subscribe to what Jeesen refers to as the Airway service; however, many of Jeesen s large customers, including major airlines such as United, American, and Delta and ackage delivery services such as FedEx and UPS request secial subscrition ackages. These secial ackages, called Air Carrier coverages, can differ from standard coverages because they contain charts with secial information, a customized configuration of ages, or other secific features that a customer might request. Jeesen maintains over 200 different standard coverages and over 2,000 different tailored coverages, made u of over 100,000 distinct images. -16-

18 Figure 1. A tyical Jeesen chart When critical aviation information changes (such as a runway at an airort is closed or exanded), the change affects multile Jeesen charts. Tyically, a change affects one airway chart and several customized air carrier charts. When a chart is revised, Jeesen issues a new manual age to all customers subscribing to coverages containing this age within one week of the -17-

19 change. Every week Jeesen sends out between 3 and 25 million ages to over 300,000 different customers. Some weeks over 1,500 images, affecting over 1,000 different coverages must be altered. Figure 2 shows a diagram of the Jeesen revision management and roduction rocess. Aviation information data Image edited electronically Imaging and rinting Machine Collating Final Assembly Yes Is this data accurate and imortant? No STOP Shiing Figure 2. Revision management and roduction rocess at Jeesen. When information regarding a ossible change first reaches Jeesen, a decision is made as to whether this data is imortant or ermanent enough to amend a chart. Some changes do not need to be included on a chart. For examle, a runway closing for 20 minutes on a articular day would not require a revision (and will be handled with a notam ). If a change to a chart is deemed necessary, the first ste of the rocess involves electronically editing the image file. Some alterations are easy to make taking less than 5 minutes, while other changes can require as much as 8 hours of work. After an image file has been edited electronically, a new negative is rinted. This negative goes to the first ste of the roduction rocess, imaging and rinting, where it is stried onto a late containing 21 negatives, the late is rinted, cut into individual sheets, and secially bound. Sheets then go into the machine collating area, where they are collated into sections. Each section contains u to 25 sheets that will eventually all go into the same coverage. At this oint large mas, called folds, are not included into sections, because collating machines cannot handle folded material. Sections and folds go into the final assembly area, where rior to the imlementation of our work they were manually assembled into coverages and stuffed into -18-

20 enveloes. Large boxes of enveloes go on to the shiing deartment. If a coverage comletes final assembly on time it is shied using standard shiing services, but if it is late the service is ugraded to overnight delivery. Figure 3. The manual rocess The bottleneck of the roduction rocess forms in final assembly, highlighted in Figure 2. Prior to imlementation of our work, in final assembly sections and folds were arranged and stuffed manually, often by a large number of temorary emloyees. Figure 3 shows a hotograh of a tyical Jeesen assembly rocess. The use of temorary emloyees has several disadvantages for Jeesen. They are often unfamiliar with the work, and tend to be less roductive and make more mistakes than full-time emloyees. Jeesen customers do not tolerate errors, so all errors are detected and corrected at great exense rior to shiing. The availability of temerary emloyees can also be unredictable. Because of these roblems, Jeesen management wished to evaluate the urchase of new, automated technology, called folder collator, for final assembly, and asked us to hel them with this decision. The dynamic and comlex nature of the Jeesen oerating -19-

21 environment makes roerly determining the benefit of the new technology difficult, and hence the alication of our method well-warranted. 4. Emirical Comarisons In this section we demonstrate how the three aroaches to estimating the benefit of a flexible resource can yield different results. Jeesen oerates on an 8-week revision cycle involving three week-tyes with differing demand volumes: odd weeks have relatively low volume, even weeks have medium volume, and eighth weeks have the highest volume. Over time, revision characteristics in terms of overall volume (number of customers), the number of different coverages, average volume, and coverage size in terms of both, folds and flats, have been evolving. Figures 4a and 4b show historical trends in weekly revision for relevant dimensions since Total Quantity: envelo count 140, , ,000 80,000 60,000 40,000 20, Number of Different Coverages Revising

22 Average Quantity: customers er coverage 2,000 1,800 1,600 1,400 1,200 1, Average Number of Flats er Coverage Average Number of Folds er Coverage Figure 4a. Revision characteristics over time: Airway. The number of airway coverages revising increases significantly over time, while the average number of Airway customers er coverage declines, highlighting the fact that demand for customized roducts increases over time. -21-

23 Total Quantity: envelo count 200, , , , , ,000 80,000 60,000 40,000 20, Number of Different Coverages Revising 1,200 1, Average Quantity: customers er coverage

24 Average Number of Flats er Coverage Average Number of Folds er Coverage Figure 4b. Revision characteristics over time: Air Carrier. The number of air carrier coverages revising is fairly steady over time, but the average number of flats and folds er coverage is growing. So Air Carrier coverages are becoming larger over time airlines add information to their customized coverages. As we mentioned earlier, Jeesen has two tyes of customers: The Air Carrier customers, including rimarily airlines and ackage delivery services, subscribe to customized roducts, while Airway customers subscribe to standard manuals, and include rimarily cororate and rivate ilots. Historically, there are a relatively small number of airway manuals, and each has a large customer base. However, we see from Figure 4a that the number of airway coverages is increasing dramatically, and average quantity er coverage is droing. The number of air carrier coverages (Figure 4b) is growing also, but much slower, and the average quantity seems fairly steady. Air carrier coverages, however, are increasing in terms of the number of both, flat and folded charts. -23-

25 When estimating the benefit of new technology for the future, it is imortant to forecast these various trends into the future as accurately as ossible. The Jeesen roduction roblem is stochastic because roduction must begin before the entire weekly demand is known. That is, when the roduction is scheduled, rior to the first day of the week, the real demand is still a random variable. The recise moment the weekly demand is finalized at Jeesen is a matter of some debate. Jeesen assigns official close dates, but they are not always adhered to because Jeesen goes to great lengths to accommodate its customers. Therefore, for a good art of the week, demand is a moving target. 4.1 Modeling the Jeesen Problem In this section we recast the Jeesen roblem as a Stochastic Production-Planning Recourse Problem (SPP). The set of roducts = {, 1 P} P m include all finished roducts, as well as intermediate sub-assemblies. At Jeesen, the notion of a roduct changes as the material moves through the roduction system. In the rinting area, and as far as rinting vendors are concerned, roducts are individual charts and folds. In machine collating area roducts are sections, comosed of grous of 25 or 36 flat charts. For final assembly, roducts are coverages. The set S defines the bill of materials (BOM) structure for roduct, and Figure 5 shows the BOM for Jeesen revision roducts. coverage sections folded charts flat charts Figure 5. The BOM for Jeesen revision roducts. -24-

26 In general, S when is a coverage = when is a section or a fold when is a flat chart { coverages} { sections} For Jeesen k = 1,j since coverages never contain multile coies of charts. The set of j roduction resources = { r 1, mr} R includes four different tyes of rinting resses, a bindery, several outside rinting vendors, two tyes of collating machines, manual assembly, the new folder collator, and a fold-collating vendor. Caacities of those resources -25- t, c r are well-known at Jeesen, and are measured in hours a resource is available for oeration during a articular day. Jeesen revision assembly lanning is done on a weekly basis, with no major interactions between weeks. Due to the airway community s 8-week oerating cycle, there are a large number of charts scheduled to revise in intervals that are multiles of 8 weeks. So generally, every 8 th week Jeesen faces a very large revision. Even weeks (weeks 2, 4 and 6 of each cycle) are mediumsized, and odd weeks (weeks 1, 3, 5 and 7) are comaratively small. A one-week roblem is a comlete lanning roblem because of the lack of interaction among weeks, so the set of time blocks J for Jeesen consists of a single one-week time block, running from Friday afternoon to the following Friday morning. Revision information, however, is only artially known at the beginning of the week, and changes every day, with the main information udate occurring each Monday, but minor udates occurring daily. So effectively each weekly time block is broken into eight daily time eriods t (where the Friday time eriods are actually shorter than one day). The backorder structure for the Jeesen roblem is very simle. If there is not enough caacity to meet demand, the roduct is late. Late roducts incur a large enalty in the objective t, function for each day of lateness. This enalty, λ reresents not only the increased shiing

27 costs (because late roducts are automatically ugraded to overnight shiing) but also the loss of good will. Although in ractice a Jeesen revision is occasionally late, lateness is generally avoided at all costs, and only haens due to extraordinary circumstances (a machine break-down at t, a critical time, or vendor error, for examle). When t is the last eriod of a time block, λ actually reresents the cost of meeting the demand through outside vendor of last resort, so it is very high. The demand d t, exists only for coverages, and the demand for most coverages occurs on the second Friday of the week (t = 8), but some coverages that have long shiing times, such as Australian coverages, are due earlier (t = 6, for examle). To create a realistic samle of demand scenarios we used 173 weeks of demand data that started on 6 January 1995 to estimate relevant attributes of the demand. System load deends on: the total quantity demanded, number of different coverages, number of customers er coverage, number of flats er coverage, and number of folds er coverage. Historical trends for those five demand characteristics for Airway and Air Carrier are shown in Figures 4a and 4b, and we forecast all of them to generate realistic demand scenarios. Figures 4a and 4b show that there is a clear cyclical comonent to revision, and in most cases there is also a trend comonent. We fit a forecasting model to the historical data, of the form in (15), using Ordinary Least Square (OLS) estimate, ˆ t Load = Intercet + Trend time dummy variable + Even even dummy variable + Eight eighth dummy variable + ε t (15) where time dummy variable is a week number starting with week 1 being 6 January 1995, even dummy variable is 1 for revision cycle weeks 2, 4, 6, and 8, and eighth dummy variable is 1 for week 8. Table 1 shows the regression results for the five relevant demand attributes. -26-

28 Air Carrier Load (units are coverages or charts) (-value) Airway Load (units are coverages or charts) (-value) 2 2 Intercet Trend Even Eight r Intercet Trend Even Eight r Total Quantity 86, ,430 51, , ,618 81, (envelo count) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Number of Coverages in Revision (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Customers 120 (0.04) ,228 (2.00) (436.10) (373.45) 0.67 er Coverage (0.0000) (0.0423) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Flats er Coverage (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.3890) (0.0000) (0.0000) Folds er Coverage (0.0000) (0.0029) (0.0251) (0.0000) (0.0000) (0.3287) (0.1054) (0.0059) Table 1. Forecasting assembly load. Note that in most cases (13) does a good job of exlaining the variability in the data, and generally all variables are significant. One excetion is the folds er coverage in airway, where the only significant variable is eight. Also, the time trend is not significant in the flats er coverage for airway. The model in Table 2 imlies that during an odd week in 1995 around 87,000 enveloes were sent out to air carrier customers and around 32,000 to airway customers. Since then, this quantity has been growing steadily at a weekly rate of about 93 enveloes for air carrier and 42 enveloes for airway (the historical data sans 173 weeks, so during an odd week in 1999 about 103,000 envelos are sent out to air carrier customers and almost 40,000 to airway customers). During an even week, on average additional 52,000 enveloes are sent out to air carrier customers, and 61,000 to airway customers (bringing 1999 s even week total to 155,000 for airway and 101,000 air carriers). During an eighth week (which is also an even week) additional 52,000 enveloes are mailed out to air carrier customers, and 81,000 to airway on average (bringing an eighth week total in 1999 to over 200,000 air carrier envelos and over 180,000 airway enveloes). -27-

29 The forecasting model works in a similar way for all five dimensions, so the average number of subscribers er coverage, for examle, decreases from week to week. To generate a demand scenario for a articular week we use estimates for exected values of the five demand attributes and their standard errors, and draw a demand scenario from the resulting distribution. Parameter w r reresents the labor cost on resource r, and it is generally well known. Unfortunately, accurate rocessing times for the resources t, a,r were not as readily available. There were standard rocessing rates, but they did not reresent reality. The roblem of determining accurate rocessing times is interesting, because the time it takes to assemble a coverage, for examle, deends on several variables: coverage quantity, the number of sections, the number of folds, and on whether a temorary or a ermanent emloyee erforms the work. Using the manual assembly rocess as an examle, we determined the total rocessing times by systematically tracking actual rocessing and setu times for each coverage over a one week eriod. We then fit the following model: a = α + β folds + β sections + β d + ε (16), assembly where α is the intercet term, folds reresents the number of folds in coverage, sections reresents the number of sections in coverage, d reresents the quantity of coverage demanded, and ε is an unobservable random error. Equation (16) gives us an aroximation of the total time in assembly. We fit the model using ordinary least squares. All coefficients were significant, and the resulting r 2 was 78.1%. We determined rocessing times for other resources using the same method. Estimating rocessing times with the new collator was a more difficult task because we did not have the oortunity to observe the collator s erformance in the Jeesen roduction -28-

30 environment. Instead, a team of Jeesen managers observed the collator s erformance at the vendor s site. They collected the roduction data that we ultimately used to estimate collator rocessing times. For the urose of the tests, we assume inter-stage indeendence for the vector of random arameters ξ t,. Although demand information at Jeesen is udated daily, new information significantly imacts lanning only once, on Monday (t = 4) of every week. So a one-week lanning roblem is a two stage stochastic model with recourse, where the initial lan is made on Friday ( = 1,t = 1), roduction starts and roceeds for three days, demand information is revised on Monday ( = 1,t = 4 ), and the lan is adjusted given the new information. 4.2 Comarative Results To begin our emirical comarison of the three flexibility evaluation aroaches, we icked 17 actual consecutive weeks (two comlete 8-week cycles, and one additional week following the second cycle): 21 August 1998 through 11 December The date 21 August 1998 is the Friday of week 3, 28 August is the Friday of week 4, 4 Setember is the Friday of week 5, and so forth. We had the actual data that went into the revision, and the sequence in which this data was becoming available to the lanning grou. We ran the three models on the 17 weeks of data, which involved solving 17 searate twostage stochastic roblems. We modeled the current state of the system and the hyothetical system configuration with the new folder-collator. To determine the benefit of the collator each week, we take the difference of revision cost with and without the collator. Table 2 comares the SPP estimates of collator benefit with those of APP and WS. All numbers are resented as ercentage difference with SPP. We estimate V ( R' ) SPP by running 30 relications of Algorithm 1 on a 2-stage roblem. -29-

31 Week Percentage Deviation from SPP Date in Cycle WS APP 21-Aug 28-Aug 4-Se 11-Se 18-Se 25-Se 2-Oct 9-Oct 16-Oct 23-Oct 30-Oct 6-Nov 13-Nov 20-Nov 27-Nov 4-Dec 11-Dec Table 2. Comarative system erformance for 17 weeks. We learn several things from Table 2. First, we see that in every case ( R' ) ( R' ) ( R ') V V V APP WS SPP We also observe that in many of the weeks APP and WS models underestimate the benefit of the collator relative to SPP. All three models give the same solution in several of the weeks. Those are all small odd weeks, with low load. Most of the savings from the collator are due to the two 8 th weeks, since the 8 th weeks are the only weeks where internal caacity is insufficient to fulfill demand and an outside vendor is used for fold assembly. The outside vendor is much more exensive than internal fold assembly, even if overtime and temorary emloyees are used. With the new collator, the use of the outside vendor can be avoided. -30-

32 4.3 Estimating the Total Collator Imact We now comare how the three models estimate the benefit of the new collator over the three-year lanning horizon. The revious section showed that the benefit of this resource increases with system load. We estimate the benefit of the collator by simulating the three-year Jeesen roduction environment, based on (15). In other words, we generated 156 weeks of demand consistent with demand characteristics as resented in Table 1. Each week is a searate two-stage stochastic model, and the SPP estimates were obtained for each week searately by running 30 relications of Algorithm 1. Table 3 summarizes average weekly benefit estimates for all three models, along with their standard errors. Figure 6 resents our analysis grahically. Week SPP WS APP Average weekly Average weekly Average weekly benefit ($) benefit Deviation benefit ($) (standard error) (standard error) from SPP (standard error) Deviation from SPP Odd 1, , % 1, % Eight 190, , % 88, % 73, , , Even 7, , % 4, % 1, , Table 3. Summary of solution results for estimating collator benefit In the odd week roblems, WS and APP results are generally quite close because odd weeks have low volume, on average, so caacity virtually never becomes an issue. Results look very different in even and eighth weeks. We clearly see that WS and APP underestimate the benefit of the collator relative to SPP. In even weeks, APP undervalues the collator by about 36% and WS by 26%. In eighth weeks, APP undervalues the collator by 54% and WS by 41%. In odd weeks, both APP and WS undervalue the collator by about 26%. All of these differences are highly statistically significant, using a two samle t-test assuming unequal variances. -31-

33 Flexibility Value 13, , , , , SP P DPP APP 3, , Cycle (a) Tyical even week Flexibility Value 351, , , , , , , , SP P DPP APP Cycle (b) Tyical eighth week Flexibility Value 3, , , , , SP P DPP APP Cycle (c) Tyical odd week Figure 6. Grahical reresentation of results. -32-

34 Our analysis showed that the annual discounted savings a new collator will generate are around $1.4 million. Jeesen acceted our analysis and recommendation, and urchased the collator in July Collator Imact The collator, referred to as Longford at Jeesen, after its manufacturer, was custom-built and delivered in late December Figure 7 shows the collator in action at Jeesen. After an initial training eriod, Jeesen s assembly area started using the collator on 6 January We asked the assembly oerations manager to systematically kee track of all work erformed on the collator, which he has accomlished. Table 4 summarizes this data for the eriod of 8 January 1999 through 21 May 1999, and comares the actual savings with savings forecasted by the three alternate models. We determine actual savings each week by considering that week s entire revision and determining which coverages should be assembled using the collator. We then comare the actual cost of assemblying these coverages on the collator, and what it would have cost to manually assemble them. The difference is in the column labeled internal in Table 4. On 8 th weeks internal caacity without the collator is not sufficient to meet the load, so a large number of folds would have been assembled by an outside vendor. We use the vendor s actual rice schedule to determine the outsourcing cost that would have been incurred if the collator was not available. This figure is shown in the column labeled external in Table

35 Actual Savings Forecasted Savings Week of Cycle Internal External Total APP WS SPP 08-Jan ,657 5,145 5, Jan ,322 3,322 1,045 1,085 1, Jan , , ,687 28,537 56,365 92, Jan ,478 1,527 1, Feb ,076 6,378 7, Feb ,178 1,261 1, Feb ,572 4,825 5, Feb ,300 1,365 1, Mar ,178 4,178 3,537 5,316 6, Mar ,583 2,583 1,183 1,200 1, Mar ,425 89,679 97,104 28,888 60, , Mar ,465 1,538 1, Ar ,021 2,021 4,086 5,559 6, Ar ,550 1,630 2, Ar ,647 2,647 4,912 6,334 7, Ar ,040 1,127 1, Ar ,854 3,854 6,406 7,314 8, May ,625 3,625 1,471 1,531 2, May ,317 82,808 93,125 27,737 53,555 93, May ,010 1, May ,915 5,640 7, Jun ,141 19, , Jun ,130 3,130 5,409 5,815 7, Jun ,846 7, , Jun ,582 1,582 4,005 4,541 5, Jul ,936 7,936 1,299 1,331 1, Jul ,994 68,529 75,523 27,542 51,905 90, Jul ,961 14, , Jul ,758 6,758 6,310 6,294 8, Jul ,825 8, , Aug ,901 5,889 7, Aug ,600 7, , Aug ,891 4,843 6, Aug ,567 1,567 1,013 1,108 1, Se ,615 72,560 81,175 27,394 69, , Se ,706 19,706 1,381 1,464 2, Se ,909 3,909 4,968 4,937 6, Se ,482 18,482 1,063 1,110 1, Oct ,415 3,415 5,317 5,286 7, Oct ,481 10,481 1,194 1,230 1, Oct ,874 4,874 5,552 5,517 8, Oct ,955 5,955 1,579 1,601 2, Oct ,510 72,820 86,330 46,190 75, , Nov ,136 24,136 1,344 1,365 2, Nov ,180 24,180 3,915 3,842 5, Nov ,117 4,117 1,160 1,221 1, Nov ,223 3,223 5,350 5,321 7, Dec ,439 2,439 1,160 1,222 1, Dec ,413 5,372 8, Dec ,513 8,513 1,160 1,223 1, Dec ,542 78,761 92,303 46,984 84, , Dec ,716 13,716 1,160 1,224 2, Jan ,623 7,623 5,681 5,667 8,836 Total Savings to date: 323, , , , , ,812 Table 4. Savings and forecasts over the test eriod The internal savings are due to lower cost, in terms of man-hours, for using the collator instead of the manual assembly method, the increased caacity the collator rovides, and increased decision flexibility the collator offers. The vast majority of the savings, however, are external, -34-

36 meaning that the collator allowed Jeesen to bring much of the work in-house that was reviously sub-contracted out to a vendor. These external savings illustrate how the collator increased Jeesen s volume flexibility. In an internal memo dated 14 May 1999, Paul Vaughn, the assembly oerations manager wrote: The bottom line is that the Longford continues to meet exectations and we are saving dollars! Figure 7. Collator in use at Jeesen The 53 weeks of data resented in Table 4 demonstrate that our new method for determining equiment benefit (SPP) is much more accurate in forecasting actual savings than the other two common methods. SPP estimates are determined by running 30 relications of Algorithm 1 for each week s roblem, while APP and WS estimates are determined by solving corresonding deterministic roblems. Using a matched air t-test, we cannot reject with the null hyothesis that -35-