LIDS-R-1408 September, 1984 Revision of March, 1985 A CONTROL THEORIST'S PERSPECTIVE ON RECENT TRENDS IN MANUFACTURING SYSTEMS by Stanley B. Gershwin *, Richard R. Hildebrant **, Rajan Suri ***, and Sanjoy K. Mitter * * Laboratory for Information and Decision Systems Massachusetts Institute of Technology Cambridge, Massachusetts 02139 ** C. S. Draper Laboratory Cambridge, Massachusetts 02139 Division of Applied Science Harvard University Cambridge, Massachusetts 02138 *** presented at the 23rd IEEE Conference on Decision and Control Las Vegas, Nevada December, 1984 page 2 PERSPECTIVE 1. Introduction the While basic understanding plete. These the ments), large uncertainty multiple machine to due require- production in variability hierarchies, level and studied have theorists and by randomness complicated control and systems in in manufacturing a control a view and in essential that issues to present certain theorists systems thinking. current in a that way establish, in issues that We then research will summarize of that framework. make that 4) facilitate This influenced of this area. impact, it paper is and become intended is to control this. for manufacturing by control practice and critique of has a vocabu- management 2, a framework current in perspective and terminology in manufacturing heavily (Section helpful directions. Section is to the problems research with recent familiar We learn they group the this community that can be of systems order for this However, systems from believe that theorist. We interpretation an to present is this paper of purpose progress recent view scheduling, planning, contexts. The lary and incom- remains issues (particularly requirements, data that other issues are They rapidly, a improving is production environment manufacturing and systems the include issues failures other of -- including processes -- manufacturing software of work-in-process. control in and hardware computer and of technology and systems (Section them from the 3) and point of page 3 PERSPECTIVE 2.0 General Perspective The purpose of manufacturing system control is not different in essence from many other control problems: it is to ensure that a complex system behave in a desirable way. from control realization areas. theory are relevant is quite different here, although their specific from more traditional The standard control theory techniques have not yet seen a manufacturing represented by a linear system is not surprising; application do not apply: we system that can be usefully with quadratic standard techniques what have been standard problems. Many notions have objectives. This been developed for Manufacturing systems can be an important area for the future of control; new standard techniques will be developed. Some central plexity, back. hierarchy, Important control issues in manufacturing discipline, capacity, notions of control variables, the objective model, and constraints. It systems uncertainty, theory include function, include and comfeed- state and the dynamics or plant would be premature to try to identify these with all the issues outlined in this paper; it would even go against the purpose of the paper, which is to stimulate such modeling activity. 2.1 Complexity Manufacturing systems are large scale systems. volumes of data are required to describe them. impossible; suboptimal strategies for planning archical being Enormous Optimization is based on hier- decomposition are the only ones that have any hope of practical. page 4 PERSPECTIVE 2.2 Hierarchy There are many time scales over which planning and scheduling decisions must be made. The longest term decisions involve capital expenditure or redeployment. The shortest involve the times to load individual parts, or even robot arm trajectories. While these decisions are made separately, they are related. particular, each long term decision presents an assignment next shorter term decision-maker. scale. the capacity--explicitly into The definition of the capacity depends on the time For example, short-time-scale the set of machines operational capacity is Machine capacity is at any instant. an average of short-time-scale Level trol. This includes the calculation robot arm trajectories:; semiconductors. control of There capacity. the machine level con- Other and other steps short-scale-issues tool: of optimal "ladder diagrams" motors, and hydraulics of furnaces a cutting in particular, in machine for tools, in the fabrication include adaptive of the detailed machining. is no rule that determines exactly what this shortest time scale is. semiconductor The issue A robot arm movement can take seconds while a oxidation step can take hours. at this time scale individual operation. is the optimization of each Here, one can focus on minimizing or other cost of each separate material. Long-time-scale and implementation the design of microswitches, and the control a function of Control At the very shortest time scale is relays, to the The decision must be made in a way that takes the resources--ie, account. In movement the time or transformation of One can also treat the detailed relationships among page 5 PERSPECTIVE operations. problem. An example of this is the Here, a large set of operations to be performed line is grouped at stations along a production tive is to minimize balancing line. into tasks The objec- the maximum time at a station, which results in maximum production rate. Additional control problems detection of wear and breakage temperatures control at this time scale include the of machine tools, the control of and partial pressures in furnaces, of the insertion of electronic the automatic components into printed circuit boards, and a vast variety of others. Cel I Level At the next time scale, one must consider of a small number of machines. includes This is the interactions cell level control and the operation of small flexible manufacturing The important issues include routing and scheduling. problem is The control that of ensuring that the specified volumes are actu- ally produced. At this level, operations are taken as given. operations systems. the detailed specifications of the In fact, for many purposes, the themselves may be treated as black boxes. The issue here is to move parts to machines in a way that reduces unnecessary idle time of both parts and machines. The loading problem is that of choosing the times at which the parts are loaded into the system or subsystem. The routing problem is to choose the sequence of machines the part visits and the scheduling problem is to choose the times at which the parts visit the machines. The important considerations in routing include the set of PERSPECTIVE machines page 6 available not desirable that can do the required to use a flexible machine done by a dedicated one, since to do jobs for which there In scheduling, required machines quirements are met. system resources the flexible failures, machine may be able that parts visited also guaranteeing in an efficient way. operator is allocating These resources and storage problem at this level is their that production re- At this level, the issue disruptions on factory operations. chine is often are no dedicated machines. machines, transportation elements, A control It to do a job that can be one must guarantee while tasks. include space. to limit the effect of Disruptions are absences, material due to ma- unavailability, sur- ges in demand, or other effects that may not be specified in advance but which are inevitable. This problem may be viewed as analogous to the problem of making an airplane robust to sudden wind gusts, or even to loss of power in one of three or more engines. Factory Level At each higher level, the time scale lengthens under concern grows. several cells. first stage At the next higher level, one must treat For example, in printed is a set of operations Metal is tion. Later, and the area circuit fabrication, the that prepare the boards. removed, and holes are drilled. At the next stage, components are inserted. The next stage is the soldering operaStill later, the boards are tested and reworked if they are assembled takes much time and a good deal Issues of routing into the product. necessary. This process of floor space. and scheduling remain important here. page 7 PERSPECTIVE However, setup times become crucial. That is, after a machine or cell completes work on one set of parts of the same or a small number of types, it is often necessary to change the system configuration in some way. the cutters in a machine must remove the remaining For example, one may have to change tool. In printed circuit assembly, one components from the insertion machines and replace them with a new set for the next set of part types to be made. The scheduling problem is now one of choosing at which these major setups must take place. called the tooling This is the times often problem. Other issues are important at still longer time scales. One is to integrate new production demands with production already scheduled in a way that does not disrupt the system. class of decisions are those pertaining Another to medium-term capacity, such as the number of shifts to operate, and the number of contract employees to hire, for the next few months. sion, at a still longer capital time equipment of the consider such strategic scale, is factory. goals duct quality, and responsiveness Another deci- the expansion of the At this time scale, as market share, one must sales, pro- to customers. 2.3 Discipline Specified Manufacturing, tems degenerate operating rules are required for complex communication, and other large sys- into chaos when these rules are disregarded when the rules are inadequate. participants transportation systems. In the manufacturing must be bound by the operating or context, all discipline. This includes the shop floor workers, who must perform tasks when re- page 8 PERSPECTIVE quired; and managers, who must not demand more than the system can produce. It is essential that constraints on allowable control actions constraints be imposed on all levels of the hierarchy. must allow sufficient freedom for the These decision-makers at each level to make choices that are good for the system as a whole, but they must not be allowed to disrupt its orderly operation. 2.4 Capacity An important element in the discipline capacity. will of a system is its Demands must be within capacity or excessive queuing occur, leading to excessive effective capacity. costs, and possibly High level managers must not be allowed to make requirements on their subordinates city; subordinates must be obliged to accurately capacities to reduced to those All operations higher which exceed their capareport their up. at machines take a finite amount of time. This implies that the rate at which parts can be introduced into the system is limited. Otherwise, parts would be introduced into the system faster than they could be processed. would then be stored in buffers system) while (or These parts worse, in the transportation waiting for the machines to become available, resulting in undesirably large work-in-process and reduced effec- tive capacity. that throughput (parts actually produced) rate is The effect is may drop with increasing beyond capacity. system carefully loading Thus, defining rate, when loading the capacity of the is a very important first step for on-line scheduling. An additional complication is that manufacturing systems page 9 PERSPECTIVE involve people. It is than machine capacity, aspects. since it Human harder particularly capacity measure at all outside may be harder to define of its capacity the capacity hierarchy. and it is No system may be possible Just-in-Time place over a relatively long time scale. essential, therefore, a discipline can futile at best and dama- On the other hand, it This is a goal of the Japanese then to develop well, are expand the capacity of a given system by a learning is as such as whether the environ- and respecting levels ging at worst to try. It capacity rapid changes. Defining, measuring, important human when the work has creative can depend on circumstances ment is undergoing produce to process. approach, which takes See Section to determine for staying to 4.4. what capacity within it, is, and finally to expand it. 2.5 Uncertainty All real systems are subject to random disturbances. precise time or extent of such disturbances some statistical The may not be known, but measures are often available. For a system to function properly, some means must be found to desensitize it to these phenomena. Control theorists often distinguish between random events and unknown parameters, and different methods have been developed to treat them. In a manufacturing operator absences, examples of random parameters system, machine failures, material shortages, and changing demands events. Machine that are often unknown. reliabilities are are examples Desensitization of to uncertain- page 10 PERSPECTIVE ties is one of the functions of the operating discipline. In particular, the system's capacity must be computed while taking disturbances into account, and the discipline quirements to within that capacity. must restrict re- The kinds of disturbances that must be treated differ at different levels of the time scale hierarchy: at the shortest time scale, a machine failure influen- ces which part is loaded next; at the longest scale, economic trends and technological changes influence marketing decisions and capital investments. It is our belief that such disturbances effect on the operation of a plant. take these events ty in doing into account, in can have a major Scheduling spite and planning must of the evident difficul- so. 2.6 Feedback In order to make good decisions under uncertainty, necessary to know something amount is of material already processed. goals. strategies essential, especially decisions At the shortest time the conditions of the machines and the designing good feedback It is about the current state of the system and to use this information effectively. scale, this includes it is Control theorists know that generally a hard problem. at the short time scale, are calculated quickly and be relevant that these to long term The tradeoff between optimality and computation gives rise to many interesting research directions. page 11 PERSPECTIVE 3.0 Survey and Critique of Practical Methods The manufacturing ennvironment of important and challenging aware. is one of the richest sources control problems of which we are Until recently, however, the classical and modern control community has not been attracted to this opportunity. is undoubtedly that the manufacturing area has never been per- ceived as needing the help. manufacturers One reason Extreme competition from overseas has, more than anything else, changed this percep- ti on. Another reason why the manufacturing the attention of control theorists and some argue This is is theory has to a large automatically manual the effort. of more fully automated flexibility, manufacturing beginning inexpensive that A wide variety of methods and planning. availCon- assumed plant that systems run largely on to change, however, due to computation, the installation systems, and the additional quality, etc. with scheduling information current or even correct. extent, implicitly controlled; All this is availablity not been, large complex systems is Also, there has not been sufficient able for feedback control that is is that the area has still not, amenable to the their techniques. because, in part, modeling difficult. trol is area has not enjoyed are placed requirements on these are available of of systems. to industry to deal The purpose of this section is to survey these methods and to critique them according to the outline of the previous practice section. section. in controlling A representative manufacturing The intention is to give current state of manufacturing systems survey is of current provided the reader perspective control. in this on the PERSPECTIVE page 12 3.1 Factory Level Control 3.1.1 Traditional Framework The manufacturing community is accustomed to thinking about production control within a particular mature framework. All of the functions necessary for planning and executing manufacturing activities in order to make product most efficiently have been grouped into a few large areas. These areas and the general interrelationships among them are shown in Figure 1. This diagram shows a tremendous amount of interaction, where information is fed forward and back, among the different areas. Also, the diagram deals mainly with the resource allocation aspect of the production control problem. Other important traditional areas that are integral to a successful control system are receiving, cost planning and control, and such financial functions as accounts receivable, accounts payable, etc. Function Descriptions Brief descriptions of the functions major areas are given below: performed within the FORECASTING. Demand is projected over time horizons of various lengths. Different forecast models are maintained by this function. MASTER SCHEDULING provides a "rough-cut" capacity requirements analysis in order to determine the impact of production plans on plant capacity. Comparisons are made between the forecast and actual sales order rate, sales orders and production, and finally, scheduled and actual production. MATERIAL REQUIREMENTS PLANNING (MRP) determines quantity and PERSPECTIVE page 13 timing of each item required -- both manufactured and purchased. For each end-item, the quantity of all components and subassemblies is determined, and, by working backwards of final assembly, MRP determines when production or ordering these subassemblies should occur. A more detailed capacity quirements analysis is made and operation termined. Also, lot-sizing is from the the date performed tends to be highly detailed, there is sequencing at this re- is de- stage. no mechanism of While MRP to take random events and unknowns into account, so it can lead to excess compu- ter usage, in providing and misleading precision, delays schedules, rigidity. CAPACITY REQUIREMENTS PLANNING forecasts work-center load for both released and planned orders and compares this figure with available capacity. The user specifies the number and duration of time periods over which the analysis is ORDER RELEASE is manufacturing planning and execution. function the connection between When an order is creates performed. scheduled the documentation required for release, this for initiating pro- duc ti on. SHOP FLOOR CONTROL that is responsible function performs well as ancillary is a lower level, for carrying INVENTORY tions efficiency, and and tooling. Data is This collected on the of work centers (uti- productivity). CONTROL performs general accounting as well plan. and tracking of product as disposition of product and the performance lization, control function out the production priority dispatching material "real-time" as controlling the storage and valuation func- location of materials. page 14 PERSPECTIVE It also often supports priority allocation ducts or orders, and aids in filling of material to pro- order requisitions. Procedure Planning Many companies These functions have always been performed. For example, separate are organized according to these areas. dedicated groups of people are often given the responsibility of controlling inventory, planning master schedules, etc. The advent of the computer age brought software products that mirror almost exactly the functional above. It is tional area. framework outlined possible to buy software that addresses each funcIn fact, the software is usually modularized so that the system may be acquired piecemeal. Whether or not a manufacturing company has software that and control, the following helps perform production planning procedure very simplified form) is (in STEP 1: Forecast is determined. STEP 2: usually used. - A forecast of future production requirements Master Schedule Planning cast, current inventory, - Using the production fore- and other data, the master schedule planning function makes a rough estimate of short and long term demand on resources. versus demand, Depending a largely upon available manual procedure resource capacity is employed to make adjustments to insure a feasible STEP 3. Requirements Planning and Capacity Planning Material master production schedule. - This function takes the master schedule for finished product and estimated manufacturing lead times, and then determines an "order release" date for each of all materials and components that comprise the finished product. At this step, a more detailed page 15 PERSPECTIVE analysis of resource supply and demand is performed. STEP 4: Order Release - The Order Release function initiates the production plan, as determined by the MRP function. STEP 5: Shop Floor Control - Once the material has reached the floor, the Shop Floor Control function takes over and insures that the production schedule is met. Materials are requisitioned, jobs are assigned and dispatched, and data is collected. STEP 6: Inventory Control - Coincident with the Shop Floor Control function, inventory is controlled in order to store material and fill requests 3.1.2 Recent Finite efficiently. Trends in Production Capacity Material Control Requirements Planning Traditional MRP has offered little more than a computerized method of keeping voluminous records on material, and the resulting resource, requirements. There has never been an attempt, in any but the most superficial way, to account for the actual resource capacity in production planning and control. It has always been handled in an iterative, ad hoc, manual fashion. The manual approach is often a frustrating and impossible task. There is growing interest in devising models that integrate actual resource better factory level capacity with production requirements. In fact, one or two products that claim this capability have come onto the market within the past few years. Products that attempt to perform finite capacity planning often meet mixed reviews because ther treatment cannot be comprehensive. A model formulation and its associated optimization procedure can be specified in a relatively straightforward way, page 16 PERSPECTIVE but solving the problem with finite impossible. Practical approaches computational must reduce king, what often turn out to be, limiting or Kanban The Just-in-Time turing control is a Japanese The objective above. and sub-assemblies. pliers deliver just Just-in-Time of this recent trend in material at just the the right items, philosophy. actually installed inventories control implementation A kanban is is sup- right place, at for forcing a a job ticket that accompa- the assembly process. in an assembly variable of material may be met. When the part is or subassembly, the kanban is sent back to its source to trigger the production The control control is By requiring that external and internal a particular nies a part through approach to manufac- refinement to the approach discussed just the right time, this objective Kanban is the problem by ma- Approach the need for large, expensive to reduce is assumptions. (JIT), or Kanban (Kb), The Just-in-Time resources of a new part. the number of kanban tickets in the system. The high risks of interrupted production due to low inventories are somewhat mitigated by imposing pline on all scheduling facets of manufacturing. producing, Maintenance to reduce the need for inspections. elaborate, and and inventory- Also, very good predictability suppliers that participate of procedures Outside suppliers must insure high quali- portation times and strong communication benefits of disci- must be tightened up, lest the flow of parts that are needed downstream stop. ty in order a great degree this discipline of trans- ties are required of in a JIT program. The long term can lead to productivity increases page 17 PERSPECTIVE beyond the simple reduction of inventory carrying costs berger, This point is 1982; Hall, 1983). elaborated (Schon- in Section 4.4. JIT usually results in smaller and more fre- Implementing of material. quent deliveries the still This can exacerbate necessary task of inventory management. Although zero inventories goal, it to the extent that is an appealing should be moderated required to achieve costs it increase. The JIT philosophy for production control where production requirements and where effects buffering of process are is most applicable known and fixed far in advance, is not required to smooth the unavoidable time variations. This last point is illus- with FMS's, job shops, or trated by the material flow associated any other system where a variety of parts with wide variations in process times share the same resources. failures, buffers are required Even without machine to reap the maximum production. such as the The JIT approach works best in applications assembly process for products tors, automobiles, The etc.). with uncertainties tions are not high enough to require order to achieve predictable sales (refrigera- in these intermediate applica- buffering in the maximum production rate possible. 3.2 Cell Level Control 3.2.1 Traditional There have the activities to determine planning Approach been very few successful in a cell. scheduling problems. Simulation strategies, However, it is approaches one that is floor layout, is expensive to scheduling widely used and for other in both human and page 18 PERSPECTIVE computer time since simulations, to be credible, complex and require a great deal of data. Many simulation the decision are required to make a decision; parameter optimal, or at least satisfactory, "tuned" until tend to be runs must be is behavior found. in Cell Design and Control 3.2.2 Recent Developments Recent developments process are beginning of the manufacturing "flexibility" concept the traditional One in automation and new constraints on the of direction Flowlines, of a cell and how it Cellular or called development, Manuf acturing, in a single cell. This is is and greater Technology was manufactured from intended of product flow and scheduling, less inventory, to be controlled. Cells, stimulated A family of products with very by reports of Japanese successes. similar operation sequences Group is to alter start to finish to lead to a simplification of operations, tighter coupling worker coordination. A second, stimulated by advances in automation and control technology, described is below. the Flexible Manufacturing A good overview cepts can be found in Black Flexible Manufacturing System, of cellular manufacturing (1983) or Schonberger System which is con- (1983). Control A modern example of a cell is a flexible manufacturing system (FMS) which consists of several storage elements, connected system. ters. It machines and associated by an automated materials handling is controlled by a computer or a network of compu- The purpose configuration of the flexibility is to meet production and versatility of the targets for a variety of part page 19 PERSPECTIVE types in the face of disruptions machine such as demand variations and failures. In an FMS, individual of two factors: part processing is practical because system; and the the automated transportation setup or changeover time (the time required to change a machine which is small in from doing one operation to doing another), The combination of these with operation times. comparison tures enables redistribute the FMS to rapidly different parts. fea- its capacity among Thus, a properly scheduled FMS can cope effec- tively with a variety of dynamically changing situations. The size of these systems range from approximately 5 to more than 25 machines. They are also specifically designed for the concurrent processing of a number of different parts (5 - 10 unique parts-types variety of is not unusual), each of which may require (milling, processing A Flexible Manufacturing The operational cell whose main that must be made include: decisions that the following, may be met: (and often conflicting, sub-objectives are evenly balanced among Workload requirements Machine other such tools) to machines the machines and material handling * etc.). master production schedule. Allocation of operations v boring, System is a simple to meet a predefined objective is * drilling, system. failures have a minimum effect on machines' · Work-in-process * Processing work availability. requirements redundancy are (duplicate minimized. tooling) is page 20 PERSPECTIVE maximized. Re-allocation * machines of operations fail such that, in tives listed above, * and tools to machines Real-time allocation tool when addition to those objec- changing of resources effort is minimized. for processing piece- parts such that: * Workload requirements are evenly balanced among the machines and material handling v Quality of processed parts The first two areas are generally is system. maximized. not well addressed by the vendor community and, in fact, often cause the users major operational problems when trying to run an FMS. Because of the diffi- culty in juggling the conflicting under sometimes severe constraints machine (limits on the number of tool pockets per and on the weight the tool chain may bear), it is difficult recent objectives to manually strides have capability The real-time to resources. been taken in solving is not yet widespread into the operating addressed allocate processing very Some this problem, but the and has not yet been integrated software that controls FMS's. scheduling of parts to machines, however, is directly by the vendors that supply Each vendor usually takes a unique problem (this is motivated, each vendor's design) and, because edge, is often reluctant tation. Nevertheless, approach "turn-key" FMS's. to the scheduling in part, by the unique aspects of to divulge of a perceived the after analyzing proprietary details of its implemen- the behavior many FMS's page 21 PERSPECTIVE over a period of time, we can make the following The decision the classes, activity in an FMS are listed below. a detailed schedule construct however, the complexity possible decision source choices availability) that of are needed. these questions v related In practice, the fact. of FMS control taken by the practioners scheduling. Decisions and constraints to dispatch scheduling into FMS are as made they considered when making The criteria Part Sequence and re- in material this. Very little information is decisions. In principle, one can and uncertainty prevents dispatch before for scheduling of the problem (the large number of The general approach is or control variables, observations: - for a variety of are: Since an FMS can pro- cess a number of different parts and since these parts are required active v in certain ratios control Sequencing of relative to one another, of the part input sequence is Fixturings - Many parts required. must make a number of passes through the system in order to process different sides. The sequence for these separate passes could be chosen to enhance the performance of the system. l the system, part must often visit a number of different machines Sequencing before of Operations - processing is complete. Once in a The sequence of these separate machine visits could be chosen to enhance the performance of the system. page 22 PERSPECTIVE v Machine Choice - Of ten operation a particular may be performed at more than one machine. When this is true, a choice must must be made among the posibili- ties. + Cart Choice a Many FMS's employ - number of sepa- rate carts for transporting parts from machine to maWhen the need arises for transporting chine. a part, a choice among the carts of the system must be made. * moving except cart movement: Carts are always while undergoing unload operation load/ queuing at an occupied node. chosen when there is checked for a a destination. Backlogs are of parts are The closest cart with the correct chosen for loading parts. pallet is Deadlocks computed and Intervals are parts are introduced. is routes are periodically. requests for carts: tracked. Shortest or while pallet The closest empty chosen for the each part coming off a shuttle. * Operation Check - and Frequency FMS's Many Selection being built are a Coordinate Measuring Machine (CMM). this machine is Quality for equipped with The purpose of to monitor the quality of being processed as well as the processes the parts themselves. Through the measurement of part dimensions, the nature PERSPECTIVE page 23 of process errors fixture the (tool misalignment, CMM resource of operations is quality standards are misalignment, etc.) may be inferred. limited, the to measure to measure them is errors wear, machine Because intelligent selection and the frequency with which required in order to insure that are satisfied quickly and that processing identified. 3.3 Machine Level Control The bottom tier of the manufacturing of individual work stations, even lone workers Control at the machine or other logistical that have rather for instrumenting include material considerations. These is- at this level are sometimes more in been traditionally and modern control continuous, or for at the cell and factory level. The problems encountered classical comprised be actual machinery level does not really been accounted line with those is (as is the case with manual assembly systems). flow, scheduling, sues have which may structure framework. treated within the The domain is often than discrete, and there is often opportunity the machinery for full automatic control and feedback. 3.3.1 The Traditional In Approach the beginning, there and feedback was accomplished were just hand tools. through eye-hand All control coordination. This continued to be the case, for the most part, up until recently (1950's). and more computers was The tools complex, were the result. (lathes, drill but the principle presses, etc.) remained the applied and Numerically Controlled Here the position, feed, became same. (NC) and speed of larger Then machinery the tool PERSPECTIVE page 24 relative to the part is techniques. In controlled through standard addition, the different operations feedback a part required could be programmed to occur automatically on one machine in the proper sequence. Operation sequencing has not been sufficient changing is generally performed open-loop; reason to alter the sequence. in some environments there This is where there is full automation. It may happen that a tool breaks part way through a "tape segment". If the part has to leave the machine and come back for any reason (quality control check, cult to pick up where extract broken the processing tool, etc.), it is diffi- left off. 3.3.2 Recent Developments Until recently, the position of the tool, and its feed rate and speed relative to the part has been controlled in an entirely open-loop manner. tools, anomolies variables Regardless of what was happening in casting would remain constant. monitoring the power requirements tion of the casting/tooling adjustments and quality, etc.), these This is beginning to change. can be determined and vision is another means by which feedback is for the presence at the machine level. or absence for measuring By of a particular cut, the condi- combination being used in control employed. of made. Electronic techniques dimensions (wearing These systems check of tools in the spindle. Other the wear on these tools are also being page 25 PERSPECTIVE 4.0 Survey and Critique of Recent Research This section reviews recent developments optimization of manufacturing systems. the outset that there is traditional approaches a large in analysis and We should emphasize at body of literature to manufacturing. of Production and Operations Management available on For example, the area (POM) occupies a signifi- cant place in most business schools, and many textbooks exist for this well-developed area (e.g. restrict to the ourselves problems, Tersine, 1980). systems could significantly In addition, proposed luative aspects of manufacturing to areas relevant to the framework as developed in Section 2, and to recent developments believe Here we will in affect the progress we find it Suri (1984a), techniques or in these areas useful to adopt between models. A of set of decisions An evaluative and predicts generative generative system under those decisions. tive, and descriptive, are technique (evaluates) and as eva- technique is and generates is one which takes the performance a of a (The terms prescriptive or norma- also used for these Separation of these two categories means to clarify existing the field. the distinction, one which takes a set of criteria and constraints, a set of decisions. which we two categories.) is useful not just as a research, but also from a practitio- ner's point of view, since solutions based on one or the other type of technique have quite different behavior actual manufacturing found in Suri (1984a). systems. when applied to Details on this latter point can be We recognize that not all techniques fit easily into these two broad categories: some may be a hybrid of both, others may truly fall in between the two; examples of all PERSPECTIVE of these page 26 follow. 4.1 Early Research Early research in manufacturing management science and operations research this was directed at production (Planning quirements ling involves over systems determining the detailed the aggregate planning often refers to decisions refers to low-level In particular, techniques with the mathematical (e.g. Dzielinski zation often problems resource while these re- scheduresources Thus high in the hierarchy and decisions.) problem of fitting of a large number and Gomory, 1965). are and planning very difficult together the produc- Such combinatorial in the sense amount of computer machine failures was concerned of discrete, distinct parts more, they are limited to deterministic including of problems. time period at hand. made scheduling require an impractical effects, Much of a great deal of the work on generative for production tion requirements periods, allocation to particular tasks for the immediate scheduling literature. planning and scheduling a set of future time determines can be found in the optimi- that they time. problems Further- so that random and demand uncertainties, cannot be analyzed. In order to deal with alternative approaches extensive investigation complex have these of large problems. difficulties, been explored. of heuristics systems under realistic development practical hierarchical An excellent The first involves for conditions. approaches two use in scheduling The second involves to solving review of production these scheduling page 27 PERSPECTIVE both exact and heuristic methods, can be found in approaches, Graves (1981). Typical hierarchical Hax and Meal (1975), and Graves approaches are described in (1982). The early work on evaluative models was mainly an attempt to represent the random nature of the production process by using models. queueing-theoretic Most industrial engineers are taught the basics of single server queueing and M/G/1 queues the classic M/M/1 nately, however, that is theory, such as Unfortu- 1975). (Kleinrock, most industrial engineering where courses stop, and IEs today often have the impression models (IEs) today of queueing simple, and impractical. as esoteric, What is less well known in the manufacturing community is that the applicability of queueing theory to manufacturing was enhanced by the development of network-of-queues considerably theory by Jackson (1963), Gordon and Newell (1967), coupled with algorithms the more recent development of efficient computational and good approximation methods. for reasonable manufacturing "first-cut" The state-of-the-art evaluative models of today allows fairly complex systems. Another early development in the area of evaluative was the use of computer-based simulation methods, which employ a "Monte Carlo" approach to system evaluation. accessability simulation this area of computing With the growing power, the development of easy-to-use packages, and the advancement of simulation has made area "close" models major strides forward recently. It theory, is also an to control theory in many ways. In the following, we describe time-scale hierarchy recent research using the of Section 2, rather than the framework of page 28 PERSPECTIVE practical methods of Section 3. from the control amenable The former is more appealing theorist's point of view, and perhaps more to rigorous development. 4.2 Long Term Decisions In this section we consider decisions that involve consi- derable investment in plant, equipment or new manufacturing thods. me- Typically, such decisions may take over a year to imple- ment, and may have an operational lifetime of five to twenty years during which they are expected to pay back. Generative Techniques Traditional term decisions approaches systems-based approaches for generating long include the production planning and hierarchical mentioned casting, decision above, as well as strategic planning, fore- analysis, and location analysis. We do not deal with these here, but an overview and literature survey can be found in Suri (1984b). In the context of automated manufacturing, programming of equipment techniques (LP and IP) have been applied to selection and of production strategies Stecke, 1983; Whitney and Suri, 1984). (e.g. programming be very complex. (1984) has proposed sions, which is a new frame-work Graves, 1982: However, the constraints involved in the mathematical Whitney mathematical for problem formulations sequential deci- developing algorithms for solving these complex optimization can heuristic problems. It has been successfully applied to the problem of selecting parts and equipment for manufacturing Some recent approaches, in a vary large organization. which should be of interest to the page 29 PERSPECTIVE community, use dynamic control The decision to invest in alternative manufactu- (and equipment), over a period of time, is formu- decision-making. ring strategies problem (Burstein and Talbi, 1985; lated as an optimal control 1985: Kulatilaka, Gaimon, insight to help decision-makers models offer Such 1985). However, practical qualitative faced with the complex alternatives that modern manufacturing investment volve. investment models for long term application set of systems in- of these models and use remains to be seen. Another set of recent models tie in long term decision-making use control-theoretic ideas to with shorter term decisions. These are dealt with in the next section. Techniques Evaluative The evaluation of long term effects of a decision on an enterprise is difficult problem, a particularly and evaluative models for long term planning deal primarily with strategic and accounting issues Kulatilaka, 1985). (Hutchinson and Holland, 1982: Jaikumar, Strategic issues 1984: involve such questions as how improved product quality or response time to orders will affect the market trade off current streams. expenditures However, that evaluative issues are currently the issues require models to with future (uncertain) revenue Neither of these areas is of primary interest to the current audience. first is Accounting share. (relative) we should mention two factors. models for strategic undergoing failure of U.S. radical The and accounting changes, industry to make in the face of prudent invest- PERSPECTIVE page 30 ments (Abbott and Ring, 1983: Leung and Tanchoco, 1983). The second important point, which is often missed by those undertaking decisions modelling/analysis are influenced accounting issues. studies, is that the long term to a large extent by these strategic Even though we do not cover them here, it important for analysts working Many modelling efforts fail to be useful to the manufacturing they focus on minor technical overall insight that is and analysis community because points and do not provide the needed for this stage of the planning Professor Milton Smith (of Texas Technological versity) said at the recent First ORSA/TIMS ible Manufacturing is in this general area not to lose sight of the forest for the trees. process. and Conference Uni- on Flex- Systems that around a hundred man-years had been expended on solving minimum makespan scheduling problem, but he did not know of a single company that used minimum makespan to schedule their shop floor operations! 4.3 Medium Term Decisions Here we are concerned with a time period ranging anywhere from a day to primarily a year, and the scope of the decisions involves tradeoffs between different modes of operation, but with only minor investments in new equipment/resources. Traditionally, such decisions have been the domain of master scheduling, MRP, and inventory management systems, which are reviewed in Section 3. uncertainty in a direct manner, through the use of or other quantities. deterministic These systems generally do not account but rather, "safety" values, whether Master scheduling plan, which gets in an indirect for way in stocks, lead times and MRP systems work to a updated periodically (say once a page 31 PERSPECTIVE week, or once a month). Inventory uncertainty theory models to derive optimal and analyzes the stock policies. effects of Inventory stocking policies assume that each item stocked has an exogenous demand, modelled by some stochastic stocking process, and attempt to find the best policy for each item. The fact that the demand the master scheduling is and MRP decisions clear that much useful information being thrown away. manufacturing entire Of course, it is system that makes it problem simultaneously. able structures the size and complexity of a Nevertheless to solve the we feel that suit- Some attempts more in this direc- Techniques decision-making. cept of time generative Control scale planning techniques theorists are decomposition should therefore readily problem for decision making is in this section. We begin by reviewing production ignored, and thus it very difficult these components. tion are described Generative is can be developed to make the decision-making coherent across of on inventory comes as a result of familiar with the con- and hierarchical understand (See references above.). It The are many: in addition to computational approach requires less detailed data, and it structure (Meal, and the shor- imposes con- advantages of the hierarchical approach organizational partitions with successively The solution of each subproblem straints on lower subproblems. control, the idea behind hierarchical into a hierarchy of subproblems, ter time scales. for this level 1984). savings, this mimics the actual page 32 PERSPECTIVE The original structure on theorists should by ideas intuitive and find it the use of an liers, that the hierarchy, based derations, is derived by the planning manufacturing explicitly in the the around is failure the be replaced By is difficulty state using assume over of derived from but the of solving by a static (or each capacity set of policies re-solving ment what Bellman this termed failures as to a is primarily hierarchical heuristic arguments interaction of the in of problem. based This in periodically, feedback" on the To get optimization the time to do procedure between system condition). an "open loop randomly the solution. a stochastic problem includes an intractable approaches to what uncertainty (hence leads average developments uncertainty programming that the dynamics methods. of time, and then This proposed alternative feasibility consi- Recent to approximate (1980) as a natural An to represent the new showed, that demand and capaci- a period usually control multip- multiplier periodically. on a mathematical propose "open loop" state. this propose and Suri problem. formulation. considerations, based lem, they states problem Since the hierarchy tractability levels (1981) approaches the contribution Hildebrant (1982) but rather but also equipment ways that they where Suri problem capacity). problem, that Graves systems have sought not just demand, varying However, could be derived optimality known and deterministic re-solve the hierarchical formulation and Lagrange hierarchy not on based arguments. a primal optimization These hierarchical in heuristic appropriate of approach interesting Hax-Meal decomposition ty are for this prob- between failure spent in leads each to an each failure they can policy. impleThe PERSPECTIVE page technique showed reasonable for automated improvement manufacturing (Hildebrant, over existing 33 heuristics 1980). Kimemia (1982) and Kimemia and Gershwin (1983) have derived an alternative, proach sues closed loop solution to this problem. has also (the backlogs been to separate response to machine and surpluses) dispatching. the relatively failures from the short and Their ap- long term isto production term problem of part The long term problem accounts for the dynamics be- tween failure states, as well as that due to demand, in more detail than the above approach. The mic long term problem programming given problem. the control objective is to production and demand be within a capacity tional machines. as vector of process. and there minimize modeled The by a discrete Markov tor is is a continous machine states The production a specified demand the cumulative difference set that is determined is rate vec- is and the production rate is dyna- rate. The between constrained to by the set of opera- A feedback control law, which determines the next part to be loaded of current machine and when it should be loaded as a function state and current production surplus, is sought. This machine formulation, failures, Suri (1980), exactly. which reflects had previously this involves approximation disruptive proposed but they had concluded that it The contribution find a good been the was too hard to solve exact solution. two steps in a dynamic of by Olsder and of Kimemia and Gershwin to the nature has been to Essentially, programming framework: page 34 PERSPECTIVE the top level separating (the solution to problem a Bellman into a number of subproblems, obtained formally through equation) a constraint-relaxation procedure, value function for each subproblem principle and then approximating by a quadratic. the middle level of the is to a sub-optimal control u. times The maximum which gives rise hierarchy, this case, the maximum principle problem. is a linear programming dispatch In the Then, at the lower level, one to attempt chooses part rate u. Simulation experiments indicate to achieve this flow that these procedures give rise to quite good policies in the face of the various uncertainties. More recent work by Gershwin, Akella, and Choong (1984) Akella, Gershwin, and Choong computational (1984) has further simplified the Simulation results effort. system is vior of a manufacturing and indicate that the beha- highly insensitive to errors in the cost-to-go function, so the Bellman equation can be replaced by a far simpler procedure. simplifies the on-line Evaluative The quadratic approximation linear program. Models Evaluative models for this decision-making approaches both analytic the ability to represent chine failures) intractable. Koenigsberg and simulation. (1959). (1967, 1971, 1976) that give features (such work in this surveyed by were made by Buzacott who looked at various approximate insight into these off again analytically field is Notable contributions are as ma- buffer stocks, in order to trade For large systems, this is The earliest level involve Important production uncertainties and limited between the two. models further issues. analytic page 35 PERSPECTIVE Most of these analytic studies are based on Markov models of transfer lines and other production systems. An appreciation for the difficulty of the problem is seen from the fact that the largest general model for which an exact solution is available involves three machines with two buffers between them (Gershwin and Schick, 1983). of a technique for decomposing a production line into a set of (Gershwin, 19831. one-buffer subsystems two-machine for solving this system is analogous point boundary value problems. the method is problems cycle is results indicate that Numerical very accurate, and what is times however, The procedure to the idea of solving two- more, fairly large (20 machines) can be solved in reasonable technique has come out A promising recent development currently restricted to the time. case where Altiok (1984) of the machines are identical. recently developed methods for systems with more general type time processing Other evaluative the has phase- distributions. queueing techniques include and simulation. Both of these methods decision-making as well. models are The However, network models, can be used for short term we feel that queueing network best suited to more aggregated decision making, while simulation is more suited to detailed decisions. Therefore we discuss the former here and the latter in the next subsection, although the particular application may suggest the use of the other technique for either of these A fairly recent manufacturing network development systems has models in been for system levels. (analytic) the growing planning one or evaluative models of use of queueing and operation. A simple- page 36 PERSPECTIVE minded, static, capacity account the system model allocation does not take into and uncertainties dynamics, interactions, inherent in manufacturing systems. Queueing these features, able to incorporate tions, and thus enable with some restric- albeit more refined network models are evaluation of decisions for The increased use of such models stems manufacturing systems. algorithms primarily from the advances made in the computational available to solve queueing networks, both exactly and approxima- tely. (1980) Solberg tractable. made the solution of these systems algorithm (1973) Buzen's FMS, and Stecke (1981) applied this to capacity for solving production used it for planning developed a number of approximate Shanthikumar (1979) problems. planning queuing models for manufacturing. The development of the mean technique for solving these networks (Reiser value analysis (MVA) and Lavenberg) opened up a host of new extensions and approxima- Various approximate MVA algorithms have tions. (Schweitzer, 1979: Bard, 1979) solution which enable of very large networks. Hildebrant (1980) fast and accurate Hildebrant applied MVA techniques been developed and Suri (1980). and to both design and real- time operation problems in FMS. An extension (Suri and Hildebrant. 1984) solution of systems with machine-groups, enabled efficient and has been incorpo- rated in at least one on-line decision support system for an FMS (Suri and Whitney, 1984). (Shalev-Oren, features et al., 1984) Another recent extension, called PMVA allows a wide variety of operational to be modelled. An important reason for the increasing popularity of such page PERSPECTIVE models in manufacturing is that they have proved in the area of computer/communication of giving reasonable performance have given a theoretical network models The disadvantages Recent results basis to the robustness of queueing situations of queueing (Suri, represent space. (Some recent steady-state modelling way, and they certain other features, such as limited buffer and Suri and Diehl output measures 1983). network models are that they model many aspects of the system in an aggregate fail to their usefulness systems modelling, in terms predictions. for use in practical 37 developments, (1983) e.g. do allow Buzacott and Yao (1982) limited buffer sizes.) they produce are average operation of the system. transient effects The values, based on a Thus they are not good for due to infrequent but severe disrup- tions such as machine failures. However, the models tend to give reasonable system performance, require estimates of and they are very efficient: that is, relatively little input data, and do not use ter time. A typical FMS model (Suri they much compu- and Hildebrant, 1984) might require 20 to 40 items of data to be input, and run in I to 10 seconds on a microcomputer, bers for simulation. to quickly arrive can then refine in contrast to the much larger num- Thus these at preliminary these models can be used interactively decisions. More detailed models decisions. Queueing network models suggest themselves for use in the middle level models of a hierarchy. Development along with suitable control and higher levels of the hierarchy, of queuing aspects to tie network in the lower could be a useful topic of page 38 PERSPECTIVE research. Lasserre (1978) and Lasserre and Roubellat the medium term production planning program of special technique (1980) represent planning problem as a linear structure, and develop an efficient solution for it. 4.4 Short Term Decisions Decisions at this level typically have a time a few minutes up to about a month. have included those exact approaches Traditional frame of from generative models for lot sizing and scheduling, using both as well as heuristics or rules. a number of recent interesting developments There have been in this area, which are now described. Traditional lot sizing models traded off the cost of setting up a machine with the cost of holding inventory, on an individual product basis. The Japanese (Just-In-Time and Kanban) approaches have challenged these concepts as being narrow-minded and myopic in terms of the long term goals of the organization. cate operation with minimal or no inventory, not only saves inventory carrying They advo- claiming that this costs, but also gives rise to a learning process which leads to more balanced production in the long run ISchonberger, This thesis is try as well. becoming more widely accepted However, and inefficiencies 1982; Hall, 1983). reduced inventory in the short term. It ask what is the optimal rate of reducing in U.S. indus- leads to line stoppages is therefore logical to inventory, so that short term losses are traded off against long term gains due to increased learning. of this point (Suri There has been some preliminary and DeTreville, 1984). It is investigation a problem that PERSPECTIVE page 39 would fit naturally into an optimal control framework, and further investigation would be useful. There have also been some recent studies indicating that lot sizing in a multi-item environment ought to be treated as a vector optimization problem. each product affects The idea is that the production primarily through the queueing rate of other products, of each lot of parts waiting for other lots to be done at each machine. consider This integer programming computationally tem. Therefore, one ought to the joint problem of simultaneously lot sizes. intractable the lot size of optimizing problem would normally be for any realistic manufacturing However, by modelling the system as a queueing then solving a resulting nonlinear program some have been obtained development that needs Hitz (1979, duling (Karmarkar further et al., 1984). recent results This is a promising deterministic sche- systems: In these systems, parts follow a common path from machine to machine. He found that appropriately, a periodic sequence times. network and class of flexible manufacturing flexible flow shops. sys- exploration. 1980) studied the detailed, of a special all the he could design This substantially reduced by grouping parts of loading the combinatorial optimization problem. Erschler, Roubellat, nistic, combinatorial class of optimal and Thomas scheduling decisions. technique Rather part to send into the system next, it of candidate choices. (1981) describe a determi- that searches for a than deciding which presents to the user a set This flexibility is intended as a response PERSPECTIVE page 40 to the random events such as machine to represent failures that are difficult explicitly in a scheduling model. Perhaps the most widely used evaluative tool for manufacturing this systems today is simulation. context refers specifically simulation. The term "simulation" to computer-based discrete event Such a model mimics the detailed operation of the manufacturing system, through a computer program which effective- ly steps through each event that would occur in the system be more precise. simulation models can be made very accurate is the programming time to generate detailed time the model is analyst tries (or to each event that we wish to model). In principle, the price in run. time to create the model, -- the input data sets, and the computer time each In addition, the more phenomena that the to represent, the more complex the code, and the more likely there are errors, some of which may never be found. In addition, it is simulations is limited by the judgment and skill of the program- mer. Detail sometimes forgotten that the accuracy of a and complexity are not necessarily accuracy, if major classes of phenomena synomous with are left out. (While simulation can be used at any of the levels of the decisionmaking process, we choose to describe it here since it mine the most detailed operation of a manufacturing can exa- system.) Two reasons for the recent popularity of simulation are the number of software tools that have been developed to make simulation more accessible in computing costs factors make it lation before to manufacturing designers; and the decrease and the availability of microcomputers. well worth an organization's making large investments. In These effort to use simuaddition, there have page 41 PERSPECTIVE been many developments in the design and analysis of simulation experiments, which have contributed tion as a valid and scientific Recent developments categorized to the acceptance methodology development output graphics). in this field. in software tools for simulation can into simulation languages, active model of simula- "canned packages", (or graphical (or features will not be discussed further here, but good examples are SIMAN input) inter- input), and animation The two input/output graphics be (for and SEE WHY (for output). Although simulation languages have been around for a while, the last five years have seen the development languages, such as GPSS/H, SIMSCRIPT II.5, and SLAM II, as well as the development facturing user now available of languages specially tailored (e.g. SIMAN and MAP/I). on microcomputers UI.5, SIMAN, MICRONET). the manufacturing packages, of many powerful Also, most languages as well (e.g. are GPSS/PC, SIMSCRIPT Another development, designer, has to the manu- specially geared to been the development of canned which do not require programming skills, but are com- pletely data driven (e.g. GCMS, GFMS, SPEED). have a number of structural assumptions of how the manufacturing very quick analysis of they built into them, in terms system operates, but can be useful for a system. trum, for very detailed simulation to a programming Of course, At the other end of the specit may be necessary to resort language such as FORTRAN or PASCAL. Clearly then, there are tradeoffs (1982) involved among these op- tions. See Bevans for a discussion. Even though simula- tion is perhaps the most widely used computer-based performance PERSPECTIVE page 42 evaluation tool for manufacturing systems, we would recommend greater use of analytic and queueing conducting the more expensive to the numbers network models simulation studies -- prior to in comparison quoted for queueing network models, a simulation model might require 100 to 1000 data items and 15 to 10,000 seconds to run on a microcomputer. In the area of simulation design and analysis there have been several developments control community. tion, detection control The analysis of simulation outputs -identification, parameter involves -- spectral that should be of interest to the bias of has used many techniques from time series analysis and methods (see Law, 1983, for a survey). mization in simulations involves stochastic niques 1984), (e.g. Meketon, 1983; Parameter opti- approximation Ho and Cao, 1983; tech- Suri and Zazanis, which again are familiar ground to our community. A recent development, crete and run length transients, and initial genera- interval confidence which event systems, meters in simulations technique is enables very efficient (see Ho et al., 1984 related to linearization again has parallels Cassandras, called perturbation with conventional 1982; Ho, 1985). analysis optimization of para- for a survey). This of dynamic dynamic Essentially of dis- it systems, and systems (Ho and enables the gra- dient vector of system output with respect to a number of parameters to be estimated sense it is an evaluative not only evaluates improving by observing only one sample path. and "semi-generative" decisions but also suggests In this tool, since it directions for the decisions. While much of the original work on perturbation analysis PERSPECTIVE page 43 relied on experimental results to demonstrate its accuracy Ho et al., 1979 and 1983), recent analyses have given it rigorous it is foundation (Suri, 1983a and 1983b), probabilistically Cao, 1983: Suri and correct for certain Zazanis, better than repeated systems (Cao, 1984; a more and also proved that 1984; Cao, 1985), simulation (e.g. (e.g. as Ho and well as Zazanis and Suri, 1984). Another recent interesting development tion of Petri net theory to the performance turing systems (Dubois and Stecke, use of Petri nets (in computer titive questions as: will 1983). science) was to answer such quali- there be any deadlocks? of in the However, there theory of timed nets. Following al. analysis of manufacIn the past, the main have been some important recent advances Petri has been the applica- [11983, some work by Cunningham-Greene 1984) have developed production processes. complex set of e.g. performance control-theoretical transfer functions, main disadvantages with completely evaluative, it It controllability, to some situations, However, it events systems, observability. has currently are that it not generative. answers also gives rise to a parallel concepts for discrete determinstic Cohen et a linear system theoretic view This enables efficient questions. (1982), can only deal and that it is The is only a promising new development. One that of level. of the most useful areas that require more research is real time control Little theoretical of manufacturing research systems, at a detailed has been done on this, apart PERSPECTIVE page 44 from the large body that exist of heuristics A few researchers (Graves, 1981). formal way (Buzacott, Akella, Choong, have treated the issues in a 1982: Gershwin, and Gershwin, for scheduling 1984; Akella, and Choong, Cassandras, 1985). 1984: This seems to be an area for control theorists to apply their experti se. Indeed, Ho et al. (1984) Dynamic have coined System, the term DEDS, for Discrete Event to emphasize that manu- facturing systems are a class of dynamic systems, and that there are concepts from dynamic system theory that need to be developed or applied for DEDS as well. analyzed either by purely chains, queues) In the past we have seen DEDS probabilistic approaches or by purely deterministic and other combinatorial methods). well as the perturbation (e.g. Markov approaches (scheduling The work by Cohen et al. analysis approach, have shown the as use of a dynamic systems view of the world. As an example, Suri and Zazanis tion analysis ly optimize combined with stochastic while a facility is have used perturba- approximation to adaptive- This could be used, for example, a queuing system. for improving the choice parts, (1984) of lot sizes for a number operating -- of different the approach is to implement and has obvious applications in real systems. ever, many interesting questions be answered for this adaptive of convergence, method. etc. simple How- remain to page 45 PERSPECTIVE 5.0 Conclusion We have described a framework for many of the important problems in manufacturing people systems that need the attentions of trained in control and systems theory. existing practical fall short. methods solve those problems, and where they We have also shown how recent and on-going research fits into that framework. been to encourage analysis We have shown how efforts very important An important goal of this effort has control theorists to make the modeling and that will lead to substantial field. progress in this page 46 PERSPECTIVE REFERENCES Abbott, R. A. Effects on and E. A. Ring (1983, "The MAPI Method Productivity: An Alternative is -- Needed, Its " 2, 1, 1983, 15 -30. Akella, R., Y. Choong, Hierarchical and S. B. Gershwin Production on Components, Scheduling (1984), Policy, " IEEE and Manufacturing Hybrids, "Performance of Transactions Technology, Sep- tember, 1984. Altiok, T., "Approximate with Blocking, " Eur. Analysis of Exponential J. Bard, Y., "Some Extensions sis, " in Performance Oper. Res. 11, 390-398, Queueing to Multiclass of Tandem Queues Systems, Computer 1982. Network Analy(M. Arato, ed. North Holland, Amsterdam, 1979. Bevans, J. P., "First, Choose Machinist, pages Black, J. "An Overview T., Comparison Vol. 15, No. to 143-145, an FMS Simulator, May 1982. of Cellular Mnaufacturing Conventional Systems, 11, November, 1983, Talbi, C. and M. Introduction of Flexible Manufacturing "Economic a Dynamics Operations Research, Buzacott, J. A., " Industrial Systems and Engineering, pp. 36-40. Burstein, M. Procedures Versus " American Justification Technology: Based Approach", to appear, "Automatic Transfer for the Traditional Annals of 1985. Lines with Buffer Stocks", ), page 47 PERSPECTIVE Int. J. Prod. 5, Res. Systems, " 1 8 3 - 2 0 0. of Inventory Banks in Flow-Line Pro- Buzacott, J. A., "The Role duction 1 9 6 7, 3. J. Int. Prod. Res. Vol. 9, No. 4., pp. 1971. 425-436, Buzacott, J. Limited A., "The Production Capacity of Job Shops Storage 1976, pages Space, " Int. J. Prod. Res. , with 14, Vol. No. 5, 597-605. Buzacott, J. A., "Optimal Operating Rules for Automated Manufacturing Systems, trol AC-27, Vol. Buzacott, J. Flexible on Automatic " IEEE Transactions 1, No. pages 80-86, A. and Shanthikumar, Manufacturing J. Systems, February G., "Models " AIIE Trans. Con- 1982. for Understanding , pages 339- 350, December 1980. Buzacott, J. A. tems! A Review and Yao, D. D. W., "Flexible Manufacturing of Models, "Working Paper No. 7, Department IndustrialEngineering, University Buzen, J. P., "Computational works with Exponential of Toronto, March, 1982. of Algorithms Servers, Sys- " for Closed Queueing C. ACM, Vol. 16, Net- #9, September 1973, 527-531. Cassandras, rial C., "A Hierarchical Handling appear, 1985. Systems", Routing Control Annals of Scheme for Mate- Operations Research, to page 48 PERSPECTIVE Cao, X. R., "Convergence of Parameter Sensitivity Estimates Stochastic Proc. trol, Experiment", Dec. Cohen, Conf. to Function Analysis of Queuing Operations View IEEE and Conf. Research, 1985. Dec. of Discrete-Event Control, G., D. Moller, J. P. Quadrat and M. Theory For Discrete Event Systems", Control, December Cunninghame-Green, Approximating Vol. Erschler, Decision J., R. A., F. June, Draper and dans L'Ordancement Toulouse, Lab. Viot, Proc. "Linear System 23rd IEEE Conf. Dec. Behaviour", Processes Op. Res. and Quar- 1982. Roubellat, Scientifiques 22nd 1983. "Describing Industrial 95-100, Journ6es Proc. 1984. Their Steady-State 13, "A Linear- Processes", December Cohen, terly, and Con- Networks," G., D. Dubois, J. P. Quadrat and M. Viot, System-Theoretic and Decision 1984. Cao, X. R., "Sample submitted IEEE in a et V. Thomas d'Atelier (1981) en Temps Techniques "Aide R6el, Automatis6e, a " 3e Manufacturing Systems Handbook, The Charles Stark Draper Laboratory, Cambridge, Massachusetts, 1982. me ADEPA, 1981. The Flexible la page 49 PERSPECTIVE Dubois, D. and K. E. Stecke, duction Processes", Proc. "Using Petri Nets 22nd IEEE Conf. to Represent Pro- Dec. and Control, December 1983. Dzielinski, Sizes, Vol. B. P. and Gomory, R. Inventory 11, No. and Labor 9, 874-890, E., "Optimal Programming Allocations", Manag. of Lot Sci. July 1965. Gaimon, C., "The Dynamic, Optimal Mix of Labor and Automation", Annals of Gershwin, Operations S. Research, to appear, 1985. B., "An Efficient Decomposition Method for the Approximate Evaluation of Production Lines with Finite Storage Space, " Massachusetts Institute of Technology Laboratory for Information and Decision Systems Report LIDS-R-1309, December, 1983. Gershwin, S. B., R. Akella, and Y. C. Choong, "Short Term Produc- tion Scheduling of an Automated Manufacturing chusetts Institute of Technology Facility, " Massa- Laboratory for Information and Decision Systems Report LIDS-FR-1356, February, 1984. Gordon, W. J. and G. F. Newell, "Closed Queueing Systems with Exponential Servers", Op. Res. 15, No. 2, April 1967, 254 -265. Graves, S. C., "A Review of Production Scheduling, " Ope- pp. page 50 PERSPECTIVE rations Research, Vol. 29, No. 4, 1981, pages Graves, S. C., "Using Lagrangean Techniques Production Problems, " Management 28, No. Planning to Solve Hierarchical Procedure for Manufacturing IEEE t e r s , Hall, (ed. R. International W., Zero Programming System Design and Evaluation, " Conference Inventories, C. and Meal, H. Planning ), Vol. on Circuits and Compu- 1 9 8 0. Hax, A. tion Science, 3, March 1982, pages 260-275. Graves, S. C. and Lamar, B. W., "A Mathematical Proc. 646-675. and C., "Hierarchical Scheduling, North Holland, Dow-Jones " in Irwin, 1983. of Produc- Integration Logistics, M. A. Geisler, 1975. Heidelberger, P. and P. D. Welch, "A Spectral Method for Confidence Interval C. ACM 24, Hildebrant, Machines Generation and Run Length Control 4, 233-245, April R. R., "Scheduling are Prone xible Manufacturing sium on Large 1981. Flexible Machining Systems when to Failure", Ph. D. Thesis, MIT, Hildebrant, R. R. and Suri, R., Algorithm Structure in Simulations", "Methodology for Scheduling Systems, Engineering and Multi-Level and Real-Time " Proc. Systems, 1980. Control of Fle- 3rd International Sympo- Memorial University of page 51 PERSPECTIVE Newfoundland, July 1980, pages 239-244. Hitz, K. L. (1979), Laboratory for Information 879, January, Laboratory Ho, Y. C., "A "Scheduling of Flexible Flow Shops--II, " MIT and Decision Systems Report LIDS-R- 1980. Survey of the Perturbation Analysis of Discrete Dynamic Systems, " Annals of Operations Research, to 1985. Ho, Y. C. of and Decision Systems Report LIDS-R- for Information 1049, October, appear, Shops, " MIT 1979. Hitz, K. L. (1980), Event "Scheduling of Flexible Flow and X. R. Queuing Cao, Networks, "Perturbation JOTA, Vol. Analysis 40, No. 4, and Optimization 1983, pp. 559- 582. Ho, Y. C. and C. Cassandras, Discrete trol, Event Dec. Systems", "Computing Proc. 19th Determine Parameter Science, Vol. Conf. Dec. and Con- 29, Chien, "A New Approach to Sensitivies of No. 6, 1983, Ho, Y. C., M. A. Eyler, and T. T. Buffer Storage International IEEE for 1980. Ho, Y. C., M. A. Eyler, and T. T. for General Costate Variables Journal 6, 1979, pp. 557-580. of Design Transfer pp. Chien, Lines, " Management 700-714. "A Gradient Technique in a Serial Production Line," Production Research, Vol. 17, No. PERSPECTIVE Ho, page 52 Suri, R., Y. C., Cao, X. R., Diehl, W., Dille, G. Form) Queueing Networks Using Perturbation Scale Systems, to Hutchinson, G. Flexible appear, Volume 1, No. 2, pages " Journal J. R., "Jobshop-like 10, 1, October 1963, of Manufacturing Systems, Systems", Manag. Sci., Systems: A Managerial Per- Queueing 131-142. spective", Harvard Business School, working U. S., Kekre, S. and Kekre, Item Multi-Machine Large R., "The Economic Value of Jaikumar, R., "Flexible Manufacturing Karmarkar, Analysis", 215-228, 1982. Jackson, No. and 1984. K. and Holland, J. Automation, W. (Non-Product- of Large Multiclass Zazanis, M. A., "Optimization J. paper, 1984. S., "Lotsizing Job Shops", Univ. of Rochester, in Multi- Grad. School of Mgt., Working Paper QM8402, March 1984. Kimemia, J. G. (1982), Flexible Manufacturing "Hierarchical Control Systems, Massachusetts of Production in Institute of Tech- nology Laboratory for Information and Decision Systems Report LIDS-TH-1215. Kimemia, J. G. and S. B. Gershwin Computer Control tems, " IEE pp. 353-362. (1983), "An Algorithm for the of Production in Flexible Manufacturing Transactions, Volume 15, No. 4, Sys- December, 1983, page PERSPECTIVE in Investments Annals FMSs", 1975. Wiley, Support System to Evaluate N., "A Managerial Decision Kulatilaka, appear, John Systems, Queueing L., Kleinrock, 53 to Research, of Operations 1985. Research, Operations 6, 31, No. Vol. Output Data, " of Simulation Law, A. M., "Statistical Analysis 983- 1983, December, 102 9. J. Lasserre, de "Etude (1978), B. la Planification Moyen a Terme d'une Unite de Fabrication, " thesis presented to Universite Sabatier Paul Lasserre, for degree the J. B. and F. Roubellat Docteur of (1980), "Une Resolution de Certaines Programmes Lineaires lier, " RAIRO 191, May, Operations Research, Volume a Structure en Esca14, No. 2, pp 171- 1980. on Productivity Analysis Manuf. , Sys. Vol. -2, An Alternative No. 2, Business Review, Vol. 62, 175-188, No. Decision Based to the MAPI Method", 1983. Meal, H. C., "Putting Production Decisions Where Harvard de Methode Rapide Leung, L. C. and J. M. A. Tanchoco, "Replacement J. Ingenieur. They Belong, 2, March, 1984, 102- 111. Meketon, M. S., "Optimization in Simulation: A Tutorial", pre- page 54 PERSPECTIVE sented at Winter Simulation Conference, WSC83, December t1sder, G. J. and Suri, R., "Time-Optimal Manufacturing IEEE Conf. Dec. Operations No. Systems with Failure and Research, Control, Special Issue Multichain Networks," Schonberger, Press, R. 19th 1980. on Simulation, Vol. 31, J., S. S., "Mean J. ACM Japanese 27, Value Analysis pages 313-323, Manufacturing of Closed April 1980. Techniques, Free 1982. Schonberger, R. Just-In-Time J., "Integration Production, 11, November, 1983, Schweitzer, P. " of Cellular Manufacturing Industrial Engineering, Stochastic Seidman, J., "Approximate Analysis Control and Optimization, A. and Schweitzer, a FMS Cell, Working of Multiclass Vol. 15, P. Paper J., Amsterdam, Real-Time QM8217, On-Line Graduate S., A. Seidman and P. Schweitzer, Control School Shalev-Oren, Systems with Priority on 1979. University of Rochester, New York, 1982. J. Closed Conference Management, Flexible Manufacturing and pp. 66-71. Networks of Queues," Presented at International of Proc. 6, 1983. Reiser, M. and Lavenberg, No. Control of Flexible Prone Machines", Dec. 1983. "Analysis of of Scheduling: PMVA", page 55 PERSPECTIVE Annals of Research, Operations Shanthikumar, to J. Solberg, J., CAN-Q School of Industrial IN, User's Guide, Purdue Engineering, facturing Ph. D. Systems, Engineering, Production Problems Sci. R., No. 9 (Revised), University, W. Lafa- School W. Lafayette, IN, 29, No. Industrial of 1981. and Solution of Nonlinear for Flexible Manufacturing Integer Systems", 3, 273-288, March Management Concepts for Large , Vol. Resource for Flexible Manu- Problems Dissertation, Purdue University, Stecke, K. E., "Formulation Suri, Report 1980. Stecke, K. E., Production Planning Manag. Engineering, Uni- Industrial Toronto, 1979. versity of yette, 1985. Queueing Models of Dynamic Job J. G., "Approximate " Ph. D. Thesis, Department of Shops, appear, 1983. Systems, Pergamon Press, 1981. Suri, R., ACM, "Robustness Vol. 30, No. of 3, Suri, R., "Implementation Carlo 1983b. Experiment", J. Queueing 564-594, of Network July Formulae, J. 1983a. Sensitivity Calculations Optim. " Theory Applic. on a Monte 40, 4, August, PERSPECTIVE Suri, R., Dynamic and page 56 "Infinitesimal Systems, Control, Perturbation Analysis A General Dec. Theory", Proc. 22nd IEEE pear, Event Conf. Dec. 1983c. Suri, R., "An Overview of Evaluative Models facturing of Discrete Systems", Annals of for Flexible Manu- Operations Research, to ap- 1985. Suri, R., "Quantitative Techniques Chapter of in Handbook for Robotic Industrial Suri, R. and DeTreville, S., "Getting Systems Robotics, Wiley from Just-In-Case Analysis", 1984b. to Just- In-Time: Insights from a Simple Model", Harvard Business School, Division of Research Suri, R. and Diehl, Working Paper, September, G. W., "A Variable Use in Analyzing Manufacturing ber, 1983, Surf, R. and Hildebrant, Systems 3, No. submitted Using 1, 27-38, Mean to Buffer Size Model and Its Networks with Blocking", Septem- Management R. R., Value 1984. Science. "Modelling Analysis, " Flexible J. Mfg. Manufacturing Systems. , 1984. Suri, R. and C. K. Whitney, "Decision Support Requirements in Flexible Journal No. 1, Manufacturing", 61-69, Vol. 1984. of Manuf. Sys. , Vol. 3, page 57 PERSPECTIVE Suri, R. and Zazanis, M. A., "Perturbation Analysis Gives Strong- ly Consistent Estimates for the M/G/L Queue", submitted to Management Suri, R. Science, and Zazanis, in Manufacturing Sept. 1984. M. A., "Adaptive Simulations Using Perturbation king paper, Harvard University, Tersine, R. J., Production Optimization of Lot Sizes Analysis", wor- 1984. and Operations Management, North Holland 1980. Whitney, C. K., submitted Whitney, Selection tions to "Control Principles J. Manufacturing in Flexible Systems, Manufacturing," 1984. C. K. and R. Suri, "Decision Aids for Part and Machine in Flexible Manufacturing Research, to appear, Yao, D. D. W. and Buzacott, Manufacturing Systems J. Systems", Annals of Opera- 1985. A., "Modeling a Class with Reversible Routing, of Flexible " Tech. Rep. 8302, Dept. of IE, Univ. of Toronto, August 1983. Zazanis, M. A. and R. Suri, "Comparison with Conventional tic Systems," Sensitivity submitted to of Perturbation Analysis Estimates for Regenerative Operations Research, Stochas- 1985. Customers ) Forecasting ~Orderorders OrdrOrders Delivery Dates Processing ,tMasteria Available Capacity Scheduling Availability Schecdr ......... .. Manufacturing Capacity Requirements Planning .... Resource Planning Sd. ......... inventory Floor Shop Control s hedu/e PFloorolhManufacturing Acrx .eeds Management i.... :.: : Inventory HaiMntengance Product Figure 1 Traditional Production Control Framework Customers ) (Forecasting Oraders al,/very Order Processing Dates Available Capacity terial _ Availability Master Scheduling ::::.:..:. ....... :::: ............-.-.-. . ... '"""""":-:' ........ .....-:..... ::- :::::.- Manufacturing Resource Planning Capacity Requirements .. Planning Needs ::...:::. " ".:: .... fl.rder R!elease Release .. Purchasing :I. nventory Shop Floor Management schiedu/e < jI ::::i. . / Manufacturing Services _ , V.ee.i..' AInual Prdo Schedule Results Control *... .. . . . . . . . . . .ii. IProductionl I Too/s MJaintenance ProdUct OCSS M r Figure 1 Traditional Production Control Framework