Conclusions
A Dynamic Approach to Operations Management:
An Alternative to Static Optimization
Ramchandran Jaikumar and Roger E. Bohn
1. INTRODUCTION
Advances in technology and changes in the nature of competition have affected the structure of manufacturing and service operations. Yet, we have seen no concomitant change in the paradigms for modeling and managing operations, despite acknowledged dissatisfaction with them. The paradigm that remains in use can be crudely characterized as "static optimization of a known objective function, subject to known and stationary constraints." This paradigm holds manufacturing separate from knowledge-creating activities, such as product design, process design, and scientific research. Thus research on R&D and engineering have traditionally been a separate subdiscipline from research on manufacturing.
We propose a dynamic approach that explicitly treats important elements of modern manufacturing operations that are ignored by static paradigms-- specifically, knowledge, learning, problem solving, and contingencies. We show that static and dynamic perspectives are effective in different manufacturing contexts and offer evidence that the domain in which a dynamic approach is applicable is growing. It is no longer sufficient to ask "How can we make this operation more efficient at its existing tasks?" Research must also be directed at "How can we become better at recognizing and dealing with contingencies, learning from their resolution, and accumulating a broader base of knowledge?" Some managers are asking this question, but the use of dynamic and knowledge oriented approaches remains an ad hoc pursuit, without much theoretical foundation.
We begin by articulating some of the implicit assumptions in operations management research. Section 3 describes our suggested alternative, a "dynamic approach" that emphasizes the importance of knowledge in production systems. Section 4 discusses the design and operation of an automated assembly line. We show that traditional static models of assembly lines ignore key activities of managers and workers, which center on dynamic rather than static issues. In Section 5, we consider other applications of a dynamic approach. We conclude, in Section 6, with a reexamination of the comparison between static and dynamic approaches, and a discussion of the research that will be needed to build a useful and rigorous dynamic view of manufacturing.
A single paper cannot fully establish the validity or usefulness of a dynamic approach. We emphasize operations research, the manufacturing domain, and the academic study of operations management. Management practice in some industries is well ahead of research in incorporating dynamic issues into day to day manufacturing activities and is the key motivation for seeking new academic paradigms. We discuss some practical techniques in the penultimate section of the paper, but defer a general dynamic analysis of present day operations management.
2. OPERATIONS MANAGEMENT IN A STATIC WORLD
It is useful to review standard production and operations management before discussing alternatives. When he coined the term "scientific management" in the 1900's, Frederick W. Taylor[1] started the field of industrial engineering (Taylor, 1947). The assumptions, models, and thought patterns of Taylorism, so influential in the tremendously successful development of American mass production, persist to this day. One of those assumptions was that there exists "one best way" to undertake each task.
The principal assumptions of the static paradigm are that: production technology is known, that labor's role is solely to perform procedures, that the environment is known and stationary, that inputs are homogenous, and that there is a single, known goal. Working under these assumptions, the manager or analyst selects the type, amount, or use of assets. Lower level management or workers, such as manufacturing engineers, then implement the decision. Choice and execution are separate stages, and neither feeds back to the other. The job of the manager is once-and-for-all decision making, as opposed to incremental problem solving or ongoing learning.
Taylor was himself an avid experimenter, with a strong belief in the importance of knowledge about technological processes. For example, his model of cutting speeds versus tool wear, developed through exhaustive experimentation, is still used today. However in Taylor's paradigm, knowledge was developed off-line by specialists and then passed in a one way information flow to workers. The role of feedback (actual performance compared with expected performance) was limited to a punitive one, not as a source of knowledge. This view that R&D can and should be completely separate from execution is a crucial difference between dynamic and static approaches.
We will now review each of the major implicit assumptions in the static operations management paradigm. We do not claim that managers or modelers defend the literal truth of these assumptions, but lack of appropriate dynamic frameworks leads them to make decisions and build models as if the assumptions were approximately correct.
Assumption 1: Known Production Technology
This assumption states that manufacturing knowledge is complete. Each realistically possible production technique is known, well defined, and fixed over time. This includes hardware (equipment), operating procedures, and other aspects of the technology. In situations involving a new plant or new equipment, the relevant knowledge is assumed to be available from vendors for a price. Knowledge about the technology is thus not a key issue, and learning is not a goal. Design of production methods is hence a "choice of technique" problem: which of the known available machines and procedures should be used? This technique becomes the "optimal" way to produce.
The only role of learning in this view of the world is as training--the transfer of procedure from manuals or instructors to employees who are taking on an unfamiliar task (e.g., new assembly line workers who need to be taught procedures). There is no need for organizational learning or research.
Assumption 2: Labor's Role Is Solely to Perform Procedures
This assumption holds that the task of each worker is to carry out assigned, fully specified procedures in response to unambiguous signals from machines, other workers, and the environment. By "procedure" we mean a well defined set of actions, analogous to a computer program. Procedures, like computer programs, may contain some conditional instructions, but all contingencies are assumed to have been anticipated and appropriate responses specified. Management's role is to specify these procedures (the "one best way") and monitor their execution. Since tasks are well defined, performance can be monitored simply by "looking over the worker's shoulder," i.e., by observing inputs and whether operator behavior follows the correct procedures. Learning and modification occur external to the production unit, with new procedures communicated back to it.
This assumption is powerful and useful when the necessary tasks have been reduced to appropriate procedures. However, reduction to procedure is not desirable for some environments and tasks characterized by large and important contingencies, high complexity, or ambiguity about key variables and their relationships. This includes many aspects of product design and problem solving.
Assumption 3: Known and Stationary Environment
This assumption holds that the environment, like the technology, is known and stationary. Usually it is static (i.e., deterministic and constant). If not static, it is stationary (i.e. drawn from a stationary probability distribution with known parameters). For example, when capital investment choices are being made, demand and factor costs are assumed to be known for the life of the equipment. So is the nature and performance of the production technology. At the extreme of this assumption, product markets, input markets, workers, and machines are all deterministic and unchanging. Choice is easy under such a strong assumption: just optimize for the current environment. If it is not deterministic, the environment is assumed to be stationary.
Assumption 4: Homogeneous Inputs
Factors of production such as labor, raw materials, machinery, and energy, are assumed to be homogeneous, with exogenously determined standardized characteristics, and available in complete markets. The markets are usually assumed to be efficient.
Assumption 5: Known Goal
This assumption holds that the purpose or goal (objective function) of the organization is known and well defined, and is uniform throughout the organization. Typically this goal is profit maximization or a temporary sub-goal such as maximizing output.
These five assumptions lead to a consistent view of a world that may be complex, but is fully specified. The task of operations management research is static optimization to select the best way to produce despite complexity. The task of managers is to select a rigid procedure for workers, then monitor and ensure their compliance.
This static paradigm has no place for explicit consideration of dynamic issues such as knowledge, learning, or problem solving. It deals with contingencies only by invoking predetermined conditional procedures and by using stationary stochastic models which treat the contingencies as exogenous.
Previous Literature
We are not the first to argue that standard academic approaches to manufacturing management are deficient for the modern environment. The criticisms in Ackoff's article, "The Future of Operations Research is Past" [2] (Ackoff, 1979), evoked a variety of responses [3](Dando and Bennett, 1981). One of Ackoff's criticisms called for a dynamic approach (p. 98).
- The structure and the parameters of problematic situations continuously change, particularly in turbulent environments. Because optimal solutions are very seldom made adaptive to such changes, their optimality is generally of short duration.... With the accelerating rate of technological and social change dramatized by Alvin Toffler and others, the expected life of optimal solutions and the problems to which they apply can be expected to become increasingly negative.
- For these reasons there is a greater need for decision-making systems that can learn and adapt quickly and effectively in rapidly changing situations than there is for systems that produce optimal solutions that deteriorate with change.
In their critical examination of "the standard economic theory of production functions", Murnane and Nelson (1984) list assumptions implicit in the economic theory of production functions. Their list is very similar to our statement of the assumptions of the static paradigm of operations management. Murnane and Nelson hold that these assumptions, and therefore the standard theory of production, do not work for education and other sectors "where techniques are poorly articulated and idiosyncratic." They theorize that other factors, including experimentation and creative problem solving, are important for explaining the success of different teachers. Nelson (1980) and Nelson and Winter (1982) also discuss difficulties with the standard economic theory of production, and propose an evolutionary model.
3. A DYNAMIC APPROACH
One approach to dealing with dynamic elements of the manufacturing environment is to construct models that relax the five assumptions of the static paradigm one at a time. For example, decision tree models can be used to handle uncertain and non-stationary environments. Another approach is to reject formal modeling and quantification in favor of intuition. A third approach, which we propose here, is to develop new concepts for dealing with critical issues that are not treated by static approaches.
A. Elements of a Dynamic Approach
For several years, we have been pursuing an alternative view of operations that takes account of four elements critical for effectively analyzing dynamic manufacturing situations (Jaikumar and Bohn, 1984). This approach aims at managing and improving operations in a world with a high rate of change, where full proceduralized methods are widely practiced but are incomplete and subject to continuous improvement.
Knowledge. Knowledge is an explicit input to the production process. Knowledge of the best ways to produce is always incomplete. Firms with more knowledge about products, processes, and environment have, in effect, better production technology. Relevant knowledge includes how production should be done, the sources and types of common contingencies, how effective problem solving and learning can be done, and what further learning is needed. Knowledge, unlike most inputs to production, is not consumed by use. Neither is it automatically generated by experience. Managing, exploiting, and augmenting a firm's stock of useful knowledge are key operating tasks.
Learning. Because knowledge is a key element of processes, a foundation of competitive advantage, and is always incomplete, organizational learning is a key task. Learning must be interwoven with production, not be conducted entirely in laboratories and other non-production facilities. Manufacturing facilities should be designed and operated to enhance the rate of learning. A variety of methods for effective learning exist.
Contingencies. Contingencies arise due to gaps in knowledge about the internal and external worlds. We have a contingency when a realized event does not match an anticipated event. Both favorable and unfavorable contingencies occur. Contingencies should be considered explicitly during the design of processes and operating methods, as well as during start-up and ongoing operations.
Problem solving. Contingencies define problems. A fundamental task in production operations is to identify and solve the problems that lead to contingencies. Problem solving that is neglected or performed ad hoc will reduce performance in the short run and retard learning over the long run.
Fig. 1. Static versus dynamic approach to contingencies
Figure 1 contrasts static and dynamic approaches to contingencies. Gaps in knowledge give rise to contingencies, which give rise to the opportunity for problem solving. In a static paradigm, contingencies are "patched" or worked around. In a dynamic approach, contingencies lead to deeper problem solving ("root cause analysis") which investigates why a contingency occurred, how to better predict and detect it, and how to prevent it (if undesirable) or make it more frequent (if desirable). Such problem solving is a form of learning. When properly fed back to change actions, the new knowledge alters future events and capabilities in ways favorable to a firm.
B. Stages of Knowledge
New knowledge drives both progress in old technologies and the development of new technologies (Sahal, 1985). Since knowledge is so important, we have developed methods to analyze and even model the extent of a firm's knowledge. We postulate eight stages of knowledge that describe how much a firm knows about a process (Figure 2). Knowledge progresses over time from Stage 1 towards Stage 8. The firm's ability to avoid or respond to contingencies becomes progressively higher with each stage.
Fig. 2. Stages of knowledge about production
Process knowledge can be decomposed into knowledge about primary and secondary variables. Primary variables have a significant impact relative to an allowed tolerance, secondary variables a moderate or negligible impact. For example, in the assembly of a child's plastic toy by a robot arm, thermal distortion and vibration of the arm are secondary variables. The design and operation of the line can ignore them. But for the assembly of a watch mechanism, the same conditions are probably primary variables. In addition, a secondary variable may change and become a primary variable. An example is a production process that suffers a loss of precision when moved to a different site because of humidity.
Knowledge about process variables is very uneven. If we are capable of recognizing good output, without any sense of how to obtain it, we are at Stage 1. We move to Stage 2 when we begin to recognize variables, and to Stage 3 when we begin to perceive their relevance. We have Stage 4 knowledge when we can measure a variable, and Stage 5 knowledge when we can exert local control over it. When we understand how local changes in a variable affect output, we have arrived at Stage 6, and when we develop a complete understanding of how primary variables affect output within allowed tolerances, but subject to controlling secondary variables, we are at Stage 7. Stage 8 means complete knowledge of a process for all values of all relevant variables. A firm's ability to effectively respond to or avoid contingencies improves progressively with each successive stage.
The stage of knowledge has a critical effect on the degree to which a process should be automated. A purely procedural process executes as an algorithm, with zero judgement or human intervention. At the other extreme, pure expertise, individuals carry out tasks without the use of any repetitive procedures; each execution is completely novel. Figure 3 shows the relationship between degree of procedure and stage of knowledge.
Fig. 3. Degree of procedure versus stage of knowledge
Such "backsliding" can be seen in many areas. Competitive pressure and customer needs force vendors to put most technology on the market tions of a process. Thus, in many industries, new knowledge about a process allows a momentary advance to a higher stage of knowledge, but carries the seeds of change that force a regression to earlier stages. Learning becomes a continuous task of the design and manufacturing operations. Management and control methods must be able to manage effectively at multiple stages of knowledge.
4. AN EXAMPLE: STATIC AND DYNAMIC APPROACHES TO AUTOMATED WATCH ASSEMBLY
This section presents a detailed discussion of an assembly line process. It illustrates the differences between static and dynamic approaches in a practical context and suggests how operations research methods might be developed to handle some of the dynamic variables and new problems raised by a dynamic approach. We have deliberately chosen a process that appears to fit the static assumptions well: high volume assembly of wrist watch mechanical movements by robots on an assembly line. Yet, as we will show, many tasks crucial to competitive survival are highly dynamic and cannot be described solely on the basis of static issues.
A. Description of the Assembly Line
The automated watch assembly line consists of about 50 modular stations performing 150 operations1. The line is adaptable to a wide variety of models (250 models in production in any month), high volume (a pallet of 100 watches every 600 seconds), and high precision (2 micron accuracy for critical operations). Each of the 50 stations is configured with a pick and place robot, an assembly robot with parts feeders, an inspection station, and a rotating carrier (Figure 4). All stations are modularly configured. Parts feeders and arms of the assembly robots are idiosyncratic, designed for the particular task to be performed at a specific station.
Fig. 4. Assembly workstation on the automated watch assembly line.
The sequence of events in each module is as follows. The pick and place robot retrieves a carrier containing a partially assembled device and places it in an empty slot on the index table, which rotates the carrier under the robot. The robot adds new components to the assembly, and the table rotates the carrier to an inspection station and then back within reach of the pick and place robot, which either returns it to the pa deposits it in the reject chute. This sequence is repeated for all 100 devices on the pallet, after which the pallet moves to the next station on the assembly line. As the conveyor line is asynchronous, the next station need not be physically adjacent to the prior station. The conveyor forms a buffer of pallets in process.
Each assembly robot at each station contains the tooling and software required for a particular operation. Most of the inspection stations are physically similar, but use software appropriate to the detection of specific task objectives. Stations are programmed such that all tasks take a fixed amount of time, less than or equal to six seconds. The modular assembly unit is functionally similar to a person on an assembly line, but faster and more repeatable.
B. Static Approach to the Design of Assembly Lines
Assembly line models in the OR literature cover either design or operation. Design is the problem of determining line layout and deciding which tasks to assign to each robot station. Baybars [1985] surveys exact algorithms for simple assembly line balancing. The principle assumptions are: all input parameters are known with certainty; a task cannot be processed in arbitrary sequence due to technological requirements; and all tasks must be performed. Other constraints may be formulated as a mathematical program. The objective function is usually either to minimize the number of stations along the line, keeping the cycle time the same, or, equivalently, to minimize the cycle time for a fixed number of stations. The static assumptions in Section 2 are implicit in the models.
Assembly lines used to produce two or more models of the same product may do so in sequential batches (the multi-model case) or intermixed (themixed-models case). For the latter, see Dar-El (1978), Dar-El and Cother (1975), Macaskill (1972), and Thomopoulous (1967). Constraints might include: zoning restrictions that limit the grouping of tasks at one station (Mitchell, 1957 and Tonge, 1961); tasks that must be done at a particular station; parallel stations (Pinto et al, 1981); other forms of positional restrictions; buffer stocks and other generalities such as feeder lines or parallel subassembly lines (Nanda and Scher, 1976).
The optimization problem for all of these formulations is one of search through a large space to find optimal task assignments and line layouts. The search is carried out only once, during the line design phase. The line layout and balancing problem is usually modeled deterministically, exceptions being Silverman and Carter (1986) and the articles cited therein.
C. Treatment of Contingencies in Static Models of Assembly Line Operation
We now examine, from a static perspective, contingencies in the minute by minute operation of the line. Contingencies are introduced in two related groups of models: those concerned with buffering between stations, and those concerned with inspection and rework.
Typically, each machine is assumed to have a stationary and independent distribution of failure probabilities. A failure is assumed to be visible, and to shut the machine down completely. With no buffers between stations, a failure or slowdown at one station would shut down the entire line. Therefore buffers are inserted, and the optimization goal is to set their size and location. Models for buffer sizing optimize the cost of buffer inventory versus the value of lost production time when other stations are starved or blocked. Even determining the frequency of blocking is difficult (Suri and Diehl, 1986). Failure probabilities and repair times are assumed to be stationary and outside the control of management.2
The other common contingency in static models is rework or scrapping of units. If an operation is performed incorrectly, a subsequent inspection will detect the problem and the unit can be routed to a rework station. Optimal placement of inspection stations and the optimal capacity and buffering for rework are determined by dynamic programming and queuing models of work under repair. Again, analysis is by static economic optimization incorporating variables such as the costs of: rework, scrapping, idle repair capacity, and inadequate capacity. Probabilities of flawed operation of machines or test stations are assumed to be exogenous, stationary, and not subject to controlled learning. Furthermore, the models neglect the roles of rework and testing as likely sources of information, noise, and confusion. (Slade and Mohindra 1985).
Thus, though the static paradigm accepts contingencies as important to line design, the only response it allows is to work around them. All probabilities are constant and exogenous, the goal of static models being to optimize static efficiency despite the presence of contingencies. Contingencies are treated neither as sources of information for learning nor as targets for improvement.
D. Performance of the Watch Line
By the standards of the static paradigm, performance of the assembly line was excellent. Each machine had an average uptime of about 99.8%, with a mean time to repair of 4 minutes and a standard deviation of repair of 2 minutes. A single pallet (100 watches, 10 minutes of work) buffer between each station would be sufficient to prevent starvation of any station. A one pallet buffer for each of the 50 stations would total less than a day of work-in-process inventory, which is minimal by any standards.
Similarly, the cost of scrap plus rework was about l% of total material cost. Problems were concentrated in about 5 of the 50 stations; an approximate model would be 2 watches rejected per thousand per station, or one watch rejected every 50 minutes. A single operator could visit each station once an hour to rework or scrap the defective movements.
Viewed in the static paradigm, this automated line is remarkably free of contingencies, and those that do arise can be accommodated by one or two people. The line is also very flexible (250 models on a line at one time, with more than half of these new each year). Finally, it has low costs. As the line is run 24 hours per day, 5 days a week, the robots, which are mostly standard models with few axes, are used intensively. Capital cost is well below $.25 per watch. In terms of the static paradigm, the line is already very efficient.
E. Dynamic Approach to the Watch Line
Given such performance, why bother with a dynamic approach? Primarily because despite its excellent performance, the line was continually evolving and process improvement was a high priority. For example, actual equipment turned over frequently, with 40% of the machines being replaced by better ones each year. In addition, as the locus of planned learning shifted from one station to another, the tasks assigned to the existing machines were reconfigured weekly. This dynamic evolution had been going on continuously for more than five years, since the first robotic line was built. About 60% of the labor cost of the entire plant (both direct and indirect labor) was associated with process improvement.
Paradoxically, it is precisely the efficiency and programmed nature of automation that causes competitive pressure. Fully proceduralized methods can be easily copied by competitors, and the watch company has several strong competitors. Unless it constantly creates new products or new process innovations embodying new knowledge, the firm will not be able to generate profits beyond economic rent. Given the rapid evolution of the world watch market, with thousands of new watch styles each year, including ten or more new families requiring new tooling and new programming, the firm would quickly lose market share if it failed to innovate. Consequently, the production line makes a number of product families, each in a different stage of the product life cycle. The most crucial period is early in the life of the family.
We observed one new watch family from ramp-up to full-scale assembly. A critical component of the new watch is stamped and formed from sheet metal, then heat treated and fed directly to the assembly station. Of the 19 specified dimensions of the part, three, all involving bending the sheet metal at an angle, were considered critical. Due to springback effects, it was very difficult to meet the required tolerances for these operations.
In early production of the watch, frequent rejects occurred during assembly of this component, indicating one or more out-of-tolerance conditions. As part of their problem solving, operators took a sample of 1,000 of these components and measured each of the 3 critical dimensions on each part. The results were plotted and the mean, standard deviation, probability of rejection, and coefficient of performance (defined as the ratio of the measured standard deviation times 6, to the allowed tolerance band) calculated. This allowed them to compare actual process precision to what is required. For two of the three dimensions it was well above one, suggesting that process precision is adequate. (Both of these dimensions were highly shifted to one edge of the tolerance band, however, indicating that the stamping die needed to be changed.)
For the third dimension, the coefficient of performance was about 1.0, with a process standard deviation of 27 microns and an allowed tolerance of 160 microns. This was unacceptable.
One solution, commonly employed in the static world, would be to inspect and weed out all bad parts. This is often the easiest, fastest to implement, and cheapest solution, since it requires no further experiments or changes. Numerous existing OR models deal with optimal frequency of inspection for weeding out.
A dynamic approach rejects such weeding out for a variety of reasons, the most immediate being that the module where this part was inserted would thereafter be unstable and problematic, requiring a large buffer and frequent human intervention to deal with contingencies. Over the long run, this solution would impede further development of new watches, since all would have to confront the same problem. In contrast, a root cause solution to this problem would add to the base of knowledge available for later products.
One solution considered was to reduce the process standard deviation. Another was to redesign the watch mechanism to make the acceptable tolerances broader. For other types of problems, a solution might be to alter the assembly station to make it more forgiving. That is not an option here, as this dimension is important not just to watch assembly but to watch performance.
Problem solving activity like this also occurred after ramp-up of a model. New equipment was continually being installed, and small improvements in module speed or conformance were frequent targets of experimentation. In addition, it occasionally happened that a previously well behaved module developed problems for some watch models, as indicated by the size and composition of its reject pile. Continual effort was made to detect and focus attention on such problems, and then to solve and learn from them.
One might think that eventually all 50 modules of the watch line would reach Stage 7 of knowledge, where all primary variables are well understood and all secondary variables are adequately controlled. If they did, the process could be optimized and stabilized, and the learning resources (60% of the people plus some surplus equipment) removed. However, the business environment of the firm does not allow this. Competition pushes toward smaller watches with tighter tolerances. With increased precision, less is known about how some of the control variables affect results, and variables that were previously secondary become primary. This is evident when old recipes no longer work. Thus, the process is pushed back to Stage 5 for some modules, and the cycle repeats.
F. Dynamic Approach to the Design of Assembly-Lines
Our dynamic approach takes an opposite tack from the static paradigm, regarding all contingencies as endogenous and key foci during design and subsequent management of the line. Much of the management of the line focused on learning, i.e., on deliberate improvement in the operating behavior of the line, brought about by new knowledge. Primarily because line layout is dynamic, and evolves in response to the location and status of various contingencies, the design of the line does not fit the patterns predicted by static models. At all times, a few of the modules of the line are undergoing intensive work. This may take the form of experiments to speed up the station or improve its conformance, the introduction of a new piece of specialized equipment, or even the replacement of an entire module. In other cases, it fits the previous example, in which the fabrication process for one part is being intensively evaluated, and interactions between fabrication and assembly are under study.
At almost no time is this line well"balanced" in the static sense. A few of the stations are run at the six second cycle time, but many of the other robotic operations can be done in less than six seconds. The conformance quality of the module interacts with its speed because of the micron tolerances used to fit parts together, so that higher speeds can sometimes be achieved only at the risk of increasing the need for retries. To speed up a module, its robot arm must be accelerated, causing vibration, which, if not damped, will be greater than two microns. Various methods for reducing the vibration are known, but must be applied and tested on a case by case basis.
Alternate modules were sometimes introduced into the process flow during these periods of experimentation so that the module under study could be isolated. At other times, the in-process buffer on both sides of the module was increased to allow more latitude for contingencies without affecting the rest of the process. Following a period of intensive investigation, experimentation, and change, buffer size might be kept larger than normal for a few days while the frequency and nature of contingencies was carefully tracked.
G. Relevant Models for a Dynamic Approach
Useful models can be developed to describe dynamic activities. In some cases, existing classes of models are relevant, with changes to capture certain variables. In other cases, entirely new models are needed. We see three categories of models as useful and feasible to develop from existing operations tools/research methods: attention focusing mechanisms, problem solving methods and strategies, and physical/economic models of processes.
Attention-Focusing Mechanisms. JIT is an attention focusing mechanism designed to highlight contingencies as soon as they appear. The extreme form of JIT is kanban control with no buffers. With this system, when a machine breaks down or there are quality problems, the assembly line comes to a halt. For instance, if the percentage defective at each station in the watch line was l% (or when uptime for each machine was only 99%), a kanban production system with no buffers would yield a system uptime of less than 5%. Clearly, pure kanban forms of JIT are applicable only to processes under very tight control.
A related attention focusing mechanism is the E-lot system of manufacture (Jaikumar, 1988a), which locates small lots of buffer inventories at specific points on the line. When a contingency arises, a component from the buffer lot is used to maintain the normal flow and the rejected item is shifted to the E-lot for analysis. The existence and use of the E-lots is the attention focusing mechanism. The size of the lot is set to bring the system to a halt for systematic problems but keep it moving for random errors. The size of the E-lots can be reduced to focus management attention on a specific problem. The most significant difference between E-lots and traditional inventory systems is that E-lot inventories, used specifically to control contingent conditions, are the exception rather than the rule. As problems are permanently resolved and contingencies are reduced, the E-lot system approaches a kanban system.
In the watch line, E-lot buffers are maintained at the five stations where problems occur, and rejected lots are analyzed for defects. When systematic errors occur, the E-lot inventories are used up before they are replenished, bringing the system to a halt. In the systems we studied, E-lots ranged between 1 unit and 100 units.
Such a system raises a host of questions: Where should E-lots be located? What should be the size of the inventories? How should they be replenished? When should the size of the inventories be reduced? These questions can be modeled by applying OR tools to dynamic activities.
Problem-Solving Strategies. The process of assembly on the watch line is divided into 50 modules. Processes are better understood in some modules than in others. A module in which a robot repeatedly executes an algorithmic procedure flawlessly may be at Stage 7, while processes in other operations, such as bonding components with an adhesive, may be at much lower stages. Environmental conditions may change and cause a usually reliable Stage 5 process to fail. The process of fault diagnosis in a rework station might be only at Stage 3, requiring pure expertise.
When a system detects a number of problems, a decision must be made where to allocate limited resources. Fault tree analysis provides one modeling construct for laying out contingencies and their possible causes. This can be complemented by statistical techniques and industrial engineering methods of analysis and by experiments constructed to identify process functions more precisely. In future, the operations research models capable of examining a system as a whole could develop problem-solving strategies in the large. In formulating the assembly-line problem at the ramp-up stage, for example, the objective when assigning tasks to stations would be to gather the most information on effective process parameters for each operation, not to minimize initial cycle time.
To be most effective, experiments should be tied to models of physical phenomena and related to the economics of production. For instance, in punch press models, the physics of deformation and control should be studied. One can construct sequential Bayesian models of the value of information from experiments and the economics of process change.
Which problem-solving strategy to use depends on how much one knows about the problem already. Each stage of knowledge requires a different kind of experiment with its own economics. In the out of tolerance problem described above, for instance, the variance of the dimension falls with heat treatment, while the mean shifts one way or the other. The physics of why and how variance falls is not well understood; this effect is at a low stage of knowledge. The reductions seem to be related to the shape, thickness, and material composition of the component, and to the heat treatment process. Beyond this, not much is known. The process of improvement in control is thus one of rudimentary controlled trials. Expertise and judgment enable us to assess similarities and differences between components, relate control parameters to output, and decide what kinds of experiments to conduct.
By contrast, the stamping process before heat treatment is well understood. The relationship between process parameters and process variance is sufficiently well understood that statistical relationships can be built and, for simple shapes, functional relationships estimated. Thus, the problem solving methods should depend on what we already know about the processes we are attempting to improve (Bohn, 1987).
Physical/Economic Models of Processes. The physical behavior of a process is critical to its operation and economics. For physical operations at Stage 7 of knowledge, a model can be written to predict the occurrence of endogenous contingencies. For example, engineers in the watch line used computer-aided engineering (CAE) to simulate vibration at the end of a robot arm at different speeds and accelerations, using different tool designs and damping methods. This allowed them to design the robot procedures. Had knowledge of the robot only been at Stage 6, pure simulation would not have been adequate, but a combination of experiments, algebraic models, and simulations could still have found and described key relationships quite effectively, providing guidance for the design and software of the assembly robot.
The goal of such work is to develop a science of manufacturing methods. We believe that the necessary tools are now available to do this. It is no longer necessary to use pure expertise to design manufacturing methods, as if we had only a Stage 4 knowledge. So far, such modeling has been done most extensively by domain specific engineers during product design, such as stress and vibration calculations for airplanes and hard disk drives. CAE tools are only beginning to be used for manufacturing engineering, but it is already possible to incorporate operations research methods directly into the CAE tools, for example to conduct searches for lower cost or higher performance configurations. A few fields, such as chemical engineering, have already begun to use OR methods.
5. APPLICATIONS OF A DYNAMIC APPROACH
We now illustrate how a dynamic approach can illuminate diverse manufacturing phenomena.
Evolution of Process Control
A recent historical analysis of process control in machine tool-based industries found six epochs, each characterized by an intellectual shift and the development of an entire new system of manufacture, spanning machines, the nature of work, and the organization (Jaikumar, 1988b).
The first three epochs involved increased mechanization: substitution of capital for human labor; progress through economies of scale; and increasing mechanical constraint to increase precision and control despite higher energy intensity. The last three epochs reversed these trends, fostering increased versatility, substitution of intelligence (both human and machine) for capital, and economies of scope. Today, machines are increasingly used as extensions of the human mind, and both human and machine discretion and versatility are growing.
There is a correspondence between the six epochs of process control and the eight stages of knowledge. In the first three epochs, as in the first five stages of knowledge, the emphasis is on identifying, differentiating, measuring, and gaining localized control of a process. In the last three epochs of process control, and in the latter stages of knowledge, system developers study and gain control of process contingencies until they are able to extract general principles and technologies that can be applied in a variety of domains (Jaikumar, 1988b, p. 90). Problem solving and development of new knowledge assets such as software and parts descriptions become dominant activities. Thus key activities shift from static to dynamic tasks.
Development of Flexible Manufacturing Systems
An analysis of flexible manufacturing systems (FMSs) in Japanese and U.S. machine tool companies found them to be markedly different in ability to run untended, in versatility (number of different parts made), and in effective metal cutting time (Jaikumar, 1986). For example, none of the U.S. systems was able to run untended, while 18 of the 60 Japanese systems had been developed to control process contingencies to the extent that they could run untended. This required Stage 7 knowledge.
The differences are traceable in part to system development practices. In the United States, system development projects were treated as a one-time activity, at the end of which the development team was disbanded. These workers and engineers who run the system were told not to experiment and lacked the expertise to fix even known bugs. This is consistent with the static paradigm's separation of knowledge from work, but is highly inappropriate in a CIM environment. In contrast, the successful Japanese efforts kept the original teams small and required them to remain with the system until it attained 90% uptime. In fact, Japanese developers took on day-to-day operational management, and continued to innovate and develop new applications for the systems. Learning was integrated with routine activities, most of which could be reliably performed by the machines themselves.
Comparison with the Experience Curve Model of Learning
The standard economic model of learning is the experience curve, also known as the manufacturing progress function or learning curve (Dutton, Thomas, and Butler, 1984; Alchian, 1963). Experience curve models state that improvement (usually measured as cost reduction) is an inevitable by-product of normal production, and that manufacturing performance improves with the log of cumulative volume produced. Various studies have used empirical data to estimate "the" slope of the experience curve in different industries.
Although the experience curve concept is useful for diverse purposes, including setting performance targets and prices, much of the existing literature on experience curves is seriously incomplete and misleading. It implicitly treats the rate of learning as beyond the control of management. This follows from the assumption that the slope of the experience curve is fixed, approximately constant over time, and constant across all firms in an industry. Although Alchian's early article provided extensive evidence that all of these assertions are wrong, subsequent research has not followed up on this aspect of his findings.3
Models constructed within the dynamic paradigm suggest that there are many ways for management to directly and indirectly alter the rate of learning. For example, (Bohn, 1991) documents very different process variability levels in different plants and shows how this will retard learning. In this model, the key independent variable is the cumulative number of experiments performed, rather than cumulative production volume.
Analysis of the learning process in electronics ramp-ups suggests that four forces determine the effectiveness of experimental learning in producing useful knowledge (Bohn, 1987):
oinformation turnaround time (the time needed to design, execute and analyze a single experiment);
osignal to noise ratio of the experiment (which is affected by processvariability, how the experiment is conducted, and sample size, among other things);
ocost per experiment, both direct and indirect; and
ofidelity of the experimental environment to the true process.
All four are affected by management of learning per se, and also by management of the routine manufacturing process. Thus two plants using the same process can exhibit different abilities to learn.
6. CONCLUSIONS
Taylor believed that learning through experimentation was crucial. But in his world, it was to be done off-line, by specialized personnel, usually in a lab or pilot line. Information fed back from normal manufacturing was used to reward high output but not as a source of new knowledge. Thus the activities of execution and knowledge creation were both present, but highly separated. This separation is found today in static approaches which assume progress comes from outside the manufacturing plant, from vendors, research labs, and development efforts. Within manufacturing, knowledge is assumed complete. Long-term competitive success will go to the firms that improve the fastest over a sustained period. In principle they can do this by (1) purchasing outside knowledge, (2) intensive R&D outside manufacturing, and (3) learning within existing manufacturing. We will look at each in turn.
With world equipment markets becoming global, any knowledge embodied in purchased equipment or software is widely available to competitors. As well understood procedural tasks are increasingly turned over to machines and software available from vendors, competitive advantage lies more and more in expertise based tasks such as design and process improvement beyond the original capability of purchased equipment. This is strengthened by the trend toward flexibility in automation. In short, economic rents are today available mainly from knowledge which extends what is directly available from vendors.
In principle, development of knowledge within the firm might be done within separate research and development organizations. When is it effective to completely separate learning from manufacturing, as static approaches do? The difficulty with such separation is that non-manufacturing environments have inherently low fidelity for some key issues. Fidelity is the similarity between the location where learning occurs and the manufacturing floor where it is used. For variables at stages six or seven of knowledge, an artificial environment (pilot line) can be created with adequate fidelity. Experiments run in such an environment can be extrapolated to predict accurately what will happen in actual manufacturing. But pilot lines cannot have complete fidelity for important issues at early stages of knowledge. For example, when tolerances tighten, subtle previously unimportant disturbances on the manufacturing floor may become significant and have to be analyzed in order to devise countermeasures. Furthermore, interactions among people, machines, and materials in high volume manufacturing cannot be duplicated realistically in pilot lines.
Therefore, fully effective learning for process improvement must use information from the manufacturing process itself. Speed is also a factor, since it is time consuming and expensive to create and run pilot lines. Thus, for learning about certain issues, it is both more effective and more economical to use the manufacturing process as the laboratory for experiment and observation. These arguments apply also to research environments, which have even lower fidelity.
In short, firms that maintain a rigid separation between learning in R&D and execution in manufacturing, with a one-way flow of knowledge and information between them, reduce their ability to learn and their rate of improvement. Competition and other external factors force continual change in the technology of manufacture and the products offered by firms. Schumpeter (1939) argued this more than 40 years ago, and recent competition from the Far East has painfully exacerbated the pressure for such change. We argue that the production situation is dynamic; therefore knowledge is key to competitive success and must be explicitly considered in operations management. Managers in some industries have known and responded to this for decades. However the formal analytical tools and concepts available to them have their roots in Taylor's separation of knowledge from execution, and are thereby limited in dealing with dynamic issues.
A protected monopolist might be able to ignore dynamic issues and maintain a strictly static approach. Historically, though, monopolies that have assumed they were immune to competition have eventually been supplanted.
Usefulness of Operations Research in the Dynamic World
We have argued that static models of operations are useful only for those problems for which relevant knowledge is at or close to Stage 7, and that this is only a subset of manufacturing in general. Nevertheless, the various standard tools of operations research are still usable, even when static paradigms and models may not apply. We can divide problems into three categories, depending on what types of tools and variables are suitable.
1. Problems that can be handled using standard tools applied to stationary variables; that is, problems that fit the static paradigm. Efficient execution in stationary situations is still necessary, even if it is not sufficient for success in most industries. As technology, competition, and operating practices progress dynamically, new problems amenable to stationary analysis arise. The advent of Flexible Manufacturing Systems, for example, hwinteresting new scheduling problems. Yet, it can be dangerous to ignore the dynamic aspects of such problems. Schedules should be designed to enhance the rate of learning, not just to reduce immediate costs.
2. Problems that can be handled using standard tools, but applied to dynamic variables and issues. The dynamic paradigm covers many problems that, although they remain to be researched, seem amenable to existing operations research methods. Search theory and dynamic programming, for example, have wide application in conducting sequences of experiments over long periods of time. Examples of this class of problem were discussed in the preceding section.
3. Problems attended by dynamic issues that are not fully amenable to standard tools. We see many dynamic problems that are empirically important but cannot be fully modeled. For example, just-in-time inventory control seems to have an impact on the pressure experienced by workers and managers to do root cause problem solving. Both the success and the value of JIT are crucially tied to whether problems are solved superficially, once-and-for-all (removal of the root cause), or not at all. Yet this issue may be mediated by psychological factors we do not yet know how to model. Suri and deTreville (1986) model some relevant effects of JIT, but they are unable to model the underlying driving forces. Another example is the management of product design, which occupies an increasingly important role in modern manufacturing, but which has so far proved quite hard to model.
It is our hope that this paper will stimulate new research, both into the development of new tools and models , and into the application of known tools to dynamic problems. We believe that framing situations explicitly in a dynamic paradigm will yield powerful practical and theoretical insights. Although dozens of practical books contribute to effective management of dynamic situations (Schonberger, 1982; Slade and Mohindra, 1986), with some exceptions operations management research has not yet contributed enough.
ACKNOWLEDGEMENTS
Our thanks to many colleagues for vital discussions and comments, and especially to John Bishop, Kim Clark, Richard Rosenbloom, Don Rosenfield, Earl Sasser, Gordon Shirley, and Michael Watkins. John Simon and Liz Bohn provided invaluable editorial assistance. We also benefited from the penetrating comments of our anonymous referees. We alone are responsible for remaining errors, omissions, and bad writing.
References
- R. L. Ackoff, 1979. "The Future of Operational Research is Past," Journal of the Operational Research Society, vol. 30, pp. 93-104.
- A. Alchian, 1963. "Reliability of Progress Curves in Airfare Production," Econometrica, vol. 31, pp. 679-693.
- I. Baybars, 1985. "On Currently Practiced Formulations of the Assembly Line Balance Problem," Journal of Operations Management, vol. 5, pp. 449-453.
- W. J. Bell et al., 1983. "Improving the Distribution of Industrial Gases with an On-Line Computerized Routing and Scheduling Optimizer," Interfaces, vol. 13, p. 4.
- R. E. Bohn, 1986. "An Informal Note on Knowledge and How to Manage It," Working paper 9-686-132. Boston: Harvard Business School.
- R. E. Bohn, 1987. "Learning by Experimentation in Manufacturing," Working paper 88-001. Boston: Harvard Business School.
- R. E. Bohn, 1991. "Noise and Learning in Semiconductor Manufacture."
- R. E. Bohn and R. Jaikumar, 1986. "The Development of Intelligent Systems for Industrial Use: An Empirical Investigation," in Research on Technological Innovation, Management and Policy, vol. 3, R. Rosenbloom, Ed. Greenwich, Conn.: JAI Press, pp. 213-262.
- M. R. Dando and P. G. Bennett, 1981. "A Kuhnian Crisis in Management Science," J. of the Operational Res. Society, vol. 32, pp. 91-103.
- E. M. Dar-El, 1978. "Mixed-Model Assembly Line Sequencing Problems," Omega, vol. 6, pp. 317-322.
- E. M. Dar-El and R. F. Cother, 1975. "Assembly Line Sequencing for Model Mix," International Journal of Production Research, vol. l3, pp. 463-477.
- J. M. Dutton and A. Thomas, 1984. "Treating Progress Functions as a Managerial Opportunity," Academy of Management Review, vol. 9, no. 2, pp. 235-247.
- J. M. Dutton, A. Thomas and J. E. Butler, 1984. "The History of Progress Functions as Managerial Technology," Business History Review, vol. 58, pp. 204-233.
- R. Jaikumar, 1988. "From Filing and Fitting to Flexible Manufacturing: A Study in the Evolution of Process Control," Working paper 88-045. Boston: Harvard Business School.
- R. Jaikumar, 1986. "Post Industrial Manufacturing," Harvard Business Review, vol. 64, no. 6, pp. 69-76.
- R. Jaikumar and R. E. Bohn, 1986. "The Development of Intelligent Systems for Industrial Use: A Conceptual Framework," in Research on Technological Innovation, Management and Policy, vol. 3, R. Rosenbloom, Ed. Greenwich, Conn.: Jai Press, pp. 169-211.
- R. Jaikumar and R. Bohn, 1984. "Production Management: A Dynamic Approach," Working paper 9-784-066. Boston: Harvard Business School.
- J. L. C. Macaskill, 1972. "Production Line Balances for Mixed Model Lines," Management Science, vol. 19, pp. 423-434.
- J. Mitchell, 1957. "Computational Procedure for Balancing Zoned Assembly Lines," Research report 6-94801-1-R3. Pittsburgh: Westinghouse Research Laboratories.
- R. Murnane and R. R. Nelson, 1984. "Production and Innovation When Techniques are Tacit: The Case of Education," Journal of Economic Behavior and Organization, vol. 5, pp. 353-373.
- R. Nanda and J. M. Scher, 1976. "Non-parallelability Constraints in Assembly Lines with Overlapping Work Stations," AIIE Transactions, vol. 8, pp. 343-349.
- R. R. Nelson, 1980. "Production Sets, Technological Knowledge, and R&D: Fragile and Overworked Constructs for Analysis of Productivity Growth?," Innovation and Technological Progress, vol. 70, pp. 62-67.
- R. R. Nelson and S. G. Winter, 1982. An Evolutionary Theory of Economic Change. Cambridge, Mass.: Belknap Press.
- P. A. Pinto, D. G. Dannenbring and B. M. Khumawala, 1981. "Branch and Bound and Heuristic Procedures for Assembly Line Balancing with Parallel Stations," International Journal of Production Research, vol. 19, pp. 565-576.
- D. Sahal, 1985. Patterns of Technological Innovation. Reading, Mass.: Addison-Wesley.
- J. A. Schumpeter, 1939. Business Cycles: A Theoretical, Historical, and Statistical Analysis of the Capitalist Process. New York: McGraw-Hill.
- F. N. Silverman and J. C. Carter, 1986. "A Cost-Based Methodology for Stochastic Line Balancing with Intermittent Line Stoppages," Management Science, vol. 32, pp. 455-463.
- B. N. Slade and R. Mohindra, 1985. Winning the Productivity Race. Lexington, Mass.: D. C. Heath & Co.
- R. Suri and S. deTreville, 1986. "Getting from 'Just in Case' to 'Just in Time': Insights from a Simple Model," Journal of Operations Management.
- R. Suri and G. W. Diehl, 1986. "A Variable Buffer-Size Model and its Use in Analyzing Closed Queuing Networks with Blocking," Management Science, vol. 32.
- F. W. Taylor, 1947. Scientific Management (Comprising Shop Management. The Principles of Scientific Management, and Testimony Before the Special House Committee). New York: Harper & Brothers, condensed.
- N. T. Thomopoulous, 1967. "Line Balancing--Sequencing for Mixed Model Assembly," Management Science, vol. 14, pp. B59-B75.
- F. M. Tonge, 1961. A Heuristic Program for Assembly Line Balancing. Englewood Cliffs, N.J.: Prentice-Hall.
- D. A. Wren, 1979. The Evolution of Management Thought, 2nd ed. New York: John Wiley, Ch. 21.
Notes
Research funding was provided over many years by the Division of Research, Harvard Business School.
1.This description is a simplification derived from observations of several lines of large Japanese watch companies made in 1986 and 1987.
2.Some models of repair, especially simulation models, explicitly model repair people as a finite resource. Hence, time to repair depends on how many people have been assigned, a management decision. However, the parameters of these models are still assumed to be exogenous.
3.Dutton and Thomas [1984], whose survey of more than 200 learning curve studies shows that the rate of learning is far from fixed, are notable exceptions.
