In this article, Mark L. Spearman explains how to develop a system that works in rich mix/low volume environments and, just like the Toyota Production System, minimises waste and stock while maximising throughput.


Modern efforts dating back to the early 20th century dealt with inventory control, efficiency, and timely delivery. Computers were first used by Joseph Orlicky in the development of material requirements planning (MRP) at IBM in the mid 1960s. MRP lived up to its name and planned material requirements without considering the capacity available. Thus, the basic problem with MRP is twofold:

  • 1. It assumes the time to traverse the production facility depends only on the part being produced and not on the current work in process within or utilisation of the facility. Consequently, planned lead times must be inflated to account for the times when the facility is overloaded resulting in inflexible, long lead times along with high inventories.
  • 2. Lot sizes are determined without considering capacity. Instead, the classic lot sizing problem considered the trade-off between an inventory carrying cost and a setup cost. The setup cost provides a clumsy alternative to a capacity constraint and was typically overstated, resulting in high inventories and a loss of flexibility.

MRP II (manufacturing resources planning) attempted to address these problems by adding additional modules to MRP in the form of rough-cut capacity planning and capacity requirements planning. Sadly, these additional modules were essentially a patch that could identify a capacity deficit after the plan was produced by MRP and offered few suggestions of how to remedy the problem. Even so, MRP and its variants continue to form the core of most modern ERP systems, including the high-end offerings of Oracle and SAP. For this reason, almost all companies that use MRP (in ERP) also have ad hoc spreadsheets to massage the MRP output before using the results on the shop floor.

More recently, ERP companies have offered a “solution” to the deficiencies of MRP in the form of advanced planning and scheduling (APS) or advanced planning and optimisation (APO). APS and APO are impressive in their ability to model complex situations and provide a means to generate a deterministic schedule for hundreds of jobs running on scores of machines. Nonetheless, there are at least three problems with this approach:

  • 1. APO/APS systems do not directly consider a lack of knowledge (e.g., forecast) nor do they consider randomness in production or demand.
  • 2. Such systems require an enormous amount of basic data to specify run rates, setup times, labor availability, shop schedules, etc.
  • 3. Because scheduling problems are typically large and almost always NP hard, it is impossible for the fastest supercomputer imaginable to find an optimal solution within a reasonable time (i.e., less than a year). For this reason, many of the offerings do not attempt to develop a schedule but instead leave it to the planner to use the APO as a “what-if” tool to evaluate different alternatives.

While using APO has helped with some particularly difficult scheduling problems, many companies we have dealt with have found few benefits from a system that requires an enormous effort to keep running.


While America was increasingly experimenting with the use of the computer in the factory, Toyota was developing means to control production that were simple, effective, and did not use computers. It is now known as the Toyota Production System, the focus of which is to operate with the minimum resource required to consistently deliver just what is needed, in just the required amount, at just the right time. A number of elements are key to making TPS work:

1. Takt time production. The takt time is simply the average time between outputs of the line. For an assembly line, it is the cycle time of each station on the line – something that is easy to control in an assembly environment. It is considerably less straight forward to implement such a takt time in an asynchronous line consisting of separate process centers. Much mystery is removed, however, when one realises that for any given production period there are only so many takts. If the takt time is 30 seconds then in 8 hours there should be 960 units produced. A production quota may not sound as fashionable as a takt time, but the two are equivalent. Nonetheless, it is much easier to implement a takt time/production quota on a paced assembly line than it is on a multi-flow job shop. Indeed, this is one of the primary challenges facing the use of the TPS in manufacturing environments facing a high mix/low volume demand. The use of a takt time or a production quota implies the existence of a virtual queue. Such a queue need not contain material but only the information regarding jobs to be released. We are guaranteed such a queue because, under random demand, production never exactly equals demand. Consequently, when demand exceeds production, the virtual queue increases and when it falls below production, the queue falls. Recognising the existence of this queue and then providing direct controls is an important constituent of dynamic resource scheduling.

2. One-piece flow. A goal of the TPS is to be able to produce one part of a kind at a time. This is a worthy goal because, as Factory Physics shows, one piece flow with no variability results in minimum cycle time with maximum throughput. Of course, when there is variability, as there always is in the real world, there must be some form of buffer. Thus, almost all assembly lines run with a takt time that is significantly longer than the average process time of the bottleneck operation. Longer takt times imply lower production rates. So if a process that requires 25 seconds on average has a 30 second takt time, it is effectively running at 83% utilisation. This is equivalent to scheduling five days of work in a six day period. When understood in this way, one-piece flow becomes more understandable and less of a magic trick. When there are significant setups, it gets even trickier. The “solution” is to reduce setup times to 10 minutes or less. When this can be done, onepiece flow can be very good subject to the qualifications given above. However, there are many situations in which this simply is not possible (for example, a furnace). In such a case, one-piece flow would result in a severe increase in waste in the form of under-utilised equipment.

3. Recourse capacity. An important benefit of a production quota is the triggering of make-up time. The classic formulation of the Toyota Production System would schedule 10 hours of work for every 12 hour period. If any problems occurred during the 10 hours, there would be two extra hours to catch up. Of course, such a setup implies a maximum utilisation of only 83%. But the use of this capacity buffer does a great deal to make achieving production goals more feasible without requiring enormous amounts of inventory. The DRS system will employ a similar capacity buffer but one that is dynamic in that it does not require idle time if the make-up time is not needed.


The basic idea is this:

Instead of creating a detailed schedule for a single situation that will never occur, create a set of dynamic policy parameters that works for a range of situations.

The TPS uses one such system. The dynamic policy parameters are the takt time, the kanban levels, and the make-up time. No schedule is created for lines feeding the assembly line. Those lines are controlled by kanban squares. Whenever a square is unfilled, fill it! Likewise, there is not a schedule of the use of make-up times. When make-up time is needed, use it! However, to make it work in a low volume/rich mix environment one needs to understand the essence of pull production.

Pull production. The word “pull” has caused vast confusion, but what it means is the control of work-in-process (WIP) as opposed to trying to control the start of new jobs. Knowing this, it becomes easy to implement a generalised pull system known as CONWIP that will work in a wide variety of manufacturing situations – even those with high mix and low volumes.

CONWIP also provides many of the benefits of takt time production. This is because the entire line will run at the rate of the bottleneck naturally, that is, without scheduling. Consider the example of the dynamically controlled system shown in figure 2, where we can sell all we can make. The second machine from the left in the top fab line is the bottleneck for the entire factory. Thus that fab line is controlled with a bottleneck CONWIP loop that initiates the start of a new job as soon as the WIP falls below the CONWIP level regardless of how much stock is present. The bottom line has abundant capacity and can always keep up with the bottleneck. For this reason it is controlled with a non-bottleneck CONWIP loop which starts the next job only when the WIP plus stock level falls below the CONWIP level.

The result is a factory in which every machine runs at the bottleneck rate. Obviously the machines in the top fab line must run at the bottleneck rate since they are directly linked to it. Assembly must run at the bottleneck rate since it cannot assembly parts faster than Fab Line 1 can create them. Therefore, the bottom fab line also runs at the bottleneck rate because it is receiving a start signal from the assembly operation that is running at the rate of the bottleneck. Thus, every machine is running at the bottleneck rate, with no scheduling.

Regrettably, simple configurations like the Toyota Production System and the one described above have limited applicability and are not generally applicable to a wide variety of manufacturing environments such as job shops and instances with a high mix of parts that may have low volumes. Moreover, some of the dynamic parameters are not optimised for performance. For instance, kanban levels are usually set with a simple heuristic formula and lot sizes are set as close to one as is possible. Neither approach is optimal and can be very important in high mix, low volume environments.

The good news is that we have all the elements to create a generalised TPS that considers risk and randomness – dynamic risk-based scheduling. The elements are:

  • 1. Production planning is performed using traditional MRP to generate planned work orders with some important differences. Although MRP has a poor reputation, it does provide a hierarchical planning and scheduling system. Its deficiencies are addressed by periodically optimising the traditional parameters such as lot size, safety stock and planned lead time in ways that consider capacity, customer service, and inventory.


a. Instead of insisting on a lot size of one, we compute optimal lot sizes that minimise inventory and out-of-pocket setup costs subject to capacity and service constraints.

b. Instead of a fixed assignment of jobs to lines, we dynamically assign jobs to minimise the maximum utilisation of any machine in the system. This results in lower overall cycle times and less WIP.

c. Rather than using heuristic inventory policies such as those used in kanban, we compute optimal levels of safety stock, days-of-supply, and planned lead times that require the minimum amount of inventory for the available capacity and desired customer service level.

d. Instead of establishing a WIP level for every part at every work station as in kanban, WIP levels for an entire flow are optimised resulting in minimum WIP and cycle time while maintaining a given throughput.

e. As opposed to the fixed lead times in traditional MRP systems, we periodically compute planned lead times based on anticipated load levels. Process centres with higher loads have longer lead times while those with low loads have much shorter lead times. The planned lead times include time spent in the virtual queue.

  • 2. Dynamic execution is performed without a published schedule.

a. Planned work orders from the optimised MRP are not started on the “start date” (i.e., due date less planned lead time) unless allowed by the generalised pull system, CONWIP. Instead these are placed in a virtual queue awaiting the CONWIP pull signal. Thus, when demand drops, production goes down and when demand increases, production rises, within limits.

b. Instead of takt time production we apply the CONWIP release strategy. CONWIP naturally provides a smooth flow at the rate of the bottleneck as well as limiting WIP. This means that planned work orders become active work orders (WIP) only when the WIP level in the line falls below the maximum CONWIP level. WIP is calculated in terms of “equivalent WIP units” in order to account for routings with different lengths in a single flow.

c. Recourse capacity is managed using a virtual queue and a capacity trigger. When planned work orders exceed a pre-determined maximum virtual queue level, the make-up time is authorised. This is done in such a way as to prevent late jobs without requiring a large capacity buffer at all times. In other words, instead of keeping an extra two hours for every 12 hour shift, we provide for a weekend shift for every week. Thus, if the weekend shift is not needed, it is not used.

If demand rises to the point where the system can no longer meet it, the virtual queue will increase to the point where the make-up time is needed. If this happens too often, the system will be re-optimised (shift schedules, workforce, lot size, etc.). If demand falls, the virtual queue will be empty indicating the need to reduce capacity. If there is no make-up time available (for example in a 24/7 schedule), then the only remaining buffer is time. Consequently, due dates will be pushed out to match what the system is capable of meeting. The use of dynamically quoted due dates is a key part of the dynamic risk-based scheduling strategy.

The result is a system that: „„

  • Achieves the goals of the Toyota Production System in that it eliminates as much waste as is possible by minimising WIP, maximising throughput, maintaining on-time delivery and minimising inventory. „„
  • Eliminates the need to create detailed schedules of every job on every process during its time through the facility. „„
  • Works in a rich mix, low volume environment.

The system has been successfully implemented in a number of installations including one in textiles that managed more than 120,000 SKUs in which dye lot integrity must be maintained. The result has been lower inventory, better on-time delivery, and a higher utilisation of labor and equipment.

It appears that the dynamic risk-based scheduling system can be applied to a much wider range of environments than the traditional Toyota Production System.