So, we thought we’d try to sort out what’s what here. We won’t get
it all finished this week, but think we can make some headway. Look
for a longer report in a few weeks.
Improving Decisions :
There are hundreds of places where supply chain software goes beyond
automation of processes to helping users make better decisions. The
most basic approach to this is the use of “heuristics,” in which
business rules or algorithms are executed by the system to make or
recommend a decision.
In a distribution center, for example, the decision about where to
put a product in a storage location is made based on some algorithms
that factor in the product type, velocity category, and other
attributes. “Mathematical” optimality is not used, and is not
required or likely even feasible. Optimization almost always takes
at least some minutes to process (and in some cases hours), and
hence isn’t generally usable in an execution environment. There can
be gray areas, however. To use another distribution center example,
“slotting” systems sometimes use heuristics, while a few use true
optimization technology. There is no right and wrong, just make sure
you understand the vendor’s approach. Each has its own trade-offs.
Relatedly, heuristics are often used as a pre-process, even in a
true optimization-based program, to break the problem down a bit to
make it easier for the optimizer to find a solution.
Optimization-base programs, such as those usually found in supply
chain network planning, transportation planning, inventory
optimization, factory scheduling, etc. use well-known mathematical
techniques such as linear programming and its cousin
constraint-based optimization to scientifically determine the “best”
That “best solution” is usually defined as minimizing or maximizing
a single, specific variable, such as cost or profit. Other factors,
such as customer service, can be included, but as a “constraint” to
the optimization run, eliminating certain answers from the potential
solution set (e.g., if the highest profit design results in fill
rates averaging below 95%, don’t include it).
When using optimization processing over a very large data set, such
as a complex global supply chain network or huge transportation
plan, heuristics are often used (as noted above) to reduce the size
of the problem that the optimizer is working against. This enables
it to complete faster, or to ensure it doesn’t produce theoretically
optimal, but practically impossible solutions.
Software vendors will sometimes quibble about this, with one side
claiming the other is using too much heuristics up front and not
truly optimizing the solution. Operations research types recognize
there are always trade-offs, and that there is no universal right or
wrong, just what makes the most sense for the specific
problem/decision that a company is trying to improve.
Supply Chain optimization technologies are in use in thousands of
companies, and we’ve actually noted a bit of an upsurge in interest
over the past year or so (more on that soon). But there are some
things optimization isn’t so good at.
Optimization is generally based on some fixed estimate of demand
over a given time frame. You can alter that demand estimate and run
a different scenario to compare the impact on the recommended
solution, but optimization in general is not good at handling highly
variable demand or system inputs.
Optimization also tends to be a “black box” approach, taking inputs,
crunching the numbers, and presenting a solution. It’s often hard
for the user to really understand the interplay of various factors,
and how the supply chain “system” (whether that’s a network or a
factory) works as a whole.
That is where “simulation” can come into play. In simulation, a
model of the system is built (again, whether it’s a conveyor system
in a DC or a supply chain network). Rules are created (often still
through programming, but increasingly with at least some level of
system configuration) that describe how the system should work.
The key is that demand (or other key inputs) aren’t static, but are
more dynamic. Demand can be estimated (or based on actual history)
at a daily level. For individual plants or DCs, it could be on an
almost minute- by-minute basis. It is also possible to use
techniques such as “Monte Carlo” analysis to have demand or other
variable populated more or less randomly over some period.
Running the simulation then allows the analyst to see the behavior
of the supply chain system over time, as these inputs change. It may
allow bottlenecks to be identified that would be missed in an
optimization program that gives the best total answer but misses
supply chain or operational glitches along the way. But to find the
“optimal” answer, the analyst has to observe what has occurred, make
some hypotheses about the dynamics of the “system,” change a factor
(for example, add some more inventory, or another packing station),
and see what happens.
The benefits: better ability to understand the impact of dynamic
events, better total system understanding, and (increasingly
important today) risk mitigation. But these benefits come at some
cost, as we’ll explore next time.
Optimization and simulation can be used together, as is increasingly
common in supply chain network design. In the next few weeks, we’ll
also get into the pros and cons in more detail, and provide some
additional examples of use cases and vendors.
Hope this clarified things a bit, and I would welcome your thoughts
What are your thoughts on optimization versus simulation? Is this
whole area something only for the Operations Research Experts, or
should supply chain practitioners and executives get knowledgeable? How