Abstract:
This study presents the results of using common two or three-parameter "default"
distributions in place of "best fit distributions" in simulations of serial production lines
with finite buffers and blocking. The default distributions used instead of the best-fit
distribution are chosen such that they are non-negative, unbounded, and can match either
the first two moments or the first three moments of the collected data. Furthermore, the
selected default distributions must be commonly available (or easily constructed from)
distributions in simulation software packages. The lognormal is used as the two-parameter
distribution to match the first two moments of the data. The two-level hyper-exponential
and three-parameter lognormal are used as three-parameter distributions to
match the first three moments of the data. To test the use of these distributions in
simulations, production lines have been separated into two major classes: automated and
manual. In automated systems the workstations have fixed processing times and random
time between failures, and random repair times. In manual systems, the workstations are
reliable but have random processing times. Results for both classes of lines show that the
differences in throughput from simulations using best-fit distributions and two parameter
lognormal is small in some cases and can be reduced in others by matching the first three
moments of the data. Also, different scenarios are identified which lead to higher
differences in throughput when using a two-parameter default distribution.
