The Latin Hypercube Sampling option in the Run Preferences
dialog changes the way Crystal Ball generates assumption values during
a simulation. This method works by segmenting the assumption's probability
distribution into a number of non-overlapping intervals, each having
equal probability. Then, from each interval, Crystal Ball selects a
value at random according to the probability distribution within the
interval. The collection of values forms the Latin Hypercube sample.
The Sample Size option (also in the Run Preferences dialog)
controls the number of intervals. The figure above represents how a
distribution might be broken into 10 parts, if the Sample Size was set
to 10.
Latin Hypercube sampling is generally more precise for
producing random samples than conventional Monte Carlo sampling, because
the full range of the distribution is sampled more evenly and and consistently.
Thus, with Latin Hypercube sampling, a smaller number of trials achieves
the same accuracy as a larger number of Monte Carlo trials. The added
expense of this method is the extra memory required to hold the full
sample for each assumption while the simulation runs.
When Crystal Ball has used all the values from each sample,
Crystal Ball generates a new batch of values. For this reason, the simulation
appears to stop while Crystal Ball calculates the new values.
As of version 4.0c, Crystal Ball supports Latin Hypercube
sampling for all its distributions.
