Resimulation does not produce new results

[todd-skinner]

I'm copying an excerpt of Blake Anderson's email to me describing this bug.
See attachment.

During our troubleshooting session, Ryan and I experience several strange things. The cells in question are E10 (producing “0”) and F10 (producing “#DIV/0). I think this is an official bug.

Note E10 and F10 contain op formulas to calculate the mean and standard deviations of a result Argo distribution.

OS: Windows 7
System Info: Excel 2013
Argo Version: 3.0.0

Observation: Given the file as-is, hitting simulate recalculates the workbook and places values in the cells in question. However, re-simulating again does not change those cell values.

Observation: Any change to the options of the simulation (random seed, native excel, number of trials, or latin hypercube sampling) produces values in the cells in question. However, re-simulating again does not change those cell values.

Observation: Using the OP function to extract the individual trial results works. However, re-simulating again does not change the trials.

Conclusion: It would seem to be 1 of 2 things. 1) Could be that the random number generator’s seed is locked accidentally (unlikely) 2) A trigger that calls for a recalculation isn’t firing properly and does not re-calculate the workbook unless absolutely necessary (simulation options are changed or values themselves are changed)

[boz]

Some of the things described above are actually features (which I'll detail below), but first to address the issue: E10 producing 0, and F10 producing #DIV/0.

Issue

Upon opening this workbook in Excel 2010, E10 produced 0 and F10 produced #DIV/0. After running a simulation the values updated to E10 ~ 10.23 and F10 ~ 2.85. Saving and reopening the workbook resulted in E10 ~ 10.23 and F10 ~ 2.85.

Does the series of outcomes described above replicate the behavior? Or is it that E10 = 0 and F1 = #DIV/0 always? If its the latter case then there is a bug unrelated to the observations.

Argo Simulation Process (Features)

The below should be documented in Argo help material after review. Also, these are referred to as features because breaking up the simulation workload in this manner is what enables faster run times and dynamic simulations

Distributions

Argo distributions have 3 distinct components. Sampling, Inverse CDF, Correlation.

• Sampling: All distributions have a random uniform sample (U) between [0, 1] that is generated when the distribution is defined.
• Inverse CDF: The inverse CDF for a distribution is calculated based on U and the distribution's parameters, CDF(U, ...). This happens when the distribution is defined.
• Correlation: Correlation is applied to distributions by sorting their U values to approximate the applied correlation coefficient.

The above are independent of running a simulation. What this means in practice:

• When a simulation is run, U is not re-sampled. Each U is just as random as any other U so it does not need to be updated.
• If a distribution's parameter is changed, U is not re-sampled. CDF(U, ...) recalculates immediately based on the existing U and the new parameter.
• If correlation is applied to a set of distributions, U is not re-sampled. The elements of each U are reordered to approximate the applied correlation.
• Changes to 'global' settings such as number of trials, method of simulation (native vs rta), sample seed, or sampling method (mc vs lhc) will cause U to be re-sampled because these settings impact the construction of samples.

Simulation

The simulation process translates an Excel workbook into a function, SF, that includes distributions as arguments, i.e. SF(CDF(U, ...), ...)

• If a cell containing a distribution parameter is changed then the SF recalculates on worksheet recalculation
• If a constant value that SF depends on is updated then SF recalculates on worksheet recalculation

If a simulation is rerun then a new SF(CDF(U, ...), ...) is constructed. Note that U is not updated. If no other aspect of SF was changed then the exact same SF will be constructed.

If a formula is changed, then clicking simulation will generate a new SF that is constructed differently to reflect the formula change.

Observations

Taking the above into account, the three observations are the expected behavior.