StressByIterationSlow.jpg (55509 bytes)

The example shows the results from slow autopilot for a particular data set.
Because it had the "slow" setting of autopilot, it attempts 6D down through 1D solutions.

Several points to notice about these graphs:

  • Curves for higher dimensionalities have lower final stress than curves for lower dimensionality.
    That is because it is easier to fit the data with more dimensions in the solution.

  • Curves for the randomized runs have higher final stress values than the curves for the real runs.
    This is because the real data have a correlation structure among the variables that allows a lower stress solution.

  • Curves for the real runs are more variable for a given dimensionality than the randomized runs.

  • A few real curves have rather high instability, as shown by their jagged shape.