Convergence Tolerance
Optimizations are involved in many different aspects of SpatialAnalyzer’s computations. While these optimizations usually have a single “best” answer in theory, in reality often they can only be found through iterative analysis. As a result, finding the “correct” answer would theoretically take an infinite amount of time.
Consider a best fit optimization: once the computer finds what it perceives as the solution that best fits a series of measurements to a CAD surface, it will stop. However, if you were to increase the internal decimal precision and “jiggle” the points around with even more fine movements and rotations, it could find a better solution than the solution initially proposed.
Continuously increasing the decimal precision and making finer adjustments becomes unwieldy after a short time, and it would most certainly try your patience. While a best fit showing 1x10-12 inches of RMS error is better than a best fit showing 1x10-11 inches of error, we don’t particularly care, because the difference is much less than the accuracy of the measurement in the first place. The additional computation required to reach the better solution is just not worth our time.
As a result, SA follows a set of rules that tell it when an optimization is “good enough”. When SA notices that the difference between two iterations in a solution is smaller than these thresholds, the solution terminates, because the result is considered to be “good enough”. The default values provided for the convergence tolerances have been found to work well for a wide array of measurement devices and geo- metric optimizations.
Convergence Tolerance Settings
-
Length (object fits). When the RMS error of a fit, such as a geometry fit, surface fit, or relationship fit, does not change by more than this value, the optimization process terminates. Using a smaller value here will theoretically give a better solution, but will increase the computation time required for the fit.
-
Angular (bundle adj). When the RMS pointing error in a Bundle Adjustment does not change by more than this value from one iteration to the next, the bundle adjustment process terminates. Using a smaller value here will theoretically give a better solution, but will increase the computation time required for the fit.
Stop optimizations when the RMS of the residuals does not change by this tolerance from one iteration to the next.