Interpreting Results of Robust Tuning
When you tune a control system with
systune or Control System
Tuner, the software reports on the tuning progress and results as described in Interpret Numeric Tuning Results. When you tune a control system with parameter
uncertainty, the results contain additional information about the progress of the tuning
algorithm toward tuning for the worst-case parameter values.
Robust Tuning Algorithm
The software begins the robust tuning process by tuning for the nominal plant model. Then, the software performs the following steps iteratively:
Identify a parameter combination within the uncertainty ranges that violates the design requirements (analysis step).
Adds a model evaluated at these parameter values to the set of models over which the software is tuning.
Repeats tuning for the expanded model set (tuning step).
This process terminates when the analysis step is unable to find a parameter
combination that yields a significantly worse performance index than the value
obtained in the last iteration of the tuning step. The performance index is a
weighted combination of the soft constraint value
fSoft and the
hard constraint value
gHard. (See Interpret Numeric Tuning Results for more information.)
The result is that on each iteration of this process, the algorithm returns a
range of values for each of
The minimum is the best achieved value for that iteration, tuning the controller
parameters over all the models in the expanded model set. The maximum is the worst
value the software can find in the uncertainty range, using that design (set of
tuned controller-parameter values). This range is reflected in the default display
at the command line or in the Tuning Report in Control System Tuner. For example,
the following is a typical report for robust tuning of an uncertain system using
only soft constraints.
Soft: [0.906,18.3], Hard: [-Inf,-Inf], Iterations = 106 Soft: [1.02,3.77], Hard: [-Inf,-Inf], Iterations = 55 Soft: [1.25,1.85], Hard: [-Inf,-Inf], Iterations = 67 Soft: [1.26,1.26], Hard: [-Inf,-Inf], Iterations = 24 Final: Soft = 1.26, Hard = -Inf, Iterations = 252
Each of the first four lines corresponds to one iteration in the robust tuning
process. In the first iteration, the soft goals are satisfied for the nominal system
fSoft < 1). That design is not robust against the entire
uncertainty range, as shown by the worst-case
fSoft = 18.3.
Adding that worst-case model to the expanded model set, the algorithm finds a new
fSoft = 1.02. Testing that design over the
uncertainty range yields a worst case of
fSoft = 3.77. With each
iteration, the gap between the performance of the model set used for tuning and the
worst-case performance narrows. In the final iteration, the worst-case performance
matches the multi-model performance. The multi-model values typically increase as
the algorithm tunes the controller against a larger set of models, so that the
gHard values are typically
larger than the nominal values.
systune returns the final
values as output arguments.
Robust Tuning With Random Starts
When you use
systuneOptions to set
0, the tuning software performs nominal tuning from each of the random
starting points. It then performs the robust tuning process on each nominal design,
starting with the best design. The “robustification” of any particular
design is aborted when the minimum value of
fSoft (the lower
bound on robust performance) becomes much higher than the best robust performance
achieved so far.
The default display includes the
gHard values for all the nominal designs and the results of
each robust-tuning iteration. The software selects the best result of robust tuning
from among the randomly started designs.
The robust-tuning algorithm finds locally optimal designs that meet your design requirements. However, identifying the worst-case parameter combinations for a given design is a difficult process. Although it rarely happens in practice, it is possible for the algorithm to miss a worst-case parameter combination. Therefore, independent confirmation of robustness, such as using μ-analysis, is recommended.