Tikhonov Regularization
When activated, Tikhonov regularization can implement three different aspects of regularization in the PEST setup.
Automatic regularization mechanisms
FePEST provides three automatic options as regularization definitions for the Tikhonov regularization. Advanced users can provide other regularization definitions modifying either manually or automatically (utility GENREG) the PEST Control File.
Assuming that the initial parameter value represent the expected values of the expert modeller, equations of prior information are created that will penalize departures of parameter values from initial values.
As a result, calibration adjusted parameters will be close to the values preferred from a modeller’s point of view.
This level of regularization suggest a preferred homogeneity in the parameter distributions between adjacent pilot points. PEST creates prior information equations, which basically contains each two adjacent pilot points, and indicate a preferred difference of zero.
Typically, such a regularization approach is intended for zones of the model that belong to the same geological unit.
Pilot-point type parameter values for each definition are regularized taking into account the expected correlation (i.e., using a covariance matrix). Differences in parameter values of pilot points closer than their correlation length will be penalized. As a result, smoother parameter definitions will be preferred over heterogeneous ones.
The range parameter in the Kriging Settings indicated for each pilot-point parameter definition is used to compute the covariance matrix. The larger the range value is, the larger the number of pilot points influencing the regularization weights.
FePEST provides the ability to use the PEST Utility GENREG for defining advantage settings of regularization. GENREG can automatically fill the Prior Information section in the PEST Control file (*.pst) based on user-defined rules. Typical examples are preferred homogeneity, preferred ratio (e.g. for anisotropy settings of conductivity), etc.
Objective Function Limits
The Tikhonov regularization aims at finding a minimum of the regularization objective function while observing user-defined limits for the measurement objective function (for which the model is still considered calibrated). The following settings determine this limit:
This is the limit for the measurement objective function below which the model is considered calibrated.
The value can be calculated by summing up the (weighted) measurement noise associated with the observations. If this is not possible, a value somewhat higher than the objective function that results from a calibration run without Tikhonov regularization can be chosen.
This additional threshold is usually set slightly (5% to 10%) above PHIMLIM. PHIMLIM and PHIMACCEPT define a buffer zone for the measurement objective function value for stability reasons. In this zone, an objective function value is tolerated even though it does not meet the target value. This is necessary as the parameter upgrade vector will often "miss" the exact limit because it relies on a linearity assumption.
Figure below illustrates how the iteration adapts to these limits.
Development of the regularization and measurement objective function during Tikhonov-regularized parameter estimation.
Successive reduction of the objective function Limit (FRACPHIM)
This option allows a different strategy to determine the target objective function limits. If set to a nonzero value (allowed values are between 0 and 1), PHIMLIM is determined by multiplying the last achieved value of the measurement objective function with FRACPHIM.
In this way, PHIMLIM decreases from iteration to iteration. It will however never be smaller than the value defined in PHIMLIM.
Optimal values for FRACPHIM are normally in the range 0.1 to 0.3.
FePEST activates this option by default in combination with a very low objective-function value (these are the same defaults as applied by the PEST tool ADDREG1). This improves the well-posedness of the optimization and thereby leads to a more stable behaviour of the GLMA optimization. The limits should however be adapted by the user as these settings might not yield the desired plausible parameter values.
Further reading: PEST Manual: Chapter 7.3.3. |
Weight Factors and Adjustment
Settings to determine the weight factor can be used for fine-tuning or trouble- shooting. Their default values follow general recommendations that are suitable for most applications of PEST.