The Prediction Analysis Controls pane is on the PEST Properties dialog box under Prediction Analysis.
The variables specified on the Prediction Analysis Controls pane are described in the PEST User Manual, Part I, Section 8.3. More extensive descriptions of these variables are in the PEST user manual. The variables specified on this pane appear in the Predictive analysis section of the PEST control file.
NPREDMAXMIN When PEST is used in “predictive analysis” mode, its task is to maximise or minimise the single model prediction while maintaining the objective function at or below ϕmin+ (i.e. δϕ0). NPREDMAXMIN determines whether it maximizes or minimizes the prediction.
PREDNOISE determines whether or not "predictive noise" is taken into account when maximizing or minimizing the prediction
PD0 is a value for the objective function which, under calibration conditions, is considered sufficient to “just calibrate” the model. It is equal to ϕmin+δ, i.e. ϕ0 in the above discussion. A PEST predictive analysis run should be preceded by a parameter estimation run in which ϕmin is determined. The user then decides on a suitable value for δ and hence ϕ0 before supplying the latter as PD0 for a PEST predictive analysis run.
PD1 While PD0 represents the target objective function, matching it exactly can be difficult. PD1 is a value slightly larger than PD0 that PEST can accept.
PD2 When PEST is run in “predictive analysis” mode, as the objective function approaches PD0 the relative change in the objective function, ϕ, between Marquardt lambdas may be small; however the relative reduction in (ϕ - ϕ0) (i.e. the objective function minus PD0) may be sufficient to warrant testing the efficacy of another Marquardt lambda. The objective function value at which PEST stops testing for a relative objective function reduction, and begins testing for a relative reduction in (ϕ - ϕ0) is PD2. Generally this should be set at 1.5 to 2 times PD0. In either case the decision as to whether to try another lambda or move on to the next optimization iteration is made through comparison with PHIREDLAM.
ABSPREDLAM and RELPREDLAM If the objective function is below PD1 and successive Marquardt lambdas have not succeeded in raising (lowering) the model prediction by a relative value of more than RELPREDLAM or by an absolute value of more than ABSPREDLAM, PEST will move on to the next optimisation iteration. Due to the fact that the approach to the critical point is often slow, these values may need to be set low. A value of 0.005 for RELPREDLAM is often suitable; the value for ABSPREDLAM depends on the context. If you would like one of these variables to have no effect on the predictive analysis process (which is mostly the case for ABSPREDLAM), use a value of 0.0.
INITSCHFAC When undertaking the line search, the initial model run is undertaken at that point along the parameter upgrade vector which is a factor of INITSCHFAC along the line of the distance that PEST would have chosen using the theory presented in Doherty (2015) alone. It has been found from experience that if complexities of model behaviour dictate that a line search is necessary, then it is worth doing it properly. So unless there is a good reason to do otherwise, a value of 0.2 to 0.3 is appropriate here
MULSCHFAC In undertaking the line search, PEST moves along the parameter upgrade vector, increasing or decreasing the distance along this vector by a factor of MULSCHFAC as appropriate. A value of 1.3 to 1.7 is suitable for this variable in most cases.
NSEARCH A line search is undertaken for each trial value of the Marquardt lambda. The maximum number of model runs that PEST will devote to this line search for any value of lambda is equal to the user-supplied value of NSEARCH; set NSEARCH to 1 if you wish that no line search be undertaken. Otherwise, a good value is 15.
ABSPREDSWH and RELPREDSWH When used in “predictive analysis” mode, the role of PHIREDSWH is unchanged if the current objective function is above PD1. However if it is below PD1, PEST’s decision to switch from two point derivatives calculation to higher order derivatives calculation is based on improvements to the model prediction. If, between two successive optimisation iterations, the model prediction is raised (lowered) by no more than a relative amount of RELPREDSWH or by an absolute amount of ABSPREPSWH, PEST makes the switch to higher order derivatives calculation. A setting of 0.05 is often appropriate for RELPREDSWH. The setting for ABSPREDSWH is context-dependent. Supply a value of 0.0 for either of these variables if you wish that it has no effect on the optimisation process. (On most occasions ABSPREDSWH should be set to 0.0.)
