alternative loss functions

[jbensabat]

Hello I am trying to identify parameters of surface flow using WRF-HYDRO as simulator. I have a case with ~10 hydrometric stations. in 50% of the stations the discharges are substantially higher than in the remaining 50%. When I run GSA or GLM all the results are driven by the stations with the higher discharges unless I set a weight to each residual (possible but complex if data is abundant). then I have two questions from my experience running Google Ceres: 1) Is it possible to use a loss function that is more adaptive (HUBER, CAUCHY); 2) we could also integrate an automatic procedure for defining weights based on the value of the initial residual. is it possible for a novice (like me) to supplement the code with these two features ?
best
jac

[jtwhite79]

If you are wanting a loss function to handle outlier residuals, then It seems like a post processor that runs after the model would be the easiest way to deal with that. As for the automatic weight adjustment, I've tinkered with that myself its pretty gnarly both mechanically and conceptually

[jbensabat]

hi and thanks for your reply
actually i am not talking about outliers. Each discharge time series has no outliers be each time series has a peak discharge that can be substantially different from the others. so pest++ will tend to find the parameters that reproduce the time series with the highest discharges at the expense of the not less important time series with the smaller peak discharge. A huber type of loss function would scale this properly
best
jac

[jtwhite79]

Oh in that case, using the weights seems to be the most general and generic approach...

[jbensabat]

thanks will try now to find a weight definition strategy