Closed FelicienLL closed 3 years ago
To sum it up, it seems like L-BFGS-B sometimes thinks the optimal parameter values are the initial values. Convergence achieved without changing the parameter values, nor the OFV. No error thrown in the optimx output (convcode = 0).
When we run sensitivity analysis - I mean changing the initial values (instead of 0, the default with L-BFGS-B) - the problem is solved in most of cases. Except for a "space of parameter", for instance negative values of ETA1&3 +positive value for ETA2&4. I am pretty sure that a 4-dimension representation would show that the blue "TRUE" point are all all stuck together. It means that initial values must not be taken from this zone (Bermuda triangle...).
Possible solution : decrease the tolerance of the optimizer ? I am not sure of what I am doing here, or if my vocabulary is correct. I tried to tweak the L-BFGS-B arguments (ptol, lmm etc...) very empirically, but nothing happens (not shown).
Conclusion:
For now, I think I will just add a "while" loop that re-run the optimization if there is no change in OFV. New initial value sampled from a rather "large interval" (i.e. runif(n, -1,1)
) to be sure to get out of the "Bermuda triangle".
Created on 2020-12-22 by the reprex package (v0.3.0)