Once default optimizer fails, try all optimizers but with increased number of iterations
Opened this issue · 2 comments
danielinteractive commented
This is a first step towards improving convergence behavior on difficult data sets.
See #380 for a motivating example.
To do:
- Define
optimizer_control
lists to use for the second try optimization- e.g.
eval.max = 1000, iter.max = 1000
fornlminb
- e.g.
maxit = 50000
foroptim
- e.g.
- For
refit_multiple_optimizers
default to standardoptimizer_control
if not provided otherwise by the user
clarkliming commented
idea: provide additional default optimizers that has larger maxit or eval.max, if needed they will be executed after the first failure. or, provide batches of optimizer?
with this it can be a bit cleaner, and we don't need to mess up the defaults and user control (e.g. what if a user already provides a optimizer with maxeval set to a larger number? shall we retry with a even larger max eval?)
danielinteractive commented
thanks @clarkliming , yeah additional optimizers with higher thresholds would also be fine