I was thinking, admitting I don't know too much about how Zorro internally handles the training, but was thinking, if an additional train Parameter could make sense (or maybe it is already how training is handled).
Suggestion, find within the standard ascend train option the most stable one, in terms of neighbouring train parameter values.
To illustrate, using optimize(5,3,10,1) e.g. value 3 gives good results and would be picked.
So suggestion, if 6 gives maybe worse results than 3, but 5 & 7 give just sligthly worse than 6 (in all terms that are evaluated to choose the parameter) algorithm should choose 6 instead of three, since the neighbouring values don't deviate too much.
Although, don't know if that isn't already the way or it really makes sense.
If it does, it should of course be explicitly activated.
Could be helping against overfittig, maybe even a warning in the log could be issued if the sourrounding values are delivering too drastic differences from the one picked.
Last edited by danatrader; 06/29/20 19:30.