since we're now using ensembling to pick which algos to use, we want as many as possible. so consider training the same total number of estimators, but making each individual estimator less tuned, and just doing more rounds of parameter searching.
we'd also have to make a similar change in our logic to stop training more of a certain type of algo if it hasn't been competitive in it's first 3 attempts. maybe bump that up to 5 or something larger.
since we're now using ensembling to pick which algos to use, we want as many as possible. so consider training the same total number of estimators, but making each individual estimator less tuned, and just doing more rounds of parameter searching.
we'd also have to make a similar change in our logic to stop training more of a certain type of algo if it hasn't been competitive in it's first 3 attempts. maybe bump that up to 5 or something larger.