Ensemble Decision Tree with Gradient Boosted Tree

That's probably not a good idea. Both RF and GB are just combinations of decision trees. It doesn't make sense to have a single "best" decision tree because that one tree is most likely severely overfit, so while it may look good on the public LB or your CV, it should generalize very poorly. Including such a tree into the ensemble is counterproductive. Furthermore, the creation of a decision tree is inherently an NP-hard problem so no matter what you can't really find a perfect tree even on the training data in a reasonable amount of time. The ensembling of RF and GB is actually not great. I've rarely found that combination to much value.