Boosting is powerful, but touchy–it is easy to overfit.
Avoid by regularizing: penalize complex models to reduce the size of the component trees.