Overestimation learning with guarantees