Generating complex movements in redundant robots like humanoids is usually done by means of multi-task controllers based on quadratic programming, where a multitude of tasks is organized according to strict or soft priorities. Time-consuming tuning and expertise are required to choose suitable task priorities, and to optimize their gains. Here, we automatically learn the controller configuration (soft and strict task priorities and Convergence Gains), looking for solutions that track a variety of desired task trajectories efficiently while preserving the robot's balance. We use multi-objective optimization to compare and choose among Pareto-optimal solutions that represent a trade-off of performance and robustness and can be transferred onto the real robot. We experimentally validate our method by learning a control configuration for the iCub humanoid, to perform different whole-body tasks, such as picking up objects, reaching and opening doors.