Systems and methods for tuning multi-input complex dynamic systems in order to automatically obtain optimal performance are provided. Training is performed by measuring performance of the complex system using an objective function for a sparse sampling of input values over a variety of dynamic regimes. A sparse representation of the performance for each dynamic regime is stored in a library. At run-time, performance is measured and matched to a sparse representation in the library, and the complex system is configured with the optimal input values associated with the matching sparse representation from the library. Performance may then be optimized using an extremum-seeking controller. In some embodiments, the disclosed techniques are applied to a self-tuning mode-locked laser. In some embodiments, the disclosed techniques are applied to other complex systems such as phased array antennas and neurostimulation systems.