Combining Open-box Simulation and Importance Sampling for Tuning Large-Scale Recommenders