UvA-DARE Safe Exploration for Optimizing Contextual Bandits