Towards Simplicity in Machine Learning for Live Systems with Adaptive Experimentation