Improving Predictive State Representations via Gradient Descent