Practical Learning of Predictive State Representations