Learning Prediction And Abstraction In Partially Observable Models