Hyperbolic Discounting and Learning over Multiple Horizons