Outperforming Good-Turing: Preliminary Report