Progress has been made recently in developing tech- niques to automatically generate effective heuristics. These techniques typically aim to reduce the size of the search tree, usually by combining more primitive heuristics. However, simply reducing search tree size is not enough to guarantee that problems will be solved more quickly. We describe a new approach to auto- matic heuristic generation that combines more primi- tive heuristics in a way that can produce better heuris- tics than current methods. We report on experiments us- ing 14 planning domains that show our system leads to a much greater reduction in search time than previous methods. In closing, we discuss avenues for extending this promising approach to combining heuristics.