When Optimal Is Just Not Good Enough: Learning Fast Informative Action Cost Partitionings