Exploiting First-Order Regression in Inductive Policy Selection