Maximizing classifier utility for a given accuracy