Using TAR effectively (sample-based vs. continuous learning)