One thoughts I had when first I skimmed the paper is that it should be straightforward to train a small, fast DNN to approximate the technique, and then use the DNN for applications that have more stringent performance requirements, including games.