Earlier experiments suggested similar conclusions to the ones above. It appeared as if the approach resulted in slower progress and worse results, as deduced from the full model of what was said to be registered data.
Later on, and quite recently in fact, it was shown that values go lower (i.e. registration is improved) by using the subset approach. Many iteration though were required to show this. The subset-driven function caught up with its full-set equivalent and sank well below it. It was not clear though what had happened to the data, which could as well drift away. It is indeed possible that it got eroded more quickly for reasons that were earlier explained.
To summarise, subset-driven function are yet to be investigated, but they do not seem as powerful as adaptive precision in problems of shapes and images, for example. They often showed to be worse in terms of time, as well as worse in terms of performance.