The hypothesis of this paper is that the greater flexibility and reduction of vertical node interference in MBPR outweighs the possible negative effects of increased sibling coordination interference and the possible cost of introducing memory.
The experimental results presented show that the intron effects from MBPR are too damaging for those problems (e.g., regression) that are very homogeneous. This correlates well with whether or not those problems benefit from the introduction of memory; When introducing memory to a problem does not help or even hurts, MBPR is unlikely to provide a benefit since memory must be added for MBPR to be accomplished.
When memory is necessary, the experimental results indicated that, for the test problem, MBPR did not have a dramatic performance impact. Given the negligible cost, the flexibility and generality of MBPR certainly makes it worth consideration for these problem types. However, if the problem is more "horizontal", the likelihood of introns is already present, and vertical node interference is a frequent occurrence then (e.g., the Parity problem) MBPR is likely to give a boost to the search process.
The take home messages are two fold. First, MBPR does not seem to be a problem independent improvement for tree representations, although it seems to be an advantage when memory use is simple, helpful, or required. In addition, the MBPR paradigm, as described in section 2.1, has several other points in its favor. To verify these conclusions, for both tree-GP and non-tree-GP representations, further experiments should be done with more complex problems.
The second conclusion of this paper is that introns are probably damaging. This goes against conventional wisdom, but in the data presented in this paper, the control experiments identify intron effects as the only possible cause of significant performance degradation under a variety of conditions and on a variety of problems. Results were presented that showed this change, to varying degrees, with local, hierarchical, and sibling introns.
The corollary to this conclusion is that simple changes in the function set can have large effects on computational effort. In particular, the addition of one intron causing function can make it difficult or impossible to solve what otherwise would have been an easy problem in genetic programming.