It depends on what and how you test. What I was talking about was to compare move ordering schemes by looking at node counts for fixed depth searches over a big number of positions. For such tests, the results are almost completely worthless if the test is being run with LMR enabled, for reasons explained previously in the thread. The nodes count with LMR on mean nothing.bob wrote:Since I use LMR, I test with it. Otherwise something can hurt more with LMR and help without LMR. So occasionally disabling LMR to see if the program gets weaker is a good sanity check. But I would never test by adding new ideas while disabling others, because the interactions between those two sets of things is critical to capture when trying to decide which is better.
There is also another reason for disabling LMR when running this kind of test, which I haven't mentioned previously in the thread: LMR makes the assumption that the best move is almost always near the beginning of the move list. If the best move is always among the first two or three moves searched, LMR is completely safe and stable. When the quality of the move ordering as measured with LMR off increases, so does the accuracy of LMR.