Milos wrote:bob wrote:All of that is simply 100% nonsense. But don't let that stop you. If I thought you had even a tiny bit of 'understanding of scientific approach and method' our conversations might take a different path. But you butt in with nonsense, then after getting kicked around, you leave with your tail tucked. repeatedly. This will be another one of those cases, probably sooner rather than later...
You are the only one here writing nonsense.
Things like:
in reality, once you have searched one move, you almost _always_ want to split there. Except for the occasional case where you think another move might be best, and know that it is better to search such moves one at a time, using all available threads.
just prove that you have no clue what a scientific method is.
Unfortunately (for the education) there are university professors like that. And that is exactly what makes a difference between excellent and mediocre university programs, not the students that are attending those...
So you _really_ don't understand what is going on? If so, you would realize that (a) the above is the basis for YBW. The difference at the root is that you _do_ get a pretty good hint that a second-best, or third-best move might become the best move, by remembering the node count for each root move from the previous iteration. In a one-best-move root position, you might see the first move with a node count of 25x to 100x or even more compared to the second-best move. But in a position where a couple of moves are "close" those two moves will have a node count way higher than the other moves. Of course, had you been doing this stuff, rather than trying to become a legend in your own mind, you would have either (a) discovered this yourself or (b) find that this is the way Crafty splits at the root and is one of the reasons its parallel search is highly effective.
In your weak mind, "scientific method" is apparently "whatever you say is fact, whatever anyone says is junk." Not exactly the "accepted definition." I actually run these kinds of tests, by the hundreds of thousands. You might try testing rather than posting. You will learn more.
BTW, for two examples of my "nonsense:"
example 1, one clear best move at the root, but a couple of other moves are threatening to replace that move:
Code: Select all
move nodes R/P
Bxe4 15396401 0 0
d5 9510064 0 0
Nh5 5469105 0 0
Nxe4 1912204 0 1
Bd5 952451 0 1
Ne5 866838 0 1
Rfd8 519152 0 1
e5 393799 0 1
Nc5 140114 1 1
Qb8 132144 1 1
Bc6 126426 1 1
Ng4 104296 1 1
Nd5 103276 1 1
Qc5 100265 1 1
Kh7 86073 1 1
g6 82893 1 1
Ra8 77554 1 1
Rfe8 74845 1 1
g5 69785 1 1
Qxc4 63719 1 1
h5 54724 1 1
Nb8 52359 1 1
Ba8 48347 1 1
Ba6 45300 1 1
Ne8 43563 1 1
Qd8 42437 1 1
Nh7 41668 1 1
Rcd8 39294 1 1
Rce8 38811 1 1
Rb8 38811 1 1
Bd8 33900 1 1
Kh8 29350 1 1
Qc6 15577 1 1
total 36705545
First column is a move at the root, second column is the nodes in that sub-tree only. The next two are flags the first says reduce if 1, else not, the second says "this move can be searched in parallel with others if 1, else it has to be searched by itself by all processors."
One clear best move. Two potential replacements. I reduce neither at the root, although I reduce others unless something else causes some to not be reduced, and I sic all cpus on each of those 3 moves, one at a time, to avoid giving one of them to one cpu where the search could take an abnormally long time.
(the above is kopec #22, Bxe4 is the proposed best move...)
Here's another one from an old Belle vs Cray Blitz game. One clear best move, Bxh6. Check the node counts and flags.
Code: Select all
move nodes R/P
Bxh6 184605726 0 0
Bf4 24200673 0 1
Qd6 22641716 0 1
a3 13915861 0 1
b4 3124291 0 1
b3 2711741 0 1
h4 2112311 0 1
a4 1003096 1 1
Qxb6 433634 1 1
Bd2 346845 1 1
Rb1 312422 1 1
Qg6 299091 1 1
Be3 286610 1 1
Qc6 215689 1 1
g4 196193 1 1
Qg4 154844 1 1
Qe7 138909 1 1
Bg5 125354 1 1
Qd5 71011 1 1
Qb3 44163 1 1
Qf7 43624 1 1
Qxh6+ 25994 1 1
g3 24814 1 1
Qc8 24591 1 1
Qc4 23722 1 1
Qg8+ 19774 1 1
Qxe5 17483 1 1
Qd7 15264 1 1
Qf5 11680 1 1
Qe8 10041 1 1
Qf6 8902 1 1
total 257166069
It can search all (but the first move) in parallel, which is classic YBW. A few of the first group have node counts high enough to mark them "interesting" so they do not get reduced, but they do get searched in parallel to improve search efficiency.
Too bad you don't understand the basic mechanics. But that is just one of many things you don't understand. now feel free to once again, tuck your tail and run away. This is not nonsense. It is a _very_ sound approach, with supporting mathematics proving that it works...
BTW, I suppose that the "you think" that you highlighted was a bit beyond your ability to grasp. The program is making these decisions. Based on the ratio between two node counts, the first move and any one of the others. If the ratio is above a carefully tested threshold, that move can be searched in parallel with others that meet that same test. If the ratio is below the threshold, that move is classified as a "potential new best move" and it is searched by itself, using all processors.
Again, just like the roof at your house or wherever you live, that is also over your head. But others might get it and want to give it a try. It _does_ work. It was developed scientifically and with careful thought and accurate testing...