Yes, but with some care. I simply don't have so many engine families, most of them are commercial or non-UCI. By the way, how did you include Crafty 23.4? I would love to see some Fritzes, as I have some ideas about them (I put Ruffian because it's probably very similar to a UCI Fritz 6 ). I don't have them, as I don't have Hiarcses, Juniors, Sjengs, etc., etc., etc. Putting many fruitish, rybkish and glaurungish engines along with very few other could be a little skewed.Adam Hair wrote:It is instructive, I agree.Laskos wrote:Thanks Adam, the first graph looks a lot like mine. It's also true that adding a lot of Fruitsih 2.1 close derivatives like Toga, etc. and Rybkish 1.0 like Rybka 2, etc. scrambles the upper lineages. I may include some to show that your second graph is also correct. But I would intuitively pay more attention to the first graph, because I am not quite so interested in Rybka 1,2,3,4 obvious lineage, which will scramble Rybka 1 - Fruit 2.1 relation (Rybka 4 being a totally different animal). I even might suppress some of the over-represented Ippos in my graph. This is an art-science what we are doing, but it's at least instructive .Adam Hair wrote:Kai, I think you maybe saying " all obvious things are confirmed" is a bitLaskos wrote:Using my data, in your tree (not the dendrite), Fruit appears as an isolated engine, which, I am sorry to say, is unlikely, Fruit being an inspiration to many. We are using different clustering methods, for now I trust my graph better, all obvious things are confirmed, and there are some surprises too (you know ).michiguel wrote:
In the data you sent me before, Glaurung is closer to Fruit than R1 or Strelka to Fruit, which makes me doubt of the "stylistic" relationship between R1 or S to F. In Adam's data and Michael Hart's, when they included more engines, the relation between S and F fell apart. More engines got in between (but I am saying this without checking details). I always saw distinct branches.
Of course, this does not mean absence of cloning or confirmation of it (the material tables may be a dominant factor in changing the "style" of the engine) but I am not sure we can say Strelka's style is close to Fruit more than Glaurung's.
It is very clear, Strelka's "style" is w/o question Rybka's.
Miguel
Disregarding this,
The fact is, my point (1) stands, _therefore_ it is up to Fabien to make claims about the relation between Strelka and Fruit (with implications on Rybka 1.0 Beta).
Kai
premature.
I created a dendrogram from my data using Systat. I used the same
engines that you used ( except for Shredder 10 and 12 ). Complete
linkage and normalized Euclidean distance.
Looks alot like yours, maybe Fruit and Rybka 1.0 Beta/Strelka not quite
as close as your data, but no big difference I think.
Then I added 10 more engines.
Rybka 3 and Houdini/Ivanhoe are now closer than Fruit and Rybka 1.0
Beta/Strelka.
Kai
I believe the best approach is to include as many engine families as
possible, at least for an overall picture.
Did you need the diagonal in the matrix? I did need it, I put 100% by hand. I tried with 70% and 50%, it didn't change anything, in my case I am not worried about that.
Did you try another linkage method rather than Complete? I did try, the shape changes. I will give details later. "Complete" gave the most expected and interesting results, although, for example, the problem with Naum is independent of any method, as the relations of Ippos too (Fruit 2.1/Rybka 1.0 is more fragile, but fainter or not, it's usually there).
The normalization is probably different in our methods, as your individual "leaves" are much longer. But it doesn't change anything important.
Kai