What the heck happens here?

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
Laskos
Posts: 9543
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: What the heck happens here?

Post by Laskos » Thu Aug 22, 2019 12:59 pm

Komodo 13.1 doesn't seem to migrate significantly to Lc0 branch, regular version or no-LMR version. The numbers in the matrix are probably more interesting than the dendrogram, but I will post both. I do not understand why is that, it could not be that "Komodo 13.1 is too weak" to show that migration. Komodo 13.1 MCTS shows itself as a completely different engine compared to other engines, but I am not sure it outputted meaningful PV to Sim analyzer at 1s/position. Can somebody compile a Windows binary of SF_dev with LMR disabled and post the link? I am unable to do that. Here are the results:

Code: Select all

sim version 3

  Key:

  1) Andscacs 0.95 (time: 100 ms  scale: 10.0)
  2) Ethereal 11.50 (time: 100 ms  scale: 10.0)
  3) Fire 7.1 (time: 100 ms  scale: 10.0)
  4) Fruit 2.1 (time: 100 ms  scale: 10.0)
  5) Komodo 13.1 (time: 100 ms  scale: 10.0)
  6) Komodo 13.1 MCTS (time: 100 ms  scale: 10.0)
  7) Komodo 13.1 No LMR x30 (time: 100 ms  scale: 300.0)
  8) Komodo 13.1 x10 (time: 100 ms  scale: 100.0)
  9) Komodo 13.1 x30 (time: 100 ms  scale: 300.0)
 10) Lc0 11261 (time: 100 ms  scale: 10.0)
 11) Lc0 320x24b (time: 100 ms  scale: 10.0)
 12) Lc0 320x24b x0.2 (time: 100 ms  scale: 2.0)
 13) Lc0 32930 (time: 100 ms  scale: 10.0)
 14) Lc0 42184 (time: 100 ms  scale: 10.0)
 15) Lc0 42850 (time: 100 ms  scale: 10.0)
 16) Senpai 1.0 (time: 100 ms  scale: 10.0)
 17) SF 10 (time: 100 ms  scale: 10.0)
 18) SF 8 (time: 100 ms  scale: 10.0)
 19) SF dev (time: 100 ms  scale: 10.0)
 20) SF dev x10 (time: 100 ms  scale: 100.0)
 21) SF dev x30 (time: 100 ms  scale: 300.0)

         1     2     3     4     5     6     7     8     9    10    11    12	13    14    15    16    17    18    19    20    21
  1.  ----- 49.19 45.69 37.95 49.07 34.47 49.28 48.69 49.78 45.19 44.57 44.68 43.86 44.14 44.25 46.88 50.36 52.22 49.93 46.58 45.47
  2.  49.19 ----- 48.05 39.58 50.17 35.08 49.73 49.56 49.76 47.01 45.51 45.65 45.40 45.88 46.07 48.66 52.15 52.48 52.09 47.49 47.15
  3.  45.69 48.05 ----- 40.17 47.56 33.03 46.03 45.93 46.50 42.74 42.17 43.06 41.85 42.69 42.35 45.35 48.36 50.24 47.69 43.76 42.91
  4.  37.95 39.58 40.17 ----- 38.70 28.31 37.21 37.98 37.76 35.34 35.45 35.71 35.08 35.31 35.07 46.55 37.81 39.88 37.50 34.78 34.69
  5.  49.07 50.17 47.56 38.70 ----- 38.76 59.91 62.61 61.99 48.96 48.09 47.67 47.57 48.57 48.31 47.96 52.83 53.99 53.14 50.70 49.78
  6.  34.47 35.08 33.03 28.31 38.76 ----- 39.90 38.48 38.40 35.96 35.87 34.90 35.79 35.70 35.97 35.01 36.76 37.42 36.94 38.27 37.16
  7.  49.28 49.73 46.03 37.21 59.91 39.90 ----- 61.02 61.57 54.92 54.18 51.61 53.46 53.90 53.91 46.61 54.54 54.58 55.06 58.59 57.08
  8.  48.69 49.56 45.93 37.98 62.61 38.48 61.02 ----- 62.66 49.87 48.90 48.13 48.37 49.04 49.17 46.97 52.78 53.06 52.73 51.19 50.85
  9.  49.78 49.76 46.50 37.76 61.99 38.40 61.57 62.66 ----- 49.22 48.71 48.00 48.23 48.92 48.94 47.37 52.88 53.78 53.34 51.68 50.45
 10.  45.19 47.01 42.74 35.34 48.96 35.96 54.92 49.87 49.22 ----- 74.13 66.64 73.45 73.46 73.95 42.40 51.41 49.22 50.66 59.04 58.92
 11.  44.57 45.51 42.17 35.45 48.09 35.87 54.18 48.90 48.71 74.13 ----- 72.29 78.42 81.80 82.58 42.11 50.28 48.05 50.35 59.25 59.07
 12.  44.68 45.65 43.06 35.71 47.67 34.90 51.61 48.13 48.00 66.64 72.29 ----- 68.35 70.64 70.72 42.61 49.78 47.16 49.42 54.14 54.62
 13.  43.86 45.40 41.85 35.08 47.57 35.79 53.46 48.37 48.23 73.45 78.42 68.35 ----- 77.70 77.43 42.02 50.12 47.50 49.58 57.71 58.05
 14.  44.14 45.88 42.69 35.31 48.57 35.70 53.90 49.04 48.92 73.46 81.80 70.64 77.70 ----- 84.94 43.01 49.70 48.18 49.56 58.30 58.30
 15.  44.25 46.07 42.35 35.07 48.31 35.97 53.91 49.17 48.94 73.95 82.58 70.72 77.43 84.94 ----- 42.61 50.08 48.49 49.73 58.45 58.82
 16.  46.88 48.66 45.35 46.55 47.96 35.01 46.61 46.97 47.37 42.40 42.11 42.61 42.02 43.01 42.61 ----- 46.42 48.07 46.56 42.70 41.94
 17.  50.36 52.15 48.36 37.81 52.83 36.76 54.54 52.78 52.88 51.41 50.28 49.78 50.12 49.70 50.08 46.42 ----- 58.76 63.17 55.01 53.95
 18.  52.22 52.48 50.24 39.88 53.99 37.42 54.58 53.06 53.78 49.22 48.05 47.16 47.50 48.18 48.49 48.07 58.76 ----- 57.13 52.22 51.29
 19.  49.93 52.09 47.69 37.50 53.14 36.94 55.06 52.73 53.34 50.66 50.35 49.42 49.58 49.56 49.73 46.56 63.17 57.13 ----- 55.39 53.79
 20.  46.58 47.49 43.76 34.78 50.70 38.27 58.59 51.19 51.68 59.04 59.25 54.14 57.71 58.30 58.45 42.70 55.01 52.22 55.39 ----- 72.27
 21.  45.47 47.15 42.91 34.69 49.78 37.16 57.08 50.85 50.45 58.92 59.07 54.62 58.05 58.30 58.82 41.94 53.95 51.29 53.79 72.27 -----
 
Here is the clustering (method is correlation):


-
Lc0_04_dendr.jpg
Lc0_04_dendr.jpg (79.69 KiB) Viewed 790 times

jp
Posts: 841
Joined: Mon Apr 23, 2018 5:54 am

Re: What the heck happens here?

Post by jp » Thu Aug 22, 2019 3:14 pm

Can someone please explain what dendrograms show?

Pio
Posts: 120
Joined: Sat Feb 25, 2012 9:42 pm
Location: Stockholm
Contact:

Re: What the heck happens here?

Post by Pio » Thu Aug 22, 2019 8:50 pm

jp wrote:
Thu Aug 22, 2019 3:14 pm
Can someone please explain what dendrograms show?
They try to show the similarities between different engines. The more similar they are the closer they should be in the diagram. But you should not take the picture as the truth because the similarities are dependent on the metric used and the clustering algorithm. In the most basic clustering algorithm you always identify the two most similar nodes/engines according to your metric and merge them together to a fictive ancestor node/engine whose length to the other engines are defined as the average length of the children subtracted with half of the length between the two identified children (so that the sums of the lengths walking in the tree from one leaf node to another should hopefully be quite close to the original metrics distances). You then identify more nodes in a greedy way until you come to the root. The algorithm I think is used has many flaws as it is dependent in the order it merges nodes and that it is very greedy in its nature. For example it will identify two nodes that are close to each other but that has very little shared resemblance with the other nodes. If that happens the tree will be very bad. For a more thorough understanding of the problem you can read my master thesis https://www.math.kth.se/xComb/x2.pdf

Best regards
Pio

Post Reply