Playing with ponder hits

[Laskos]

Just for fun I let 9 engines (many of them Chessbase proprietary, on them I cannot use Don's Sim tool) to play a RR of about 1,000 ponder-on games at move/1s TC. There are about 2,000 game positions for each pair of engines. The ponder-hit matrix (the equivalent of the Sim matrix, in fact the methods are very similar, only the noise is larger in the present case) is

Code: Select all

              Fri.10  Fri.11  Jun.12  Jun.13  Rybk.4  Fri.12  Fri.13  Hou.15  Stre.2

Fritz 10        ----   56.43   55.38   53.83   51.95   60.94   48.52   51.97   54.79
Fritz 11       56.43    ----   53.32   45.65   54.71   66.07   61.77   59.71   68.99
Junior 12      55.38   54.32    ----   59.25   55.10   51.19   49.95   52.99   55.98
Junior 13      53.83   45.65   59.25    ----   52.85   50.38   50.86   52.38   51.92
Rybka 4        51.95   54.71   55.10   52.85    ----   61.51   55.10   58.97   60.05
Fritz 12       60.94   66.07   51.19   50.38   61.51    ----   65.34   56.26   63.83
Fritz 13       48.52   61.77   49.95   50.86   55.10   65.34    ----   55.15   58.07
Houdini 1.5    51.97   59.71   52.99   52.38   58.97   56.26   55.15    ----   60.59
Strelka 2      54.79   68.99   55.98   51.92   60.05   63.83   58.07   60.59    ----

The most similar pairs are:

Code: Select all

Fritz 11   Strelka 2    68.99
Fritz 11    Fritz 12    66.07
Fritz 12    Fritz 13    65.34
Fritz 12   Strelka 2    63.83
Fritz 11    Fritz 13    61.77

The dendrogram using average linkage and the correlation as measure is:




The length on the horizontal axis from origin to the common ancestor is relevant, shorter - more related.
Hope some folks find it interesting.

Kai