Discussion of anything and everything relating to chess playing software and machines.
Moderators: hgm, Harvey Williamson, bob
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
- Posts: 599
- Joined: Tue Nov 24, 2015 8:11 pm
- Location: upstate
kasinp wrote:Here is what I see:
a) 4-CPU direct comparison between 0.921 and 0.92 is not available (ver. 0.92 not having been tested with 4CPUs)
This happened because v0.921 came out shortly (less than two months) after v0.92. There were only a couple of dozen 40/40 games completed with v0.92 at that point; further testing was abandoned in favor of the new version.
kasinp wrote:b) However, a comparison between single CPU performance indicates that ver. 0.921 (based on some 300 games) is 23 ELO *below* that of 0.92.
After 850 games the gap has been reduced to 15 Elo, which is still well within the +/-19 error bars.
kasinp wrote:This would support the hypothesis that 0.921 may have have introduced an issue, irrespective of the illegal move played in one of the TCEC games.
Such an issue might be very hard to confirm. According to the author the gain from v0.92 to v0.921 is only 5 Elo, and that would take more than 10,000 games to verify -- an unrealistic proposal at 40/40. If there is indeed an issue (which I doubt) it certainly doesn't show up on the single-core 40/4 list where both versions reached exactly the same rating after 1000+ games.
Thank you, Daniel.