hgm wrote:
NightmareA unfortunately was unstable, and kept crashing, so that it had to be withdrawn.
It was not really crashing but something went wrong with the HPTC timer it uses.
My Atom computer has been offline for many months and it took hours to have all the updates installed (it runs windows server 2012 v2), after this I replaced VS2013 with VS2017 and recompiled the engine, by that time it was already 20.30. I let the engine play one quick game and everything seemed normal so I decided to enter it in the tourney.
Next time I will take more time to make sure everything runs like it should.
It seems Fizbo wants to repeat winning opening lines. In the october 2016 tournament WaDuuttie (with white) lost badly against Fizbo from ths position:
[d]2r2rk1/2q1bp1p/p3b1P1/N3p1P1/p7/P3BP2/1PPQ3R/1K1R4 w - - 0 24
He then continued with Rdh1 but could not find the right plan.
The same position was reached again in this tournament when we met in the 7th round. This time WaDuuttie played gxh7+ and managed to draw. Maybe this was caused by a slightly better evaluation or by the book- and position-learning WaDuuttie uses or by a combination of both. Anyway this was an improvement.
Schooner had its best ics tournament so far defeating top programs Bliep, rpiStockfish and Goldbar. Unfortunately, it still had a problem checkmating.
I have been experimenting with some mate finding accelerators. There is a flaw where if a strong engine follows the best PV, then there is no problem. The PV is read easily from the hash table. However, if a weak engine breaks from the main line, the accelerators cannot properly pick up the PV and the engine gets lost. That is why Schooner can checkmate Goldbar, but not RookieMonster (good game Rookie Monster!). That is also why it was so difficult to replicate the problem. It is part of the reason the same mate score kept appearing in the output display after every move. The accelerator ignores mate distance. The current fix is to turn off the mate finding accelerators AFTER a mate has already been found.
I've been looking into the timing problem somewhat deeper and actually it seems to be a communication problem between the engine and Winboard 4.8.0. Strange is that on FICS the problem does not occur and that the same engine runs fine on an other PC with Windows 10 and with the same Winboard version.
Now I'm wondering if there is a slight incompatibility between Winboard 4.8.0/Polyglot and Windows server 2012-R2 or that it has something to do with the slowness of the Atom processor it runs on.
When I have some time I will dive deeper into it because I want to have it solved, but it has not a very high priority.
I finally found the culprit, accessing the Nalimov code breaks the communication between the engine and Winboard, I guess it is a bandwidth problem, the EGDB sits on an SSD and I don't access it at very low depths, there is still enough free memory left, there is no paging going on but it blocks communication via the pipe completely.
I didn't notice this problem at FICS, the engine probably won already before hitting the EGDB, it depends upon who you play.
D Sceviour wrote:
Schooner had its best ics tournament so far defeating top programs Bliep, rpiStockfish and Goldbar. Unfortunately, it still had a problem checkmating.
I have been experimenting with some mate finding accelerators. There is a flaw where if a strong engine follows the best PV, then there is no problem. The PV is read easily from the hash table. However, if a weak engine breaks from the main line, the accelerators cannot properly pick up the PV and the engine gets lost. That is why Schooner can checkmate Goldbar, but not RookieMonster (good game Rookie Monster!).
Thank you. It was an interesting game indeed. RM was a bit too optimistic in shedding its pawns for piece play, but not bad for getting half the ply of Schooner! Hopefully it wasn't too far off the PV in the KQK endgame since it does well in finishing those off on the winning side (as well as KRK and KBBK).
WaDuuttie lost his first game (against Kingslayer) because I could not switch of a follow-command. At that time noone could reproduce the problem. Now I know what happened: I used namecompletion to add an engine to my followlist (eg. "follow rpis" to add rpiStockfish). Now giving the command "follow rpis" again does not toggle but prints "[rpiStockfish] is already on your follow list.". Now you have to give the full name to toggle: "follow rpistockfish" prints "[rpiStockfish] removed from your follow list.".
Did you have the option 'Background observe' switched on in WinBoard? I think that with that option on you should be immune against observed games trying to break in on the game that you are playing.
Speaking about ICS options that could wreck things:
It dawned on me that the bug I fixed would only be activated when there were users logged in that had set their ICS variable avail=1. So the fact that this suddenly became a problem after years of trouble-fee running could have been caused by someone setting his avail variable, which by default is 0. First I though mamer would do this, but when I printed mamer's vars, it did have avail=0.