Suggestion thread for TCEC

[101]

In this thread, which should ideally have been just in the TCEC subforum and not in the season 8 sub-subforum, I invite everyone to share their suggestions for TCEC, for season 8 or for future seasons if impossible to implement mid-season. To post such suggestions in a permanent medium like a forum should prove superior to posting it in the Chatwing chat.

Now to one suggestion I have for Martin and TCEC for season 9. We have already seen some very interesting endings this season. Some of them have unfortunately, in my opinion, been adjudicated when five pieces remained. One of these is the KBBKN ending from Gull-Hannibal (game 12 of stage 3), where Gull would only win within the 50 moves rule by three plies with perfect play. This is one of the situation where two engines without tablebases could have produced an interesting endgame lesson, which I believe many would have enjoyed watching.

Here are some other examples of endings which I would like to see "played through", i.e. converted, in TCEC:
(KQKR) - interesting for humans, and some weeker engines might fail to win, especially against a tablebase-equipped defender
(KQKRP)
KQPKQ - often almost impossible to play perfectly, even for strong engines, without tablebases
KQNKQ - a win or a difficult-to-defend draw in some exceptional situations (general result: draw)
KQBKQ - a win or a difficult-to-defend draw in some exceptional situations (general result: draw)
KRPKR - in some special cases difficult (but possible) to win or defend
KRBKR - a win or a difficult-to-defend draw in some exceptional situations (general result: draw)
KRNKR - a win or a difficult-to-defend draw in some exceptional situations (general result: draw)
KBPKB - long and difficult wins are few and far between for engines, but this should still be an exception both for those cases and for the often still instructive value of this endgame in the simpler cases
KNPKN - long and difficult wins are few and far between for engines, but this should still be an exception both for those cases and for the often still instructive value of this endgame in the simpler cases
KBPKN - in some cases instructive to see converted
KNPKB - in some cases instructive to see converted
KNNKP - often, not always (but then sometimes difficult to draw), a win disregarding the 50 moves rule. Sometimes a win, sometimes a draw, under FIDE rules; obviously, extremely precise play is often required to convert within the 50 moves rule.
KPPKP - can transpose into one of the above-mentioned endings

There are even some extremely exceptional cases in addition to these, which I don't mention (like the longest win of KBNKN). My primary suggestion is this: Replace 5 piece adjudication with 4 piece adjudication, possibly with an exception for the KQKR ending. A secondary suggestion is to use the list above, a solution which is good but not perfect (there will always be some extremely exceptional cases in addition which one would have liked to include but would make the list much longer and further reduce the point of even having a 5 piece adjudication instead of 4).

[H.G.Muller]

I think adjudication is always a very bad idea. You should always play the engines until they resign, or the game legally ends.

If this is not considered interesting for the audience, there are several solutions all much preferable over adjudication.
* you can reduce the remaining time (and increment) 10-fold at the point where you otherwise would adjudicate, to force a quick termination of the game
* you could adjourn the game, and finish it on another computer (which people still interested could watch on another web page), while the main computer would continue with the next game.

[101]

I don't think your alternative solutions are practical for TCEC. It would be a bit half-hearted to force a finish with a different TC or on a different hardware than the ones used for the rest of the game. I think an evaluation based adjudication is good, since it is based on both engines' evaluation, but that tablebase adjudication should be either abolished or restricted to four pieces.

By the way, it seems that the discussion has evolved much more quickly in the Rybka forum, so you (and others) might want to have a look there.

[Mark Bennet]

Here is a different modest suggestion.

Some engines move instantly without posting a position evaluation. In some games this resets the TCEC win or draw rules. I think there was one game in stage 1 where an engine lost on time or by stalling after the win rule was reset in this way.

If the engine posts no evaluation, the evaluation from the previous move should be posted to avoid this effect.

One minor issue would be that the distance to mate will be wrong in the case of M evaluations.

[Mark Bennet]

Here is another suggestion. Supplement the draw rule by 10 moves (20 plies) - or some other amount - at agreed zero score regardless of pawn moves or captures.

[101]

Two suggestions from parallell thread
Two new suggestions have been posted in the Rybka forum, one of which is an alternative to the one suggested by Mark Bennet regarding adjudication.

See "Suggestion thread for TCEC" in the Computer Chess subforum: http://rybkaforum.net/cgi-bin/rybkaforu ... pid=557475.