To tablebase or not to tablebase, that is the question.

[drahacikfm]

Richard has mentioned adding endgame tablebases to the analysis engine that generates problems.

50961 is an interesting case.  Right now, these are the top three moves:

1.Kf4 is mate in 2.
1.Kg6 is +10.53 and is an allowed alternate
1.g3+ is +3.33 and is an allowed alternate

If 5-piece tablebases were added to the analysis, 1.Kg6 reaches a tablebase position on move 5, and 1.g3+ reaches a tablebase position on move 3.  So the evaluation of these moves would change from the numbers above to mates in about 20 moves.

Using the current site rules, these two moves would no longer be alternates, because when you have a mate in 2 (1.Kf4) you are not allowed to play mates that are longer than 5 moves.

Maybe this is not necessarily bad, because if you don't play 1.Kf4 mating in two moves, it's reasonable to fail the problem.  You should see that mate.

[uri blass]

I think that the current rules do not make sense and they need to be changed.
 
If you decide to allow +10 evaluations as alternatives and if you use common sense then you have to allow every mate to be an alternative.
 
The current rules simply sometimes encourage people who choose inferior moves and
even without tablebases
mate in 8 that can be result of the fact that the solver saw winning a lot of material should be better than +10 that is result of the fact that the solver saw winning less material.
 
Uri

[richard]

Currently there is a big problem with mates with material winning lines as the generator can completely miss them as the engine only returns the "mate in N" eval without the normal board eval based on material/position.  I make sure the main line in a mate doesn't also win a lot of material prior to the mate before rejecting them as alternatives when they are long mates but other lines can still have important material wins for the moving side.  The only way of completely solving this is either to allow all mates or to modify the engine to return a material eval when facing a mating line. I don't like the latter as it restricts me to open source engines and would require a change for each engine I'd like to use.

I've been considering using tablebases in the next problem set verification run, but I'm hesitating due to the fact that it will make the above problem worse without changing other items.

Note that alternatives don't directly reward the user for making inferior moves. Currently the user still has to find the "best move", but yes sometimes the alternatives can stop the user from shooting themselves in the foot where maybe the user might learn more by feeling the pain.  Overall the alternatives are mostly (but not often enough for tomohawk :-) ) doing their job which is to prevent users getting angry when they play a good move and fail the problem. The problem is if you allow every mate as Uri suggests this solves the "material captures hiding in mates" problem but does not solve the "people who choose inferior moves" problem, in fact it makes it worse as it gives the user much more scope to play moves they don't understand the final result of and still not get punished. If there is a mate in 5 and a mate in 20 in the endgame and the user played the mate in 20 move it may be because they saw a material win but it might also be because they simply blundered, it is unlikely they played the move because they understood that particular endgame well.  I'm still tempted to allow longer mates to avoid the "hidden material win" problem but judging from user comments after weeding out material wins in the main line of mates it doesn't very often occur that a material win is hiding in a mate that wasn't an alternative (except in the mate in 1 situation which is a whole other thread :-) ).

The tablebase situation is tricky as I'd prefer to have problems in the tactics page being well...tactical.  I'm happy with endgame positions that require tactics but I'd prefer not to have positions that are not outright tactical problems but based on endgame theory (I'd like to have another mode that targets endgame positions). Having the tablebases and allowing mates of all lengths is going to create positions where endgame theory plays a much larger role in move selection (it already does in some positions in the set).

The current rules are certainly not consistent, that is because they are attempting to serve several purposes.  There is not just the issue of what is fair and logical there is also  'protect users from generator' and problem set composition issues. Some changes that look logical can reduce problem set diversity. For example as I produce more problems it would be possible to remove all problems that have alternatives, however the reality of real positions is that the player is often choosing between several reasonable looking lines and if all positions have one good move and all other moves are very obviously bad (or at best even) then the position lacks something that many real games positions do.

I agree that the current situation is not optimal but at the moment the lack of highly rated problems is higher on my priority list as it can compromise rating fairness in the long run.

Richard.

[oded ross]

Unrelated to the topic's title question, why are those two moves allowed as alternates?
I remember reading a post by Richard in the lines of:
1) No alternatives are allowed for a mate in 1.
2) Mates in 3-4 moves are allowed as alternates for a mate in 2.
3) For mates of 3 or more moves, longer mates are allowed if not more than 3 moves longer than the solution.

This seems fair and logical to me. For longer mates (at least mate in 3 in my opinion, preferably 4), a +10 can be a valid alternative, but not for a two-mover.
I agree with drahacikfm - every move but 1.Kf4 should fail this problem.

[richard]

Hi Oded,

The situation described there WAS the case but mate in 1's now allow mate in 2's as alternatives, but nothing longer.

The reason the two alternatives shown are allowed is that the rules you've quoted only apply to alternative mates, alternative material winning lines that don't end in mate (within the computers vision) are allowed as alternatives for mates >1.  This probably looks a bit illogical but is designed to deal with situations where there is say a mate in 2 or longer, but the user sees a queen winning move and takes that instead.  Ideally I'd have a length check to see at what point the material winning line actually won the material and not allow it as an alt if it was too long after the mate would have occurred, but currently this isn't done.

Richard.

[oded ross]

Unrelated to the topic's title question, why are those two moves allowed as alternates?
I remember reading a post by Richard in the lines of:
1) No alternatives are allowed for a mate in 1.
2) Mates in 3-4 moves are allowed as alternates for a mate in 2.
3) For mates of 3 or more moves, longer mates are allowed if not more than 3 moves longer than the solution.

This seems fair and logical to me. For longer mates (at least mate in 3 in my opinion, preferably 4), a +10 can be a valid alternative, but not for a two-mover.
I agree with drahacikfm - every move but 1.Kf4 should fail this problem.

Hi Oded,

The situation described there WAS the case but mate in 1's now allow mate in 2's as alternatives, but nothing longer.

As said I think the no-ALTs allowed for a mate in 1 was in place, but this is a fair compromise.
However, I do think rules for two-movers should be more strict, as to not except material gains nor long mates as ALTs (and I stand corrected about #2 in my previous post - IINM in the major upgrade in May you've thrown out any mates in 2 that also has mates in 3 or 4 instead of keeping those as ALTs; I'm not sure but maybe the ALT concept was only introduced shortly thereafter?).
Another related question - what is the longest mate detected as a valid alternate in any mate problem? Mate in 9 just like the longest mate problem on the site?

[richard]

Hi Oded,

Your points 2) and 3) are both still accurate (at least from memory - I've messed about with that code a few times in the last few months, so my memory may fail me).

I can't easily answer your question on alt mate length as at the moment the alternatives are all stored in a format not easily searchable.  In theory the longest possible alt would be mate in 12 (9+3), but I suspect that the only mate in 9 problem is unlikely to have a mate in 12 (haven't checked though). Before having much stricter ambiguity checking the longest mate was 15 (problem 1394).

Richard.