idea for generating defensive problems

[argentum]

Hi,

As we know now the problem generation is based on idea that opponent's last move caused a jump in evaluation, for example from 0 to 3. So we are tasked to find the move to gain the advantage.

How about a defensive situation...

If opponent's last move caused a jump from -3 to 0.  And you can get that gain (and save the game) if you find the right move.. wouldn't that be a sort of defensive kind of tactical problem to solve?

regs,
argentum

[drahacikfm]

Yes, those problems would be good to, where you can go from -3 to 0.

But they are not necessarily defensive.  They could be "offensive" problems like we have now, except that you start with a piece down and then you have a combination to win a piece.

[argentum]

ok,

maybe we can't call them generally defensive but you got the idea.

if CT doesn't have that sort of problems yet those would be nice addition to the problem mix, if there is no problem of identifying and generating such problems.

it might also complicate the problem set (which I consider good) that users wouldn't always know what to look for. now it is always look for a win or winning situation but that sort of 'defensive' problems would add new aspect that you should be looking for to equalise or crawl back from the hole. user should really understand the situation and learn to read the board that maybe there is not a win but see way to escape from losing situation.

[drahacikfm]

Yes, I agree, these problems should be added to the next generator that Richard writes.  A combination to win a piece is a good combination whether you start from equal material or you start from a piece down, doesn't matter.

Because the best move in these situations is going to evaluate at less than +1.75, there won't be any alternate moves.  Probably the gap between first and second moves, currently 1.0, would need to be larger for these problems.  Maybe 2.0 to be safe.

[chesseric]

Heh, I brought up a wish for defensive problems in a post a long time ago. What the search engine has to do is to search for positions that require a unique move irrespective of evaluation jumps. The problem with this on the other hand is to find a criterium for which of those situations are interesting, that is, those situations where you still have a shot and the next move is not obvious. But I think formalizing this last criterium for computer search purposes might be extremely hard.

[angrypirate]

I like the idea of having a problem where the eval is more or less zero (maybe +/- .75 or something) and having a problem where EVERY move loses badly except one.

For example the material evaluation is more or less equal but you are under huge positional pressure or some kind of attack, and you must spot the one and only move that mnaintains equality.


To me that is the idea of a  good defensive problem.

[argentum]

Hi,

Angrypirate, that sounds good.

Considering this a little bit more I thought that kind of sequence is probably quite common in forced capture sequences which might be quite trivial usually(?) (these kind of problems might be good anyway but maybe lower rating if they are easy to see). What if one criteria of such sequences would be that not all of the moves are captures? of course that would limit the number of problems found.

Maybe it would be a good idea to give this a generator run and put the found problems in some set that could be tested if the found problems are good. 

[drahacikfm]

Yes, the trick with these problems is too teach the computer the difference between a simple recapture, and a tactic.  For example, in the Ruy Lopez 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Bxc6 ... now Black must recapture the bishop with bxc6 or dxc6 to get a position that is roughly equal.  Every other move loses badly.  But this is not an interesting tactic.  It's hard to teach that to computers though.

We don't want hundreds of problems where we are just recapturing when the opponent took our piece.  Even though the recapture might be the only move that does not lose.

[argentum]

Hi,

I think that sort of situation would not be recognized as a problem. In that case there is not (should not be) evaluation difference between before the Bxc6 and after it. Thus that should not be recognized as a tactical problem situation in the first place.

If I understand correctly the problem identification sees a positive jump in evaluation before the opponent move and after which gives the chance to improve your position with right move. That is not the case in that kind of opening move which should evaluate nearly same (0.00 in this case) before and after the move.

Right?

Regs,
argentum

[chesseric]

Yes, but the point with a defensive problem is that there is no evaluation jump, since the purpose with a defense is to avoid worsening your position. This is why the current search engine that searches for evaluation jumps very seldomly finds defensive tasks. They will only exist in a few problems as an aspect of the solution, never as the main point.

[argentum]

'Yes, but the point with a defensive problem is that there is no evaluation jump, since the purpose with a defense is to avoid worsening your position.'

ok, maybe it is little bit missleading to call the target problems defensive problems.

what I meant and write about the criteria is that if we can find evaluation jump from -3 to 0 for example, it would differentiate from the current problem set in the sense that the end position is not winning but you have come from lost position to even position. i think there is no such problems at the moment and those would be interesting. maybe they are not defensive problems but something else.

[chesseric]

Oh, so that is what you meant. Actually I never wondered over the evaluation of the end position, since in training tactics you are intersted in the best continuation, irrespective of if the postition is objectively won or lost. What I was refering to was that sometimes the best continuation is not about winning a piece, but instead about not losing a piece; not about giving mate, but avoiding to be mated. Such defensive problems would make the problem set more realistic.

[argentum]

Yes, I think that kind of problems would be interesting and realistic.

The problem is how to find them. The computer doesn't understand positions, it understands only the evaluation. That's why I approached the problem from the evaluation point of view. Setting the question as how to find defensive problems via analysis of evaluation results, and also type and availability of the candidate moves.

In defensive problems we should see the evaluation is not dropping. This kind of criteria alone would produce a huge number of situations which are not strictly defensive problems but a normal well played middlegame for example.

What would differentiate a defensive problem from other situations?

for example: Forced capture sequences should have quite constant evaluation but are not really defensive problems. So maybe one criteria of defensive problem should be that the identified sequence of moves can't consist only of captures.

Also if there is checks in the sequence and there is only one possible reply move it is not a defensive problem to solve. So the criteria should be that in such case there must be more reply moves than one, leading to the fact that user must choose the good move.

In a defensive move sequence there should be at least one reply that is not trivial, ie. not only captures but also a quiet move or escape from a check when possible number of replies is more than one and only 1 move out of N (N>1) really holds the evaluation.

That kind of criteria might find a defensive sequence where there is critical path of replies that are not clearly forced replies.

Maybe my idea is not yet refined to be complete but the solution for finding defensive problems might be along those lines. What do you think?

[cyanfish]

I've been reading the ideas here, and here is my idea for a set of criteria that I think could result in defensive problems:

The starting position must:
-Have at least one move with a losing evaluation (< -1.75)
-Have one move that results in a drawish position or a slight advantage (> -0.5 and < 1.25)
-Have no moves that result in an unclear disadvantage (< -0.5 but > -1.75)

Also, to filter out arbitrary problems, one of the following two criteria (or perhaps a combination) should be applied:
a) The material count at the starting position should be roughly equal (e.g. advantage of -0.5 pawns to 1 pawn)
b) If the previous move was a capture, the first move must not be a recapture on the same square.

Option A stops most arbitrary recaptures (since the objective evaluation would be equal but the starting material would be unequal). However, it wouldn't include positions that require a tactic to achieve material equality (which I think would be a good addition, though perhaps this could be achieved separately?).
Option B directly stops recaptures. It would allow situations in which you must take an undefended piece or regain material via some sort of tactic in order to achieve a roughly equal position. This would create more lower rated problems than the first option (whether this is good or bad, I don't know).

Problems would be pruned at the point when multiple moves don't lose. Forced moves should be included if (and only if) they are immediately followed by a situation in which there is more than one move possible, but only one doesn't lose/lead to an unclear disadvantage.

What do you think of this idea?

[chesseric]

Sounds as a good start, but Cyanfish's b-option could be refined, since sometimes the defense is a recapture on the same square but the problem is to see with what piece to recapture.