The strongest passer

[Ralph Stoesser]

Carl, of course I agree that black's tactical chances after c2-c3 should easily outweight the pointy chain bonus, but that's mainly because of search not because of eval. Second, and this is much more important, forget black's counter chances for a minute and only look at the white's pieces.

Do you really think the move c2-c3 does change the character of white's position on the king side enough to justify such a big, abrupt bonus? This is my point. In one case, with the pawn on c2, we award nothing for white's king attacking chances with respect to the closed center and white's "pointy" d4/e5 pawns. In the other case we give a big bonus, only because of the pawn now being on c3? Makes not much sense to me and it is called eval discontinuity, a really big one in this case!

I bet that rule will never work this way. You would have to refine and refine and again refine the rule to a point where it's just impossible to implement. Eval is about simplicity and basic things, not about hundreds of lines for a specific rule that even couldn't have an effect in the mayority of positions.

Our task would be to find much more simple principles behind such a rule. Every simple rule should bring us a step forward in helping engines to understanding closed positions better. Every step should provable result in +Elo.

[Lyudmil Tsvetkov]

Carl, of course I agree that black's tactical chances after c2-c3 should easily outweight the pointy chain bonus, but that's mainly because of search not because of eval. Second, and this is much more important, forget black's counter chances for a minute and only look at the white's pieces.

Do you really think the move c2-c3 does change the character of white's position on the king side enough to justify such a big, abrupt bonus? This is my point. In one case, with the pawn on c2, we award nothing for white's king attacking chances with respect to the closed center and white's "pointy" d4/e5 pawns. In the other case we give a big bonus, only because of the pawn now being on c3? Makes not much sense to me and it is called eval discontinuity, a really big one in this case!

I bet that rule will never work this way. You would have to refine and refine and again refine the rule to a point where it's just impossible to implement. Eval is about simplicity and basic things, not about hundreds of lines for a specific rule that even couldn't have an effect in the mayority of positions.

Our task would be to find much more simple principles behind such a rule. Every simple rule should bring us a step forward in helping engines to understanding closed positions better. Every step should provable result in +Elo.

[d][d]r4rk1/2nqbppp/4p3/p2pP3/n1pP1B2/1pP2N1P/3QBPP1/R2NR1K1 w - - 0 1

Precisely.

White is worth the pointed chain bonus for the c3,d4,e5 pawns above.

The problems is that c3 b3 at the same time creates an even longer black chain, f7-b3, consisting of 5 pawns, one of which is a passer additionally, and additionally connecting to another own passer on adjacent file, a5.

So that no eval discontinuity here at all.
The bonus for c3,d4,e5 pointed chain is due, but black receives much larger bonus at the same time, if you somehow consider longer chains and passers connecting to longer chains.

I bet the bonus should work as it is, with 4 or 5 basic types of pointed chains consisting of 3 pawns each, without caring where the king is.

BUT, the problem is, as you could have seen if you followed some of the game moves, that SF, as well as other engines, does not search well in a range of positions, involving more closed setups. Its search is biased mainly towards open setups, so in more closed setups it simply does not know what to do, so considers first the moves it should not consider, and last the moves it should.

Therefore it does not find the strongest, but the weakest moves instead.
And therefore, even if you apply the correct eval parameter, as Joerg for example is trying to do currently, SF will simply not make any use of it, as its search tweaked specifically to open positions will fail it. I think that is the root cause.

Tweak the search parameters at the same time as you introduce new eval terms, and you will see that it might work.

[carldaman]

Carl, of course I agree that black's tactical chances after c2-c3 should easily outweight the pointy chain bonus, but that's mainly because of search not because of eval. Second, and this is much more important, forget black's counter chances for a minute and only look at the white's pieces.

Do you really think the move c2-c3 does change the character of white's position on the king side enough to justify such a big, abrupt bonus? This is my point. In one case, with the pawn on c2, we award nothing for white's king attacking chances with respect to the closed center and white's "pointy" d4/e5 pawns. In the other case we give a big bonus, only because of the pawn now being on c3? Makes not much sense to me and it is called eval discontinuity, a really big one in this case!

I bet that rule will never work this way. You would have to refine and refine and again refine the rule to a point where it's just impossible to implement. Eval is about simplicity and basic things, not about hundreds of lines for a specific rule that even couldn't have an effect in the mayority of positions.

Our task would be to find much more simple principles behind such a rule. Every simple rule should bring us a step forward in helping engines to understanding closed positions better. Every step should provable result in +Elo.

OK, I see your 'point' clearly now, no pun intended.

I agree with you that's not how you build a strong "pointy" chain - as the name implies, you extend the chain at its point, by pushing a pawn that becomes the most advanced pawn in the chain. I prefer to call this pawn the spearhead.

[Ralph Stoesser]

Eval has no clue about the last move made. That's search, but because of the hash table we have to stay path independent. If you teach a human about this chain pattern it will work because a human is able to recognize the pattern even when there is no exact match. He will "smell" something. Engine is totally blind in this regard and probably would try to move the pawn to c3 to grab the bonus even if it wouldn't be necessary nor any good. In SF framework we tried to focus on the 2 blocked center pawns. There were some experiments in that direction, not only from me. Nothing worked so far ELO wise. I'm not exactly sure why. One reason might be that we hadn't the 2 move book for testing at that time but only the 8 move book.

[carldaman]

Eval has no clue about the last move made. That's search, but because of the hash table we have to stay path independent. If you teach a human about this chain pattern it will work because a human is able to recognize the pattern even when there is no exact match. He will "smell" something. Engine is totally blind in this regard and probably would try to move the pawn to c3 to grab the bonus even if it wouldn't be necessary nor any good. In SF framework we tried to focus on the 2 blocked center pawns. There were some experiments in that direction, not only from me. Nothing worked so far ELO wise. I'm not exactly sure why. One reason might be that we hadn't the 2 move book for testing at that time but only the 8 move book.

What if you greatly reduce the bonus for chains built from the bottom? I prefer that chains directed at the Kingside be extended by advancing a pawn to the foremost position (the point) of the chain. Sometimes a capture will achieve this, but that's rare.

This is the reason I like to refer to this sort of pawn push as "extending the chain" - a logical choice has to be made by the engine to actually extend the chain because it sees it as beneficial eval-wise, even though the search may be blind to any benefits.

Please check the separate thread I started on this, to try to break the concept down for greater clarity. I can picture how hard it must be to get it right, both conceptually and in its application. One misdefined or wrongly weighted term can sink the whole effort.

Regards,
CL

[Ralph Stoesser]

It is not possible, because of technical reasons. The transposition table requires to stay path independent. In TT we store a position and a value for the position, independent of the move sequence (=path) that leads to the position, hence the name transposition table. I'm not sure you understand it but it would be very important to understand this! ...OK, this was my last comment about it here in this thread.

[Lyudmil Tsvetkov]

Eval has no clue about the last move made. That's search, but because of the hash table we have to stay path independent. If you teach a human about this chain pattern it will work because a human is able to recognize the pattern even when there is no exact match. He will "smell" something. Engine is totally blind in this regard and probably would try to move the pawn to c3 to grab the bonus even if it wouldn't be necessary nor any good. In SF framework we tried to focus on the 2 blocked center pawns. There were some experiments in that direction, not only from me. Nothing worked so far ELO wise. I'm not exactly sure why. One reason might be that we hadn't the 2 move book for testing at that time but only the 8 move book.

Absolutely no distinction between how a human evaluates, and how an engine evaluates, and between how a human searches, and an engine searches.

The entities with more complex and finetuned eval and search routines win the day in the end.

An engine will play c3 to create a pointed chain and get the bonus, only if does not know that a longer f7-b3 black chain, created at the same time (actually a ply later, if we want to be precise), is due an even larger bonus.

In this case it will shun away from c2-c3.

So I bet a bonus for pointed chain will easily work in SF and elsewhere, had not it been for the obvious reason that all search parameters of modern engines are specifically tuned to perform well in open positions.

Do you remember, Ralph, when was the last time when SF tuned their search on a whole scale so that it is also responsible for picking the best moves in closed positions?

Because, best move choices in closed and open positions differ widely.
While in open positions you would prefer direct threats/direct attacks, in closed positions you would look at preparatory/regrouping moves first, having nothing to do with direct threats and attacks.

So, if your search is not tuned for play in closed positions, you will simply look at the wrong portions of the tree first, which greatly diminishes the positive effect any new reasonable positional eval term might have on engine performance.

So, do you rememebr when was the last time SF tuned their search parameters for optimal performance in closed positions?

[Ralph Stoesser]

Search tuned for closed positions? Of course never, since nobody even knows what a closed position is. I believe the search should be more selective and deep in closed positions, but how to achive this, by eval or by search or both, is not obvious to me. But, since the weather is so fine here in Berlin today, i will go out now to grab a massive sun bonus.

[Lyudmil Tsvetkov]

Search tuned for closed positions? Of course never, since nobody even knows what a closed position is. I believe the search should be more selective and deep in closed positions, but how to achive this, by eval or by search or both, is not obvious to me. But, since the weather is so fine here in Berlin today, i will go out now to grab a massive sun bonus.

Closed position is a position where you have at least 3 pairs of central blocked pawns on the c,d,e,f files.
Adding an additional condition of having 2 more pairs of blocked pawns elsewhere on the board might be even more precise.

So, for those positions you would like to search in a different way: not look first at attacking moves and moves improving your psqt, as for open positions, but at other possible moves, even moves that worsen your psqt and mobility significantly, but would instead achieve other positional goals.

I know it from my games. I feed a closed position to SF, with a pointed chain for example, or other blocked and closed position features, and, even though one side is strongly favoured by certain positional elements, and there are obvious continuations leading to decisive advantage or straight win for that side, SF does not see those good moves at all.

All the time it prefers to play something else, completely irrelevant for this sort of position. It wants to play all the time moves, that would be very relevant, if the position had been open. Therefore my conclusion: SF searches first the wrong moves in a closed position.


As a side note, Uri recently talked about ply=120 not being the limit for searching, and this is absolutely so, in closed positions you need at least 2 or 3 times deeper search, as there are lots of preparatory/regrouping moves. So that searching 80 plies deep/ 40 moves in closed positions is nothing - quite often the change in score and win is to be found well past the 60th-80th move, and that will make a whopping 120-160 plies, starting from a random position!

What are SF 35-40 usual plies in a tournament game? Or even 50-60 plies after a search lasting hours?

I do not know, you are the programmers, you can tell me what can be improved.

Nevermind, good luck with your daily sun bonus.

Ich bin ein Berliner.

[neelbasant]

[d]r1r1n1k1/3qbppp/1n2p3/p2pP3/1ppP1BNP/5N2/2PQBPP1/R3R1K1 w - - 0 1

Actually, black wins even easier here.

For example: Bg5 Bf8, followed by Na4-c3, and black pushes the a passer forward.

White can do nothing to prevent that.

I can't believe there are engines that do not recognize an easy win for black on the above diagram.

So, quite probably, on the original diagram, black has at least 70-80cps advantage in terms of eval at the root.

Here, SF proposes h5 or Bh6 with about a pawn advantage for white but let's go with the Bg5 line. So, Bg5 Bf8 (SF proposes a4) h5 Na4 (SF proposes b3) h6 Nc3? (SF proposes g6 +1.6) hxg7 with 4.5-5 pawns white advantage.

What if played Bh6 Line...Which forces bishops otherwise.