Crazy exchange sacrifices

[michiguel]

In this postion (Gaviota vs Tinker, FICS) Gaviota played f4!?!?! with a complete disregard for the exchange thinking that there is a positional compensation for it.

[D]1r1qr1k1/1pp2pp1/p2b1n1p/3Pn3/1P1B4/P1NbN1P1/R4PBP/3Q1RK1 w - - 6 19

Interestingly, Gaviota built up the postion to reach

[D]1r1qrnk1/1pp2ppn/p6p/3PP3/1P1B3P/P1NB2P1/4R1K1/1Q6 b - - 2 28

Black pieces are completely immobilized and gaviota went on to win. This is becoming a very common pattern. Gaviota likes to sacrifice the exchange when she thinks there is positional compensation. I believe that this is part of an unbalanced evaluation, but it is making a heck of entertaining games. But... sometimes, it goes down on flames.

Look at this against Toga when it played f5!?

[D]r2q1rk1/3b1ppp/p2p2n1/1ppP4/P3PP1b/5N1P/1PB3P1/R1BQR1K1 w - - 3 19

There is something interesting here. Rooks will have no files to play and white managed to dominate the board and reached the following position (against Toga, which is no pushover). Look at the game below (I was so happy when it played 24 a4! and 30 Be4!)

[D]2b1r2k/3q1prp/p7/Pp1PpPP1/1Bp1B2P/4Q2K/1P6/3R4 w - - 7 42

But gaviota was outsearched trying to opening the position and lost. Both games were one night after the other (15 min + 3 sec inc).

The problem is sometimes the exchange sacrifice is not well justified and It is going to be a challenge to tune this without sacrificing excitement. Positional sacrifices are the most fun to watch, but gaviota sometimes sacrifices like Petrosian, and sometimes like Patzerian.

Here are the games:

[Event "ICS rated standard match"]
[Site "69.36.243.188"]
[Date "2009.08.30"]
[Round "-"]
[White "Gaviota"]
[Black "TinkerFICS"]
[Result "1-0"]
[WhiteElo "2484"]
[BlackElo "2512"]
[TimeControl "900+3"]

1.d4 d5 2.c4 e6 3.Nf3 Nf6 4.g3 dxc4 5.Bg2 a6 6.O-O Nc6 7.Nc3 Rb8 8.e4 Be7 9.d5
Nb4 10.Ne5 exd5 11.exd5 Bf5 12.a3 Nd3 13.Nxc4 O-O 14.b4 Re8 15.Be3 h6 16.Ra2
Bd6 17.Bd4 Ne5 18.Ne3 Bd3 19.f4 Bxf1 20.Bxf1 Ned7 21.Nc4 Nh7 22.Bd3 Be7 23.Re2
Bf6 24.Ne5 Ng5 25.Kg2 Bxe5 26.fxe5 Nf8 27.h4 Ngh7 28.Qb1 Ra8 29.a4 Rb8 30.Kh2
Ra8 31.Be4 Kh8 32.Bf5 Qe7 33.d6 cxd6 34.Nd5 Qd8 35.Bb6 Qb8 36.Bc7 Qa7 37.Bxd6
Rab8 38.Nc7 Red8 39.Qa2 Qd4 40.Rd2 Qe3 41.Nd5 Qe1 42.Bxb8 Rxb8 43.Re2 Qd1 44.
Nf4 g5 45.Rd2 Qc1 46.Nd3 Qc7 47.Rf2 Kg7 48.h5 Re8 49.Rc2 Qd8 50.Qc4 Qb8 51.Qe4
Kg8 52.Qf3 Kg7 53.b5 axb5 54.axb5 Re7 55.Qe4 Rc7 56.Ra2 Qe8 57.b6 Rc6 58.Nb4
Rxb6 59.Nd5 Rb5 60.Nc7 Qxe5 61.Nxb5 Qxb5 62.Qd4+ Kg8 63.Rf2 Qa6 64.Bd3 Qc6 65.
Be4 Qc8 66.Qd5 Qe6 67.Qxe6 fxe6 68.Bxb7 Kg7 69.Ra2 Nf6 70.Bf3 N8d7 71.Ra5 Kf7
72.Be2 Ke7 73.Ra8 Ne5 74.Ra7+ Kd6 75.Ra4 Nd5 76.Bd1 Nc3 77.Ra6+ Ke7 78.Bb3 Nf3+
79.Kg2 Nd4 80.Bc4 g4 81.Kf2 Nd1+ 82.Ke1 Ne3 83.Bd3 Kf6 84.Ra4 Ke5 85.Ra8 Kf6
86.Rg8 Ke5 87.Rg6 Nf3+ 88.Ke2 Nd5 89.Rxh6 Nd4+ 90.Kf2 Nc3 91.Rg6 Nd1+ 92.Ke1
Nb2 93.Bb1 Nf3+ 94.Ke2 Nd4+ 95.Kf1 Nc4 96.h6 Nd2+ 97.Kf2 1-0

[Event "ICS rated standard match"]
[Site "69.36.243.188"]
[Date "2009.08.29"]
[Round "-"]
[White "Gaviota"]
[Black "TogaII"]
[Result "0-1"]
[WhiteElo "2533"]
[BlackElo "2583"]
[TimeControl "900+3"]

1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Be7 6.Re1 b5 7.Bb3 d6 8.c3 O-O 9.h3
Na5 10.Bc2 c5 11.d4 Nd7 12.Nbd2 exd4 13.cxd4 Nc6 14.d5 Nce5 15.Nxe5 Nxe5 16.a4
Bd7 17.f4 Ng6 18.Nf3 Bh4 19.f5 Bxe1 20.Qxe1 Ne5 21.Nxe5 dxe5 22.Be3 c4 23.Bc5
Re8 24.a5 Qg5 25.Qf2 Qf6 26.Bb4 g6 27.Re1 Kh8 28.g3 gxf5 29.exf5 Rad8 30.Be4
Bc8 31.Rd1 Rd7 32.Kh2 Bb7 33.Qf3 Rg8 34.h4 Rdd8 35.Rg1 Ba8 36.Kh3 Rde8 37.g4
Qd8 38.g5 Qd7 39.Qf2 Rg7 40.Rd1 Bb7 41.Qe3 Bc8 42.Qb6 Rg6 43.Qe3 f6 44.Kh2 Rg7
45.Rg1 Bb7 46.g6 Qc7 47.Rd1 c3 48.bxc3 Qc4 49.Qd3 Qxd3 50.Rxd3 Rd7 51.h5 Kg8
52.Kh3 Red8 53.Rg3 h6 54.Bc5 Bxd5 55.Bc2 Bc4 56.Be3 Kg7 57.Bb6 Bf1+ 58.Kh2 Rd2+
59.Kg1 R8d7 60.Be4 Ra2 61.Kh1 Rd1 62.Bg1 Ra4 63.Ba8 Rh4+ 64.Bh2 Rd2 65.Bg2 e4
66.Kg1 Bxg2 67.Rxg2 e3 68.Kf1 Rxg2 69.Kxg2 Rxh2+ 70.Kxh2 e2 71.Kg3 e1=Q+ 0-1

[Eraserheads]

Cool games from Gaviota! It is interesting to see yet another engine who plays with such 'disregard' for its material count! In my observation engines who have this kind of style and understanding have the potential to become top class, exciting performers on the board. Are you planning to develop Gaviota even further in this direction?

Where can I find the latest Gaviota?

[swami]

Where can I find the latest Gaviota?

https://www.msu.edu/~ballicor/gav/

[BubbaTough]

Cool games from Gaviota! It is interesting to see yet another engine who plays with such 'disregard' for its material count! In my observation engines who have this kind of style and understanding have the potential to become top class, exciting performers on the board. Are you planning to develop Gaviota even further in this direction?

This is probably because crazy sacrificing programs usually have eval functions written from scratch by strong chess players . Those that borrow eval ideas and weights from other programs end up with something well tuned not to sacrifice too much, and those that are not as strong at chess often underweight the factors that would cause one to sacrifice.


Miguel,

I have found tuning this kind of issue is very difficult. I had a version of LearningLemming that would do positional queen for rook sacrifices....awful but mesmerizing games. The last version sacrificed pieces for pawns way too often. My 2nd to last tournament for example it sacrificed a piece in over half its games. Tuning this stuff out is so tough because it is rare there is just one factor you can change...instead you have to change lots of things. For example, if you boost the value of a rook, it changes its opinion of 2 rooks vs. a queen....not just exchange saqs. If you decrease the value of minor pieces it starts unjustified piece saqs for attacks or endgame passed pawns. If you increase the value of a rook but decrease the value of rook activation, it may stop developing the rook well. Anyway, good luck. If I spot any good / bad exchange saqs in our ICC games I will message you (to help you build a test set, which is usually my first step when trying to tune this type of issue).

-Sam

[jdart]

It's a difficult issue but not insoluble. Larry Kaufman's article:

http://home.comcast.net/~danheisman/Art ... alance.htm

has been the starting point for many programmers, although I don't think it is 100% reliable.

[Graham Banks]

Where can I find the latest Gaviota?

https://www.msu.edu/~ballicor/gav/

Time for a new one. Hint hint Miguel.

[fern]

Miguel:
Tu programa parece realmente interesante y creo que lo voy a descargar...si encuentro dónde. ¿Está disponible en algún lado?
Abrazos desde el otro lado de la cordillera or whatever
Fern

[BubbaTough]

It's a difficult issue but not insoluble. Larry Kaufman's article:

http://home.comcast.net/~danheisman/Art ... alance.htm

has been the starting point for many programmers, although I don't think it is 100% reliable.

I think the issue is he has a host of other factors (center control, mobility, king safety etc.) that add up to the difference between rook and bishop faster than he would like. Kaufman's stuff does not really address that. Generally there is no quick fix for that kind of problem.

-Sam

[BubbaTough]

Are you also getting a lot of other sacrifices, like pawn sacrifices, or 2 pieces for a rook, or bishop+rook for queen? If not, then my assumption might be wrong, and you may have some serious bug or something.

-Sam

[michiguel]

Are you also getting a lot of other sacrifices, like pawn sacrifices, or 2 pieces for a rook, or bishop+rook for queen? If not, then my assumption might be wrong, and you may have some serious bug or something.

-Sam

I think you assumption is correct. Several positional terms pile up to become bigger than a certain difference in material. I see pawn sacrifices, but not many BN x R or RB x Q. Gaviota may not like rooks so much . I have seen her sacrificing a N for two pawns and positional compensation. I posted some time ago a sacrifice right out of book that looked suicidal (N x P) but not so much... In fact, Gaviota won two games and drew another against engines of similar strength. A typical sacrifice I have seen is a pawn for the pair of bishops and good general piece activity. I am happy with that most of the time.

I read the Larry Kaufman article already and I have to say that it is not the solution, it may be part of the problem I think that it is a good starting point to think about different material imbalances but I do not think that the weights are accurate, particularly the ones related to pawns. I have implemented the concept though. I have a sort of array of "interactions" between different pieces with different bonus. For instance, if I have two bishops, that is a bonus. If I have two rooks, I have a negative bonus and so on. Part of this interactions create sometimes a situation in which they accumulate favorably for one side and negatively for the other. Add positional advantage and you may justify a sacrifice.

Miguel