Opening selection for the TCEC 14 superfinal

[Vassilis]

Here are the details of the opening selection for the match, according to Jeroen Noomen

http://blogchess2016.blogspot.com/2018/12/opening-selection-tcec-14-superfinal.html

But I am sceptical about this:

[...In each line one side typically has a slight advantage, with book exits mainly between +0.25 and +0.65 (note: this is with contempt = 0)]

And this:

[...My expectations
The TCEC 9 superfinal saw 25% wins, the TCEC 10 superfinal 24% wins, the TCEC 11 superfinal 22% wins, the TCEC 12 superfinal 38% wins and finally the TCEC 13 superfinal 22% wins. It will be very hard to beat the 38% of season 12, but I'd like to see a minimum of at least 20% wins. Of course I hope to beat the 25% wins of the TCEC 9 superfinal.]

Gives me the impression, that his goal is to increase the win percentage by any means!
An opening advantage of 0.65 (in my opinion, always) is very favorable for these engines, in a complex position. Most of the cases, this advantage can be converted to a win on my modest hardware (Desktop and Laptop). I would prefer openings (dynamic and positional) near 0.1-0.3 but far from drawish, for an exciting battle.

Are we sure these lines are dynamically or positionally balanced with lots of play ahead and lots of opportunities for creativity?
Then fine, I like it.

But it's a pity for a top engine to lose, because it's simply lost out of the opening. Even if this engine is Stockfish :)

I want to note here, that I don't question Jeroen's judgment, in any case. He knows what he is doing
The guy is an expert in the openings, for years!
I'm not sure about his "intensions" though.

What is your opinion on this? I would like to hear...

[Matt Blakely]

As long as all played from both sides I'm not worried

[Vassilis]

Matt, yes.  It appears so.

But this advantage (+0.65) usually favors tactically inclined engines, especially in complex positions.

That means, Stockfish might win his white game, but Leela might draw it.

OK, one could argue here that: "Stockfish is a better engine then".

But let's assume two engines are of equal strength, and one is superior tactically while the other is superior positionally... this opening choice favors the former.

The point , from my humble experience is, it is easier to draw a superior, positional game than a tactical one - in high-level computer chess.

And given the fact Leela that "blunders" more often than Stockfish, I'm worried :)

I don't know if I make my point clear ...

I'm a bit worried of this match, and hope to be wrong. Really

We'll see

[Arlet]

I would assume only a few openings will have lopsided +0.65 eval.  Also, +0.65 does not mean a tactical advantage, it could also mean a material imbalance, possibly misjudged by materialistic engines.

[Vassilis]

Yes, possibly.

In their first game, Stockfish seems to have no problem equalizing a bigger advantage, with the black pieces...

Let's see if Leela can do the same in the second game with colors reversed!

[Lee Sailer]

It does read like the goal is to maximize wins, but I think that is not what he means.  One could do that with just outrageous unequal openings.

I try to always think about these as pairs of games.  The goal is to reduce the number of pairs that end up equal, such as two draws (==) or a win and a loss (10 or 01).  A win and a draw (1= or =1) or better yet two wins (11) show superiority.

[Vassilis]

What I was afraid of!

The king's Gambit. A dubious opening, very tactical one, where Stockfish has no trouble to convert.
Now, how probable is for Leela to transpose an opening she is not familiar with, and reach a favorable position is something we'll find out soon.

I think it's time to train the future test 50, not from the starting position, but from randomly chosen chess960 ones. Something suggested by many people here. I guess she wouldn't have any trouble playing any opening then

Regards
Vas

[Jeff Wads]

This is one reason why I don't take these things seriously.  Putting them in a losing position to see if they can recover plays to the crowd alone and amounts to just horsing around.

[Vassilis]

Yes Jeff.

[Warren D Smith]

1. Obviously, if the book gamestart is very unbalanced (one side way ahead) it will lead to one win & one loss --

even result 1-1, useless.

2. If on the other hand gamestart is very balanced, then the draw percentage will be high and the spectators will

be bored with draw after draw 1-1 results.

3. The best gamestarts lie between these two extremes.  Unbalanced just enough so that wins become frequent,

but not unbalanced enough so that we enter case (1). 

And that is what TCEC aimed for, while also trying to cover all subjective "kinds" of chess.

4. A different approach than (3) is to go for positions in which the chances do seem exactly balanced, but nevertheless

draws are artificially unlikely (for example opposite-side castling).   

(4) is superior to (3) in the sense that win+win 2-0 results will be common, whereas with (3) 1.5-0.5 results will

be common, which is less powerful.

But I think that going purely for design method (4) is

likely to distort the nature of chess, so approach (3) is also needed.

QUESTION:

Does anybody have the collected TCEC gamestart positions for various past TCEC seasons in (say) FEN format?

Reason I ask is, I just talked myself into believing that designing these position-sets was a useful contribution -- and

it would be nice for me to have the data that is the fruit of that effort.

[Vassilis]

Warren, I'll only ask you this...

In what category, the king's gambit (a bad opening, excluded from top level -human- tournaments) falls?

And why such an opening, should be included in a top level -computer- tournament?

Sorry... Two questions!

And by the way.

From someone who has played chess, Case 1. does not necessarily apply. We just saw it.

[Warren D Smith]

On Wednesday, February 6, 2019 at 1:10:33 AM UTC-5, Vassilis wrote:

Warren, I'll only ask you this...

In what category, the king's gambit (a bad opening, excluded from top level -human- tournaments) falls?

And why such an opening, should be included in a top level -computer- tournament?

--King's gambit is NOT "a bad opening excluded from top human tournaments."

It is a perfectly good opening, especially for blitz.

For classical time control, Spassky used to play it, he was world champ. Nakamura and NIgel Short still play

it occasionally.

Wikipedia: "[KGA] is infrequently seen at master level today, as Black can obtain a reasonable position by 

returning the extra pawn to consolidate."

This is not to say KGA is a "bad" opening, merely that it is "not as strong as white wishes it were"

because "black can equalize by (nowadays worked out) lines which it took 200 years to find"!

In the case of TCEC 14, they put a rather bad KGA line for white as a gamestart.

This was intentional because they wanted it to be unbalanced in favor of black, but

not so hugely in favor of black that white could not

hope to hold.   (My "category 3.") Obviously this choice by TCEC was a big success! -- Because it

succeeded in creating a game-pair where SF won as black, but Lc0 could

not win as black despite eventually reaching an early-endgame

position where Lc0 thought it was ahead by about +5.5.

(SF also thought Leela was ahead, but only by about +1.8.  And by the way, my weak human evaluation

thinks SF was a lot closer to the truth in that assessment, than leela.  I thought it was about +1.5.)

The reason Lc0 drew that game was: it sucks at endgames.  

It sucks so bad that I could see it was a dead draw, apparently considerably before Lc0 could,

and I am somewhere between 1200 and 2000 strength I think.

And sorry leela, if you suck at endgames, you are going to pay!!

[Vassilis]

Dear Warren...

The way you put it I agree with the most of what you said.

Somehow, I let the fan inside me speak (I'm a huge Leela fan, cannot hide it) and therefore I lost my objectivity...

Sorry for this. I felt like the fan whose favorite football team lost the championship in one game :)

Yes, the king's gambit is not a bad opening, just the specific line chosen. 

And the reason it is not preferred in high-level chess is the one you pointed out.

Leela lost as white, after a model play by Stockfish. I give credit to Stockfish for this

Since I'm about the same level as you (1000-1900 depending on my mood...) I'll share with you my impression for Leela's draw as black.

Leela also played very well, choosing of course a different line (it was expected) and reached a very favorable position.

Actually at some point I had a feeling she was winning, but I could not find how. She had all the positional elements on her side.

Stockfish thought it also, as its evaluation was rising from move to move, and at some point was above 2.00.

(You know, when the evaluation, remains constant no matter how high, then there could be a theoretical draw, but not when it is rising all the time)

And the "the blunder". A different move than Bluefish's suggestion, and the evaluation dropped to equality.

Note this endgame was not purely technical, one, but a highly tactical one. One mistake, and the tables would turned.

This is Leela's weak spot, and not the endgame. Excluding the classical "Leela's trolling" at simple endgames, she actually has played some superb ones.

If we want to improve Leela tremendously, she must become tactical aware. (That means training her from randomly chosen chess960 initial positions, in my opinion. It's an idea, worth trying)

Regards

[Dietrich Kappe]

Chess960 requires all new logic and net architecture for castling.

[Vassilis]

You mean for move generation, yes totally new logic.

But why new net architecture. Aren't Nets trained for evaluating positions and not move rules? I have this impression.

I mean couldn't we use a net trained from chess960 positions, directly to the normal chess variant? The move generation is another part of the program (I have in mind a typical AB engine. Correct me where I'm wrong...)

Thx

[jimgl...@gmail.com]

The opening issue cuts both ways. Stockfish tactically is clearly better than Leela, while positionally Leela is better than Stockfish.

When tactics mattered most, Stockfish destroyed Alpha0 a few times over. But generally Alpha0 smothered and outplanned Stockfish. Lc0 is in the Alpha0 mode, just not as strong (yet).

That King's Gambit was tactical from the start. Stockfish won those two games. No surprise.

Leela is a master of the semi-closed French defense. An observer at CCCC 5 said Leela understands the French as Einstein understood relativity.

Leela and Stockfish just played both sides of the French. Leela won.

They are playing 50 different openings. Some will doubtless favor one competitor over the other. But I'll be surprised if all 50 as a whole do - though I'll be watching!

[Vassilis]

As long as those openings favor both engines in equal terms, I don't mind.
So far it looks this is indeed the case. Stockfish and Leela are about equal in strength, in my opinion, and it's a pity to see one or another getting destroyed, because of an unfortunate opening suit selection. But like I said, so far so good!