Free online Chess server. Play Chess now in a clean interface. No registration, no ads, no plugin required. Play Chess with the computer, friends or random opponents.
Sign in
Reconnecting
  1.  Improving Horde
  2. Forum
  3. SF Horde Strength

I played 5 games against SF8. 3 times with black and two times with white.
I didn't thought much about my moves and ended up winning two of the Black games quite easy. The third black game I blundered in a winning position and lost.
With white I had zero chances.

I'll try to play against SF8 again, but more serious, with more thinking. :)
Maybe i can get a win with white...

@Stubenfisch
Thanks for the info. It would be very nice if you can give a summary of Stockfish's main weaknesses from your point of view after the games with more thinking time.

Alright I can do that.
I think I will play SF8 tomorrow as I am busy today.

It's been some time now but I played against SF8 with a subaccount I rarely play on and it's not only mine... Anyway I managed to win 2 White games back then:
lichess.org/LOTgYG1f
lichess.org/vzhcGFeo

Probably that SF8 was not as good then as it is today, so it would be interesting to basically relaunch a SF analysis on these games in order to see if the new SF shows improvements. The evaluation curve in the first game suggests that SF either made mistakes or realized too late what was actually going on. For example on move 60 in the first game, the eval goes from +3.6 a few moves ago to around equal and then quickly shifts to +4.3... And the suggested moves in the analysis don't seem to give Black the expected outcome.

PS: Of course I lost many many games too as White ;)

@svenos
SF's play at the beginning of the year and now is indeed hardly comparable, because it has improved by around 400 Elo since then (see http://35.161.250.236:6543/regression ).

By the way, I do not remember seeing a title on your account, so have you achieved the FM title recently? If so, congrats.

@ubdip Thank you for the information.

It's been a year now that I have the FM title and I just requested the mark on lichess so that I could play in the Titled Arena. I still take your congrats, thanks :)

Very nice, thanks ubdip ! Lots of info to digest, but the eval makes a lot more sense to me now. horde_breakthrough3 is promising :)

The "Our pieces / Their pieces" thing is complex. I suppose the values are appropriately reversed for "when we are the other side" ? This is all about slightly tweaking the pieces value to adapt to RRNB vs QBB not being exactly equal (i.e. the Imbalance factor), is that right ?

Why is pawn structure worth a lot less for the horde in endgame than middlegame ? Relatedly, in the second position, why is Pawn structure favorable to black ? (also, Threats : backrank bonus is in which ?)
Overall I would think pawn structure is a strong factor in horde throughout the game, i.e. horde somewhat blurs the middlegame-endgame distinction, especially re Pawns.

Is the total eval the simple sum of all the factors (and then it doesn't matter in which category you insert a given bonus), or is there some more complex aggregation function ?

Can we see / do we know how the eval function varies across the game ? e.g. is it always flat at first and then a slow straight move to the end result, or does it happen (often) that white (I mean the challenger) thinks it is ahead (say +2) but then can't prevent losing (so slowly (or fast) goes to -10) ? Those would be interesting to analyse "by hand" (I mean assisted to understand what's going on, of course).

@lecw
Yes, RRNB vs QBB is an example of what the imbalance factor is useful for in standard chess. In horde chess it of course has a much more significant impact and the evaluation is very sensitive especially to changes of the values that are multiplied by the number of pawns in the horde. Automated tuning often does not work very well for horde because the inherent asymmetry destabilizes the evaluation.

"Why is pawn structure worth a lot less for the horde in endgame than middlegame ?"
That is a good point. It is mainly because of the bonus for connected pawns (either next to each other or one protecting the other), which is calculated in a different way for middle- and endgame. For horde it might indeed make sense to just use the same calculation in both cases. I am not sure whether have I already tried that before, but the result will anyway be interesting, since it should be significant (in either direction). Edit: Test submitted - http://35.161.250.236:6543/tests/view/5a3c31c76e23db35f28bb07e

"Is the total eval the simple sum of all the factors"
It is, except that in some cases the evaluation is multiplied by a scale factor, e.g. to make evaluation drawish if white is behind in material but has a queen. But it never matters in which category an evaluation term is, since the multiplication is done after adding all values.

Analyzing games where SF's evaluation significantly differs from the correct evaluation or where it blunders is of course one of the main sources for ideas.

By the way, are you a programmer? You seem to dig into the code, so if you have ideas for improvements, then you are also welcome to submit tests yourself if you like.

I'm a CS grad student, yes I can look around and catch stuff. Don't really want / have the time to understand a better picture of the codebase though (test submitting platform especially). Maybe later or if this gets more fascinating to me.

The factor affected in horde_connected4 is really important, I think, though I don't understand it very well. I assume a phalanx is two pawns side by side ?

Is there any kind of bonus for pawns being forward rather than behind ? Like a phalanx on 6th rank is worth more than a phalanx on 4th rank, and unsupported pawns on 5th rank are a bigger malus than unsupported pawns on 2nd rank. This is a lot more important in horde races than in classical rook endgames.

What is already in place to favor having the same number of pawns on each file ?

I see there's a value for pawn-sheltering the king ? seems a bit irrelevant to horde, what happens if it's disabled ? (I suspect it wouldn't change much but maybe it wouldn't try to attack the king early with white as I think it does).

http://35.161.250.236:6543/tests/view/5a3b92ac6e23db35f28bb06e is so sad ! I had 200 wins more than losses at some point, and then it hit a bad streak ><

Yes, phalanx means two pawns side by side. The "connected" bonus depends on the rank, see github.com/ddugovic/Stockfish/blob/c403fc9d6ae38ea93a580e5b3518756c648ae29d/src/pawns.cpp#L662-L713. The piece-square table and the passed pawn bonus also depend on the rank, and they are the same for connected and unconnected pawns.

There is a penalty for imbalanced pawn counts on neighboring files using a quite simple formula, see github.com/ddugovic/Stockfish/blob/c403fc9d6ae38ea93a580e5b3518756c648ae29d/src/pawns.cpp#L552-L562. I tweaked the formula several times, and in total it gained about 50 Elo, I think. The improvement by the "breakthrough" bonus was more than 100 Elo if I remember correctly, so it is very clear that these kind of evaluation terms are very important in horde chess, and Stockfish was almost blind to such ideas before we added those bonuses.

It did not yield an improvement to remove the pawn-shelter term, but since it is useful to get rid of paramter values that are not gaining Elo, I am rerunning it with less strict bounds.

This topic has been archived and can no longer be replied to.