Why there is no interest in Computer with odds Vs Humans match?

[jp]

Well, I'm not one to push for material handicaps, but I wouldn't feel that a handicap game must have a crazy back-story or history that connects it the chess starting position to justify itself.

But chess games have a "history", they don't begin in the middle of the game, you get there from the opening positions, and in chess, players have infinite freedom to choose what they play, and how they arrive to it.

We just have to accept that the material-handicap game is a different game (with a different starting position), not standard chess, instead of trying to convince ourselves it's a standard game and having to invent a pre-game history.

[lkaufman]

the GM (Perelshteyn) played normal opening moves

No, he didn't, those moves were played for him (I'm advocating for freedom of choice, which to me is a core part of chess).

Looks like 1.d4 g5?? gets very close, though a part of chess is not knowing what your opponent is going to play, so this would work if the GM played 1.d4 by himself, and didn't know the engine would play g5.

So yeah, I hold that the best handicap is the one that carries all the elements from normal chess, which has all material from the beginning, the engine playing with the strongest hardware available, both players start with the same time on the clock, there's no takebacks, the GM doesn't know what the engine will play, and the GM can choose all of her moves from the opening position (no thematic chess).

What if we build an opening book that plays into a hole no matter what the GM plays? Playing the g pawn two moves at inappropriate times, the Knight dancing back to beginning square, exchanging Knight for a protected pawn, moving the King early, are all good ideas, others could be opening a center pawn two steps and move the Bishop to the 3rd / 6th rank to be captured, the a / h pawns two moves forward and the rook up to be captured by the opposing Bishop, and probably many other things could be tried (like unsound openings like a suicidal variation of the King's Gambit, or spending several moves to get the Queen to a3/h3/a6/h6, etc.)

The GM wouldn't know what the engine is going to do, and a stubborn GM could decline the "sacrifice", but that's part of chess, so it could bring some excitement back into the match, as not even the Komodo side would know what kind of handicap will the human run into, because they'd not know in advance the GM's moves, and would have to "prepare" for everything.

Since traditionally Human v Machine matches have an opening book anyway, I think it'd make sense to take advantage of it to implement the handicap on there, and it may be the closest we can get the match to resemble pure chess.

It isn't very practical to spend our time writing an opening book to respond with blunders to every move sequence the opponent could make, and in any case this wouldn't provide for varying handicaps depending on the strength of the opponent or time limit. However, you did give me a practical idea. When playing a handicap game, the opponent (or operator) would input the desired handicap, in centipawns. Komodo would start off at a low level (the exact level depending on the handicap chosen) which means short search and random errors. Once the errors brought the score down to the desired handicap (plus or minus some small amount) Komodo would revert to full strength. This would do what you want; freedom of choice for the player, no predicitibility of Komodo, and yet the handicap would be defined. I'm not convinced this is better than tradtional handicaps, but it does have its advantages.

[dkappe]

nobody cares how many shots of vodka Mariya Muzychuk would need to drink to give me an even chess game.

That I would pay to see.

Yes, the chess engines are very good, but there’s also an element to these contests that’s like having an axe murderer spellcheck your emails. One little slip up and it’s over.

[lkaufman]

nobody cares how many shots of vodka Mariya Muzychuk would need to drink to give me an even chess game.

That I would pay to see.

Yes, the chess engines are very good, but there’s also an element to these contests that’s like having an axe murderer spellcheck your emails. One little slip up and it’s over.

Based on your work with Lc0, what would have to be done to make Lc0 play handicap chess well, other than revert to an old network like 11248? Can you explain why 11248 could beat GM Naroditsky at piece odds, while the new networks just blunder left and right when trying to give knight or more odds? I understand that they don't train on positions where the result is already 99% certain, but then why did 11248 play so well down a piece?

[dkappe]

Based on your work with Lc0, what would have to be done to make Lc0 play handicap chess well, other than revert to an old network like 11248? Can you explain why 11248 could beat GM Naroditsky at piece odds, while the new networks just blunder left and right when trying to give knight or more odds? I understand that they don't train on positions where the result is already 99% certain, but then why did 11248 play so well down a piece?

Not training on types of positions is a huge factor.

Currently I’m training a net on a mix of lichess games with stockfish policy and q — essentially sleazy human play with a little bit of ab engine corrective.

Now most human games end when someone’s a piece up, and certainly very few make it to the endgame, especially in blitz. The result is that the net stinks in the endgame, and also starts moving randomly when it’s up material, until it drops a rook and can think straight again. As a corrective I’m mixing in endgames played out to terminal and kingbase games that I’d also played out to terminal. This is starting to make a difference (incidentally, CCRL games are a bit better in this regard — humans are weak).

So why the piece odds problem post t10? They introduced resignation in self-play. More games/sec., but the nets hardly ever see severely imbalanced positions. One approach to correct this would be to start from piece odds type positions in self-play. But it would have to be done by a third party, as I don’t think there’s much interest in the core team.

[Ovyron]

However, you did give me a practical idea. When playing a handicap game, the opponent (or operator) would input the desired handicap, in centipawns. Komodo would start off at a low level (the exact level depending on the handicap chosen) which means short search and random errors. Once the errors brought the score down to the desired handicap (plus or minus some small amount) Komodo would revert to full strength. This would do what you want; freedom of choice for the player, no predicitibility of Komodo, and yet the handicap would be defined. I'm not convinced this is better than tradtional handicaps, but it does have its advantages.

That's awesome! The reason this is better than traditional handicaps is that those are chess variants (in the same league as Chess 960), while this approach is still pure chess (i.e. some strong GM could play random errors to give a weaker player an even game, in a chess game that otherwise would be normal)

[lkaufman]

I think I figured out something that is important for any knight odds computer vs human games. When Morphy gave knight odds to strong players (by the standards of his time), he always gave the b1 knight, and usually opened 1.e4. Most of his opponents replied 1...e5. But even back then, strong players knew that 1...d5! was the best move and pretty much destroyed any hope for a White initiative. According to an article I read recently, when an opponent was too strong for knight odds (perhaps due to knowing this), Morphy would propose knight odds with the condition that Black had to meet 1.e4 with 1...e5, when at least White retains the first move initiative for a few moves. This was basically an intermediate handicap between unconditional knight odds and a lesser handicap like pawn and two or three moves. That explains the prevalence of 1...e5 in Morphy's knight odds games. It seems that without that condition, White is better off giving the g1 knight rather than the b1 knight, as Staunton and Lasker did in some recorded games, so that 1...d5 2.exd5 Qxd5 can be met by the normal move 3.Nc3. So we should either give the conditional knight odds, or just the g1 knight.

[carldaman]

Based on your work with Lc0, what would have to be done to make Lc0 play handicap chess well, other than revert to an old network like 11248? Can you explain why 11248 could beat GM Naroditsky at piece odds, while the new networks just blunder left and right when trying to give knight or more odds? I understand that they don't train on positions where the result is already 99% certain, but then why did 11248 play so well down a piece?

Not training on types of positions is a huge factor.

Currently I’m training a net on a mix of lichess games with stockfish policy and q — essentially sleazy human play with a little bit of ab engine corrective.

Now most human games end when someone’s a piece up, and certainly very few make it to the endgame, especially in blitz. The result is that the net stinks in the endgame, and also starts moving randomly when it’s up material, until it drops a rook and can think straight again. As a corrective I’m mixing in endgames played out to terminal and kingbase games that I’d also played out to terminal. This is starting to make a difference (incidentally, CCRL games are a bit better in this regard — humans are weak).

So why the piece odds problem post t10? They introduced resignation in self-play. More games/sec., but the nets hardly ever see severely imbalanced positions. One approach to correct this would be to start from piece odds type positions in self-play. But it would have to be done by a third party, as I don’t think there’s much interest in the core team.

Quantity over quality. No wonder.

Thanks for the keen observations + your shared nets. Better approaches are possible. Without them, we'd all be prisoners of the Leela Team.

They've done great work, to their credit, no doubt, but we can't expect perfection (at least not yet).

[jp]

Consider also that chess with odds has a long history.

Yes, but that probably dates back to exactly the times when humans enjoyed e.g. running races of humans vs trains or horses, etc., which we'd find very strange today.

Fischer played tons of handicap chess, Kasparov played one public match, and even Carlsen and MVL have played blitz handicap matches with IM Lawrence Trent. It's not quite ancient history.

I wouldn't want to discourage anyone from playing handicap games if they enjoy them, but along those lines I think it's far more common to have simuls, blindfold chess and blindfold simuls, and I assume those have an equally long history.

Simuls are splitting one human's mental resources, so it's like limiting or splitting hardware resources. I'm not sure what would be analogous to blindfold play for engines. Maybe having no RAM.

[lkaufman]

Consider also that chess with odds has a long history.

Yes, but that probably dates back to exactly the times when humans enjoyed e.g. running races of humans vs trains or horses, etc., which we'd find very strange today.

Fischer played tons of handicap chess, Kasparov played one public match, and even Carlsen and MVL have played blitz handicap matches with IM Lawrence Trent. It's not quite ancient history.

I wouldn't want to discourage anyone from playing handicap games if they enjoy them, but along those lines I think it's far more common to have simuls, blindfold chess and blindfold simuls, and I assume those have an equally long history.

Simuls are splitting one human's mental resources, so it's like limiting or splitting hardware resources. I'm not sure what would be analogous to blindfold play for engines. Maybe having no RAM.

Using only a single core is essentially like a simul, since the machine could play as many games as it has cores (or one less than cores) at once with almost no loss in strength below its single core level on one game. But we already know that single core Komodo (or SF or obviously Lc0) would easily beat even Carlsen in a match, though perhaps with enough other special conditions (White every game, draw odds every game, minimal book, no TBs, no ponder, double time) one core limit might be enough for a good match. Blindfold doesn't have a plausible analog. One big difference between playing handicap games (whether material, time, threads, or whatever) between computers and people as compared to handicap races between trains/horses vs. people is that with races, you can just measure the MPH of the train or horse vs. human and calculate who will win at any given handicap, but with Komodo vs. Carlsen at f7 odds (just as an example) no one can calculate the result, you can only try to predict it via very inaccurate simulations. One thing I may do when I have a bit more time (I'm nearly finished writing my new opening book) is to set a weekly time for Komodo to play games at various handicaps on chess.com, initially against myself and/or my son IM Raymond Kaufman, later against any worthy challengers. As long as it only requires an initial start position, no special rules, I can set it up myself. Even chess 960 with handicaps is possible if I limit it to the 18 positions with the kings and rooks on normal squares so no special castling rules are needed.