Opening book for game with handicap

[Marcus9]

I would like to know if someone has created a book for handicap games, I am interested in the handicap f7 and those of knight b1.
I need some variety but still choose good lines, it doesn't need to be too deep.

[pohl4711]

Use my Armageddon openings:

https://www.sp-cc.de/armageddon-openings.htm

They contain 3 Levels. Level 3 means, Black has no pawn on a7. Each level contains 4 opening-sets: SuperGM_4moves, IM_4moves, FEOBOS, 6 pawnplies. As PGN- and EPD-file. I built opening-books out of the 2 bigger files (IM_4moves and FEOBOS) for Fritz, Arena and polyglot.

[lkaufman]

I would like to know if someone has created a book for handicap games, I am interested in the handicap f7 and those of knight b1.
I need some variety but still choose good lines, it doesn't need to be too deep.

Actually I did make such books for those two handicaps a decade ago for Rybka, I think they were available for either the ChessBase or Aquarium Rybka versions (I don't recall which one, or both?) Probably too old to get now. For f7 handicap White should always start 1e4, then 1...e6, 1...d6, 1...Nc6, and 1...c5 (gambit) and perhaps 1...Nh6 are all decent replies, but I suppose you want a book with much more than five lines.

[mmt]

For f7 handicap: LC0 has d4 +2.23, e4 +2.22, e3 +2.19, Nf3 +2.19 after 1.329 billion nodes total.

[lkaufman]

For f7 handicap: LC0 has d4 +2.23, e4 +2.22, e3 +2.19, Nf3 +2.19 after 1.329 billion nodes total.

Something is wacky with Stockfish; 1.e4 is obviously stronger than other first moves for White with f7 missing, and analysis with any strong engine (perhaps even Stockfish) should confirm this. Just try Lc0 versions, or Houdini, or normal Komodo, or almost any engine.

[lkaufman]

I would like to know if someone has created a book for handicap games, I am interested in the handicap f7 and those of knight b1.
I need some variety but still choose good lines, it doesn't need to be too deep.

In order to understand how Komodo MCTS with Contempt 150 managed a 6 to 0 sweep giving knight odds to FM Gilden at 15' + 10", I made an opening book with 20 positions for knight odds (some b1, some g1) and am trying to pick a suitable computer opponent for that time limit and handicap. I'm using my fast (4.9 GHz) I7 laptop, seven threads for Komodo MCTS, one for single core opponent getting the handicap. I picked Pawny 0.2 64 bit for the first match, which is rated 2415 CCRL 40/40. Since the CCRL ratings are presumably scaled to estimate (conservatively, I would say) human FIDE ratings using the specified reference software for the engines, and since my computer is perhaps triple the speed of the old reference hardware, I think that Pawny 0.2 would be at least as strong as a 2500 FIDE GM at standard time limits, presumably stronger at rapid. Engines have a tougher time exploiting handicaps than humans do, since they are not usually programmed to play safe and simplify when up a piece (at least not the old engines). So I thought this would be a reasonable starting point. Very much to my surprise, Komodo is now leading by a ridiculous 13.5 to 1.5, despite the knight odds!! I'll have to look at the games afterward to see how this is possible, it makes no sense to me. But it is consistent with the Gilden match. I'll have to run other pairings.

[lkaufman]

I would like to know if someone has created a book for handicap games, I am interested in the handicap f7 and those of knight b1.
I need some variety but still choose good lines, it doesn't need to be too deep.

In order to understand how Komodo MCTS with Contempt 150 managed a 6 to 0 sweep giving knight odds to FM Gilden at 15' + 10", I made an opening book with 20 positions for knight odds (some b1, some g1) and am trying to pick a suitable computer opponent for that time limit and handicap. I'm using my fast (4.9 GHz) I7 laptop, seven threads for Komodo MCTS, one for single core opponent getting the handicap. I picked Pawny 0.2 64 bit for the first match, which is rated 2415 CCRL 40/40. Since the CCRL ratings are presumably scaled to estimate (conservatively, I would say) human FIDE ratings using the specified reference software for the engines, and since my computer is perhaps triple the speed of the old reference hardware, I think that Pawny 0.2 would be at least as strong as a 2500 FIDE GM at standard time limits, presumably stronger at rapid. Engines have a tougher time exploiting handicaps than humans do, since they are not usually programmed to play safe and simplify when up a piece (at least not the old engines). So I thought this would be a reasonable starting point. Very much to my surprise, Komodo is now leading by a ridiculous 13.5 to 1.5, despite the knight odds!! I'll have to look at the games afterward to see how this is possible, it makes no sense to me. But it is consistent with the Gilden match. I'll have to run other pairings.

Final score for Komodo MCTS giving knight odds to Pawny 0.2 was 16.5 to 3.5. Curiously the only draw looked exactly like the games "Father" draws with Stockfish by locking up the position; at least Komodo had the excuse of being a knight down! Most of the games Komodo won were by direct attacks on the king; I suppose a human IM or GM would take more care to avoid such attacks. Next I'll try regular Komodo against the same opponent, again using Contempt 150. Let's see which Komodo scores better giving knight odds.

[mmt]

Actually I did make such books for those two handicaps a decade ago for Rybka, I think they were available for either the ChessBase or Aquarium Rybka versions (I don't recall which one, or both?) Probably too old to get now. For f7 handicap White should always start 1e4, then 1...e6, 1...d6, 1...Nc6, and 1...c5 (gambit) and perhaps 1...Nh6 are all decent replies, but I suppose you want a book with much more than five lines.

After ~1.5 billion nodes total, LC0 likes e6 -2.21, Nc6 -2.30, c5 -2.33, d6 -2.36 as responses to 1. e4, similar to Rybka. It doesn't much like Nh6 -2.85, though.

[lkaufman]

Actually I did make such books for those two handicaps a decade ago for Rybka, I think they were available for either the ChessBase or Aquarium Rybka versions (I don't recall which one, or both?) Probably too old to get now. For f7 handicap White should always start 1e4, then 1...e6, 1...d6, 1...Nc6, and 1...c5 (gambit) and perhaps 1...Nh6 are all decent replies, but I suppose you want a book with much more than five lines.

After ~1.5 billion nodes total, LC0 likes e6 -2.21, Nc6 -2.30, c5 -2.33, d6 -2.36 as responses to 1. e4, similar to Rybka. It doesn't much like Nh6 -2.85, though.

Yes, 1...Nh6 is somewhat weaker than the other four, that's why I said "perhaps". I'm a bit surprised that the other four are so close in score. It is an indication that f7 is a good handicap, if there are four fairly equal moves on move 1. I think Lc0 (presumably a current net) is okay for f7 handicap, but it is totally useless for knight odds or more, the analysis is ridiculous (old net like 11248 or 11258 is ok). Meanwhile my test of normal Komodo Contempt 150 vs. Pawny 0.2 finished just 13 to 7 for Komodo, much closer than the 16.5 to 3.5 MCTS match. So it looks like I have some solid basis (always subject to margin of error) for claiming that Komodo MCTS is stronger than normal Komodo when giving knight odds, whether to engines or to humans, as long as both are using similar Contempt. So the Gilden match may not have been a total fluke.

[mmt]

I think Lc0 (presumably a current net) is okay for f7 handicap, but it is totally useless for knight odds or more, the analysis is ridiculous (old net like 11248 or 11258 is ok).

Interesting. Is the reason you're saying it's ridiculous for b1 handicap is because of a very negative evaluation? That's what seems to be happening at the beginning of the analysis but then the evals become somewhat less negative. I'll run it for longer (net 42850). A few games vs other top engines would also be interesting.