top engines vs 2400 engine knight odds matches

[lkaufman]

The eval should be something like twice as bad when down two knights as when down one. The version I tested of your engine thinks it's actually better to be down two knights than one knight!!

The raw eval in the starting position is allright (-1.6 vs -2.4). I wonder if it gets confused wrt development and castling rights from both positions.

Castling rights can be checked by inputting a move string to lose the knight, it doesn't appear to matter. If I do a "1 ply search", which apparently is still a pretty deep search based on long PV (even though just a fraction of a second), at one knight odds I get -2.59 with a sensible PV, but at 2 knights odds I get -2.49 with a totally ridiculous, blunder-filled PV. Hope this helps.

[Gian-Carlo Pascutto]

My first thought was that perhaps it is due to the training setup not having any way to arrive in these kind of "opening but immediately 2 knights down" situations, combined with a failure of the network to generalize. But it happens in middlegame positions too.

[D}r1b2rk1/pp1nqppp/2p5/4p3/3P4/2R1P3/PP3PPP/3Q1RK1 w - -

The PV here is d5 cxd5 Qxd5 Nb6 Qc5 Qe6 which looks reasonable enough, but then it wants Qxf8, which is the preferred move from the network. So it looks like the bad quality of move suggestions is causing the score to be off. I'm not clear why the move suggestions should be off though. It's an interesting thing to investigate.

[Uri Blass]

[d]r1b2rk1/pp1nqppp/2p5/4p3/3P4/2R1P3/PP3PPP/3Q1RK1 w - -

Note that D does not work for diagram now and you need to click on fen

[lkaufman]

My first thought was that perhaps it is due to the training setup not having any way to arrive in these kind of "opening but immediately 2 knights down" situations, combined with a failure of the network to generalize. But it happens in middlegame positions too.

[D}r1b2rk1/pp1nqppp/2p5/4p3/3P4/2R1P3/PP3PPP/3Q1RK1 w - -

The PV here is d5 cxd5 Qxd5 Nb6 Qc5 Qe6 which looks reasonable enough, but then it wants Qxf8, which is the preferred move from the network. So it looks like the bad quality of move suggestions is causing the score to be off. I'm not clear why the move suggestions should be off though. It's an interesting thing to investigate.

Your "first thought" above promted me to check what happens if instead of removing a knight, I blunder one in a way that would actually happen in random training, 1.Nc3 Nf6 2.Nd5?? Nxd5 (or the mirror image 1.Nf3 Nc6 2.Ne5?? Nxe5). It turns out that your engine and even Lc0 can play this position pretty well. It's not quite as bad for White as normal knight odds, for example after 3e4 Nf6 it's the same as if we played normal knight odds but required Black to meet 1.e4 with the Alekhine (1...Nf6), which although not so bad is far from the best move. But familiarity is probably more important. I'm thinking this might be a better way to do knight odds in the modern world, partly so NNs can participate fairly, and also because Black can't immediately exploit the missing knight the way he can with standard knight odds. All we need is a name for this way of giving the knight! But as you point out, clearly familiarity is not the only problem, these NNs still degenerate with totally won/lost games. For "pawn and move" the new way would be 1.e4 b5? or 1.d4 g5?? (clearly the larger handicap of the two).

[lkaufman]

I don't know anything about that specific hardware, but since Hiarcs on a cellphone got a performance rating something like 2900 a decade ago I think it's safe to say that Komodo on a modern cellphone would be way above 2600, and a fast time limit favors the computer over the human. But you don't say exactly what you mean by the same opening three times. If he can choose the opening, he might have a chance. But if you just play one random game and the computer has to repeat the same moves to move 8 as long as he does too, then this isn't a big problem for you unless it just happens to choose a very bad opening the first time. I would guess that he has a decent chance to draw one of the three games, but is quite unlikely to win one. Computers are much better at not losing than they are at winning.

[Chessqueen]

I have an interesting question about the strength of Komodo 11 on my Samsung Galaxy S8+? I have downloaded it and installed it on Chess for Android GUI, My friend who is rated 2575 by USCF wants to bet $100.00 to his $ 25.00 that he can beat it with a time control of game in 15 Minutes but I would have t play 3 games playing the same opening up to 8 moves. My question to all of you is what is my chances that after three games playing the same opening my friend would NOT end up winning at least 1 games and after he loses two games he wold NOT win the third 3rd game and my $100.00 whereas he only has lost $50.00 and get away with $50.00.

Note: Is Komodo 11 on My Galaxy S8+ rated at least 2600 USCF ? I forgot to mention that my friend wants to play 3 games using the same opening and playing with the white pieces.
[/quote]

I don't know anything about that specific hardware, but since Hiarcs on a cellphone got a performance rating something like 2900 a decade ago I think it's safe to say that Komodo on a modern cellphone would be way above 2600, and a fast time limit favors the computer over the human. But you don't say exactly what you mean by the same opening three times. If he can choose the opening, he might have a chance. But if you just play one random game and the computer has to repeat the same moves to move 8 as long as he does too, then this isn't a big problem for you unless it just happens to choose a very bad opening the first time. I would guess that he has a decent chance to draw one of the three games, but is quite unlikely to win one. Computers are much better at not losing than they are at winning.

The opening would not be favoring him but a drawish opening in closed position where he is good at, since he played thousands of games in the internet using the same opening, he might have a chance, he will repeat the same 8 moves and the computer as well, he said that if he loses twice he can improve on the third game, hope the Komodo 11 can select better moves from that point on too, since human can learn fast from previous lost games specially using the same opening against the same engine.

My worry is that if it was three different openings, he would not have a chance, but after the 3rd game he could easily improve and get my $100.00, after he only has lost two games equal to $25.00 + $25.00 = $50.00 and on the 3rd game win $100.00 a draw is equal to $50.00 for him

[lkaufman]

You can increase your chances by setting Contempt to a high value, perhaps 100, as this encourages Komodo to avoid closed positions. Also, I don't know how many threads you can use on that hardware, but as long as it is more than 1 this will make for more varied play. You can also set "variety" to a higher number, like maybe 10, with only a couple points elo loss. I think that by improving on his previous play he may manage to get a draw with luck, but a win is unlikely, especially if you can use Contempt = 100.