"LeelaPieceOdds"

[lkaufman]

A new Leela odds bot is now featured on LiChess, which allows the user to select any reasonable combination of pieces to be removed by Leela as a handicap (except for removal of just one bishop, which is deemed to distort play too much by creating a color complex). So odds of two knights, two bishops, rook and knight, two rooks, two bishops and knight, rook and two knights, queen and knight, four minor pieces, and even larger handicaps are supported! There is not yet a leaderboard for it as there is for queen odds, but one is planned. The only issue with it now is that it uses the queen odds net as there is not yet a net trained for all these unusual handicaps, but I expect at some point such a net will be trained. Even using the queen odds net, Leela plays these handicaps far better than any other engine or player could play them, but there is room for further improvement. Already one anonymous GM has lost a game at 5'3" with TWO knights handicap, something which seemed unimaginable a year ago, but now seems quite plausible given the success of the queen odds bot. On LiChess go to community, then players, then online bots, and currently it's number 7 on the list. Players rated below 1500 may need as much as queen and two rooks handicap to win, depending on the chosen time control.

[Uri Blass]

I wonder if Leela can achieve higher rating than the best humans with very big odds or not.

It is not obvious if Leela is better.

[lkaufman]

I wonder if Leela can achieve higher rating than the best humans with very big odds or not.

It is not obvious if Leela is better.

The best data that we have is that Nakamura eventually reached a 2300 chess.com blitz ratings playing 3'0" giving away his queen for a minimum of a minor piece, sometimes for piece and pawn or a rook. So slightly less than queen for knight odds, alternating colors. Leela QueenForKnight has performed, according to the calculations I posted, at about 2850 LiChess blitz (averaging White and Black) at 5'3", which implies about 2950 at 3'2" and about 3120 at 3'0". However Lichess blitz ratings are clearly lower than chess.com blitz ratings around the 3000 level (though higher around the 2000 level), so maybe about 3200 performance compared to 2300 for Nakamura, a 900 elo difference despite Nakamura giving on average a slightly smaller handicap! Note that this means that the bot can almost give that handicap to Nakamura himself (usually around 3300 chess.com blitz), probably could so at 2'0" or 1'1". Even Leela giving queen odds performs much better than Nakamura giving queen for piece or more. Nakamura is probably the best in the world at doing this (maybe Carlsen would be better with similar experience). But of course if you mean odds like queen and two rooks, then we simply don't know, there is almost no data for either human or engine. Probably we would need a net trained for the odds to get good performance, although even with the current net I would expect it to outperform any human at queen plus two rooks, the maximum odds for the bot.