What You Can Learn From Odds Chess

Odds (or handicap) chess,
where the stronger player removes one or more pieces or pawns before play starts, was very popular in the 1800s, when travel was difficult and closely matched opponents were hard to find. I learned chess from a book of
Paul Morphy'
s games from this period, of which about a quarter were odds games, so for me odds chess has always been an important part of the chess tradition, and a vital tool for teaching how to exploit a material advantage.
Paul Morphy
Over the years, the popularity of odds chess has receded and rebounded, with engines taking the place of Morphy and his fellow odds-givers. The new era of amazing machines giving increasingly impressive odds to chess players is a good moment to examine the history of odds chess, as well as to figure out the best ways to learn from it.
The History Of Odds Chess
The Rise Of Leela Odds Bots
How To Learn From Odds Chess
What Can We Learn From Leela Odds Bots?
The History Of Odds Chess
Historically, there were several standard handicaps that stronger players gave their opponents.
Pawn and move: odds-giver plays Black, usually only for (f7) pawn odds
Pawn and two moves: same handicap but 1.e4 already played with White to move
Knight odds: odds-giver White, b1 or g1 removed
Rook odds: always a1-rook to preserve usual kingside castling
Queen odds
Other combinations: rook plus knight, piece and move odds...
Morphy was able to give b1-knight odds successfully to three of America's top-10 players and the champion of Cuba. Later,
George Mackenzie
and
Wilhelm Steinitz
also excelled, giving knight odds to strong players.
Odds play faded in the 1900s when the level of play advanced and travel became easier, although IM
Al Horowitz
and later the incredible GM
Bobby Fischer
kept the tradition alive, both demonstrating that they could give knight odds in fast games to master-level players. Fischer played hundreds of games giving the f7-pawn and a move, two moves, three moves, or even four moves (not crossing midline) to masters and experts, notably still-active FM
Asa Hoffmann
. In 2001, GM
Garry Kasparov
won a Classical charity match 2.5-1.5 vs.
Terence Chapman
, a British tycoon who became an FM soon after, giving him a two-pawn handicap.
In 2007, odds chess made a comeback once engines clearly surpassed even world champions in standard chess, with odds matches of engines vs. grandmasters, mostly organized by me. In 2007-2008 it was
Rybka
giving pawn odds or rook for knight to strong GMs (
Jaan Ehlvest
,
Joel Benjamin
,
Roman Dzindzichashvili
, and
Vadim Milov
) in over-the-board fast classical.
Two years of watching these matches apparently helped my own chess, as I earned the grandmaster title at age 60, immediately afterward! Then in 2015, it was
Komodo
giving usually two pawns or rook for knight to GMs, including
Hikaru Nakamura
and
Maxime Vachier-Lagrave
, in rapid on Chess.com. By 2021, Komodo Dragon was giving knight odds to “par” (around 2500 FIDE) GMs in 15+10 rapid, though only scoring 33.3 percent. Meanwhile streamers, most notably Nakamura, often gave informal queen for knight odds to amateurs in speed runs by deliberately trading queen for minor piece as early as possible in blitz games. Nakamura reached 2300 at 3+0 blitz doing so, though it’s much harder to do this with an increment when you actually need to win on the board.
The Rise Of Leela Odds Bots
The really big development in odds chess was in the last three years, when a fork of
Leela
, the best GPU-based engine, was developed specifically for odds play. Initially it used a new version of "contempt" to improve odds play, but two years ago a huge improvement came from training neural nets just for odds play, against other neural nets which were simulations of the play of strong human players.
Leela's logo
A further big jump came from applying the idea of contempt to the search, not just to the evaluation; this means "assume all humans are morons, by engine standards!" The results were amazing, and the bots which use them are just incredibly strong, beyond anything thought possible, and still improving. Roughly speaking, they are even with par GMs at 120+30 classical giving knight odds, with GMs at 15+10 rapid giving rook odds, at fast rapid (10+0) giving queen for the b8-knight odds, and with par GMs at 3+0 blitz giving queen odds! They all run on my home computer with just one 5090 GPU, which can support up to about 20 games at once. They have played over 1.6 million games against human opponents already.
Assume all humans are morons, by engine standards!
—Chess engine "contempt" setting
Many people have contributed to this Leela odds-bot project, either by creating the bots, by training the nets they use or the human-simulating nets they train against, or by modifying the basic Leela engine to make it more suitable for odds play. My own role is to tune parameter values, to select which net to use for each handicap, and to create and update an opening book for each handicap for adequate variety and good results.
"LeelaQueenOdds" is by far the most popular bot, with over a million games, followed by "LeelaPieceOdds," for removal of two or more pieces (which most players need!) plus the rare option of bishop odds. The remaining bots, intended for very strong players, are "LeelaRookOdds," "LeelaKnightOdds," and "LeelaQueenForKnight" (a step between rook and two knights). In blitz, or even fast rapid, almost no one makes a plus score at knight or even rook odds.
Here's an example of LeelaQueenOdds taking down a master-level opponent.
Here's LeelaKnightOdds, defeating a GM.
It's worth noting that in both of the attacking masterpieces above, it's difficult to determine the losing move. If you thought that odds games were boring or were primarily just about waiting for the weaker player to make a simple blunder, please reconsider!
Most games aren’t this spectacular, but Leela rarely just waits for a gift (LeelaKnightOdds scored 89 percent in blitz this year against GMs, while LeelaRookOdds scored 87 percent; surely that would not happen just from blunders). It plays incredibly actively to maximize the chance of human error, as it was trained to do. If you play Black against Leela with suitable odds, you will be forced to improve your defensive skills. Leela will often attack like a maniac; there is usually some way to defend and win, but it is rarely obvious, as Leela has just learned what its "fair" human opponents are usually capable of seeing. It may bluff!
How To Learn From Odds Chess
So why play odds chess? If you are the odds-receiver, you must improve your "technique," or the ability to win won games. No better way to learn this than to start with a won game against a stronger player who will do their best to thwart you! You learn to play correctly rather than "hope chess," as playing for tricks is pointless against a much stronger player. Another benefit is to see how the stronger player resists from a losing position; you might need that skill yourself! Just seeing how a much stronger opponent thwarts your plans should improve your game; if you played an even position you would likely never have any chance to do more than grovel.
Of course, it’s also just fun to beat stronger players, even starting out with an advantage. If you are the odds-giver, you learn to put up maximum resistance when behind, a very valuable skill, and how to trick weaker players, which is often vital in positions that should be drawn with correct play. Giving odds also builds confidence, if you win! A few years ago, I fell into an opening trap in an over-the-board classical tournament game with a higher-rated master, losing knight for pawn. I think most masters would have resigned, but due to my vast experience both giving odds and seeing the engines do it, I chose to play on, and by channeling my inner "engine" I managed to reach rook vs. rook + bishop. I claimed the 50-move draw just as he was about to deliver checkmate!
Just before winning the Candidates Tournament this spring, GM
Javokhir Sindarov
played a 17-game training blitz match with LeelaKnightOdds, losing 15 to 2 (no wins), scoring the same two points as Nakamura had done a year earlier in 17 blitz games, counting one recent game. Sindarov played a stronger version and played at 3+1 vs Nakamura's 3+2, but Sindarov took White in all but two games (Nakamura played all Black), so it's difficult to say who did better; let's call it a tie. Other top GMs mostly did worse. Apparently, Sindarov considered this good training, and judging by the Candidates' result it didn't hurt!
Sindarov at the board. Photo: Maria Emelianova/Chess.com.
What odds should you play? The general rule is to play at the largest handicap that still favors the stronger player, ideally by about 60-40. I generally recommend taking piece (or multi-piece) odds and playing Black, no need for pawn or move odds; play is more varied and interesting when the piece odds-giver plays White. Note that the gap between rook odds and knight odds, though very real, is less than the nominal endgame two-pawn differential would suggest, since rooks are devalued with no open files, especially in the corners. After that, the gap between rook odds and queen odds is enormous. For a guide to the initial values of the pieces in odds chess, I use knight=3.5, bishop=4, rook=5, and queen=10.
An excellent way to improve your chess, whether you are rated 200 or GM level, is to start playing the bots or your teacher at a slightly challenging handicap, then gradually reduce the handicap in small steps whenever you score seven out of 10 games. Keep in mind that one step (normally half a knight, like two knights to rook, or rook to knight) typically represents a full-class improvement. You can use "and move" as intermediate steps to make progression smoother.
Strong human players can also play bold, dynamic chess giving odds to competent opponents. Leela’s play is reminiscent of the way Morphy did it, only against vastly stronger opposition now. Good odds-givers, human or engine, will try every trick possible to win, or to draw if winning is implausible. Probably Leela bots have played more “monster rook” repeated rook sacrifices to force stalemate or repetition than all human masters combined!
What Can We Learn From Leela Odds Bots?
The first major lesson from Leela Odds Bots is that it shows that the drawish view of human chess was wrong; humans make so many errors that losing modest material doesn’t always mean a certain loss, just a probable one given that both players will err. Leela wins or draws "hopeless" positions so often against strong players that it almost makes material seem worthless! Of course, that’s an illusion, but Leela does often sacrifice material for compensation that looks quite inadequate to human eyes, and it often works.
Going outside the chess world, these bots are like the canary in the coal mine. They show that there are no known limits to what future bots might be able to do, whether for good or evil. To me, beating GMs at blitz giving them queen odds seemed more unlikely a few years ago than curing aging or cancer (for example), yet it has happened first.
Have you tried playing odds chess? Let us know in the comments!