What is the elo value of memorizing book moves based on an engine?

[Uri Blass]

The idea is simply to play games in lichess and analyze every game and decide the moves to add to the book based on the engine choice.


For example if you play in the first game
1.e4 e5 2.Nf3 Nc6 3.d4 and the engine suggest your first 2 moves but suggest 3.Bb5 after a minute then add to the book of white 1.e4 e5 2.Nf3 Nc6 3.Bb5

I believe that I can memorize average of at least 20 new moves every day without forgetting the moves that I already memorized at least for 100 days if I train with chessbook.com so I may get an opening book with 1000 moves of white and 1000 moves of black that I memorize perfectly in this way(of course I need to train my repertoire every day).

Now the question is which engine is best to use to get the maximal elo rating in this way and how many elo points chess players at different levels can earn by this way(at time control of 5+3)?

[Ciekce]

exactly 5 elo, I reckon

[Viz]

exactly 5 elo, I reckon

IIRC it was 7,2 elo, but you need to remember that SPRT elo is not reliable.

[jefk]

Why this question i wonder, because i suppose you already should have an idea about the answer(s).
Anyway imo in general, the benefit of opening knowledge, not only moves based on an engine but also on statistics
of eg. correspondence games, depends on the level of the player. At higher levels opening theory tends to become
more important especially at slower time controls. A player as Kasparov in his best times was known for his deep
opening preparation; at such GM levels i would suggest opening knowledge would account for approx 100-200
Elo; also GM's and/or their seconds tended to spend lots of time on opening preparation (don't know how
this is nowadays, at top level i suppose seconds simply show some new possibilities to the superGM).
But it's not only about memorizing, but also understanding, relating it to master games already played
(with specific plans, etc), also at lower levels.
In the past opening knowledge also contributed a lot to the possibility of getting a title in ICCF
correspondence chess; nowadays it's easier in theory but at lower levels some players still
sometimes tend to be inaccurate, which can cause an occasional loss.

[Uri Blass]

exactly 5 elo, I reckon

IIRC it was 7,2 elo, but you need to remember that SPRT elo is not reliable.

I guess you mean value to top engines.

My question is about elo value for humans(of course it may be dependent on the playing strength.
It is related to computer chess because I use chess engines to decide what moves to memorize and the question is which engine is better to use.


Note that I do not use statistics in lichess because unfortunately statistics that is written there is misleading.

For example humans got in lichess more than 50% with obviously dubious gambits like 1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Nxe5 and I suspect the main reason is that there are some human players who play this gambit only against weak players.

[chesskobra]

I had been thinking about it slightly differently, and hope the following is useful to someone. Take a database of games, sort distinct lines (say 10 plies) by frequency, and then study the lines one by one. You can try to understand each move in a line at some level before moving to the next line. If you take top 1000 10-ply lines, there will be fewer than 10000 distinct moves to study. I can't comment on the Elo gain, though.

The first line below extracts 10-ply lines from Norman Pollock's 40h database, which is a far superior quality database of high level games than lichess. The second line first removes blank lines, then sorts lines, uniquifies lines and adds frequency count, and then sorts by descending frequency. After all, GMs are doing their homework with engines. So why not trust the frequencies in their play. You can even build a polyglot book by cutting off at any frequency and play with it. Once you study a bunch of lines in 10-ply or 8-ply, you can go slightly deeper in 12-ply or 14-ply for some of the lines.

Code: Select all

pgn-extract --notags --nomovenumbers -w1000 --plylimit 10 40h.pgn -o 10-ply.pgn
sed '/^\s*$/d' 10-ply.pgn | sort | uniq -c | sort -nr > 10-ply-freq.pgn

This is how the result looks like. The first column is the frequency.

Code: Select all

  24348 e4 c5 Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 a6 *
  21046 e4 e5 Nf3 Nc6 Bb5 a6 Ba4 Nf6 O-O Be7 *
   7002 e4 c5 Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 Nc6 *
   6482 d4 Nf6 c4 g6 Nc3 Bg7 e4 d6 Nf3 O-O *
   6456 e4 c5 Nf3 Nc6 d4 cxd4 Nxd4 Nf6 Nc3 d6 *
   5183 d4 Nf6 c4 g6 Nc3 d5 cxd5 Nxd5 e4 Nxc3 *
   4954 e4 c5 Nf3 e6 d4 cxd4 Nxd4 Nc6 Nc3 Qc7 *
   4239 d4 Nf6 c4 g6 Nc3 Bg7 e4 d6 Be2 O-O *
   4223 e4 c5 Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 g6 *
   3776 e4 c5 Nf3 Nc6 d4 cxd4 Nxd4 Nf6 Nc3 e5 *
   ...

[Uri Blass]

I had been thinking about it slightly differently, and hope the following is useful to someone. Take a database of games, sort distinct lines (say 10 plies) by frequency, and then study the lines one by one. You can try to understand each move in a line at some level before moving to the next line. If you take top 1000 10-ply lines, there will be fewer than 10000 distinct moves to study. I can't comment on the Elo gain, though.

The first line below extracts 10-ply lines from Norman Pollock's 40h database, which is a far superior quality database of high level games than lichess. The second line first removes blank lines, then sorts lines, uniquifies lines and adds frequency count, and then sorts by descending frequency. After all, GMs are doing their homework with engines. So why not trust the frequencies in their play. You can even build a polyglot book by cutting off at any frequency and play with it. Once you study a bunch of lines in 10-ply or 8-ply, you can go slightly deeper in 12-ply or 14-ply for some of the lines.

Code: Select all

pgn-extract --notags --nomovenumbers -w1000 --plylimit 10 40h.pgn -o 10-ply.pgn
sed '/^\s*$/d' 10-ply.pgn | sort | uniq -c | sort -nr > 10-ply-freq.pgn

This is how the result looks like. The first column is the frequency.

Code: Select all

  24348 e4 c5 Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 a6 *
  21046 e4 e5 Nf3 Nc6 Bb5 a6 Ba4 Nf6 O-O Be7 *
   7002 e4 c5 Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 Nc6 *
   6482 d4 Nf6 c4 g6 Nc3 Bg7 e4 d6 Nf3 O-O *
   6456 e4 c5 Nf3 Nc6 d4 cxd4 Nxd4 Nf6 Nc3 d6 *
   5183 d4 Nf6 c4 g6 Nc3 d5 cxd5 Nxd5 e4 Nxc3 *
   4954 e4 c5 Nf3 e6 d4 cxd4 Nxd4 Nc6 Nc3 Qc7 *
   4239 d4 Nf6 c4 g6 Nc3 Bg7 e4 d6 Be2 O-O *
   4223 e4 c5 Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 g6 *
   3776 e4 c5 Nf3 Nc6 d4 cxd4 Nxd4 Nf6 Nc3 e5 *
   ...

1)The use of lichess is to decide about moves to prepare against them and not about moves that I play.
I do not play against GM's so the frequency of GM's is not relevant for the moves of the opponent.

2)I do not trust the frequency of GM because I guess part of the games are from
years like 2018 or 2020 and engines made a lot of progress since that time.

I prefer simply to choose the move that stockfish or another top engine suggest after few minutes of search.

3)The problem with the idea to study many lines of 10 plies is that I can waste time about lines that I do not need.
For example if I decide that I play 1.e4 c5 with black and 1.d4 with white then lines with 1.e4 e5 are a waste of time because I will never get them in games in both colors.

[jdart]

Many engines in lichess have fairly wide opening books, and will play some offbeat lines. So if the move is coming from their book it may not be the best choice. It doesn't matter for engines because most are using Stockfish, which is so strong it can overcome an opening disadvantage. But if you are learning from them, you should realize that their book moves may not be as strong as moves that come from search.

[Graham Banks]

Not much to be honest unless you understand the nuances of the opening itself - what one side is trying to achieve and how the other side is aiming to counter that.

[BrendanJNorman]

If you randomly memorise lines which have no relation to one another, there is precisely zero Elo added to a human's chess IMO.

On the other hand, if you have you already have a basic opening repertoire, and make it bulletproof by "fixing" lines with modern engine improvements in certain positions, this can be big.

I've seen estimates as high as +150 Elo for having a full no holes repertoire.

In fact, I've manufactured such results when training some of the kids I coach.

One 8 year old boy in China switched from 1...e5 to Caro Kann, learned the lines I taught him (sharpened with SF, and Patricia as backup) and went from 1500-1880 Rapid on liChess.

Such training really needs to be done properly though.

As Graham says, there needs to be underlying *understanding* of the opening structures, plans, tricks etc as well.