New opening training tool.
http://fr.lichess.org/training/opening
Find the 1 to 5 best moves... Wow, nicely done! I like it.
Thankyou.
http://fr.lichess.org/training/opening
Find the 1 to 5 best moves... Wow, nicely done! I like it.
Thankyou.
It's still to get a lot of work done and A LOT of puzzles (65k) are getting generated as we speak. (This will take a few months, but we'll get there)
I and the development team are glad you like it. :)
I and the development team are glad you like it. :)
According to which parameter are the candidate moves/lines proposed?
I think we're still narrowing it down. I believe that if you're in with ~20-30 CP of the top move it's considered good, then there's a boundary of ~20-30 CP for dubious moves, and everything else is marked as bad.
Learning from the tactics puzzles, we wanted to make the system more lenient as computer analysis isn't always perfect. We also have the ability to update the grading system in the future while still retaining current puzzles.
Learning from the tactics puzzles, we wanted to make the system more lenient as computer analysis isn't always perfect. We also have the ability to update the grading system in the future while still retaining current puzzles.
Awesome addition!
It is an interesting approach to opening training, although the way of selecting best moves may be flawed.
Firstly, are these positions generated from the lichess db, or a more definitive resource (such as all GM games in last 10 years freely available and updated weekly here: http://theweekinchess.com/twic )? Choosing 'good' or 'best' opening moves from a collection of online bullet/blitz games does not seem as useful as choosing best moves by frequency as actually played by GMs.
Secondly, engine analysis alone (and its inherent horizon effect limitations) can show good moves as bad and bad moves as good depending on what engine depth is used... which is why a more thorough short and long-term positional analysis than CP difference (as assumed when GMs actually play the line) is needed.
Of course, GM use and frequency of the line can also be problematic (they may have played it frequently, but it is then 'refuted')... so it is not as simple as choosing the most frequently played lines from a GM chess db (as done usually online e.g. http://www.chess.com/opening/eco/B22_Sicilian_Defense_Alapin_Variation )
I guess the bottom line is that - unlike with tactics - there are a lot of dynamic, static, short and long-term features of an opening position, so creating positions via CP can be problematic (although I am not sure what the best way would be). Maybe, simplest, is just let others/opening book curators do the work for you? Grab one of the many free opening books and generate positions from that? (then you know the positions have been vetted and are used by the very strongest engines) Or maybe ask one of the professional vendors if you can use their book? http://www.hiarcs.com/chess-opening-book.htm - 12,681,453 positions as of December, all from top human games (if they let you use it online, I would personally donate the yearly fee :P)
That having been said!! :).....
Feature request: an ECO/opening label toggle so that we can just train the opening we want (for example A10 to train all of the English, or drill down to A17 to train the English Opening, Hedgehog defence)
Thanks as always for the addition and considering any of the above!
It is an interesting approach to opening training, although the way of selecting best moves may be flawed.
Firstly, are these positions generated from the lichess db, or a more definitive resource (such as all GM games in last 10 years freely available and updated weekly here: http://theweekinchess.com/twic )? Choosing 'good' or 'best' opening moves from a collection of online bullet/blitz games does not seem as useful as choosing best moves by frequency as actually played by GMs.
Secondly, engine analysis alone (and its inherent horizon effect limitations) can show good moves as bad and bad moves as good depending on what engine depth is used... which is why a more thorough short and long-term positional analysis than CP difference (as assumed when GMs actually play the line) is needed.
Of course, GM use and frequency of the line can also be problematic (they may have played it frequently, but it is then 'refuted')... so it is not as simple as choosing the most frequently played lines from a GM chess db (as done usually online e.g. http://www.chess.com/opening/eco/B22_Sicilian_Defense_Alapin_Variation )
I guess the bottom line is that - unlike with tactics - there are a lot of dynamic, static, short and long-term features of an opening position, so creating positions via CP can be problematic (although I am not sure what the best way would be). Maybe, simplest, is just let others/opening book curators do the work for you? Grab one of the many free opening books and generate positions from that? (then you know the positions have been vetted and are used by the very strongest engines) Or maybe ask one of the professional vendors if you can use their book? http://www.hiarcs.com/chess-opening-book.htm - 12,681,453 positions as of December, all from top human games (if they let you use it online, I would personally donate the yearly fee :P)
That having been said!! :).....
Feature request: an ECO/opening label toggle so that we can just train the opening we want (for example A10 to train all of the English, or drill down to A17 to train the English Opening, Hedgehog defence)
Thanks as always for the addition and considering any of the above!
"Firstly, are these positions generated from the lichess db,"
No they aren't. They're pulled from TWIC and only positions that occur more than 15 times are selected. Each potential move is then analysed by SF for 12 seconds.
The aim of this is not to select the absolute best engine perfect moves. It's to give you a position and ask you what might be playable. A lot of players end up in an opening position and they can't see the potential moves, that's what this is for.
For that reason, 12 seconds (which is a fair bit of time to get a good idea of how strong a move is with SF on a pretty powerful CPU core) is enough to say "this move can be played without without significant problems out of the opening."
One of the next things we'll be doing is allowing you to request problems by opening.
No they aren't. They're pulled from TWIC and only positions that occur more than 15 times are selected. Each potential move is then analysed by SF for 12 seconds.
The aim of this is not to select the absolute best engine perfect moves. It's to give you a position and ask you what might be playable. A lot of players end up in an opening position and they can't see the potential moves, that's what this is for.
For that reason, 12 seconds (which is a fair bit of time to get a good idea of how strong a move is with SF on a pretty powerful CPU core) is enough to say "this move can be played without without significant problems out of the opening."
One of the next things we'll be doing is allowing you to request problems by opening.
This seems to be training your "candidate move selection" skills, rather than "opening book memorisation" skills -- which is essentially what opening study is all about (understanding the ideas behind the moves, and then memorising the variations).
What you have here is a "Candidate Move Trainer" and can be extended in the general case to any opening/middlegame/endgame position.
An idea for improving on the "Candidate Move Trainer", you can pull down a random selection of opening/middlegame/endgame positions, apply a basic evaluation filter to them to ensure that positions are more or less balanced, and then run them through a 12s multiPV analysis from stockfish, and finally publish them.
This can be an ongoing process, where every month, you generate new problems based on the latest TWIC games. Over time, the DB will have accumulated a vast variety of positions for "Candidate Move Training".
For proper opening training, I would love to see lichess implement something along the lines of
www.youtube.com/watch?v=4OHSJ8ZL680
That would be truly awesome.
What you have here is a "Candidate Move Trainer" and can be extended in the general case to any opening/middlegame/endgame position.
An idea for improving on the "Candidate Move Trainer", you can pull down a random selection of opening/middlegame/endgame positions, apply a basic evaluation filter to them to ensure that positions are more or less balanced, and then run them through a 12s multiPV analysis from stockfish, and finally publish them.
This can be an ongoing process, where every month, you generate new problems based on the latest TWIC games. Over time, the DB will have accumulated a vast variety of positions for "Candidate Move Training".
For proper opening training, I would love to see lichess implement something along the lines of
www.youtube.com/watch?v=4OHSJ8ZL680
That would be truly awesome.
Ok, that is good to know (about TWIC) and also about the general motivation (help players see playable moves). I think the combo of TWIC and SF largely clears up my concerns (12 seconds of SF alone would still reveal how woefully inadequate engine analysis can be for opening selection - some structural flaws are only revealed 30 or moves later!)
And great news on opening choice... I know this feature will evolve over time, one addition would then be a break down of knowledge per opening and even each opening variation (e.g. if there are 10,000 English opening positions and I know 5,000 of them (by getting the training right), I would see that I know 50% of the English opening, and maybe 90% of the hedgehog) I think seeing the % we know each opening would definitely be motivating to keep training.
Thanks again!
And great news on opening choice... I know this feature will evolve over time, one addition would then be a break down of knowledge per opening and even each opening variation (e.g. if there are 10,000 English opening positions and I know 5,000 of them (by getting the training right), I would see that I know 50% of the English opening, and maybe 90% of the hedgehog) I think seeing the % we know each opening would definitely be motivating to keep training.
Thanks again!
@runpawnrun - I think it is more than candidate move training - well, for a start, it is in the opening, so minimally it is 'opening candidate move training' and then, as a CPT user, I actually can see some benefit to lichess' approach: for CPT I deliberately choose narrow opening paths (usually just 1 move for my colour) for memorization... this lets us see all the possible moves.
After seeing the possible moves, maybe I will go back to CPT and see if I prefer one of lichess' options. Basically I can see how I could use both.
Also... (feature request) to make the exercises more didactic, can we get more verbose messaging via variation lists?
So for correct moves, maybe a clickable move list of a continuation from there.
For incorrect moves, instead of just "This move gives your opponent the advantage" - show us the variation that leads to this advantage.
I often wonder why a move is good or bad, so this would be very useful.
After seeing the possible moves, maybe I will go back to CPT and see if I prefer one of lichess' options. Basically I can see how I could use both.
Also... (feature request) to make the exercises more didactic, can we get more verbose messaging via variation lists?
So for correct moves, maybe a clickable move list of a continuation from there.
For incorrect moves, instead of just "This move gives your opponent the advantage" - show us the variation that leads to this advantage.
I often wonder why a move is good or bad, so this would be very useful.
I'm most curious how the "rating" aspect of opening training works, especially whether partial credit is awarded.
So far I solved openings 309, 355, 168, and 235; and while I see the puzzles are getting more challenging (especially 235 in which I found "a3" after about 15 guesses) I haven't missed a problem yet so I don't know how the "rating" aspect works...
So far I solved openings 309, 355, 168, and 235; and while I see the puzzles are getting more challenging (especially 235 in which I found "a3" after about 15 guesses) I haven't missed a problem yet so I don't know how the "rating" aspect works...
This topic has been archived and can no longer be replied to.