New (free) Chess App - www.TheChessTrainer.com

[jryandx]

Hello everyone! I'm curious how good I can get at chess by training myself to imitate Stockfish 13, so I built an app to find out and I hope that others might find it useful as well. At the very least, I think it will be the perfect way to memorize opening variations, but I think with enough training I'll be able to pick up on patterns and just get a feel for what Stockfish would do in a given situation.

Here's how it works:
SF13-N means Stockfish 13 with search depth set to N. I'm recording games of SF13-20 playing against SF13-N with N in {1,2,3...20} and then turning those games into a game where you have to make the exact moves that SF13-20 would make. You start out playing SF13-1 and then level up to SF13-2 etc. You choose the opening, start move, #moves/game, and #reps as your training parameters. Start move is how many moves into each game to begin your training session, not counting the opening moves. # Moves is how many moves per game to make before moving to the next level. # Reps is how many times to complete # moves before moving to the next level. You chose the opening by clicking the opening at the top of the screen and the other parameters are set with an interface to the right of the board. Drag and drop pieces isn't implemented yet so you have to click the piece and then click the square you want to move it to. It currently only works with laptop/desktop devices, but I plan on having it working as a mobile app soon.

In the near future I'm going to record games against a variety of engines, not just Stockfish with different search depths. I realized yesterday that Stockfish has a built in mechanism to adjust the strength by sampling from the distribution of possible moves, but it takes about 15 seconds per move with this setting whereas just producing the best move at search depth 20 only takes 3-4 seconds, and it's only a fraction of a second for most of the lesser search depths. Once I get all of the openings recorded with search depth 20, I'll increase the search depth and record those games. Search depth 20 seems to be a good speed/strength trade off for this initial version. There's some really interesting work being done on training neural networks to imitate specific players from the Lichess database and I'm hoping to incorporate that in the future to produce more human-like games (I'm a neural network engineer by trade)

[towforce]

A very good idea!

My concern: just as it's not possible to teach a housefly to write intellectual articles, so it is not possible to teach a human to play chess like stockfish.

However, if the human could pick up some deep patterns like GMs must have, that would be a very big positive!

[jryandx]

"My concern: just as it's not possible to teach a housefly to write intellectual articles, so it is not possible to teach a human to play chess like stockfish."

How do you know? I'm not a fan of statements that aren't backed up by data.

[towforce]

"My concern: just as it's not possible to teach a housefly to write intellectual articles, so it is not possible to teach a human to play chess like stockfish."

How do you know? I'm not a fan of statements that aren't backed up by data.

I assume you don't want me to prove that houseflies cannot write intellectual articles.

The highest rating ever achieved by a human was 2882, which was achieved by Magnus Carlsen in 2015. Stockfish has a rating of nearly 3600. That's a gap of around 700 points (and growing).

Magnus's peak rating is only 30 points higher than Gary Kasparov's peak rating, obtained in 1999, so the benefit of training with strong chess computers hasn't boosted peak human elo very much yet.

Obviously this data is over-simplified and one could easily find fault with it. It's also entirely possible that your software will enable people to improve their strength more quickly than other training programmes do - no evidence against that. But I don't see it boosting peak human elo to the level of Stockfish I'm afraid.

[jefk]

[quote]I don't see it boosting peak human elo to the level of Stockfish I'm afraid.[/quote]

you maybe forget another human chess genius coming.
SF still has holes in it's eval, maybe someone as the late
R.J. Fisher living in today world and then knowing about computer
chess and having support of a team would have a chance
against the engines (not that he would beat them often,
that's for sure); my2cnts

[jryandx]

I got the idea for the app when reading the book "Game Changer" about Alpha Zero. The initial plan was to use Alpha Zero, but then I learned about Stockfish + NNUE. It seems like with the introduction of neural networks, these engines play in a more human comprehensible way and that's what I'm hoping to take advantage of. I checked Stockfish NNUE evaluations against the positions in that book and it did not make the same mistakes as the traditional Stockfish, rather it played like Alpha Zero.

"I was quite happy to see AlphaZero’s dynamic, sacrificial style. Not only because it mirrored my own, but because it could play like this, and win, against a fearsomely accurate elite program. Instead of grinding chess into dust with tedious and incomprehensible maneuvering, AlphaZero prefers piece activity and attacking chances...........One of the great Capablanca’s nicknames was ‘the Chess Machine’, reflecting his invincible consistency. Perhaps in the AlphaZero future, someone who plays ‘like a chess machine’ will be thought of as more of an Alekhine, with dazzling sacrifices and a fondness for unbalanced positions!" - Gary Kasparov, Game Changer introduction.

[jefk]

maybe call it a 'beta' ?

i'll have a look but i'm more an oldfashioned destkop guy
and my eyes are too old to play with my smartphone
(and i won't buy a tablet, unless necessary)

[jryandx]

It's a web application, and isn't yet supported for mobile. You can use it on your desktop. I've only actually tested it on Google Chrome.