corres wrote: ↑Thu May 02, 2019 1:31 pm
I do not know who was addressed by you.
It's background for all of talkchess for my background with regards to computer chess, the topic here. Followed by my take on the newer engines. Now you have some bold claims that the engines are failures or something, the newer ones, but i'm just not convinced you have any real background beyond what I have. The point made was there was no foundation to it but instead opinion. I"d be interested in hearing more arguments the cite sources or give concrete examples than conjecture that the engines are failures as I find it unconvincing.
corres wrote: ↑Thu May 02, 2019 1:31 pm
Maybe you would be in connection to other chess programmer on the sub-site of "Programming and Technical Discussions", I think.
hgm wrote: ↑Wed May 01, 2019 2:11 pm
More important is that the NN-based engines aren't really very good at Chess in general: they seem to be quite inferior for analyzing arbitrary positions. The one thing they are good at is playing games from the opening position, because that allows them to avoid the large fraction of positions where they would suck. But who wants that?
Building an opening analysis tool from openings to middle game from such an engine would not be a bad idea at all. You start in a pure examine mode start position and see what the engine thinks is the best move. Then try your own move then try an engine move. After you exhaust the sequence user backs up. tries something else. It has effectively a better grasp of the best moves here than stockfish. This would be far from a worthless tool.
hgm wrote: ↑Wed May 01, 2019 2:11 pm
More important is that the NN-based engines aren't really very good at Chess in general: they seem to be quite inferior for analyzing arbitrary positions. The one thing they are good at is playing games from the opening position, because that allows them to avoid the large fraction of positions where they would suck. But who wants that?
Obviously the inferiority of NN engines for analyzing (that is in arbitrary position) depends on the dimension and structure of NN. In the case of an NN engine with bigger and better structured NN this issue is smaller.
We can hope the development of NN engine will reduce this issue.
Adding more layers to the tower, faster processors and bigger and better structure (your words) will get you closer to knowledge in much the same way as adding to the Tower of Babel will get you closer to heaven.
Yes, obviously you can not disregard your political prejudice and your injured vanity even when we dispute strictly technical questions either.
There are some excuse for you that you are not the only one on this site but it is not a wise thing mainly from a "smart"...
And it is not a hahaha but a sad thing.
Que, political prejudice? Do your politics on CTF, here it is a technical topic.
Look, the comments 1 and 2 above are assertions made without base, failing also on Tower of Babel theory. It’s not that you are wrong, it’s that the foundations for the assertions are lacking and therefore undiscussable.
Please, try to explain at least to yourself what is the connection between the Babel Tower and a working Neural Network.
It’s base and it stares at you in plain sight. There’s no reason a repetition will have any more effect.
I am afraid it is the effect of rolling pills also...
Uri Blass wrote: ↑Thu May 02, 2019 5:39 am
I suggest the following game to train about that is not exactly chess because of different scoring rules and the fact that you have probability of 50% not to start from the opening position.
In every game you start from the opening position with probability of 50% and from random position from a big pgn with probability of 50%
You continue games until mate.
In case that you win the number of moves is important and the score is not 1 or 0.
I suggest a score of 0.5+0.5*(0.99^number of moves) for the side that wins the game so there is going to be a big difference between winning in 100 moves and winning in 200 moves and if lc0 continue trolling instead of trying to find the fastest mate it is going to lose matches inspite of having more wins than the opponent.
How would you select the random positions?
The main problem is getting the computer resources to do this training idea. Otherwise we could try lots of different forks.
selecting the random positions is easy
You take a big data of pgn and translate it to some epd file and choose random positions from the file.
Uri Blass wrote: ↑Fri May 03, 2019 7:10 am
selecting the random positions is easy
You take a big data of pgn and translate it to some epd file and choose random positions from the file.
I agree that computer resources is a problem.
I meant you have to decide which random positions go in the file. I agree once you've decided on the total set it's easy to choose randomly from it.
Uri Blass wrote: ↑Fri May 03, 2019 7:10 am
selecting the random positions is easy
You take a big data of pgn and translate it to some epd file and choose random positions from the file.
I agree that computer resources is a problem.
I meant you have to decide which random positions go in the file. I agree once you've decided on the total set it's easy to choose randomly from it.
The random positions that I suggest to go in the file are simply from pgn of the games that we practically have(humans against humans or computer against computer).
Alternatively you can decide to take the random position from random games of computer that play random moves against itself if you want a good tool to analyze also positions that do not happen practically in games.