Also, you can play with Leela here: kontrachess.com
Exhibition match (ultrabullet):
Please mind that the version on play.lczero.org is much weaker (~500 ELO) than the one playing in TCEC, even on hard mode.
so leela doesnt study books huh? but plays herself to improve? i think i should also stop studying books and start playing hundreds of games on myself to see if i improve
@nikolajtesla
Actually, I'm quite familiar with that paper and with this entire field, even the details.
Indeed, I should have mentioned that it came with some caveats regarding how it goes up against Stockfish and the specific training systems used, but the point remains: 4 hours is still 4 hours. There are always caveats, and the paper discusses them.
The main point is: regardless of caveats, which are important, these results are still real. Stockfish got beat 20 wins to 4 losses, and many people find that sensational, even if you don't.
Actually, I'm quite familiar with that paper and with this entire field, even the details.
Indeed, I should have mentioned that it came with some caveats regarding how it goes up against Stockfish and the specific training systems used, but the point remains: 4 hours is still 4 hours. There are always caveats, and the paper discusses them.
The main point is: regardless of caveats, which are important, these results are still real. Stockfish got beat 20 wins to 4 losses, and many people find that sensational, even if you don't.
@TheWidow
Exactly, it's very cool, and it's known as reinforcement learning. It is its own field, but you may be interested in an application: a major update to the original AphaGo work can be found here:
www.nature.com/articles/nature24270
The new work uses reinforcement learning exclusively, without the major supervised machine learning component that the original work had.
Exactly, it's very cool, and it's known as reinforcement learning. It is its own field, but you may be interested in an application: a major update to the original AphaGo work can be found here:
www.nature.com/articles/nature24270
The new work uses reinforcement learning exclusively, without the major supervised machine learning component that the original work had.
@asymptoticfreedom I also work in this field and I constantly see huge amount of misinterpretation, over-excitement and hype (this is what led to AI-winter) . The caveats are pretty substantial and by the way please check about 4 hours. The system which played against stockfish was trained 9 hours.
Results are not that real. There is no code to reproduce the results. There are high level ideas (but they were available even when alpha-go was released, so most of them are not that new). And for most of the people who work in the field it is known that it is extremely hard (and a lot of time impossible) to reproduce the paper just from these high-level ideas. And the whole idea of science is reproducibility.
A lot of people find many things sensational. But as I told before, the current version of stockfish will win that version of stockfish with (20, 75, 4). On the same machine under the same settings and everyone can reproduce it. Also the newer version of stockfish was created by a small number of volunteers probably without any salary. So an unreproducible (28, 72, 0) under completely different hardware (way more powerful) under completely arbitrary time control (why 1 min per move) in completely non-reproducible settings with a reasonably big team of very highly paid full-time employees, trained on huge computer cluster not available to anyone is not really that impressive.
The impressive part was alpha-zero, because the AI-experts didn't expect that this can be possible. But when alpha-zero was created, there were no doubts that the same thing can be done to chess.
Happy to discuss this privately (to not divert this thread).
Results are not that real. There is no code to reproduce the results. There are high level ideas (but they were available even when alpha-go was released, so most of them are not that new). And for most of the people who work in the field it is known that it is extremely hard (and a lot of time impossible) to reproduce the paper just from these high-level ideas. And the whole idea of science is reproducibility.
A lot of people find many things sensational. But as I told before, the current version of stockfish will win that version of stockfish with (20, 75, 4). On the same machine under the same settings and everyone can reproduce it. Also the newer version of stockfish was created by a small number of volunteers probably without any salary. So an unreproducible (28, 72, 0) under completely different hardware (way more powerful) under completely arbitrary time control (why 1 min per move) in completely non-reproducible settings with a reasonably big team of very highly paid full-time employees, trained on huge computer cluster not available to anyone is not really that impressive.
The impressive part was alpha-zero, because the AI-experts didn't expect that this can be possible. But when alpha-zero was created, there were no doubts that the same thing can be done to chess.
Happy to discuss this privately (to not divert this thread).
@asymptoticfreedom thanks do you know any idea about machine learning? am interested to learn any idea where to start please?
This topic has been archived and can no longer be replied to.