Well, Chess is a toy model. Too toyish to talk seriously about human capacities, but until 25 years ago it was considered "a supreme achievement of human brain" by most AI visionary idiots, while my cat knows better.duncan wrote: ↑Sun Jun 17, 2018 2:21 pmThe above is of course not easy and has only been achieved in the last few months. beating humans with computers with conventional a/b software is easy in the sense, it has been achieved a long time ago.Laskos wrote:
Beating all humans by giving the rules and the goal of Chess and letting it self-play to "understand" what is Chess about and how to play it well, is easy?
Something goes wrong with lc0 since yesterday?
Moderators: hgm, Rebel, chrisw
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Something goes wrong with lc0 since yesterday?
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Something goes wrong with lc0 since yesterday?
WTH, they seem to have restarted all over again, although no ratings are given. Are they sick with their "from scratch forever methodology"?Laskos wrote: ↑Sun Jun 17, 2018 12:54 pmI don't know what you are babbling. They (the new testserver) showed by yesterday morning that in 500,000 fast self-games in half-a-day, one can train a network to beat 99% of humans, and by yesterday evening, after 1,500,000 fast self-games, in a-day-and-a-half, the network can be trained to beat ALL humans. Now one has to see the limits of this small 6x64 network, and to take care of the training parameters like CPUCT, temperature, expansion parameter, etc. The result seems remarkable to me, and I hope they won't start from scratch again, what's the point? Even with now available hardware resources, they can quickly overcome the main branch.Henk wrote: ↑Sun Jun 17, 2018 10:47 am They are just demonstrating that using neural networks is inefficient. What can be wrong with that. Maybe less human work but enormous waste of electricity. If there is no knowledge something has to pay. This only holds for training/tuning/configuring part of course.
For small problems it is do-able. But chess is not that small. Be glad with a local optimum.
-
- Posts: 12038
- Joined: Mon Jul 07, 2008 10:50 pm
Re: Something goes wrong with lc0 since yesterday?
I look at it this way. Every restart means ao has become 100 elo stronger, so worth waiting for. would you like ao on your gpu to be stronger than alphazero ?Laskos wrote: ↑Sun Jun 17, 2018 2:30 pmWTH, they seem to have restarted all over again, although no ratings are given. Are they sick with their "from scratch forever methodology"?Laskos wrote: ↑Sun Jun 17, 2018 12:54 pmI don't know what you are babbling. They (the new testserver) showed by yesterday morning that in 500,000 fast self-games in half-a-day, one can train a network to beat 99% of humans, and by yesterday evening, after 1,500,000 fast self-games, in a-day-and-a-half, the network can be trained to beat ALL humans. Now one has to see the limits of this small 6x64 network, and to take care of the training parameters like CPUCT, temperature, expansion parameter, etc. The result seems remarkable to me, and I hope they won't start from scratch again, what's the point? Even with now available hardware resources, they can quickly overcome the main branch.Henk wrote: ↑Sun Jun 17, 2018 10:47 am They are just demonstrating that using neural networks is inefficient. What can be wrong with that. Maybe less human work but enormous waste of electricity. If there is no knowledge something has to pay. This only holds for training/tuning/configuring part of course.
For small problems it is do-able. But chess is not that small. Be glad with a local optimum.
-
- Posts: 12038
- Joined: Mon Jul 07, 2008 10:50 pm
Re: Something goes wrong with lc0 since yesterday?
25 years ago it was not understood that a 7 move search with positional understanding of 2000 elo = world championship play.Laskos wrote: ↑Sun Jun 17, 2018 2:25 pmWell, Chess is a toy model. Too toyish to talk seriously about human capacities, but until 25 years ago it was considered "a supreme achievement of human brain" by most AI visionary idiots, while my cat knows better.duncan wrote: ↑Sun Jun 17, 2018 2:21 pmThe above is of course not easy and has only been achieved in the last few months. beating humans with computers with conventional a/b software is easy in the sense, it has been achieved a long time ago.Laskos wrote:
Beating all humans by giving the rules and the goal of Chess and letting it self-play to "understand" what is Chess about and how to play it well, is easy?
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Something goes wrong with lc0 since yesterday?
You, as expected, missed the main part of what I said, and that was about the cat.duncan wrote: ↑Sun Jun 17, 2018 2:44 pm25 years ago it was not understood that a 7 move search with positional understanding of 2000 elo = world championship play.
As for their re-running from scratch, the first run was as good as all the following ones (now sixth !!! one), and as training DCNN goes, it is more an alchemy than a science (not my own statement, it is that of leading specialists), their tunings and re-scratchings are pretty much worthless. If they had continued the first run to saturation, say up to 20x256 nets, I bet they would have gone very far by now as strength goes (if they didn't do something dumb along the way).
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Something goes wrong with lc0 since yesterday?
Just checked the latest ID57 before the latest reset. It came at a record for 6x64 net in this gauntlet 64.5/200. I am curious what they will get in 2-2.5 million fast games using 6x64 nets with their brand new Voodoo training settings.Laskos wrote: ↑Sun Jun 17, 2018 2:30 pmWTH, they seem to have restarted all over again, although no ratings are given. Are they sick with their "from scratch forever methodology"?Laskos wrote: ↑Sun Jun 17, 2018 12:54 pmI don't know what you are babbling. They (the new testserver) showed by yesterday morning that in 500,000 fast self-games in half-a-day, one can train a network to beat 99% of humans, and by yesterday evening, after 1,500,000 fast self-games, in a-day-and-a-half, the network can be trained to beat ALL humans. Now one has to see the limits of this small 6x64 network, and to take care of the training parameters like CPUCT, temperature, expansion parameter, etc. The result seems remarkable to me, and I hope they won't start from scratch again, what's the point? Even with now available hardware resources, they can quickly overcome the main branch.Henk wrote: ↑Sun Jun 17, 2018 10:47 am They are just demonstrating that using neural networks is inefficient. What can be wrong with that. Maybe less human work but enormous waste of electricity. If there is no knowledge something has to pay. This only holds for training/tuning/configuring part of course.
For small problems it is do-able. But chess is not that small. Be glad with a local optimum.
-
- Posts: 1796
- Joined: Thu Sep 18, 2008 10:24 pm
Re: Something goes wrong with lc0 since yesterday?
After each reset do they play new games or do they use the old games somehow?Laskos wrote: ↑Sun Jun 17, 2018 7:00 pmJust checked the latest ID57 before the latest reset. It came at a record for 6x64 net in this gauntlet 64.5/200. I am curious what they will get in 2-2.5 million fast games using 6x64 nets with their brand new Voodoo training settings.Laskos wrote: ↑Sun Jun 17, 2018 2:30 pmWTH, they seem to have restarted all over again, although no ratings are given. Are they sick with their "from scratch forever methodology"?Laskos wrote: ↑Sun Jun 17, 2018 12:54 pm
I don't know what you are babbling. They (the new testserver) showed by yesterday morning that in 500,000 fast self-games in half-a-day, one can train a network to beat 99% of humans, and by yesterday evening, after 1,500,000 fast self-games, in a-day-and-a-half, the network can be trained to beat ALL humans. Now one has to see the limits of this small 6x64 network, and to take care of the training parameters like CPUCT, temperature, expansion parameter, etc. The result seems remarkable to me, and I hope they won't start from scratch again, what's the point? Even with now available hardware resources, they can quickly overcome the main branch.
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Something goes wrong with lc0 since yesterday?
New games. Just the training parameters Voodoo tweakings. That is AFAIK. Basically the old games and nets are for the trash bin.Werewolf wrote: ↑Sun Jun 17, 2018 8:08 pmAfter each reset do they play new games or do they use the old games somehow?
-
- Posts: 177
- Joined: Wed May 23, 2018 9:29 pm
Re: Something goes wrong with lc0 since yesterday?
There was a bug in versions 20180609 and 20180614, which caused it to miss checkmates sometimes (especially mates in 1).
Update is highly recommended.
Update is highly recommended.
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Something goes wrong with lc0 since yesterday?
Thanks, I suppose it affected training too, right?
I checked the new net ID1039 with the old 20180614 engine in the same gauntlet against AB engines at short TC, it came at 47.5/200, which is not bad. It remains to see if by say ID1050 it comes to the levels 5th run of ID48 of 60.0/200 or ID57 of 64.5/200 (which was a record high for any 6x64 nets).