Hai wrote: ↑Wed Feb 12, 2020 12:00 am
After only 3 months of training it should reach 85 of 111 points.
The short Training time proves the new net strongly based on the earlier Lc0 training and because of this the weakpoints of the "new" net are the same as they are in Lc0 nets. So the new net in the reality an old net with some modification only. We can expect really improvement only from a fully new net I think.
I don't think so. You could also say Stockfish 11 in the reality is an old Engine with some modification only.
This big net has much more parameters to tune than the smaller nets. That's why it can be improved much longer before it reached his limit. That's also a reason why it will avoid most of the weakpoints we know from smaller nets.
I think 1 month, maybe + a half month, and not 3 months, will be enough to reach 85 of 111 points.
Looks like the 40x512 net needs only 2 days of training to solve one more of these extrem hard positions .
Hai wrote: ↑Wed Feb 12, 2020 5:08 pm
ERET test suite.
40x512 (320)
64 GB RAM and a limit of 300 seconds per move: 88 of 111 points.
Again a big improvement .
It is possible that the fourth 40x512 net will be better than 30x384 T40-2036, which only got 93 of 111 points .
Hai wrote: ↑Wed Feb 12, 2020 12:00 am
After only 3 months of training it should reach 85 of 111 points.
The short Training time proves the new net strongly based on the earlier Lc0 training and because of this the weakpoints of the "new" net are the same as they are in Lc0 nets. So the new net in the reality an old net with some modification only. We can expect really improvement only from a fully new net I think.
I don't think so. You could also say Stockfish 11 in the reality is an old Engine with some modification only.
This big net has much more parameters to tune than the smaller nets. That's why it can be improved much longer before it reached his limit. That's also a reason why it will avoid most of the weakpoints we know from smaller nets.
I think 1 month, maybe + a half month, and not 3 months, will be enough to reach 85 of 111 points.
Looks like the 40x512 net needs only 2 days of training to solve one more of these extrem hard positions .
Your simile is very primitive and very bad.
The power (~the knowlewdge) of a net is depend on very high measure of the self playing process. If somebody shorten the self playing process for.e.g. using other net the knowledge of the new net will not be much more higher than the knowledge of the base net. It is very obvious. The effect of those much more parameters will be really effective if they works during the self learning process also.
Hai wrote: ↑Wed Feb 12, 2020 12:00 am
After only 3 months of training it should reach 85 of 111 points.
The short Training time proves the new net strongly based on the earlier Lc0 training and because of this the weakpoints of the "new" net are the same as they are in Lc0 nets. So the new net in the reality an old net with some modification only. We can expect really improvement only from a fully new net I think.
I don't think so. You could also say Stockfish 11 in the reality is an old Engine with some modification only.
This big net has much more parameters to tune than the smaller nets. That's why it can be improved much longer before it reached his limit. That's also a reason why it will avoid most of the weakpoints we know from smaller nets.
I think 1 month, maybe + a half month, and not 3 months, will be enough to reach 85 of 111 points.
Looks like the 40x512 net needs only 2 days of training to solve one more of these extrem hard positions .
Your simile is very primitive and very bad.
The power (~the knowlewdge) of a net is depend on very high measure of the self playing process. If somebody shorten the self learning process for.e.g. using other net the knowledge of the new net will not be much more higher than the knowledge of the base net. It is very obvious. The effect of those much more parameters will be really effective if they works during the self learning process also.
Hai wrote: ↑Sat Jan 25, 2020 10:33 pm
LC0 Sergio 40x512 (175)
Hard Talkchess 2020 set
First result with 8 GB RAM and a limit of 15 seconds per move: 43 of 213 points = 20.1%.
But it is better in the ERET test suite.
LC0 Sergio 40x512 (320)
Hard Talkchess 2020 set
8 GB RAM and a limit of 15 seconds per move: 59 of 213 points = 27.7%.
From second (175) to third (320) net, from 43 to 59 solved positions means an improvement of +37.20%!!
We need some engine matches.
I think the fourth net will be ready in 2 weeks and much stronger .
LC0 Sergio 40x512 (320)
Hard Talkchess 2020 set
64 GB RAM and a limit of 300 seconds per move: 82 of 213 points = 38.4%.
.
