Leela id 216 checkmated
https://lichess.org/9Pb05CPt/white#0
Leela Zero( A Neural Network engine similar to Alpha Zero)
Sort:
drmrboss wrote:
Recent ID 210 CCRL rating =2885, There is plan to increase network size from 10x124 to 15x192 but surprisingly the larger beta network doesnt increase elo.
Maybe it is because the current network has still a lot of «resources» left to become stronger? Or it is simply because the lost ELO due to the higher computation effort is equal to the gained ELO of the better NN.
@Elroch:
As far as I know you transfer data to a larger network with net2net and retrain it with several hundred thousands of the last games. In LZ a larger network was a lot stronger with the same number of games than its smaller predecessor, and in LCZ the jump from 6x64 to 10x128 resulted in an ELO gain too. (The last tests with a larger network are still visible in the strength graph of LZ (the blue dots above the blue line)).
In Go it is known that larger NN are better, and that transferring the knowledge to a larger NN works quite well. There is less experience in chess. Maybe you are right and for the next larger NN in chess new self-play games are necessary before an improvement becomes visible.
Leela vs Houdini 6.0 1-0
Leela on 1060 GTX vs Houdini 6 single core at 30' 10"
Apr 29, 2018
0
#225
How can we taok about ratings of super GM and beating strong engines, whileLeela draws and loses to an NM and even me?
Leela Milestones
A lot of milestones happened to Leela in 2 months, a lot more than a lot of people were expected and the future is brighter than 5 years old distributed network stockfish project( initial stockfish 1.3 is 10 years old now)The three challenging hurdles of project are
1. To reach the level of A0 (estimated 3300 rating in GTX 1060)
2. To reach the level of latest stockfish on GTX 1060 ( estimated 3550+ rating on 8 cores desktop)
3. To surpass the rating of stockfish on GTX 1060.(getting 3600+)
Achievement milestones
1. Finished 10 millions games,
2. Reached 2800+, super GM level, on 1060 GTX.
3. Gofundme got €5000+ donation for the project.
Future Leela
1. Network will soon be expanded into 15x192 and may further expand into 20x256
( exactly as A0)
2. cuDNN or tensorflow implementation will increase the speed/ elo of Leela on NVIDIA cards (?50% ?100% ?200%), ( too bad for AMD cards though).
3. syzgy Tablebase
4. Auto resign will speed up training up to 30%.
Good luck Leela.
if you add syzygy to lc0 it will no longer be 0
drmrboss wrote:
Leela Milestones
A lot of milestones happened to Leela in 2 months, a lot more than a lot of people were expected and the future is brighter than 5 years old distributed network stockfish project( initial stockfish 1.3 is 10 years old now)The three challenging hurdles of project are
1. To reach the level of A0 (estimated 3300 rating in GTX 1060)
2. To reach the level of latest stockfish on GTX 1060 ( estimated 3550+ rating on 8 cores desktop)
3. To surpass the rating of stockfish on GTX 1060.(getting 3600+)
Achievement milestones
1. Finished 10 millions games,
2. Reached 2800+, super GM level, on 1060 GTX.
3. Gofundme got €5000+ donation for the project.
Future Leela
1. Network will soon be expanded into 15x192 and may further expand into 20x256
( exactly as A0)
2. cuDNN or tensorflow implementation will increase the speed/ elo of Leela on NVIDIA cards (?50% ?100% ?200%), ( too bad for AMD cards though).
3. syzgy Tablebase
4. Auto resign will speed up training up to 30%.
Good luck Leela.
Nice comment!
Does auto resign really helps? Isn't leela learning the same way while finishing the rest of a lost game?
AIM-AceMove wrote:
How can we taok about ratings of super GM and beating strong engines, whileLeela draws and loses to an NM and even me?
It is weak tactically, but strong in playing positionally. This is the opposite of classic chess engines.
And maybe you can win because LCZ has to few playouts (slow hardware, short thinking time, only 2 threads)?
afbdreds wrote:
drmrboss wrote:
Leela Milestones
A lot of milestones happened to Leela in 2 months, a lot more than a lot of people were expected and the future is brighter than 5 years old distributed network stockfish project( initial stockfish 1.3 is 10 years old now)The three challenging hurdles of project are
1. To reach the level of A0 (estimated 3300 rating in GTX 1060)
2. To reach the level of latest stockfish on GTX 1060 ( estimated 3550+ rating on 8 cores desktop)
3. To surpass the rating of stockfish on GTX 1060.(getting 3600+)
Achievement milestones
1. Finished 10 millions games,
2. Reached 2800+, super GM level, on 1060 GTX.
3. Gofundme got €5000+ donation for the project.
Future Leela
1. Network will soon be expanded into 15x192 and may further expand into 20x256
( exactly as A0)
2. cuDNN or tensorflow implementation will increase the speed/ elo of Leela on NVIDIA cards (?50% ?100% ?200%), ( too bad for AMD cards though).
3. syzgy Tablebase
4. Auto resign will speed up training up to 30%.
Good luck Leela.
Nice comment!
Does auto resign really helps? Isn't leela learning the same way while finishing the rest of a lost game?
First I thought the same. Resigning in match games makes sense, but in self-play games? But now I think, that you cannot learn much from a game in which you lead clearly. Maybe LCZ could learn faster how to win a clearly won end-game position in an optimal way, but this is not much of interest. This is even more true if end-game databases will be used. And remember that the evaluation on which the resignation is decided, comes from the MCTS evaluation and the NN.
It is the same for humans, isn't it? There is nothing to learn from simple to win end-game positions.
Apr 30, 2018
0
#236
When is the network size and hardware upgrade going to happen i assume GTX Titan V has to be the right choice, its made exactly for the purpuse, cards like gtx 1060 1080 ti are gaming cards not for a.i.
And i see number of trained games has decreased.. google colab anything to do with?
ghotrix wrote:
if you add syzygy to lc0 it will no longer be 0
It is for future specifically for TCEC tournments. Syzgy version of Leela is already in leela team. Testing showed that there was "0" elo bonus from that version. Syzgy from minimax engines (stockfish) also get about 5-10 elo only.
Leela start getting draw vs stockfish 9 in rare occassion.
Stockfish 9 (i7-6700K CPU @ 4.00GHz with 32,0 GB Memory)
LCZero v 0.7 Net 217 (GeForce GTX 980 @ 1304MHz)
No books used.
100 games result, Stockfish win 98 games, 2 draws, 0 losses= estimated elo difference= 800.
Larger net,15x192, instantly increase 100 elo in self play .
4 cores stockfish latest version drew vs leela ID 227 ,draw. (overall score in 4 games= 3.5 -0.5 in favour of stockfish, so elo difference is around 300)
A network is just a place to put learned structure, so giving it a bigger network itself causes no improvement until it has adequate training. Computational parameters may need changing as the network size changes.