Does Leela just calculate from move one on its own? Or r openings added?
She was forced to play randomly from move 1.
She learn the results of games by statistics. ( e.g 1. e4 has higher chance of winning than 1. e3. And then she choose 1. e4 and then second move and third move and so on.
Leela team will start training for a chess 960 initial positions according to volunteer's request.
For those who are interested to know the opening choice by current world best chess player Leela Zero( Both Stockfish and Leela are best two programs with equal strength in comparable hardwares but Leela is better in opening and SF is better in ending).
T60, best choice in order of popularity,
1. d4
2. Nf3
3. e4
Does leela use endgame tablebases?
Thanks for the information regarding leela choice of openings. What about Alpha Zero choice of openings?
Thanks for the information regarding leela choice of openings. What about Alpha Zero choice of openings?
It is the same. 1. d4 is A0 choice. ( based on a few games publicy available).
However, A0 opening choice should be extremely similar to T40 Leela training in 2019 in which the training was done very similar to A0 training.
I'm assuming if leela calculates from move 1 then no tablebase either, all calculation. Correct?
There is no tablebase resources to search during training. However, the games were ended whenever they reached to tablebase positions.
Master, what does "ordo elo" mean on this Leela network page, https://training.lczero.org/networks/1?
Master, what does "ordo elo" mean on this Leela network page, https://training.lczero.org/networks/1?
Previously they used "elo". After a certain amount of training, newer leela play older Leela in a match. The newer leela usually beats the older one and those "self elo" become quite exaggerated from reality.
I understood. In this case, the ordo elo of network 42850 is 3329, while the ordo elo of the most recent network, 63450, is 2969. Does this mean that the 42850 network is still much stronger than the 63450?
I understood. In this case, the ordo elo of network 42850 is 3329, while the ordo elo of the most recent network, 63450, is 2969. Does this mean that the 42850 network is still much stronger than the 63450?
No! You have to compare elo in same test series.
42×××× is Test 40 series.
63xxxxxx is Test 60 series.
Test 60 gets training data T40, not from total beginning.
Let us say when T40 start with 1500 elo( as example), in which Leela is totally beginner but when T60 start with 1500 , she is already Grandmaster Level.
T60 is totally underrated compared to T40!
So, the T60 is stronger than the T40, is that it?
What about a comparison between the T60 and the SV-20b-t40-1541 network?
So, the T60 is stronger than the T40, is that it?
What about a comparison between the T60 and the SV-20b-t40-1541 network?
SV > T60> T40 only if you have top end hardwares.
SV net were used in TCEC and CCC as they have 4x 2080 ti graphic cards.
If you have just old graphic cards like 1060 gtx, and play in bullet/ blitz T40 should be stronger. SV nets are crazy big nets ( compared to other nets and Alpha Zero)
If you have more terrible graphic cards, smaller networks.T50, T70 will be stronger.
Thank you very much for the explanation, I really understood things I didn't know!
And about the T 70, what can you tell us? If the T60 is still in development, why does the T 70 already exist? Is the T 70 stronger than the T60?
Thank you very much for the explanation, I really understood things I didn't know!
And about the T 70, what can you tell us? If the T60 is still in development, why does the T 70 already exist? Is the T 70 stronger than the T60?
T70 is smaller nets compared to T60. They have plan to make bigger nets than SV in next year which is T80. ( NN size may be approx 15 times bigger than T40).
They make smaller nets like T70, to test their idea in training. ( smaller nets learn more quickly than larger nets, so they dont need to spend too much resources in training)
In term of playing strength, larger nets are usually stronger than smaller nets as long as they get access to quality graphic cards and long time control. In bullet time control with horrible cards or without cards ( cpu only), smaller nets should be stronger than larger nets.
With analysis time of 30 seconds and using one 2080 RTX card, which network is better to use? One of the big ones, like the T60 or the smaller ones?
With analysis time of 30 seconds and using one 2080 RTX card, which network is better to use? One of the big ones, like the T60 or the smaller ones?
T60 I guess, but you will hardly notice the difference.( both are very strong)
SV should be better later but SV training are not in full strength. For more recent results, you have to check in test results in leela discords.
Ok!
@501n323, your answer to English opening! Analysis by SV net, for 1 min, on my 1060 gtx gpu.
leela think 1. ... e5 gives her the best score against 1. c4
Ya but Leela's knight sacrifice in her Ruy Lopez was phenomenal
For those who are interested to know the opening choice by current world best chess player Leela Zero( Both Stockfish and Leela are best two programs with equal strength in comparable hardwares but Leela is better in opening and SF is better in ending).
T60, best choice in order of popularity,
1. d4
2. Nf3
3. e4
Well, according to the finals of CCC 13, Lc0 is stronger than Stockfish since he is winning the match. On the other hand, in the finals of TCEC 17, the duel is tied so far.
On two sites that measure engine strength, http://www.fastgm.de/16-60-0.6.html and https://www.sp-cc.de/nn-longtime-testing.htm, Leelenstein version 14 is the strongest program today, but at CCC 13 it was in third place.
Again seeing Leelestein's results, this program easily wins against Stockfish dev (version 200418), the results are here: https://www.sp-cc.de/nn-vs-sf-testing.htm.
1 Lc0 0.24.1 LS 14.3 : 3644 513 (+311,= 0,-202), 60.6 %
2 Lc0 0.24.1 LS 14.2 : 3633 520 (+308,= 0,-212), 59.2 %
3 Lc0 0.25.1 t40-1541 : 3583 516 (+269,= 0,-247), 52.1 %
4 Lc0 0.25.1 t60-3010 : 3582 514 (+267,= 0,-247), 51.9 %
** Stockfish 200418 ************** : 3568 SPCC-Elo *******************
5 Allie 0.6 LS 14.3 : 3558 519 (+252,= 0,-267), 48.6 %
6 Lc0 0.25.1 42850 : 3556 522 (+252,= 0,-270), 48.3 %
7 Fat Fritz 1.1 : 3530 523 (+233,= 0,-290), 44.6 %
8 Lc0 0.25.1 63305 : 3530 512 (+228,= 0,-284), 44.5 %
9 Lc0 0.25.1 32930 : 3483 515 (+196,= 0,-319), 38.1 %
10 Lc0 0.25.1 11260 : 3408 521 (+149,= 0,-372), 28.6 %
For those who are interested to know the opening choice by current world best chess player Leela Zero( Both Stockfish and Leela are best two programs with equal strength in comparable hardwares but Leela is better in opening and SF is better in ending).
T60, best choice in order of popularity,
1. d4
2. Nf3
3. e4
Well, according to the finals of CCC 13, Lc0 is stronger than Stockfish since he is winning the match. On the other hand, in the finals of TCEC 17, the duel is tied so far.
On two sites that measure engine strength, http://www.fastgm.de/16-60-0.6.html and https://www.sp-cc.de/nn-longtime-testing.htm, Leelenstein version 14 is the strongest program today, but at CCC 13 it was in third place.
Because Leelenstein has trade penalty ( a kind of contempt) that means Leelenstein try to avoid trading pieces and choose a slightly worse position.
Keeping pieces will give you a good chance of beating a weaker opponent. As rating pool is made up of majority of weaker engines, it helps in better rating. However, it wont help anything and even play worse against 1 vs 1 in same level of opponent or stronger opponent.
Leela is better than Leelenstein in 1 v 1 though.
Stockfish has contempt C=24 in default, which give SF approx +30 to +50 elo by just beating better against weaker engines( although actual game play quality is worse). SF programmers agreed it is a bad idea against Leela.