Opening Choice by Leela

Leelenstein 14.3 is definately an improved net! LS is currently no 1, in long time control testing as well.

In term of elo ratings, Leelenstein is definately higher than Leela. (because Leelenstein has trade penalty which avoid trading pieces therby giving higher chance of winning vs lower rated engines, but it is slightly worse against same level engine or higher rated engine).

In term of 1v1, Leela is always better than Leelenstein. 

But Leelestein vs Stockfish stein wins by a large margin, while Lc0 does not do the same against Stockfish (based on the table I posted above)

LS 14.3: 3644 513 (+ 311, = 0.202), 60.6% against Stockfish 200418, while Lc0 0.25. 1 t40-1541: 3583 516 (+ 269, = 0, -247), 52.1% against the same version of Stockfish. If Lc0 is stronger than LS 14.3 the result against Stockfish should be the opposite, don't you think?

Good point. What is head to head score of LS vs Lc0? 

@Bong711 , 4. e5 is Leela's choice

Sorry for the delay in answering my friend, but I don't have this information

@drmrboss Thanks for the tip.

Thanks for sharing. Is it also possible to ask what are Leela's top choices against the popular openings from White (i.e. what does it play against 1.d4, 1.e4)? 

Thanks ahead!

 glad to see people care about me ;P

Her choice against 

1. e4 is e5, ( Berlin Defence)

2. d4 is Nf6 ( QGD or Nimzo depends on white's move order)

I will run Leela when I reach home.

Thanks again. It is an interesting topic, even if I am not going to change my opening repertoire, I'm curious what the best chess engines have to 'say' on this. If you don't mind, can you also share something like a comparison ? Meaning to say, not just 1.e4 e5, but what are the top few options Leela chooses against 1.e4. so sorry to make you "work" more than you may have originally intended 😁

Thanks again!

This is her initial choice as black for 1. e4

This is the variation for 1. e4 e5 continuation of her choices 

Her choices against 1. d4 

Her choice as white, network , T60 4229

Thank you very much for all your sharing. If I may just confirm, the percentages, are they reflective of how many points it scores (as White) out of 100 games (as a percentage) right?

It's interesting in that they are actually pretty close to me, much closer than what some seem to make out of it.

Yes, Leela do 1000 simulation games in playing against herself in any given position. 

And the score is displayed as WDL (win draw loss ),marked with red pen,  for example, white will get

912 win 

54 draw

34 loss 

Analysis of that above position. That WDL will be translated into percentile % or +1, +2 scores to you by GUI.

Which lines she will search is determinted by policy network that she gained statistics from her training.  For example, in previous diagram of initial position, she will deeply focus on 1. d4 move  searching 45.7 k nodes with policy 19% , however, a horrible move like 1. g4 ? she will search like 48 nodes with policy 1% ) 

Thank you very much for the sharing and explanation. I really appreciate it!

Thanks for the share!

Dutch , pretty horrible, 39% by Leela.

You can see her playout 1000 games from current position, WDL score, she had 225 win, 335 draw and 440 loss for black.