matthewlai wrote: ↑Fri Dec 07, 2018 12:49 pm
All our values are between 0 and 1, so if the best move has a value of 0.8, we would sample from all moves with values >= 0.79.
The paper says: "At the end of the game, the terminal position sT is scored according to the rules of the game to compute the game outcome z: −1 for a loss, 0 for a draw, and +1 for a win.".
So it's not like that? It's 0 for loss, 1 for win and 0.5 for a draw?
Also paper says that initial Q=0 (and pseudocode also says "if self.visit_count == 0: return 0"). Does it mean that it's initialized to "loss" value?
Whether it's -1 to 1 or 0 to 1 is also important to Cpuct scaling (or C(s) in the latest version of the paper). Do c_base and c_init values assume that Q range is -1..1 or 0..1?
OneTrickPony wrote: ↑Fri Dec 07, 2018 12:13 pm
...What is important is that the games show that there are paths in chess which SF is still unable to understand and the losses are very different than just playing against another alpha-beta engine on 4x or 10x the hardware.
That does not mean we can't improve alpa-beta engines so they could handle those paths efficiently.
Jouni wrote: ↑Fri Dec 07, 2018 2:00 pm
I only looked so far for TCEC opening games. AO seems to be sometimes like patzer and loses in 22 moves to outdated SF .
OneTrickPony wrote: ↑Fri Dec 07, 2018 12:13 pm
I am not convinced the newest SF would win against it. The ELO is calculated against a pool of similar engines. It's not clear if 50 or 100 ELO more against this pool is equal to 50-100 ELO more against an opponent of a different type.
While that's true, Lc0 with the best nets on my powerful GPU and average CPU ("Leela Ratio" of say 2.5) beats heavily SF8, but loses slightly to SF10, from regular openings. Against SF8, it's similar to what happens in this paper. My guess is that this particular "old" A0 in those TCEC conditions is somewhat weaker than SF10.
Lc0 needs a "Leela Ratio" of 2.5 to have similar results to A0 ("Leela Ratio" 1 by definition), so Lc0 (with the best nets) is still lagging pretty significantly behind A0.
In some games it becomes apparent that they are fairly similar in playing style, strengths and weaknesses.
matthewlai wrote: ↑Fri Dec 07, 2018 12:49 pm
All our values are between 0 and 1, so if the best move has a value of 0.8, we would sample from all moves with values >= 0.79.
The paper says: "At the end of the game, the terminal position sT is scored according to the rules of the game to compute the game outcome z: −1 for a loss, 0 for a draw, and +1 for a win.".
So it's not like that? It's 0 for loss, 1 for win and 0.5 for a draw?
Also paper says that initial Q=0 (and pseudocode also says "if self.visit_count == 0: return 0"). Does it mean that it's initialized to "loss" value?
Whether it's -1 to 1 or 0 to 1 is also important to Cpuct scaling (or C(s) in the latest version of the paper). Do c_base and c_init values assume that Q range is -1..1 or 0..1?
All the values in the search are [0, 1]. We store them as [-1, 1] only for network training, to have training targets centered around 0. At play time, when network evaluations come back, we shift them to [0, 1] before doing anything with them.
Yes, all values are initialized to loss value.
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.
Since you're here in these posts just wanted to say congrats on the publication and thanks for the pseudocode.
Is most of the code for alphazero python code, or is the pseudocode transcribed from a different language like C++?
Thanks!
AlphaZero is mostly in C++. Network training code is in Python (Tensorflow). Network inference is through C++ Tensorflow API.
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.
matthewlai wrote: ↑Fri Dec 07, 2018 12:49 pm
During training, we do softmax sampling by visit count up to move 30. There is no value cutoff. Temperature is 1.
This is a rather important difference and will explain a lot about Leela Chess Zero's endgame problems.
Thanks for clarifying some of these things. The 0..1 vs -1..1 range thing is a bit funny. I interpreted the paper as 0..1 initially because that's what older MCTS papers used, then people pointed out that the AZ papers work on a -1..1 range and we changed things. And now it turns out the original version was what AZ had after all.
Yes, all values are initialized to loss value.
Were other settings ever considered, notably 0.5 or parent?