TalkChess.com

matthewlai wrote: ↑Fri Dec 07, 2018 2:15 am

Daniel Shawul wrote: ↑Fri Dec 07, 2018 12:35 am While I sympathize with that statement, releasing A0 source code and networks for anyone to test sounds better.
Many will not be satisfied with in-house testing with supposedly fair conditions.

That would be good, but it would also be a lot of work for us (AZ is tightly-coupled with DM and Google's systems) for not really much value to the scientific community. We feel that it's our ideas and algorithms that are important, not our implementation. That's why we have published all the algorithms we developed in detail, with almost-runnable pseudo-code, so that they can be replicated easily.

jp wrote: ↑Fri Dec 07, 2018 2:20 am

matthewlai wrote: ↑Fri Dec 07, 2018 2:15 am

Daniel Shawul wrote: ↑Fri Dec 07, 2018 12:35 am While I sympathize with that statement, releasing A0 source code and networks for anyone to test sounds better.
Many will not be satisfied with in-house testing with supposedly fair conditions.

That would be good, but it would also be a lot of work for us (AZ is tightly-coupled with DM and Google's systems) for not really much value to the scientific community. We feel that it's our ideas and algorithms that are important, not our implementation. That's why we have published all the algorithms we developed in detail, with almost-runnable pseudo-code, so that they can be replicated easily.

What were the best values/functions for CPUCT used for playing & training?

class AlphaZeroConfig(object):

  def __init__(self):
    ### Self-Play
    self.num_actors = 5000

    self.num_sampling_moves = 30
    self.max_moves = 512  # for chess and shogi, 722 for Go.
    self.num_simulations = 800

    # Root prior exploration noise.
    self.root_dirichlet_alpha = 0.3  # for chess, 0.03 for Go and 0.15 for shogi.
    self.root_exploration_fraction = 0.25

    # UCB formula
    self.pb_c_base = 19652
    self.pb_c_init = 1.25

    ### Training
    self.training_steps = int(700e3)
    self.checkpoint_interval = int(1e3)
    self.window_size = int(1e6)
    self.batch_size = 4096

    self.weight_decay = 1e-4
    self.momentum = 0.9
    # Schedule for chess and shogi, Go starts at 2e-2 immediately.
    self.learning_rate_schedule = {
        0: 2e-1,
        100e3: 2e-2,
        300e3: 2e-3,
        500e3: 2e-4
    }

noobpwnftw wrote: ↑Fri Dec 07, 2018 1:09 am I have a few questions:
TCEC SuFi used 120' + 15" TC, Division P used 90' + 10" TC, since DM went for mimicking TCEC conditions this time like using a 44-core machine and with the same openings, why mess with time controls(again)?

Also, a year ago the NPS of A0 was 80K, now it is only around 60K, that's about 30% a nerf, what happened, people overclocked the TPUs a year ago?

matthewlai wrote: ↑Fri Dec 07, 2018 2:31 am

noobpwnftw wrote: ↑Fri Dec 07, 2018 1:09 am I have a few questions:
TCEC SuFi used 120' + 15" TC, Division P used 90' + 10" TC, since DM went for mimicking TCEC conditions this time like using a 44-core machine and with the same openings, why mess with time controls(again)?

Also, a year ago the NPS of A0 was 80K, now it is only around 60K, that's about 30% a nerf, what happened, people overclocked the TPUs a year ago?

180' + 15" is the time control for Season 9 Superfinal.

In the preprint the NPS figures I believe were taken from the start position. For the peer-reviewed final paper we looked at all moves to compute those statistics.

clumma wrote: ↑Fri Dec 07, 2018 2:26 am Matthew: Congrats on the phenomenal success.

Can you help me locate the games AZ played against Brainfish? They don't seem to have their own file, and I don't see any identifying info in alphazero_vs_stockfish_all.pgn

Thank you!

-Carl

matthewlai wrote: ↑Fri Dec 07, 2018 3:45 am

clumma wrote: ↑Fri Dec 07, 2018 2:26 am Can you help me locate the games AZ played against Brainfish? They don't seem to have their own file, and I don't see any identifying info in alphazero_vs_stockfish_all.pgn

Only games from the primary evaluation and TCEC openings have been released (no opening books).

IanO wrote: ↑Fri Dec 07, 2018 2:19 am Even more exciting: they released the full game scores of the hundred game matches for all three games, chess, shogi, and go!

https://deepmind.com/research/alphago/a ... resources/

matthewlai wrote: ↑Fri Dec 07, 2018 2:24 am

jp wrote: ↑Fri Dec 07, 2018 2:20 am

matthewlai wrote: ↑Fri Dec 07, 2018 2:15 am

Daniel Shawul wrote: ↑Fri Dec 07, 2018 12:35 am While I sympathize with that statement, releasing A0 source code and networks for anyone to test sounds better.
Many will not be satisfied with in-house testing with supposedly fair conditions.

That would be good, but it would also be a lot of work for us (AZ is tightly-coupled with DM and Google's systems) for not really much value to the scientific community. We feel that it's our ideas and algorithms that are important, not our implementation. That's why we have published all the algorithms we developed in detail, with almost-runnable pseudo-code, so that they can be replicated easily.

What were the best values/functions for CPUCT used for playing & training?

They are all in the pseudo-code in supplementary materials.

Code: Select all

class AlphaZeroConfig(object):

  def __init__(self):
    ### Self-Play
    self.num_actors = 5000

    self.num_sampling_moves = 30
    self.max_moves = 512  # for chess and shogi, 722 for Go.
    self.num_simulations = 800

    # Root prior exploration noise.
    self.root_dirichlet_alpha = 0.3  # for chess, 0.03 for Go and 0.15 for shogi.
    self.root_exploration_fraction = 0.25

    # UCB formula
    self.pb_c_base = 19652
    self.pb_c_init = 1.25

    ### Training
    self.training_steps = int(700e3)
    self.checkpoint_interval = int(1e3)
    self.window_size = int(1e6)
    self.batch_size = 4096

    self.weight_decay = 1e-4
    self.momentum = 0.9
    # Schedule for chess and shogi, Go starts at 2e-2 immediately.
    self.learning_rate_schedule = {
        0: 2e-1,
        100e3: 2e-2,
        300e3: 2e-3,
        500e3: 2e-4
    }