Type of self.output in policy.py about mugo HOT 11 OPEN

I found in previous commits that output used to be, but later due to log_likelihood_cost refactor got deleted.

output = tf.nn.softmax(tf.reshape(h_conv_final, [-1, go.N ** 2]) + b_conv_final)

from mugo.

Hm. Sorry about that - work on this repo is continuing at https://github.com/tensorflow/minigo. I'll update the README.md

from mugo.

from mugo.

@CezCz yeah, thanks for your kind answer,
actually, I fixed the line 88 with
" log_likelihood_cost = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y))" and it could work,
but I could not understand the output of mcts, why it often choose the bigger value even if it is negative ?
I am confused about the result , I would appreciate if you can tell me the reason @CezCz

from mugo.

from mugo.

just like the first picture , the location of white is R4, and I get the value is -7.5, just as the second image,

the another position is Q3,

and its value is 8.5,

and why do the white chose the R4 rather than Q3, the latter value is greater than the former value , I am just very confused about this, perhaps I do not understand the code ,or maybe this is a silly question , but I want strongly to know the reason and I am very grateful to you @CezCz , you are a very kind person and thank you very much ！

from mugo.

from mugo.

Hi , @CezCz
So the next move is chosen just because the algorithm chooses the most visited move , and the value network backpropagated the visit count and the winner predicted, the positive value represents the current player wins this game , the next move is selected is not related to the value of value network, just related to the visit count , right ?

from mugo.

@JoeyQWu
The move that is chosen to be played in the actual game yes. Not to confuse with move chosen within selection phase - this one is chosen based on some sophisticated heuristic with exploration taken into consideration.
You may want to read:
https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/ - nice mcts introduction with examples
http://www.baeldung.com/java-monte-carlo-tree-search - simple monte carlo tree search implementation
https://deepmind.com/documents/119/agz_unformatted_nature.pdf - page 25-27 MCTS implementation within alphago zero (don't be confused about temperature parameter and parent visit count, these are just another parameters to promote exploration during training, but the core is visit count)

from mugo.

@CezCz
okay, I will read more to understand , thank you very much, you are so nice , very grateful to you for your help!

from mugo.

I also wrote http://www.moderndescartes.com/essays/deep_dive_mcts/ recently

from mugo.