Hi Uberdude,
Uberdude wrote:This info is basically the raw data behind the win rate delta graph, so if you could somehow dump out the data for the whole game as text/file somewhere that'd be super useful, e.g. a CSV (I added a few bonus columns) like
Move number,Colour,Bot move,Bot winrate,Game move,Game winrate,Bot choice,Policy prob
20,W,h17,54.23,j17,53.5,2,5.12
21,B,h18,46.5,h18,46.5,1,45.32
As AlesCieply mentioned above, yes I can certainly do that
One question: What is "Bot Choice"? Is that the rank of the player's move among all Leela's proposed moves?
Also, I need to clarify the methodology for calculating the winning probability for the move played by the player.
Let's take an example:
Code: Select all
Move 17, black to play
Leela's best moves:
D15: 54.1%
O16: 53.4%
R6: 48.2%
J3: 47.5%
D5: 47.3%
D8: 46.8%
Player's move:
J3
Here, because J3 is part of Leela's best moves, then we could consider that the win rate for the move played by the player is 47.5%, but maybe it is not the best idea.
Another example:
Code: Select all
Move 17, black to play
Leela's best moves:
D15: 54.1%
O16: 53.4%
R6: 48.2%
J3: 47.5%
D5: 47.3%
D8: 46.8%
Player's move:
K4
Now, K4 is not part of Leela's best moves, so how to evaluate K4 win rate? GRP solves this issue by having Leela play K4, then asking Leela for white next best move. If Leela says that white next best move is N3 with a 61.1% win rate for white, then GRP deduces that the win rate for black K4 was 38.9% (100% - 61.1%).
In fact, even if the player's move was part of Leela's best move, I use the opposite probability of the next move to estimate the player's move win rate. It allows Leela to use all her thinking time/power on the player move this way, not only a fraction of her playout, so the win rate is more accurate.
To continue with the example above:
Code: Select all
Move 17, black to play
Leela's best moves:
D15: 54.1%
O16: 53.4%
R6: 48.2%
J3: 47.5%
D5: 47.3%
D8: 46.8%
Player's move:
D15
So here, the player's move and Leela best moves are the same. So delta will be 0
But if I now ask Leela for the best answer to black 17 at D15, maybe I will get:
Code: Select all
Move 18, white to play
Leela's best moves:
D15: 45.7%
D8: 43.8%
R6: 42.3%
D5: 42.1%
Here we can see that Leela's best answer is D15/45.7% which indicates that now, Leela re-evaluated black 17 at D15 at a win rate of 54.3% (100% - 45.7%) instead of the former estimation of 54.1%
This is because for this turn, all Leela's thinking power was focused on the continuation after black 17 at D15, so this new evaluation should be more accurate.
Because Leela's best move and the player best move are the same, then the delta is zero regardless of the probability, so I won't impact the histogram. But I just wanted to make the methodology clear.
The CSV can also incorporate both winrate values, it's not an issue.
One last precision, GRP does not rank Leela's moves based on probabilities, the moves have the same ranks as provided by Leela. So low ranked moves could have higher win rate than Leela's top move. The alphabetical order used by GRP truly reflects Leela's best moves order. I guess we are all ok with that?
) and I would like your opinions on it.