I’ll check when I get home, for more accuracy, could you please tell me the capture count and Komi used?
Uh… well I pulled it from a random live dan game on OGS. Komi was probably 6.5 and the board-state doesn’t show any sequences that would be the result of a capture.
I’m not so much interested in the actual count, or point difference, just what the bot thinks is “solid” vs “potential”.
1 Like
@Kaworu_Nagisa
From the original position, KataGo thinks R14 is the best move on the board for white, taking white’s marginal lead from 0.3 points to 0.6 points 
here is the territory estimation after that move is played:
(S16 is dark because that’s where she expects black to respond)
If white plays at S15 instead, the game turns slightly in black’s favour (with an estimated 1.3 point lead)
expected response is at R16 and here is the new estimated territory graph:
note: all numbers listed in this post assumes 0 captures to either player and 7.5 points komi to white, recommended moves and score estimations may change if either of those conditions change.
2 Likes
Cool. I take it there’s no real difference after the S15 and R14 sequences are played out? (white played s15 in the game).
I’m guessing this is a no, but does KataGo have a heatmap for where the “most contested” areas are? I guess it would kind of be the inverse of the images you posted, but obviously the centre isn’t hot. Would be cool to see something like game theory temperature visualized in go.
I think I’ve seen something like that for one of the bots, but KataGo was specifically designed with accurate score estimation in mind, and so that’s why we have these territory graphs instead.
Actually they’re pretty different…
After R14 white gets the outside and a 3 point lead, but it’s gote
After S15 white gets the inside, it’s a completely even game, but white gets sente
Er sorry, I meant for the “winrate”, since KataGo says it’s only a ~2 point difference.
1 Like
As you can see in the pictures;
after R14 white has a win rate of 59.3 (inverse of black’s 40.7)
after S15 white has a win rate of 51.9
so, similar… certainly both are playable for amateurs (even pros have more variance than that during a game)
1 Like
Ah, but the initial moves were…
white R14 -> white ahead by 0.6 points
white S15 -> black ahead by 1.3 points
But the end result of the sequence is a difference of ~10%, which some could say is pretty large. So it seems like “winrate” (below a reasonable threshold, at least 10% it seems) isn’t a good way to really determine if one move is really better than the other by itself. 2 points is pretty meaningless in the mid-game for amateurs. Probably not for bots though.
(I’m kinda just pointing this out for cross referencing the ‘joseki by winrate’ thing…)
Keep in mind, the “finished” positions are both ~10 moves deeper in their respective trees, giving the bot deeper insight into the unfolding board and its possibilites, and they both have something in the order of 40,000 more variations read, since I made sure each suggested move had at least 2k variation depth before playing it and reading the next.
I’m sure both of those factors (as well as likely my aged hardware) account for the 10% difference, which as you say, is significant for bot vs bot, but is a regular occurrence in even the best human games.
Do you know how bots judge the winrate of a move?
This may be slightly different for each bot, but to the best of my knowledge, I believe the win rate specifically is separate from move recommendation, and is calculated by playing totally random moves from the known position to the end of the game, and noting how many of those random games were won by either colour… though I’ve not been deeply involved in this part of the community for at least a year and could be remembering that poorly.
For LZ like bots (neuronal nets) the net outputs the winrate for a given board position, as well as a map of where the NN expects the next move. The AI explores then multiple paths using both to decide the probable paths the game could take.
The shown winrate is then some kind of average over all explored board positions.
3 Likes