Verdict on AI verdicts on old pros

Uberdude · Post by **Uberdude** » Tue Aug 20, 2019 11:10 pm

jann, yes that's a fair point: sometimes bots play the tighter move for less profit but better aji, because they evaluate the bad aji in those cases as being too much of a negative. If I could (facetiously) put words into Kageyama's mouth: he is saying "because 1 leaves bad aji and it is too difficult / I am too lazy to read out possibly downsides of that aji which may come far in the future of many different possibilities, and evaluate the pros and cons of that vs the aji/forcing moves left by the net which surely captures the cutting stones, I am not even going to think about it, and be proud in my ignorance (and because I'm a pro you should trust me)".

Bill, here's a screenshot of Lizzie with LZ #234 after ~100k playouts. You can see the distribution of playouts is very sharp: 100k at 48.8% for 'bad' 1, under 1k for every other move considered and just 9 for a. So the error on its 42.3% (-6%) is large, playing that move and giving 100k playouts gets a winrate of 37.9% (-11%).

: Kageyama net or not.PNG (908.15 KiB) Viewed 16488 times

Here's my commentary/interpretation of LZ 234's opinion on some continuations, to help understand why it thinks net is bad.

Following Kageyama's net. LZ's instinct (low playouts) wants to peep from the outside, and this remains strong 1st choice at 100k playouts, though the expected continuation does change. The early idea was for 4 to be h17 (mine too), but then it changes to j16, a move closer to the top right corner and thus helping a bit more there. To start with it thought black should immediately answer 4 at h16 (36.9% with 30k) giving white the h17 atari, but after several 10ks playouts it decides that black is slightly better to resist and attack on the left side first with 5 (37.9% at 56k) because that's more sente. That white needs to take gote with 8-10 is indeed rather painful (the 'need' is not 100%, but pulling out the 3 stones for 8 is -10% and black gets e12 seal in sente so e14 get strong so cutting is less useful anyway).

Click Here To Show Diagram Code: [go]$$Bc LZ punishing Kageyama's net $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X X X . 1 . . . . . . . . . O . . | $$ | . . O X O O . . 4 , . . . . O , X . . | $$ | . . O O X O 3 2 . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . 0 6 . . . . . . . . . . . . . . . . | $$ | . 8 5 7 . . . . . , . . . . . , . . . | $$ | . 9 . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Black's fix at h17 avoids giving white the atari there, and white then take the big opening move of a shimari (62.5% at 20k). Continuing in top right with q16 is not bad either (62.2 at 1.5k). Black continues there, white extends on the top and we can appreciate the h15 and j16 stones as useful moves bolstering this group, cover fire. They also served a purpose in that white was confident to play 12 as tenuki to shimari and not being scared of a black pincer at m16 instead of s17, as they provide some support in a fight that could follow.

Click Here To Show Diagram Code: [go]$$Bcm11 LZ punishing Kageyama's net contd. $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X X X . X 1 . . 6 . . . . 4 O 3 . | $$ | . . O X O O . . O , . . . . O , X . . | $$ | . . O O X O X O . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . 5 . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . O O . . . . . . . . . . . . . . . . | $$ | . O X X . . . . . , . . . . . , . . . | $$ | . X . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 2 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . 7 . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Given that LZ thought black could interpose the attack on the left with 5, I had a look what LZ thinks of black not immediately peeping the net but defending there. Here's LZ's view showing that it has a strong preference for immediate peep, but the extension and tiger mouth on the left there are the 2nd and 3rd choice moves, though a long way behind at 2k to 112k for the peep (and tiger mouth can amusingly revert with black d10, white h15).

: kage after net.PNG (933.55 KiB) Viewed 16488 times

So if extend (-3.4%), LZ wants to immediately clamp, taking advantage of the fact white extended further trying to be more efficient that tiger mouth to get a good exchange (white resist with e12 extend out is not unplayable, -3%) and then pincers on top. White now tries to get the same peep, but we get a clue why LZ wanted to peep immediately: with m16 in place already black can now reasonably resist with h14 instead of g15. Timing is so important in being able to make beneficial exchanges without the opponent reasonably resisting, and is probably behind bots tending to like exchanging early because, like here, they judge that they can't get it later. As black didn't defend, white attaches at 8 and then 10, and we have ourselves a complicated fight.

Click Here To Show Diagram Code: [go]$$Bc What if extend on left? $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X X X . 1 0 . . . . . . . . O . . | $$ | . . O X O O . . . , 8 5 . . O , X . . | $$ | . . O O X O . 6 . . . 9 . . . . . . . | $$ | . . O X X X X 7 . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . 4 O . . . . . . . . . . . . . . . | $$ | . . . 3 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . , . . . . . , . . . | $$ | . . 2 . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Further insight can be gained from asking what LZ would do if white simply tenukid to shimari after the net (-17%). Now black doesn't just attack with c12 checking extension, but b11 2nd line base attack! Such a move would have been overplay following white h15 peep, because in that case after black steals the base on the left white saves the 3 cutting stones so black f14 group is floating, so white left group being weak has a black weak group to pressure too in exchange.

Click Here To Show Diagram Code: [go]$$Bc What if tenuki? $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X X X . 1 . . . . . . . . . O . . | $$ | . . O X O O . . . , . . . . O , X . . | $$ | . . O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . 5 6 O . . . . . . . . . . . . . . . | $$ | . 3 4 . . . . . . . . . . . . . . . . | $$ | . . 7 8 . . . . . , . . . . . , . . . | $$ | . . 9 0 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . a . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 2 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

So that's how LZ as white (and then black) would punish Kageyama's net. How about what happens with the bad aji with 'greedy' (or efficient) outside move? LZ does immediately want to turn. My first idea was for black to jump down to 2nd line h18, but LZ thinks j17 better (48.7% at 71k vs 44% at 60 for h18). So why is jump worse? Because even though it captures the white stones white gets a nice kikashi at j17. Actually 5 here is slack, black can play f18, white wedges at g18, black then connects and it's complicated, but for the sake of making a simple settled position to compare black ends this at 42%, compared to 38% for the net. White getting b17 in sente is some nice extra eyespace for that group and j17 peep is some use, but not as useful as the h15 and j16 forcing moves against the net. Plus remember this is black playing my bad move of h18 instead of LZ's favoured j17, though I imagine Kageyama might have glance-read this kind of line as an example of the 'bad aji not working yet' even if he professes to have not given it a second thought.

Click Here To Show Diagram Code: [go]$$Bc Greedy then jump $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . 7 . . . 9 . 3 . . . . . . . . . . . | $$ | . 6 X X X 2 . 5 4 . . . . . . . O . . | $$ | 8 . O X O O . 1 . , . . . . O , X . . | $$ | . . O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

So how is LZ's kosumi better than my jump? If white still dies then you got more points and stopped white getting j17 peep, so that would obviously be better. But white isn't dead now, but ko-ish. This is disquieting: a complicated sequence that is hard to judge and the kind of thing soba go tris to avoid, but bots relish.

Click Here To Show Diagram Code: [go]$$Bc Greedy then kosumi? $$ +---------------------------------------+ $$ | . . . . 9 0 . . . . . . . . . . . . . | $$ | . . . 5 4 . 6 . . . . . . . . . . . . | $$ | . . X X X 2 . . 3 . . . . . . . O . . | $$ | . 7 O X O O . 1 . , . . . . O , X . . | $$ | . . O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . 8 . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Here's how LZ sees the ko fight could continue. Ignoring 12 is -20%, it's almost a picnic ko for white. 13 extend instead of corner hane seems to be minimizing white's big threats because black can't afford to lose the ko, but hane is <1% worse in winrate and might revert anyway. Black concedes the ko with connect, so white gets some profit on left side and the top group isn't dead yet.

Click Here To Show Diagram Code: [go]$$Bcm11 Don't play go if you are scared of ko (14 retakes) $$ +---------------------------------------+ $$ | . . . 5 X O . . . . . . . . . . . . . | $$ | . . . X O 1 O . . . . . . . . . . . . | $$ | . 7 X X X O . . X . . . . . . . O . . | $$ | . X O X O O . X . , . . . . O , X . . | $$ | . 6 O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 0 . . . . . . . . . . . . . . . . | $$ | . . 8 3 . . . . . . . . . . . . . . . | $$ | . . 2 X . . . . . , . . . . . , . . . | $$ | . . 9 . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

The lower left resolves, and white j18 h17 to settle top group could be coming soon. Black is keeping the best win% he can get of 48-49% in these lines.

Click Here To Show Diagram Code: [go]$$Bcm21 ko line cont. $$ +---------------------------------------+ $$ | . . . X X O . . . . . . . . . . . . . | $$ | . . . X O . O . . . . . . . . . . . . | $$ | . X X X X O . . X . . 4 . . . . O . . | $$ | . X O X O O . X . , . . . . O , X . . | $$ | . O O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . 6 5 . . . . . , . . . . . , . . . | $$ | . . . 7 . . . . . . . . . . . . . . . | $$ | . . 8 . . . . . . . . . . . . . . . . | $$ | . . 2 . . . . . . . . . . . . . . . . | $$ | . . O 1 . . . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . 9 . . | $$ | . . O X . . . . . , . . . . . , . . . | $$ | . . X 3 . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Kageyama does say that white turn is disquieting even if doesn't work yet, so he could be afraid of tenuki and the aji biting him in the arse later. It's hard to pick a particular sequence of play to illustrate this and test if that judgement was good, but in the real game Kageyama defended on the left at c10 (which means black b16 isn't sente, so white f17 more pressure on corner in future) and then black continued in the upper right, allowing white to extend at m17. That f17 turn is definitely getting more disquieting now, so if black did play the greedy move he should play differently to handle it.

Click Here To Show Diagram Code: [go]$$Bc Playing rest of board like real game $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X X X . . . . . . 8 . . . 4 O 3 . | $$ | . . O X O O . 1 . , . . . . O , X . . | $$ | . . O O X O . . . . . . . . . 7 . 5 . | $$ | . . O X X X X . . . . . . . . . 6 . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 2 , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . 9 . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

So what's LZ's view on the above. 2 is -10%, should turn immediately as discussed above. 3 is then -13%, should be L17, allowing white s16. This obviously is concerned with nullifying the aji of f17 (black would now reply at f18). I could well imagine Kageyama and some other pros might not like the fact black spends 2 moves to fix the top here, instead of the 1 with the net. But the question is 1 move with net, and the useful forcing moves for white it leaves at h15 and j16, actually better than spending 2 moves on it for a position with more territory (l17 obviously big on its own and means after white s16 black can use it when activating r16) and leaving less useful kikashi for opponent. If black 3, white 4 obviously ok as urgent local shape. 5 is then a little bad (-3%), L17 or m17 again better. 6 should then be f17, and there's lots of choices for how black can continue. One interesting one is g18, then white e18 hane, f18 cut, b17 hane in corner and black d18: white sacrifices the 4 stones for b17 not getting blocked.

So the summary from LZ in this hypothetical future is: yes 1 on outside leaves bad aji, but net leaves too good forcing moves for opponent, so better to leave the aji, be mindful of it and potentially spend another move on it soon, than accept the slack result of net.

Bill Spight · Post by **Bill Spight** » Wed Aug 21, 2019 1:09 am

Uberdude wrote:Bill, here's a screenshot of Lizzie with LZ #234 after ~100k playouts. You can see the distribution of playouts is very sharp: 100k at 48.8% for 'bad' 1, under 1k for every other move considered and just 9 for a. So the error on its 42.3% (-6%) is large, playing that move and giving 100k playouts gets a winrate of 37.9% (-11%).

Many thanks for an interesting and instructive analysis.

The procedure the bots use where there is a great disparity in playouts between candidate plays is fine for finding good enough plays.

But consider that it was quite possible for Kageyama's play to gain around 4½% with 100k playouts instead of losing that much. Then it would only have been a 2% underdog, surely within any reasonable margin of error.

And it is even possible for it to become top dog, isn't it? (That happened when I played around with Deep Leela, BTW. OC, Deep Leela is a weaker player.) That's why I would like someone to develop an analyst instead of a player. But there is no glory in that, eh?

Uberdude · Post by **Uberdude** » Wed Aug 21, 2019 1:17 am

Bill Spight wrote:
Unsurprisngly, many bots (I asked LZ 234, LZ 15b, MiniGo cormorant, Elf v2) strongly prefer Kageyama's bad 'a' over his 'only move' 1, and I am strongly inclined to believe them over him that is objectively a better move.
How strong is their preference? 7% or more? How many playouts? At least 100k for each of these two plays?

That was just a quick look on my laptop, only a few thousand. With 100k Elfv2 agrees with LZ that Kageyama's bad move is better than the net (and with immediate peep and kosumi continuation), but decides another move is even better (52.3% vs 50.3% vs 35%). Black goes even bigger with k17, white turns f17 (j16 shoulder hit is 2nd choice, then black plays g16 and white j17 and black attacks left at d10, so k17 can basically be seen as bait to make black capture the cutting stones in sente so he can attack on left), then casts the big surround of 3. Interesting white doesn't then try to live with e18, which is possible but as he'd have to answer black b16 at c11 in that sequence anyway, and it would make black stronger on the outside with exchanges like j17 for k16, just defends there immediately, lets black kill the stones and then forces from the outside with 6, which si screening kikashi against black's extension to a aiming at b in the shape white plans to build at the top right.

Click Here To Show Diagram Code: [go]$$Bc $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . 5 . . . 7 . . . . b . . . . | $$ | . . X X X 2 . . . 1 6 . a . . . O 0 . | $$ | . . O X O O . . . , . . . . O 8 X . . | $$ | . . O O X O . . 3 . . . . . . . 9 . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . 4 . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Bill Spight · Post by **Bill Spight** » Wed Aug 21, 2019 1:27 am

Many thanks, Uberdude!

Fascinating stuff, eh?

Uberdude · Post by **Uberdude** » Wed Aug 21, 2019 2:09 am

And here is MiniGo's mainline at 200k, black around 48% throughout. So it too likes Kageyama's greedy 1 (and net bad and punished with immediate peep h15), but there is an interesting difference in that it wants to play 3 in corner before kosumi at 7. There is a reason for this, and it's the tesuji at 8. Having to fall back at 9 is annoying, but if you block at c19 then white trades with b17 atari, e19 capture, b18 extend, f18 capture: white is happy to get corner profit in sente and b16 now dumb.

Click Here To Show Diagram Code: [go]$$Bc $$ +---------------------------------------+ $$ | . . . 8 . . . . . . . . . . . . . . . | $$ | . 9 . 5 4 . 0 . . . . . . . . . . . . | $$ | . . X X X 2 . . 7 . . . . . . . O . . | $$ | . 3 O X O O . 1 . , . . . . O . X . . | $$ | . . O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . 6 . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

This kind of 2nd line crawl to avoid death is something humans tend to assume is bad, but MiniGo says white is fine at 53%.

Click Here To Show Diagram Code: [go]$$Bcm11 $$ +---------------------------------------+ $$ | . . 3 O . . . 6 . . . . . . . . . . . | $$ | . X . X O . O 5 4 8 . 0 . . . . . . . | $$ | . . X X X O . 7 X . 9 . . . . . O . . | $$ | . X O X O O . X . , . . . . O . X . . | $$ | . . O O X O . . . . . . . . . . . . . | $$ | . . O X X X X . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . 1 2 O . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Bantari · Post by **Bantari** » Wed Aug 21, 2019 10:06 am

gowan wrote:When I was first learning to play go I was, as probably many of us were, told that greedy play was bad. When weak players play greedily against a stronger player, the greedy player often loses. I was also told that go is a game of balance and if you try to keep your opponent from getting anything you will lose. The kind of greedy play often exhibited by weak amateurs is not the same as John Fairbairn cited above. John described the Korean and Chinese approach to go as playing the percentages. I believe that the win rates given by AI bots are not quite exactly the percentage chances of winning. I'm not sure what they are exactly, maybe the percentage of winning in the playouts? Could the percentage play John mentioned might be just playing the most efficient moves? I'm sure that's what pros in general try to do.

I think there is a fine line between greed and efficiency, and both are very different from balance.
Greed is often thought in connection with overplays, as in 'too greedy'.
Efficiency goes in the same direction but not quite so far.
Balance is often thought of as timid play - the idea being of leaving the game result up to opponent mistakes. Not sure this is correct, though.

Ultimately, both efficiency and balance are good - if both players play equally efficient, the balance is preserved. But balance for the sake of balance should not be the goal, I think. Once the enemy makes a move that destroys this balance (like overplay or too slow) - you need to pounce and take advantage. I thought this was the general idea all the time, and I still don't see nothing wrong with that. But maybe I misunderstand.

John Fairbairn · Post by **John Fairbairn** » Wed Aug 21, 2019 11:15 am

I think there is a fine line between greed and efficiency, and both are very different from balance.

Before this gets out of hand, let me try to bring things back in, er, balance.

I used the phrase about greed being good, rashly assuming people would understand the cultural reference as a joke (Gordon Gekko, Wall Street - greed is behind the "upward surge" in human development was the key phrase, I believe). It was also shorthand. Too short, obviously.

So, trying again, and digging back into my possibly unreliable memory of this, Wang Xi (one of the most intellectual pros) argued from the premise that we are typically faced with situations, mostly in the middle game, which are too difficult to read out accurately, but we still have to play something.

He claimed that the typical Japanese response (souba=soba go) would be to make a choice along the lines of "this may not gain me anything, but it won't lose me anything." That's par. Essentially it's risk free.

But the typical Korean approach was quite different. At that time Korean go was in a clear ascendancy (Yi Ch'ang-ho, Cho Hun-hyeon, Yu Ch'ang-hyeok, Seo Pong-su etc) and Chinese players were switching from trying to copy the Japanese to copying the Koreans.

Wang said (always with the proviso I'm not misquoting him) that the Korean approach was to look for an edge by accepting risk, and so if a player felt that he had more than 50% chance of getting an edge and less than 50% chance of losing out, he should consider taking his chances. The considerations are obviously based on things like confidence in one's own play (aka experience) and the overall state of the game or match.

The requirement for this to work, however, is that your assessment of the chances is reliable enough. Underlying this is an important point about risk. There is not just risk in reading out variations accurately. There is risk also in making the risk assessment itself. But it was believed that Yi Ch'ang-ho had found a way of making such assessments reliably (Wang gave an excellent specific example). That improved enormously his chances of succeeding with percentage plays.

Parenthetically, among people who mix up yose and the endgame (i.e. the typical western player), the notion developed that Yi was fantastic at the endgame. He no doubt was, but what he really excelled at was yose, or boundary plays, and the early-game example that Wang gave illustrated just that.

Chinese players ended up copying Korean players and the rest is history. Somebody should also tell the Japanese, and that, it seems to me, is what Ohashi is trying to do when he argues that it is this area in which Japanese pros have to look at in order to catch up (with AI, he says, but by implication, I think, with Chinese and Korean humans).

You can loosely talk (like me) of playing the percentages as greed, or you can call it something else. But whatever you call it, in practical go it has nothing for most of us to do with efficiency.

We can, and do, all make an assessment of our chances whether a particular move works. That's how we play blitz games and how many of us play slow games. Our assessments may be way out of kilter, but we still make them. And if we do improve them, we notice we become stronger, so in that sense they work. So we still play the same probabilistic way most of the time. From my own experience, I have always felt that the level that marks a pro is when you stop relying almost entirely on guesses/probabilities and, instead, play precisely, for par. Ironically, though, it now seems that's a stage you have to go through before you start switching back to playing probabilistically, because that's the only way you can beat people who are also capable of playing a whole round in par.

Assuming that's anything like correct, a principal characteristic of the probabilistic way of playing is that you do your reckoning up in advance. Of course, you may read deeply and back up that information, but the measuring point is at the time you start your manoeuvre.

Efficiency, for most of us us, even pros, is quite different. It is a measure, but we can normally only apply it at the end of the manoeuvre.

Except that it seems that there are geniuses like Dosaku who can apply some sort of measure of efficiency instead at the start of a manoeuvre. If so, that seems a incredible edge.

Again with the proviso that these ramblings do describe anything real, I find it fascinating to speculate on whether AI bots are, in effect, playing probabilistically in typical human fashion, but with better numbers, or are they able to use a measurement of efficiency in advance - or both.

Either way, I think it could be useful to devise a good way of measuring efficiency. Maybe even essential, if we are to get closer to AI play. (And maybe the unit of measurement should be the dosaku

)

moha · Post by **moha** » Thu Aug 22, 2019 3:51 am

John Fairbairn wrote:... argued from the premise that we are typically faced with situations, mostly in the middle game, which are too difficult to read out accurately, but we still have to play something.

He claimed that the typical Japanese response (souba=soba go) would be to make a choice along the lines of "this may not gain me anything, but it won't lose me anything." That's par. Essentially it's risk free.
...
Wang said (always with the proviso I'm not misquoting him) that the Korean approach was to look for an edge by accepting risk, and so if a player felt that he had more than 50% chance of getting an edge and less than 50% chance of losing out, he should consider taking his chances.

Bantari wrote:Greed is often thought in connection with overplays, as in 'too greedy'.
Efficiency goes in the same direction but not quite so far.
Balance is often thought of as timid play - the idea being of leaving the game result up to opponent mistakes.

Not sure if this is the same idea, but I always thought game result SHOULD be left to opponent mistakes. Making good moves, gainful moves is not really possible for obvious reasons (symmetry). The best move is the move that loses nothing, all other are negative. This is on a different level, but when the opponent makes a move that aims to shift balance in his favor, I feel it is a weak move. I may not be able to find it, but there should be a way to exploit. Either he already had the advantage BEFORE that move, or he handed me a way to get it myself. The Korean/Chinese approach (again, not on pro level) seem to pay less attention to realistic or reasonable outcomes.

On pro level even a slight difference in what is reasonable matters, and I doubt it is possible (for humans) to asses the balance that accurately, so ultimately it is matter of style. And once a player thinks he has >50% chance to gain and <50% to lose, the position may already be considered good for him (different balance).

Tryss · Post by **Tryss** » Thu Aug 22, 2019 7:26 am

Not sure if this is the same idea, but I always thought game result SHOULD be left to opponent mistakes. Making good moves, gainful moves is not really possible for obvious reasons (symmetry). The best move is the move that loses nothing, all other are negative. This is on a different level, but when the opponent makes a move that aims to shift balance in his favor, I feel it is a weak move. Either he already had the advantage BEFORE that move, or he handed me a way to get it myself. The Korean/Chinese approach (again, not on pro level) seem to pay less attention to realistic or reasonable outcomes.

The problem with that argument, is that what we believe is a "realistic or reasonnable outcome" may be suboptimal : a mistake. And those are self punishing and harder to detect.

DoubleSente · Post by **DoubleSente** » Thu Aug 22, 2019 7:32 am

moha wrote:The best move is the move that loses nothing, all other are negative.

I think that is the most succinct summary of what is really going on at the end of the day. The perfect player would not have a style, they would simply reflect their opponents mistakes back at them. I'm reminded of Fan Hui 2P's comment in the Alphago documentary that playing against Alphago is like looking in a true mirror. All your flaws are reflected back at you.

Incidentally, at the most recent US Go Congress, Tianfeng Fang 8P made the comment that danger is the default state in go, so safety should be viewed with skepticism. This is the idea behind the recent shift way from the small knight's enclosure to the large knight's enclosure as the norm in professional play.

John Fairbairn · Post by **John Fairbairn** » Thu Aug 22, 2019 8:03 am

I'm not sure we are all on the same page in use of terms, but I sense a theoretical tinge to some comments. I think it's important to remember the practical aspect.

When Korean go was in the clear ascendency, there was a strong trend there towards much shorter time limits. China followed to some extent before backtracking.

It would seem that a riskier-in-theory style of play was considered less risky in practice in faster games because it increased the chance of a big mistake by the opponent. Although we often say the loser is the one who makes the last mistake, in pro praxis making the first mistake is probably the real killer.

Even today, when extreme positions have often been retreated from, Korean and Chinese go seems to feature far more resignations than Japanese go.

Japanese players with the luxury of longer time limits have often seemed inclined to strive for par until the endgame allows them to become more precise in their calculations.

Allied to that has been the feeling - less prevalent nowadays but still there - that a game for publication in a newspaper (which the winner might be expected to annotate) must not be spoiled by rashness. The sponsor has to be given his money's worth. In tournaments just designed to find a winner rather than newspaper copy, the style of play can be freer.

Kirby · Post by **Kirby** » Thu Aug 22, 2019 10:10 am

John Fairbairn wrote: Even today, when extreme positions have often been retreated from, Korean and Chinese go seems to feature far more resignations than Japanese go.

Japanese players with the luxury of longer time limits have often seemed inclined to strive for par until the endgame allows them to become more precise in their calculations.

Another possible explanation is that, when both players are always playing for par, neither gets much further ahead than the other until the endgame. Either explanation is a subjective generalization.

Life In 19x19

Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros

Re: Verdict on AI verdicts on old pros