Swim's Reviews

[Tryss]

All of these posts have little value is the problem.

Entertainment has some value, even of this strange kind. It's like some surrealist artist performance.

[alphaville]

All of these posts have little value is the problem.

Entertainment has some value, even of this strange kind. It's like some surrealist artist performance.

Sometimes reality is stranger than fiction.

[djhbrown]

Swim rethinks her calculations (details in icGo documentation):
picture here (url instead of image because i keep changing it as definitions become more refined)
https://sites.google.com/site/djhbrown2 ... edirects=0
to be continued

[tchan001]

Telling someone the strategy is to attack is as practical as telling someone he will know how to attack and defend by just looking at the cover of the book "Attack and Defense".

[alphaville]

Telling someone the strategy is to attack is as practical as telling someone he will know how to attack and defend by just looking at the cover of the book "Attack and Defense".

If I understand correctly, this is the goal setting phase, where the algorithm decides what to do next.

[djhbrown]

Swim's review is here:
https://papers.ssrn.com/sol3/papers.cfm ... id=3027817

[Kirby]

Swim's review is here:
https://papers.ssrn.com/sol3/papers.cfm ... id=3027817

Thanks, and welcome back... If I understand correctly, Swim recommends for white to play either: O5, P6, Q7, Q11, P11, or O11 in this position?



Is that what you mean?

In the game, I played R10. The paper says:

Attachments are excluded, following the heuristic adage: "do not contact the stones you are attacking".

Is this the only reason that Swim thinks that R10 is bad?

[djhbrown]

Is that what you mean?

The illustration is only a partial review; the fightback method proposes attacking a weak group or invading a moyo; methods of achieving the latter are discussed in "Swimming with Alphago"; disjunctive methods are not exclusive, but generate alternative areas for moves; the intersection of all proposed move areas would create a preferred list of candidates; i considered taking the analysis further, but as my main purpose was to get section 2 of the paper into print (section 3 is what Arthur Sullivan calls "mere corroborative detail, designed to lend verisimilitude to an otherwise bland and unconvincing narative"), i didn't bother. My guess is that if i did so, Swim might prefer green over red, at least until it started reading; and it is conceivable that Swim might end up preferring red over green, which at first glance looks slow, but letting black split white's two weak clusters might not be a good idea.

at the end of the day, Swim's deductions would be sent off to the jury, which might come up with a different move altogether - i considered feeding the position into Leela or Ray to see what they, as jury members, might propose as follow-ups (or as alternative initial candidates), but figured that if anyone wanted to see their probable moves, they could do their own checks. i recall that in https://www.youtube.com/watch?v=6XG9WWQ ... a3Hl1X_v-S
Swim came up with different moves to Leela.

Is this the only reason that Swim thinks that R10 is bad?

Swim doesn't think R10 is bad, merely that it wouldn't be a good way to go about attacking R9. i can envisage the methods described in "Commonsense Go" suggesting R10 as a leaning attack feint as preparation of an invasion of the upper right, but again, a thorough analysis of the position wasn't my main motive for the paper.

[djhbrown]

DeepZen (w) vs Fine Art (b), semifinal of 1st World AI Go Open, black to play.

dzfac2sgm.png (9.92 KiB) Viewed 7914 times

Swim's assessment:

"black has 107 points of influence; white has 71. So black should play safe at 'a', 7 steps from strong black group 1 along a green line, to reduce white's centre moyo".

comment:

Uberdude reports that Fine Art played at 'b'
forum/viewtopic.php?p=222078#p222078

This is a terrible move - a beginner's error - trying to save two worthless stones.

Swim observes: "black group 5 is not important, it is much smaller than black's lead and although it does separate white 4 from white 3 and 5, white 4 is not weak, so there is no point attacking it".

How on earth could a program as strong as Fine Art (FA) make such a stupid move?
Answer: because, like its parent Alphago, FA has sweet FA idea what it's doing, relying on short-sighted deep search instead of understanding.

[djhbrown]

Uberdude's review of a review of Swim's review #1
https://www.reddit.com/r/baduk/comments ... o_and_god/

[djhbrown]

white to play...(don't tell if you know the "correct" answers, of which there are two)

113.png (17.75 KiB) Viewed 7710 times

[djhbrown]

Swim's preliminary perceptions (colour and shadow maps):

113cmdes3.png (54.64 KiB) Viewed 7597 times

group qualitative strength:
a = alive = two eyes or one big eye that isn't a killing shape
s = strong = not a, but a big enough eyespace within the shadow map
w = weak = not(a or s) and without a clear path to a wide open space
- = unknown (= not(a or s or w))

Black group 1's colour/shadow map has 5 points, but it has no eyes (G15 is not an eye because H15 is not colour-controlled by black). You can read that G12 and E14 are miai to connect black 1 to black 8; Swim will have to read that too; meanwhile, it regards black 1 as weak, so its 5 points aren't counted in black's total.

White 5 doesn't have two clear eyes (black can close its potential eye on the side), but it has 6 internal colour-controlled points, so - before reading - Swim thinks it is strong.

.. to be continued

[djhbrown]

Assuming komi is 7.5, white leads on the board by 11.5 - and it's white's turn. Being on move is worth komi, so white effectively leads by 19.

Code: Select all

let lead = value(me) - value(you); 
if lead + chance > 0 then playsafe
                     else fightback

value(x) =  sum over x.g of size(area(g))*(if weak(g) then 0 else 1) 
            + captives(g) 
            + if white(x) then komi

chance =  h(size(neutral space))

Swim decides white should playsafe.

Code: Select all

to playsafe 
    do foreach me.group
            if me.group is weak and 
               me.group is important 
            then save(me.group)) 
     or foreach you.group 
              If weak (you.group) and 
                 you.group is important 
              then attack(you.group) 
     or foreach me.moyo 
              expand(me.moyo) 
     or foreach you.moyo 
         if wideopen(you.moyo) 
         then invade(you.moyo)
         else limit(you.moyo)

x is important =  big (x) or (x splits (y,z) and 
                  (y is important 
                     or 
                   z is important)) 

big (x) =  (value(save(x)) - value(tenuki(x))  > chance)

113cmdes3.png (54.64 KiB) Viewed 7564 times

white has 4 weak groups: 2,4,6 and 9

Code: Select all

if me.group is weak and 
               me.group is important 
            then save (me.group))

x is important =  big (x) or (x splits (y,z) and
                  (y is important
                     or
                   z is important))

big (x) =  (value(save(x)) - value(tenuki(x))  > chance)

white 2 doesn't split black 1 and 8 (they can connect at G12)
white 4 splits black 2 and 3, but neither is weak
white 6 splits black 4 and 5, but neither is weak
white 9 splits black 4 and 6, but neither is weak

so none of white's weak groups is doing something important and worth saving at this stage

... to be continued

[djhbrown]

Black 1 is weak; black 4,6,7,8 are unsettled. Black 1 has two paths to black 8: G12 and {E13,E14,F14]. White's playsafe purpose in attacking a weak group is not to kill, but to force a defensive move. Here, white G12 would force black to capture F13 to connect, so G12 would expand white 3 in sente.

Alternatively, a move on any of the green lines, or a leaning attack on C8, would expand white's moyo between white 3 and white 8

113cmdes3wsl.png (58.92 KiB) Viewed 7458 times

..... to be continued

[djhbrown]

review continuation:
https://papers.ssrn.com/sol3/papers.cfm ... id=3071677