Alphazero and Go strategy

[London Go Centre]

Daniel Hu, 5D, has written a paper discussing how he thinks Alphazero is contributing to Go strategy. It can be downloaded from our online resources page http://gocentre.londongo.club/resources/

[Bill Spight]

Very good paper. Many thanks too Daniel.

[sorin]

Daniel Hu, 5D, has written a paper discussing how he thinks Alphazero is contributing to Go strategy. It can be downloaded from our online resources page http://gocentre.londongo.club/resources/

The link mentioned on the page above used to work,but now it is broken: http://gocentre.londongo.club/wp-conten ... 190329.pdf

[Kirby]

Daniel Hu, 5D, has written a paper discussing how he thinks Alphazero is contributing to Go strategy. It can be downloaded from our online resources page http://gocentre.londongo.club/resources/

The link mentioned on the page above used to work,but now it is broken: http://gocentre.londongo.club/wp-conten ... 190329.pdf

pdfZero

[tchan001]

Daniel Hu, 5D, has written a paper discussing how he thinks Alphazero is contributing to Go strategy. It can be downloaded from our online resources page http://gocentre.londongo.club/resources/

The link mentioned on the page above used to work,but now it is broken: http://gocentre.londongo.club/wp-conten ... 190329.pdf

Go back to the resource page and follow the new link.

[Knotwilg]

One would think the author has understood the difference between influence and thickness for the first time, in that walls have influence but are only to be called thick if they already know where their eyes will be (eyespace, a base) and are already solidly connected (so no forcing moves like peeps).

For me this has been the capital difference all along but we've run in soulful discussions on those concepts.

Other than that, I mostly concur and didn't really learn anything new, although the author is much stronger than me. I have published my own inferences from playing/studying with bots here:

https://www.lifein19x19.com/viewtopic.php?p=240781

I'm a little more positive about the degree to which we can derive a theory from bot play. The author did just that and also we must learn that AI speaks in branches and probabilities, not in language.

[BlindGroup]

I finally got a chance to read this, and I was wondering if someone could explain this idea a bit:

"... there is the principle that attacking moves often become scattered disconnected groups that don’t make territory, and hence become worthless and hence raise the efficiency of the enemy stones. Hence most forcing moves are bad moves that are worth less than they cost. And hence most stronger players tend to prefer to take territory (or big points) directly or threaten and pressure enemy stones from afar rather than directly attack and harass."

I think the problem is that I don't understand what he mans by an attacking move. It seems to contradict one of the basic tenants of direction of play as I understand it which is to lean on weak groups for profit.

[sorin]

I finally got a chance to read this, and I was wondering if someone could explain this idea a bit:

"... there is the principle that attacking moves often become scattered disconnected groups that don’t make territory, and hence become worthless and hence raise the efficiency of the enemy stones. Hence most forcing moves are bad moves that are worth less than they cost. And hence most stronger players tend to prefer to take territory (or big points) directly or threaten and pressure enemy stones from afar rather than directly attack and harass."

I think the problem is that I don't understand what he mans by an attacking move. It seems to contradict one of the basic tenants of direction of play as I understand it which is to lean on weak groups for profit.

Attack to kill certainly fits the definition, but I am not sure the author is referring to that. And like you say, it is not that attacking groups is bad, bots certainly attack.

Maybe the author means that bots seem to attack less than humans do, overall, which can certainly be seen in the opening (very few bot hasami moves).

[Bill Spight]

I finally got a chance to read this, and I was wondering if someone could explain this idea a bit:

"... there is the principle that attacking moves often become scattered disconnected groups that don’t make territory, and hence become worthless and hence raise the efficiency of the enemy stones. Hence most forcing moves are bad moves that are worth less than they cost. And hence most stronger players tend to prefer to take territory (or big points) directly or threaten and pressure enemy stones from afar rather than directly attack and harass."

I think the problem is that I don't understand what he mans by an attacking move. It seems to contradict one of the basic tenants of direction of play as I understand it which is to lean on weak groups for profit.

I don't know what he is talking about, either. I have gotten into enough trouble saying that sente gains nothing, and here he is saying that most forcing moves are bad moves. I doubt if there is any principle here, but there may be some experience talking (not necessarily his own). IMX, attacking players are relatively easy to play against, as they tend to make overplays. But for that to be the case, you have to correctly identify the overplays and counter them accordingly. But I see no principle that, just because a player has made attacking moves to which his opponent has responded, the responses are ipso facto efficient.

To back up sorin's point, bots pincer less than humans have done (about half as much, by my estimation). And perhaps there is a principle there. That is, if a stone is pincered and it jumps out towards the center, it cuts a sector line. Then it may well threaten to counterattack on one side or the other. In such a case cutting the sector line is efficient. (Divide and conquer! )

[moha]

I think this is about the common dilemma: Opponent has some weak group. One could play a capping move or similar play to threaten it more. That would push it around, opponent would react somehow. But is it a good idea now, would it worth it?

I also noticed when started bot reviews: they don't feel "it would be a pity" to leave a weak group alone, and only act when the pressuring move is a good territorial or influental move by itself, and remain stable regardless of the attack.

[Knotwilg]

I don't think bots imply "forcing moves are bad". We see enough shoulder hits against stable positions.

For me it's just another way of saying, "if you don't have a clear immediate gain from an attack, then don't attack". We've discussed earlier how bots prefer sideways attacks, which grow territory/moyo at the other side, above attacks from above (boshi) which merely grow some influence, while often leading the attacked group into destroying potential territory or even safety of the surrounding group.

Rather than attacking directly, bots seem to be fine with a forcing move far away, which reduces the health of the weak group without creating damage around it.

[Bill Spight]

BTW, I have finished a note on Elf's evaluation of an enclosure facing enclosure extension by Takagawa.
viewtopic.php?p=242715#p242715

[BlindGroup]

I think this is about the common dilemma: Opponent has some weak group. One could play a capping move or similar play to threaten it more. That would push it around, opponent would react somehow. But is it a good idea now, would it worth it?

I also noticed when started bot reviews: they don't feel "it would be a pity" to leave a weak group alone, and only act when the pressuring move is a good territorial or influental move by itself, and remain stable regardless of the attack.

If I understand you correctly, I think this makes sense to me. The idea is that a move that only attacks rarely produces profit on it's own. So, the attack is inefficient. But attacks that are dual purpose -- they attack a weak group and provide another strategic benefits (ideally making immediate profit) -- is worth considering.

[Bill Spight]

I have finished my second note about Elf's evaluation of an enclosure facing enclosure position, here: viewtopic.php?p=242716#p242716 . This game pits Fujisawa Hideyuki (AKA Shuko) against Sakata.

[Uberdude]

The bots really hammer home the proverb: "Corners, then sides, then centre, dumbass".

Elf's suggestions match those recommended by my LZ gospel of the early opening, if I clarify that White's knight defence preventing the press is almost as valuable as the press so is at poistion 3a (but sometimes overlaps with 4)