KataGo does not show influence

[RobertJasiek]

In Deutsche Go-Zeitung 5/2019 on page 19 in the article "AI Sensei goes KataGo" is a diagram on KataGo with "influence".

However, it is NOT influence, which - according to my definitions of this strategic concept - represents (existing or expected) connection status, (existing or expected) life status and (expected) territory of the players or their (existing or playable) stones. It is also not distance (to live stones).

Rather, the black and white square markings in the diagram seem to be something else: for sampled scoring positions (and in the meaning of expected values), the represented concept seems to be EXPECTED CONTROL of each intersection by black or white two-eye-alive-formations (or independent life formations that can be transformed into such) comprising their stones and two single point eyes. (I cannot know how KataGo interprets sekis but this detail is immaterial for the fundamental understanding of what is represented at all.)

It cannot be 'current control' because the shades of the markings do not correspond to such. However, as empirical 'expected control' they do. Near 50% expected values, there are no (perceivable) shades. The shades are partially lighter where nearer to 3-3 under 4-4 than they are somewhat off the corner. A possible shoulder reduction including some counter-attack expected by the program can be recognised. All these are manifestations of expected control - not, however, of current control or even of influence.

Therefore, KataGo itself, its applying-programmer-users and end users should not mislead the, or other, go players (or themselves) by the disinformation "influence". Instead, the should use the apparently correct information "expected control" and name / refer to the concept accordingly. At least, the markings are more useful than the fake influences in the meaning of measures of distances by old programs a couple of decades ago. The narkings are, however, by far not as useful as real influence is.

[Bill Spight]

I agree that influence, however defined, is a static concept which is independent of who has the move. Projected territory or area at the end of the game is different.

[dfan]

KataGo does not call this "influence". It is referred to in the documentation as "predicted final ownership".

[RobertJasiek]

Good!

So I have identified correctly what is shown from thoroughly studying just one diagram.

The problem then is one of some users of KataGo calling its markings "influence". I have seen this mistake several times but never before seen them using the correct description.

[Gomoto]

KataGo does not call this "influence". It is referred to in the documentation as "predicted final ownership".

And because it estimates predicted final ownership quite accuratly it is able to give a rather good POINTS estimate as well

[RobertJasiek]

Gomoto, how does KataGo predict final ownership quite accuratly? Every variation can result in different ownerships. KataGo may predict empirical final ownership with KataGo's sampled variations accurately and therefore estimate points well. Predicting final ownership of specific intersections differs from predicting empirical final ownership!

Having not used KataGo yet, I wonder what kind of points does KataGo estimate: a) current territory but not points outside safe territory regions or b) an empirical estimation from sampling many variations and their resulting scores? How reliable is the points estimate and possibly why could we know the degree of reliability?

[Gomoto]

Robert, this was just meant to be a fun remark with regard to the other thread about points.

From reviewing many games with KataGo, I know that KataGo gives a good estimate in the later phases of the game, when you can play out the variations and check. In the earlier phases I have just a good feeling about the KataGo score points to compare different plays (I simply trust KataGo).

I am sure that you use the word "accurate" much more strictly than my fun remark implied.

[dfan]

Having not used KataGo yet, I wonder what kind of points does KataGo estimate: a) current territory but not points outside safe territory regions or b) an empirical estimation from sampling many variations and their resulting scores? How reliable is the points estimate and possibly why could we know the degree of reliability?

Basically b), if I understand you correctly. It tries to predict the probability of final ownership of every location on the board just as it tries to predict the probability of the final win/loss result of the game. These predictions are of course much more accurate for safe territory.

[Bill Spight]

In any event, KataGo's estimate of the final score difference should be much more accurate than its estimate of the final ownership of any point.

[lightvector]

Yeah, I consistently in the code (variable names, comments) and documentation always refer to it as a prediction of ownership. I don't know what exactly to do about if some users that I've never interacted with still call it "influence".

The prediction of ownership is separate from the prediction of win/loss or score. It's the same neural network, so it can uses what it learns about predicting one of them to help also predict the other, but they are not guaranteed to be mutually consistent.

(Thought experiment: consider if you asked a top human pro to give you a numeric probability for their belief of the ownership of each point in the board one by one, in total 361 different values. And consider if you then could wave a magic wand so that they entirely forgot you had asked them that and you then asked them to predict either the final score or the number of extra komi points they think would make the game into a fair game. Do you think that the sum of their 361 ownership predictions would be exactly equal to either their score or lead-in-komi-points predictions? Probably not.)

[xela]

I'm finding the predicted ownership useful from a positional judgement perspective, regardless of whether or not it might have any connection with "score". Often when I look at positional judgement problems in books, I get the wrong answer because I misjudge the strength or weakness of groups. I'll choose a move that defends a "weak" group, then read in the commentary that tenuki was OK because the group had a defensive resource that I overlooked. Or I'll abandon a "strong" group only to learn that it was much more vulnerable to attack than I thought.

KataGo's predicted ownership will tell me things like:

• In about a quarter of the variations KataGo looked at, those stones got captured; or
• where I thought white had some territory on the left, KataGo counts the area as dame because a running battle is about to hurtle right through the middle of it.

These are cues that an unsettled group might be weaker or stronger than it looks to my amateur eye.

Of course it's possible that I'm horribly misinterpreting things here...