Weak score estimator and Japanese rules

Shouldn’t that be the case anyway? (i.e., broken By Design)
Nobody should rely on the SE in the early game.

1 Like

well, I explained algorithm. Basically its just


creating algorithm itself is 5 minutes
only someone who not against it and understands OGS code is needed.

7 Likes

What about using shortest Manhattan distance to decide which color an intersection belongs to (with stones being opaque to influence propagation)?

3 Likes

This is not a problem, because of the separate problem that the score estimate always counts area :wink:

(although of course one shouldn’t give one player dame or other valuable moves while the other plays unnecessary moves. Just pass or fill in territory)

1 Like

If this is supposed to be a Voronoi diagram, then why is M16 (the middle Hoshi point on the upper side) painted white in this diagram? Clearly the closest stone is F16 which is black.

How does your algorithm handle ties?

I think generating some sort of discrete Voronoi diagram is a neat idea, but I’m a bit confused by these sample diagrams.

4 Likes

That’s what I’m saying!


I think if you look at the algorithm, you will see why it’s not as you expect (i.e. not totally reflective of the closest color) around the edges. Weak score estimator and Japanese rules - #38 by stone.defender for a description. The algorithm I suggested above (flood in all directions at the same time) should fix that.


@stone.defender says it’s a bad idea, but I still think ties should be marked neutral!

4 Likes

So has anyone here actually produced a tangible idea of what a SE should do?
Jumping into algorithms when you haven’t decided on the fundamentals seems a bit premature.

Uses I can imagine:

  1. Help players make strategic decisions based on sound score estimates.
  2. Help players estimate the score by themselves.
  3. Help players visualize an easier-to-understand representation of the board.

By what means?

  1. Lead indicator: Use a strong engine to give accurate but imprecise information “W leads” or “The game is about even”.
  2. Territory counter: Allows drawing rectangles on the board, tallies the sizes of each, per color.
  3. Territory visualizer: show engine’s prediction of where territory will probably be.

Et cetera.

p.s.: Feature request: Flavor text / personalities for lead indicators: “My grandma could win this.”, “You dun’ goof’d.”, “This match is oddly even.”

6 Likes

I think (2) is in large part what people want. (1) was proposed and rejected in [poll] Balancing the new AI score estimator

(3) is important in a weak score estimator because it helps you mentally calibrate which parts of the estimate are wrong.

2 Likes

I can’t think of a reason to maintain a broken visualizer. Until very recently, there simply wasn’t anything else, but now there is. Either give the best visualization available, or don’t give any. To deliberately give people errors to sort out seems quite nonsensical.

If you invite someone over, you wouldn’t necessarily give them driving instructions according to maps from the 1920, either. “But they’d know the general direction and well, here and there they’d have to adjust to the actual road situation. It trains them to think on their feet.”

2 Likes

What I’m saying is you need some kind of visualization in order to help with counting (unless you have a very strong estimator that simply knows the score)

Not adding a visualization is like giving them no directions at all. “The destination is located 2.3 km southeast, see you soon!”

1 Like

Almost. But then again the broken visualizer is more like “Here’s an address, it could be mine, but maybe it’s yours - or that guy’s.” Sure, if you go somewhere, you’ll arrive somewhere.

My point is that if you just want to help people with counting, 2 is good (as you still have to make correct judgements about territory but it’s all your own perception) and (broken) 3 is bad (because it is broken).

I think the only reason people think they’d want something that’s a combination of 2 and (broken) 3 is that they’re used to (broken) 3. That’s what we had.

I always cringe hard when I see some kyu streamer click SE and say “O look I’m still comfortably ahead” when the picture it draws is patently wrong. Streamer’s fault for believing in the broken tool, sure, but it’s not too silly to assume there are lots of people who think “I’m weak, so SE is probably right”. It actively hinders learning. If people only saw black/white dots and still had to count for themselves, that would already be a big, positive difference. Rather than mislead players and cop out with “well, just don’t believe the SE”, that is just disingenuous.

5 Likes

I think this is a good point though.

What should the estimator do?

  1. :x: Tell you the “real” score, where real is a prediction by a strong bot like katago. It can still make mistakes, but it’s probably going to be too good, and account for things you might not even see. (fairly strongly opposed)
  2. :question: Loosely mark areas of the board in some fashion only with no real knowledge of the game itself. Something like these Voronoi diagrams might do something like that.
  3. :question: Loosely mark areas as above but also give you a score counting areas based on those points marked?
  4. :question: introduce a weak “AI” that knows the rules of the game to help mark areas. It’s probably what the score estimator does now, with some Monte Carlo playouts to help figure out captured stones.
  5. :question: improve the current estimator to account for Japanese rules (which is what’s happening) and fix some other quirks like handicap and leave it like that.
  6. :question: draw tool to manually mark areas and stones to give to each player purely to help visually.
  7. :question: draw tool to manually mark areas as above but also a button to count what you’ve manually marked for you.

Basically it’s the question of what help do people want and what is too much?

  1. Help visualising the areas
  2. Help with counting (because it’s quite tricky to keep track of and add a whole bunch of numbers)
  3. :x: Help to figure out who’s winning (I think that’s really what we don’t want)
  4. :question: Help to figure out how much one players winning by when it’s “obvious” one player is winning. (Even a not so great estimator can tell you that reasonably well).

Should it be different for live/blitz vs correspondence?

Is this the right thread for this discussion or shall we continue it in a dedicated “How can we improve the score estimator” thread? :slight_smile:

3 Likes

I guess it’s slightly on-topic since if there’s a better thing coming soon, why mess with the old one? However, the discussion seems to demonstrate the opposite - it’s going to be a lot of work to even agree on what we want, let alone implement it.

So, did anyone try out the Japanese estimation fixes on the beta site yet? :grin:

3 Likes

This is the option I favour. Anything involving strong AI is a no-go for me (in game, after the game its fine).

5 Likes

It’s all about to which point OGS wants to enforce the no external help policy. A balance between being rigorous or satisfy the wish of some players. You know we allow joseki books in corr games. And even sometimes here, players ask for AI in-game …
But the truth is that giving a weak SE, which can induce hints and failures is a bit too weird flattery in my opinion.
The main advantage i still see is a help for full beginner to close boudaries and group status, that sort of things. In that case i would better offer a strong SE to them only (20-30k). Or use humans.

1 Like

What about a prominent button which appears during the agreement phase (or maybe after there have been more than a certain number of passes without the game ending?) (maybe only for TPK players?) which lets volunteers accept the request from help, and jump to the game. These volunteers would have no powers regular users do not have, but would be able to offer advice on how to end the game, and call a mod if the players are still unable to resolve it. The players could, of course, still call a mod directly. Perhaps this is providing a solution to a problem noone had, though.

2 Likes

when one of players is so far ahead that other player should resign and endgame would be real waste of time - I think to understand when that happens is the real thing that estimator should do. Estimator shouldn’t be able to help you to know if you are few points ahead or behind. But it should be able to tell you when you are really far behind. I think purpose of estimator is to know when to resign.

OK , OK, pure version:

res cr
black : 189
white : 140
neutral: 32


b 164
w 164
n 33

res st
b 164
w 169
n 28

res real
b 169
w 163
n 29

it literally has no AI or Go knowledge
…but looks like current old estimator(that ogs really use during game) is worse in every single thing
so

therefore not a problem
they just should not forget to click dead groups

5 Likes

If you click dead groups, would you run the algorithm again but just pretend that those stones aren’t there?

I guess it works well for area scoring. Would you just count the shaded areas minus the stones for territory scoring?

2 Likes

they would be just replaced with neutral points.
no need to rerun estimation from beginning, It able to just continue iterations from current modified state.
(and so, if neutral points will be surrounded by same color, they will be painted in it)