Estimate score missed

I see that ‘estimate score’ often makes a very different score that just when the game ends. Is there any way to make a closer ‘estimate score’ to the real one?

Thanks.

Score estimating is a very difficult and near-AI task, so it would need lots of work for only a little improvement.
(That’s why humans are still better in Go than computers.)

2 Likes

The current Score Est. Is a little buggy. If you are willing to manually mark some stones/groups dead then it will become very accurate. CTRL- or SHIFT- click on stones to mark them dead.
Sadly you currently have to mark bigger groups dead stone by stone.

Seems like on beta estimator makes very big mistakes, does it count komi? Looks like difference in the end is always 5.5-6.5 points less then predicted by estimator. On old OGS it was very accurate when it is not so hard to predict and count score. Few minutes ago, I had a game where it showed that difference is +7 for black, but I won (as white) by 0.5… By the way, what is estimation confidence (how should we use it when estimating score)?

No, based on what I’ve observed, it looks like the estimator is unaware of komi and just counts what’s on the board. So, you’ll have to manually adjust what it says. I guess it would be a simple change to the code to take into account komi. However, if the game is close enough where komi makes a difference and the game is not very close to being over, I would be hesitant in trusting the accuracy of the score estimator to make the correct call (even after manually marking dead stones).

In a different forum thread, a moderator mentioned that the estimator uses a monte-carlo method to estimate the score and the confidence value being larger means that the system is more confident that the score estimate is correct. He mentioning that it had to do with how many “fractional points” were awarded. I don’t know exactly what that means. However, at a conceptual level, a monte-carlo method just means that the game is randomly played out to the finish many, many times, and some sort of intelligent averaging is done to figure out what the likely result might be. The best go AI programs these days use monte-carlo methods to evaluate moves, but tuning the fine details of such a system to perform well can be tricky.

As others have pointed out, a big issue with the score estimator is that it often fails to identify large or complex dead groups, thus wildly throwing off the estimate unless you manually mark those as dead by tediously selecting each stone.

Anyways, estimating the score of a game that is not already very close to end can be a very difficult task for both humans and computers. The final score will depend entirely on how the players choose to play it out, and if they make any mistakes or discover any brilliant moves. For example, a player that is well ahead might accept reducing the margin of victory to ensure the security his position and eventual win.