# Help understanding the computer analysis output

Hello all,

I am new to the site and have been trying to use the computer analysis to review my games. I’ve been looking for documentation of the feature to help me understand the output, but have been unsuccessful at finding an explanation. (For example, I found documentation of other features here https://ogs.readme.io/docs)

I just have some basic questions like how to understand the numbers assigned the different moves and how those numbers relate to the bar that shows the chances of each player winning. I presume that positive numbers favor Black and negative numbers favor White? Any insight would be appreciated.

Thanks!

3 Likes

Update February 17, 2020

The documentation for the AI review feature is available at https://github.com/online-go/online-go.com/wiki/Understanding-the-AI-review

Hi @Cyanriddle. Welcome on OGS.

The computer analysis feature is really new and still might change in appearance or other minor ways in the coming weeks. Therefore, the feature hasn’t got in the documentation.
Additionally, the documentation you found is an old one. You can find a much more up to date one at https://github.com/online-go/online-go.com/wiki

The graph shows the win probability for black and how it changed while the game progressed.
If it is at the upper bound, black is likely to win, if it’s on the lower bound, white is likely to win.

The bar above shows the probability to win for black and white for the current board position. If you explore a variation, the probabilities won’t change and still show the numbers for the root board possition.

The 10+% moves indicate how many moves of each player reduced their chance to win by at least 10 percent points.

The numbers on the board are the change of the probability to win for the player to play the next move. If the number on one move is -8.0 for example and black to play, that move would lower blacks probability to win from 74.6% to 66.6% for example. If it would be whites turn, whites probability to win would change from 25.4% to 17.4%.
The blue move is the move the AI would play. The actual played move is represented by a small triangle and a transparent white circle. The triangle is on the AI move when AI and player agreed on the move to play.

The green background now represents the time LZ spent on the variation. Brighter green means more time spent on variations of the move.

2019-10-17T22:00:00Z Updated: Should represent the current state of the AI review.

22 Likes

This was exactly what I needed. Thanks!

2 Likes

Whoa - I thought the white background was a rendering bug!

5 Likes

Is it normal that the bar and the graph don’t match for some moves?
The graph is very useful to see sudden big changes in win rate (wich should be bad mistakes) but sometimes I go there in analysis and the bar shows a very different win rate. Which one is the good one?

In my opinion, both should be the same. Please share the games/moves in question so we can find out.

1 Like

I will. I just have to find them again

Update 2019-09-20T22:00:00Z:

The green background now represents the time LZ spent on the variation.

4 Likes

Thank you for writing this guide and keeping it up to date.

I was wondering what that meant. Originally, I thought that it was counting something good, but it’s actually counting just the biggest mistakes/blunders.

I wonder what the statistics of these counts look like when compared to rank.

By the way, does the AI analysis always assume area scoring (even for Japanese rules games)? I came across a case where one of the top three game changing moves appears to be failing to fill a dame in a very close position (which would of course be irrelevant under Japanese rules, but would matter under area scoring).

1 Like

The AI uses always Area Scoring and a Komi of 7.5.

LeelaZero isn’t capable of using other values.
Somewhere on anoeks infinite todo list is to evaluate if we could use KataGo instead, which is capable of using different rulesets and Komi

3 Likes

And handicap and very accurate score estimation

2 Likes

I guess this is pretty off topic, but…

Out of genuine curiosity, if we had very accurate score estimation, what common issues would be solved or remedied automatically?

• Mass timeout game annulment.
• Wrong bot game results (bot game results rely on score estimator)
• Stone removal phase issues
2 Likes

I think whether or not AI should be used to judge games that have been abandoned by mass timeout would be a contentious issue, which I believe was debated when it was brought up in one of the threads about mass timeouts.

From what we’ve seen of strong AI analysis of amateur games is that kyu-level players are liable to make many game changing blunders throughout a game. Using AI to call an incomplete game would rule against whoever happened to make the previous blunder. However, if the game were actually finished by the human players, the outcome would be determined by whoever made the last blunder.

On another note, I actually hope that the “ongoing game score estimator” (I only mean the one that is available to players while the game is still ongoing, since using AI as tool to help in stone status and scoring disputes is different and potentially helpful) is not improved with strong AI. Right now, the score estimator is terrible, so I don’t think too many people take it seriously and it doesn’t impact too many games. However, if players had the ability to query a strong AI to evaluate the current position, then I think that would be a form of assistance that should not be allowed.

For example, sometimes my opponent would make a blunder that I would not realize and fail to capitalize on (essentially one blunder followed by another). However, with a strong AI giving score (or even just win probability) estimates, that might tip me off that my opponent’s last move presents an opportunity that I should search for, and thus a decent score estimator could provide unfair AI assistance.

4 Likes

It is true that mass time out solution would not be immediately solved, fairer to say that case might be opened again

I wonder if we might make the improved score estimator available during game - but not available with analysis disabled, and make this the default, has has been argued for from time to time

2 Likes

I think that’s fair.

Analysis mode itself changes what a player can do way more than anything else I can think of. So if you have analysis enabled, you invite the other things like computerized score estimation without there being any sort of contention for fairness.

3 Likes