Key moves by score rather than winning chance

When looking at AI review of a game, you have the ability to display the graph as winning chance or estimated score, however the key moves underneath are always selected in terms of winning chance rather than score loss/gain. This is the case even as OGS supporter.

I feel like score is much more informative. For example let’s say I lead a game by +30. I make a huge mistake in the endgame which brings me down to +2, then a small mistake which brings me to -1 and I lose. In such situation the AI may select the second move as “key move” because it switched the result, while the actual key mistake I’m interested in is the first move.

On a similar note, is there a way to display the entire list of key moves (instead of having just 3 key moves and “+5 others”)?

Thanks, and apologies if I just missed a setting somewhere but I can’t find it!

4 Likes

Another relevant thread

and a github issue but not sure much changed yet.

1 Like

Thanks. Glad to see I’m not the only one, but sad to see it probably won’t be fixed anytime soon…

2 Likes

“key moves” mean the situation where someone blundered >50% winning position into losing one, so point-based metric doesnt work there.

And yeah, the point metric is far more useful (at least for me) than the winrate when analysing the game, but for key moves it makes more sense to get the moves which blundered away the winning position.

1 Like

Well I don’t really agree. “Key moves” are the important moves that deserve attention and may be worth commenting on.

Reasoning in terms of winrate will often provide uninteresting “key moves” which are just standard moves which so happen to occur while the game is close to 50/50.

In a game where a player dominates at +30, and then collapses at +1 before losing -1, I think the move causing the collapse is the actual blunder and key move of the game, and the one which should be highlighted and discussed.

3 Likes

^^

Yeah the last move of a 0.5 point game might be a key move, and it could just be not fighting that half point ko, or letting the opponent get the “right” dame in Japanese rules, it might not be the 20 point mistake you made when you were 30 points up.

In particular for handicap games it doesn’t make much sense the way it is at the moment.

I’m not really sure the one mistake that tipped the scales is as useful as the many blunders that might have led up to it. The exception is of course missing a move to kill/save a large group.

It’s very easy to define numerical criteria (more than n points, winrate sign changes, etc) but it’s quite harder to nail the actual key moves in a game.

Blunders jump to the eye, but they are useful just for weaker players. When a game is well balanced and the points chart span is less than say 10 points, you can’t just measure strategy on winrate or points.

Look at some ai vs ai games and find the key moves if you can! :grin:
It’s all a complete fighting mess! At least from my sdk point of view.

Ai sensei has something similar. I don’t use it, so I don’t know for sure, but I believe it’s using a points trigger, as OP is suggesting: it points out all moves that had a point loss bigger than n.

1 Like

I don’t think you need to choose it based on a fixed value of n in your “more than n points”. You could literally just choose the 3 biggest points mistakes, preferably 3 for each player, and show those instead and the system is already better.

At the moment I don’t even really like looking at the top three moves because I doubt they were actually important points of the game half the time.

Look at the transatlantic pro league and see that they look for the biggest mistakes over a certain number of points when reviewing with AIsensei (it has a feature for that). Since you mentioned it also, it’s a slider than you choose to pick out moves over a certain size, so if you only want to see ten point mistake first set it to 10, then you want to see something over 5, slide it to 5 etc.

2 Likes

I agree.
What I do is looking for big steps in the points chart.
Frequently my mistakes drag through many moves, so I just look at graph shape and then investigate few moves before and after those bigger steps.

It happens sometimes that both me and my opponent miss an important move. In the chart you see big mistakes from both players at each move until one of us plays something that settles that situation.
Just one key point in game but several mistakes in the chart. This could take all three biggest points mistakes from both players.

Bots can help, but I think we still need humans for a meaningful highlight feature. :grinning_face_with_smiling_eyes:

2 Likes

By score makes more sense. But it’s a preview of supporter perk. It makes sense it’s not as useful as might be also.

That kind of makes sense when considering the depth of the variations, number of them compared to the number of moves in the game, and compared to the recent interactive ai analysis.

That said, I think having a reasonably good free version is probably more enticing to sign up for the paid version than a terrible free version. Eg, the free version of Spotify is absolutely terrible, particularly so on mobile, and it seems like they want to take away more and more features as time goes on. I actually would prefer not to give Spotify any subscription ever because of how terrible they intentionally made the free version (of course they’ll still get some ad money, but half the ads are just for the premium version).

Maybe it works for some people though, you show them a terrible version of the product in the free version and it prompts them to pay for the full version.

Except it’s the same for supporters so that’s not really the point: you only get three uninteresting key moves highlighted.

The difference of course is that supporters can go and search themselves for the interesting moves in the game, but there is a value in having those already highlighted by the AI to save time (like AI Sensei does).

2 Likes

This topic was automatically closed 91 days after the last reply. New replies are no longer allowed.