what is wrong with the 4-5 and 3-5 joseki?

[ez4u]

Dave, your results surprise me in one way: I thought Black's win rate in modern play is greater than 50%. Do you know if that's right, and if so, what makes your results look different?

I don't have a source off hand, but I think I had read somewhere to that after komi went up from 5.5 to 6.5/7.5 it went from black winning slightly more to white winning slightly more.

Mef is correct. The sample I used was 25K+ games from 2002. Black winning percentage was 49.1%.

A quick (actually slower ) check using the real komi (sgf KM tag) shows:

Code: Select all

Komi               Black winning        Games
4.5                55.1%                2,543
5.5+2.75           52.8%               28,729
6.5+3.75           49.2%               24,481

There seems to be a complication however. Note the combination of different komi = Japanese and Korean games on the one hand (e.g. 6.5) and Chinese games on the other hand (e.g. 3.75). I always use the period since 2002 as proxy for the "6.5 komi era". However, the 3.75 Chinese komi is the equivalent of 7.5. I have seen it written that under Chinese counting there is no difference between '3.25' and '3.75'. The stats say that there is a difference between '6.5' and '3.75' though!

Code: Select all

Komi               Black winning        Games
5.5                52.7%               21,440
2.75               53.2%                7,289
6.5                50.3%               16,416
3.75               47.1%                8,065

[HermanHiddema]

I have seen it written that under Chinese counting there is no difference between '3.25' and '3.75'. The stats say that there is a difference between '6.5' and '3.75' though!

Actually, the reason that the Chinese moved from 2.75 (5.5) to 3.75 (7.5) is because there is usually no difference between 2.75 and 3.25 (due to the score difference almost always being odd). So the difference between 3.25 and 3.75 is in fact quite significant.

[hyperpape]

Today I have learned something, then. Thanks.

[gowan]

We know that from time to time there are fads in professional go openings. For example, in the late 1970's into the 80's there was a big fad for the Chinese style opening, and I have the impression, not backed up by any data base research, that in that period Black had a significantly high win percentage. However, eventually people learned how to play White against that opening and the Black win percentage dropped. In fact, now the so-called high Chinese style is completely out of fashion. Same thing with the san-ren-sei in its time. Lately the mini-Chinese has been popular. I suspect that these fads start because a few extraordinarily talented players have a lot of success with particular openings so other players imitate. Certainly Takemiya was very successful with the san-ren-sei, and Kato with the Chinese style. Anyhow, my question would be whether win/loss statistics are vulnerable to bias due to these fads. It might happen that an opening, like CHinese style, becomes popular and successful but once people figure out how to play against it it isn't played as much and the win/loss imbalance doesn't get corrected.

[wineandgolover]

Back to the original question... Nothing is wrong with them.

Winning percentages might be lower, but that is explainable because amateurs aren't nearly as good as pros at understanding and using thickness, which means it's harder to catch up when cash-players exploit the opening's weakness.

But if your goal is to learn, then these are great opening moves. Plus tey are fun. I play 3-5 in every even game. Of course, I suck at go.

[SmoothOper]

Winning percentages might be lower.

Didn't the statistics just say that white 3-5 has a winning record in Chinese counting with modern komi?

... amateurs aren't nearly as good as pros at ...

I would be interested in seeing the statistics categorized by rank for the opening to see if the winning percentages are indeed affected by rank, to answer the conjecture that there is a difference between amateurs ability to utilize a specific opening.

I would suspect that beginners and weak amateurs wouldn't do as well using 3-5 because there isn't much information regarding how to play those openings, but that there would be substantial variation in play due to so many other issues during play. I would expect a bit of an advantage for strong amateurs playing 3-5 due to the not mastering every opening and not consistently playing competition that uses it, and this effect would diminish as people get stronger and have encountered more of the openings.

[yoyoma]

I read somewhere on senseis (can't find it now) that the stats on joseki are:

3-4 47%
4-4 46%
3-5 2%
4-5 2%
3-3 2%
other 1%

well, what's wrong with 3-5 and 4-5?

Why didn't you include 3-3 in your question?

[ez4u]

We know that from time to time there are fads in professional go openings. For example, in the late 1970's into the 80's there was a big fad for the Chinese style opening, and I have the impression, not backed up by any data base research, that in that period Black had a significantly high win percentage. However, eventually people learned how to play White against that opening and the Black win percentage dropped. In fact, now the so-called high Chinese style is completely out of fashion. Same thing with the san-ren-sei in its time. Lately the mini-Chinese has been popular. I suspect that these fads start because a few extraordinarily talented players have a lot of success with particular openings so other players imitate. Certainly Takemiya was very successful with the san-ren-sei, and Kato with the Chinese style. Anyhow, my question would be whether win/loss statistics are vulnerable to bias due to these fads. It might happen that an opening, like CHinese style, becomes popular and successful but once people figure out how to play against it it isn't played as much and the win/loss imbalance doesn't get corrected.

This raised an interesting point that I can't really deal with in depth (at least yet). But I have wanted to get back to it for some time with a little factoid and to raise the question whether it is the statistics that are vulnerable to bias or just our memories?

The point is "...and Kato with the Chinese style...". This caused one of those "hmmm..." moments that tend to send me scurrying off to GoGoD/Kombilo, because I recalled that in the 2nd Kisei every game began with the Chinese.

Sure enough we see that Kato was quite successful playing the High (sorry I was too lazy to be thorough) Chinese back when it dominated the Japanese tournament scene, relative both to his overall results with Black and to other players' results with the same fuseki:

Code: Select all

                         Games B Win%  W Win%
All High Chinese (HC)    1,191  51.7%   47.8%
All Kato as Black          647  61.1%   38.5%
Kato as Black with HC       56  71.4%   28.6%

So maybe we have something! But wait. We remember specifically that Kato liked to play the HC. However, at that time so did everyone else (OK, not quite everybody but many of the top-flight players that Kato was facing, e.g. Fujisawa Shuko). What about Kato's impact on the statistics from the opposite side? How did he do against the HC?

Code: Select all

                         Games B Win%  W Win%
All High Chinese (HC)    1,191  51.7%   47.8%
All Kato as White          636  46.4%   53.6%
Kato as White vs HC         37  37.8%   62.2%

This tells us that Kato was just about as much better than his average when playing against the HC as he was playing with the HC! It also shows us that our selective memory is based on only 56 of the 93 games that Kato played involving the HC.

If we compare the stats on the HC with and without Kato we can see the effect he had on the overall stats:

Code: Select all

                       Games  B WinS   W Wins  B Win%  W Win%
All High Chinese       1,191     616      569   51.7%   47.8%
HC ex Kato             1,098     562      530   51.2%   48.3%

So we see that he did have an impact, boosting the Black winning % by 0.5 percentage points. As to the significance, I will leave that for the reader. (I did the analysis just for the fun of it.) As always YMMV.

Note: the "All" figures here are filtered for the period 1964-2005 to match Kato's professional career.

[Kirby]

...Anyhow, my question would be whether win/loss statistics are vulnerable to bias due to these fads. It might happen that an opening, like CHinese style, becomes popular and successful but once people figure out how to play against it it isn't played as much and the win/loss imbalance doesn't get corrected.

I think this is a fantastic point. My take on this is that this is simply a result of the model we are using when we simply take a win-percentage based on a given opening. But really, I think that was is happening in the general case with this type of analysis is a maximum-likelihood estimation based on the sample data.

Currently, there is only a single input variable used in our model: the opening type. To spice things up, we could use a number of different input variables: the opening type, the player, the time period, etc.

With whatever model you select, you could then use something like logistic regression, for example, to predict the probability of success given this set of input variables, eg. "If i use the chinese opening, my name is kobayashi, and the year is 1980, what's the probability that I'll win this game?". Contrast this with the current approach of a single variable, where you're saying something much more general, "If I use the chinese opening, what's the probability that I'll win this game?"

I don't think that we can say absolutely that the more general model is bad, however, sometimes having additional input variables like this can help to form more accurate hypotheses given new data.

[skydyr]

I think this is a fantastic point. My take on this is that this is simply a result of the model we are using when we simply take a win-percentage based on a given opening. But really, I think that was is happening in the general case with this type of analysis is a maximum-likelihood estimation based on the sample data.

Currently, there is only a single input variable used in our model: the opening type. To spice things up, we could use a number of different input variables: the opening type, the player, the time period, etc.

With whatever model you select, you could then use something like logistic regression, for example, to predict the probability of success given this set of input variables, eg. "If i use the chinese opening, my name is kobayashi, and the year is 1980, what's the probability that I'll win this game?". Contrast this with the current approach of a single variable, where you're saying something much more general, "If I use the chinese opening, what's the probability that I'll win this game?"

I don't think that we can say absolutely that the more general model is bad, however, sometimes having additional input variables like this can help to form more accurate hypotheses given new data.

It's important to note that more information is not necessarily better, as you still have to separate the wheat from the chaff. Malcolm Gladwell has a bit written about this here, if you go down to the answer to question 2.

The real question, I suppose, is which games are representative for statistical modeling of the openings in question, and which are not. I am not sure we will ever have a perfect answer to that.

[Kirby]

...

It's important to note that more information is not necessarily better, as you still have to separate the wheat from the chaff. Malcolm Gladwell has a bit written about this here, if you go down to the answer to question 2.

The real question, I suppose, is which games are representative for statistical modeling of the openings in question, and which are not. I am not sure we will ever have a perfect answer to that.

I agree that more information is not necessarily better. The only thing I might say differently is that the real question is which attributes of games are representative for statistical modeling. A given game could contain many pieces of information that may or may not be useful to a model.

[SmoothOper]

Its too bad the openings I play only have one or two pros to do statistics on, and for *most* of their games they could probably pass once or twice and still win.