Promotion/Demotion and Yusho discussion Nagoya 2021

[Kashunowaka 296]

4 minutes ago, Seiyashi said:

I'm not so sure it's quite that straightforward.

Interestingly, 2003 (the year of Asashoryu's promotion and Musashimaru's retirement) saw quite the dip. You would have thought that with Musashimaru mostly sidelined and Asashoryu making a charge up the ranks, there would be a higher KK % there.

The other three points of interest are the troughs of 1998 (the promotion of Wakanohana) and 1964 (the promotion of Tochinoumi), and the peak of 1970 (the double promotion of Kitanofuji and Tamanoumi). It sort of boggles the mind that there's an S/K KK boost when there's two ozeki kicking ass and taking names.

Perhaps combining the two - S/K + joi - gives a better picture? The eye test suggests that they seem to complement each other - 1970 has a trough in the joi, but that's not so clear cut for 1998 and 1964.

The global peak is in 1969. Can be hard to see that from the graph.

[TakanohanaFan 2]

It's a 8 man sanyaku. Thus the other 34 men man the ranks of 17 maegashira.

I already know how the joi stands. GL for all GTB competitors.

[Halian 7]

Attached: my first pass at 2021.09 GTB using Hālian's GTB Secret Sauce™, that is to average out the last ten instances of that rank-record combination, and then fill the bandzuke based on the results.

[Seiyashi 3,912]

38 minutes ago, Halian said:

that is to average out the last ten instances of that rank-record combination, and then fill the bandzuke based on the results.

How far back do you go? The data scientist in me says you ought to weight more recent instances.

[Halian 7]

1 minute ago, Seiyashi said:

How far back do you go? The data scientist in me says you ought to weight more recent instances.

However far back the last ten instances go, but never earlier than 1935.

[Yamanashi 3,359]

1 minute ago, Seiyashi said:

How far back do you go? The data scientist in me says you ought to weight more recent instances.

Agree.  I don't do GTB at all, but I have noticed a consistent difference in results before, say, 2010, and after 2010.  If you have a rare combination of rank and W_L and have to go back to the 90's for your 10 examples, you may end up overbalancing your average with outdated "NSK banzuke theory."

[Jakusotsu 5,294]

I tried that kind of method once, scored a huge success in my debut GTB basho, but that turned out to be an equally huge fluke.

It's inherently bound to fail because you always have to take all circumstances into account, which probably only a highly sophisticated neural network might be able to solve satisfactorily.

Edited July 21, 2021 by Jakusotsu
typos

[Seiyashi 3,912]

26 minutes ago, Jakusotsu said:

It's inherently bound to fail because you always have to take all circumstances into account, which probably only a highly sophisticated neural network might be able to solve satisfactorily.

[Gurowake 3,730]

6 hours ago, Halian said:

Attached: my first pass at 2021.09 GTB using Hālian's GTB Secret Sauce™, that is to average out the last ten instances of that rank-record combination, and then fill the bandzuke based on the results.

Good luck!  You'll need it.  Here's one hint about things that this method fails to take into consideration, something that's been mentioned around here multiple times: Ichiyamamoto will almost certainly be ranked higher than anyone being promoted from Juryo.

Edited July 22, 2021 by Gurowake

[Qwerry 25]

3 hours ago, Halian said:

Attached: my first pass at 2021.09 GTB using Hālian's GTB Secret Sauce™, that is to average out the last ten instances of that rank-record combination, and then fill the bandzuke based on the results.

 

too many blatant errors.
there is too little data to get a good model from scratch.
It was necessary to study the area and adjust the formulas and coefficients before launching

[Godango 835]

I've tried this approach before. My GTB record speaks for it's appropriateness.

[Asojima 2,848]

5 hours ago, Jakusotsu said:

 

It's inherently bound to fail because you always have to take all circumstances into account, ...

a.  Heya affiliation of the members of the Banzuke committee.
b.  Media hype given during the previous basho.
c.  Number of teenagers who held up rooting banners.
d.  Number of appearances on the spirited rikishi list.
e.  Highest previously held rank.
y.  Current rank.
z.  Win/loss record

Edited July 22, 2021 by Asojima

[neonbelly 226]

Sumo fans: run monte carlo banzuke simulations overnight.

JSA banzuke committee: throw chicken bones and laugh.

[Tsuchinoninjin 1,151]

Yeah I used to do simulated annealing based on historical by-the-numbers, it wasn't great, but it was sure fun to watch the names bubble around the banzuke violently.

[Tigerboy1966 1,220]

5 hours ago, Tsuchinoninjin said:

simulated annealing

I wonder what "simulated annealing" is? Maybe Wikipedia can help... ah, here we are... "Simulated annealing is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem".

Well, that's that sorted out!

[Jakusotsu 5,294]

6 hours ago, Tsuchinoninjin said:

Yeah I used to do simulated annealing...

Thanks for the trip down the memory lane. I vaguely remember holding a seminar lecture about that some 25 years ago in cluster physics. Of course I had no clue what I was talking about...

[yorikiried by fate 1,868]

6 minutes ago, Jakusotsu said:

Thanks for the trip down the memory lane. I vaguely remember holding a seminar lecture about that some 25 years ago in cluster physics. Of course I had no clue what I was talking about...

Hadn't we all, hadn't we all...

[Seiyashi 3,912]

9 hours ago, Asojima said:

a.  Heya affiliation of the members of the Banzuke committee.
b.  Media hype given during the previous basho.
c.  Number of teenagers who held up rooting banners.
d.  Number of appearances on the spirited rikishi list.
e.  Highest previously held rank.
y.  Current rank.
z.  Win/loss record

You missed out the amount of sake served at the meeting! 

[since_94 647]

Do forum members think Hakuho actually appears bored and looks at his Rolex often as he's being reprimanded by Hakkaku and the other oldies of the YDC? Or does he feign some level of interest in the proceedings and solemnly promise to reflect, take the criticism to heart, and mend his ways?

I hope it's the former, but I suspect it's the latter.

Hail the GOAT!

[botev1921 33]

19 hours ago, Halian said:

Attached: my first pass at 2021.09 GTB using Hālian's GTB Secret Sauce™, that is to average out the last ten instances of that rank-record combination, and then fill the bandzuke based on the results.

Lovely to see such effort. I felt Terasawa should be promoted at the expense of Daishoho, but what can you do? :)

[Tsuchinoninjin 1,151]

7 hours ago, Tigerboy1966 said:

I wonder what "simulated annealing" is? Maybe Wikipedia can help... ah, here we are... "Simulated annealing is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem".

Well, that's that sorted out!

I dunno why math related articles on wikipedia are so obtuse. Encyclopedias are suppose to be an introduction to a topic someone knows nothing about, right? But the math articles on wiki only make sense if you've already understood the concept before, which is pointless.

Anyway, simulated annealing is like baking brownies. First they are pretty liquid and can bubble up or mix due to convection. But at some point they set into a solid. Simulated annealing allows the function you are trying to optimize (in this case the banzuke) to move to a 'more wrong' state early in the optimization, but near the end the banzuke is only allowed to become better. This is so the banzuke doesn't get trapped in a state where any position swap makes it 'more wrong' but there might be a better optimum out there. I was determining what was 'more wrong' by checking the amount of promotion/demotion for certain records against historical data. You also mark it 'more wrong' if something silly happens like a M14 9-6 record passing a M13 9-6 record.

Starting from last basho's banzuke makes the annealing take too long, so I would randomize it a bunch of times using Monte Carlo to get a reasonable 'wrongness score'. Monte Carlo is like shaking the shit out of the liquid brownie goop so its mixed well enough.

[Seiyashi 3,912]

How many people here have actually tried building a neural network for the banzuke?

I would be interested in doing one as hands-on learning about NNs myself, but I got jammed very badly in terms of data representation of bouts. (But then again I could just have been trying to do too many things at one go).

[Reonito 989]

51 minutes ago, Seiyashi said:

How many people here have actually tried building a neural network for the banzuke?

I would be interested in doing one as hands-on learning about NNs myself, but I got jammed very badly in terms of data representation of bouts. (But then again I could just have been trying to do too many things at one go).

I haven't gotten past the annoyance of scraping past bout data from the db...

[Seiyashi 3,912]

2 minutes ago, Reonito said:

I haven't gotten past the annoyance of scraping past bout data from the db...

That isn't the bad part (slightly biased, speaking as someone whose first job was to attempt to scrape over a million records), but once you have data you have to decide how to store it, which is the problem. I couldn't decide if I should store it on a per-rikishi basis or on a basho basis.

Per-rikishi makes more sense, but I really didn't want the NN to screw up and learn promotion trends based on rikishi rather than result. Unless there's a solution to that in terms of restraining what the NN learns from.

If we go deeper into this is it still technically on topic, or should we open up a new thread in Ozumo Discussions?

Edited July 22, 2021 by Seiyashi

[Sue 463]

1 hour ago, Seiyashi said:

How many people here have actually tried building a neural network for the banzuke?

Building a machine learning model for A) the banzuke, and B) predicting career trajectory for rikishi was on my "to do" list when I was the VP of Engineering at an AI company, but since I left, I haven't had time.