Official: SSBPD unsupported; source code released.

[Purpletuce]

Perhaps the system could also reward victories over diverse regions, ex: a player who beats 10 1500 ELO players all in one region is rewarded less if he beats 10 1500 ELO players from many regions. Perhaps the ELO should also put arbitrary rankings into the region, based on how players from various regions compete, also weighting those players relative to their skill level. For example, region A and region B have 5 players, ranked numerically 1-5. Player 1a plays player 1b, the system begins to assume the winner comes from a stronger region. This of course would be very susceptible to skewed data, especially considering that usually better players travel more. I think that over time, the ELO could probably provide a decent ranking system for regions though...

 

[Nintendude]

In order for Elo to be as objective as possible there needs to be none of the adjustments that people are suggesting. I think you all got the wrong idea because the old system was just lacking enough data. Had the site not gone down, Kels' rating would have been brought down to earth by his performance at nationals this past fall and winter. In addition, the overall spread of points would have been greater which would have also alleviated some of the inaccuracies.

 

[Terral]

To add to the discussion as far as the old SSBPD site goes, it was using a simplified somewhat-incorrect Elo algorithm and using tournaments instead of individual matches to calculate changes. It's possible that using a standard more fine-grained Elo implementation would make a pretty substantial difference even without trying to add in custom tweaks.

My view on tweaking the algorithm is really just that people shouldn't be taking the numbers that seriously. Elo is a system for objectively rating players based solely on their performance. We also can't expect TOs and players to keep track of character and stage selections for individual matches, the most data we can expect to have is wins and losses. If we were to include characters/stages in the equation, aside from being a mathematical challenge that I don't think anyone would be up for, we would need to enforce that for every set in the database, throwing out any set that didn't have every stage and character choice recorded (we also couldn't immediately assume someone is using their main) else the results would be skewed. It's easy to suggest hypothetically including multiple factors to fix a supposed problem with the algorithm, but it isn't easy to create a system that is actually able to achieve the desired results. As has been said, we are better at determining whether a match could go either way or if there is a nearly certain winner than an objective computer is. Ratings are just about fun and competition and graphs and wonderful numbers.

 

[Doomolish]

I'm sad to read this project has ended. I was really interested on uploading my regional results.

 

[Fortress | Sveet]

If you have tio files, for the moment you can email them to me at sveetx@gmail.com. I can't say exactly when, but soon the SSBPD will be revived as the Smash Boards Tournament Database

 

[Awstintacious]

So sad that the website went down, that was like the only evidence of the my first tournament at CEO

 

[Mahie]

http://www.aligulac.com/

Meanwhile, the SC community came up with something like this.

 

[PGH Carroll]

#NeverForget

 

[Fortress | Sveet]

Don't worry carroll, people are working on this project again

 

[Mahie]

So where do things stand at the moment?

It's been forever.

 

[Fortress | Sveet]

I'm not exactly sure. I did some work last year to extract data from tio files and put it into a database. I eventually stopped working on the project because of the tier list project taking up my time, as well as other things going on in my life. Recently Dubforce has expressed interest in moving the project forward, so i've given him all my source in hopes that this can be completed. Not 100% sure what his plans are, but I hope he can pull through

If nothing else, my motivation for this project has been going up again recently. If it wasn't for the XenForo interfacing problems, I would seriously consider finishing this in my spare time. But I seriously dislike how poor the documentation is relative to the amount of overhead that has to be done. I'd probably just make my own UI if I went back to work on it (but that doesn't sound very appealing either).

 

[Terral]

I don't know who Dubforce is, but I had picked up working on the project again back when I had free time. I didn't get all that far before my son was born, so there hasn't been much progress made recently on my end.

 

[thespymachine]

*giddy giddy giddy*

As an ELO nerd, what would the 'base' rating players will get when introduced into the system? What the earliest tournament being used to start the system? What kind of K-factor are we looking at? (when I did Vegas Melee's PR Elo I gave K a 50 for our monthlies, and 25 for in-between tournaments - I realize now, 2 years removed from my work, that I should have done the ratings throughout the whole rating period instead of tournament to tournament, but anyways) Will there be a points gained/lost for attending/not attending? (it would seem easier to implement a loss of points over a certain amount of time than to have an accurate/viable point reward for playing - or am I wrong?)
[I did all my work in Excel, btw. :| ]

This is sort of directed to anyone willing to discuss. lol

Edit - 5 days later:
I studied a bunch of what I did with my ELO system for Vegas, as well as experimented with and learned a few new things.
- Excel isn't made for a 'holistic' ELO ratings system. Because there are, seemingly short, limits to the length of formulas, you have to do a lot of mitigating/buffers with various additional arrays. Which makes doing ELO for a single ranking period, instead of tournament-to-tournament, more work.
- I implemented ideas from chess ratings with K-factor, where 2400+ get a k-factor of 16, and then I just went by intervals of 300pts and increased the k-factor by 8: 2100-2399 = 24, 1800-2099 = 32, etc. This, I admit, is not the greatest of ideas, because 1) it creates a linear degradation of k-factor, when a logarithmic one would be better suited for player skill [more players are average than players are elite] and 2) I should be implemented a k-factor determined completely by the player's rating instead of which interval/tier they place in.

Solutions?:
1) The first steps would be to determine the interval lengths and the differences of k-factor per interval.
With a little thinking, I believe a logarithmic degradation of interval lengths will be good to.
The median rating for USCF is around 600, so I think the first interval should be 100-599. From there, we do an interval of 600 and reduce the interval by 100 for each interval after to preserve the 2100-2399 interval. Getting:

100-599
600-1199
1200-1699
1700-2099
2100-2399
2400-2599
2600-2699
2700-2799
2800-2899

​

With k-factor I'll keep 2400 at 16, and 2100 at 24 - with increase in pts, decrease k-factor by 1/2 for the previous interval; and with decrease of pts, increasing k-factor by 1.5 of previous interval.
So:

100-599: 112 (rounded up)
600-1199: 71
1200-1699: 54
1700-2099: 36
2100-2399: 24
2400-2599: 16
2600-2699: 12
2700-2799: 10
2800-2899: 9
2900-2999: 8
... ... ...
​

I dunno. Seems wrong.

2) Or, base it off of a statistician's calculation that 24 should be the k-factor of 2400+, put the max k-factor at 100 (@1 pt being the minimum rating), keep intervals at 300, and see what happens. lol (it worked out that if you start the k-factor differences at 13 and subtract 1 from it for ever interval, 24 ends up being the k-factor for 2400, er 2401 in this case)

rating pts - k-factor
1 - 100
301 - 87
601 - 75
901 - 64
1201 - 54
1501 - 45
1801 - 37
2101 - 30
2401 - 24
2701 - 19
3001 - 15
3301 - 12
3601 - 10
3901 - 9
... ... ...
5401 - 1
​

Yeah, I don't know what I'm doing anymore. lol

​

​

- Elimination-style tournaments don't seem to do well with ELO because better players will always get more chances to increase their score. Round-robin and swiss are far better suited for ELO. The whole goal is to give each player equal opportunity to perform (Even a slight inclusion of DE/SE ruins things: a player who makes it out of RR pools but loses their two games in DE will lose more total points than someone who didn't make it out of pools, simply because the former player had more opportunities to lose). That's why I'm glad SWF is basing their new ranking system of of ATP's: http://smashboards.com/threads/announcing-smashboards-rankings.345659/