Warrior General Games

One thing I've been doing this whole time is assuming that there is no "intercept" that is, its a linear combination without any +x at the end. I guess I could try adding that in

Guard:

AVEmaj 0.448147682
AVEmin 0.635246194

awful residuals

Guard w/ intercept:

Intercept 3.221273681
AVEmaj 0.381534855
AVEmin 0.654460054



Center:

Intercept -16.84303266
AVEmaj 0.538995537
AVEmin 0.773301063

rofl...



I don't think that's the correct model. It's basically forcing the majors to have equal weights.. Though it would have been easier for Bort to implement, it doesn't seem to work so far.


Though, I need a lot more individual position data to even think about the overall bars. Only about 15 points on each.

Originally posted by Octowned
As expected, this is just the solution of the last problem multiplied by the number of attributes..

OT:

AVEmaj 0.6910061
SUMmin 0.411311757

Which is odd to sum to 1.1.



lol I should have seen that coming.

Can you set the constraint that these must sum to 1?

Also - if you have the data points in a txt file (or even something like excel), I wouldn't mind playing with them, but I am too lazy to type them all in myself.

Originally posted by Octowned
Feel free to try the formulas suggested...

STR 0.2
BLK 0.45
AGI 0.075
VIS 0.075
CON 0.1
TKL 0.1


Pos Level Formula Actual

C 37 62.34 63
C 30 61.84 63
C 37 59.07 60
G 40 57.13 57
G 38 56.51 57
OT 34 53.76 54
C 20 52.24 54
G 21 48.11 51
C 4 26.44 30
C 4 23.39 27
C 4 (Slowbuild) 21.66 24

It seems to work better at higher levels.

lvl 16 guard
open build
stats w.o equip, no SA
39 OVR, 48 BLK

Strength: 70.05
Blocking: 51.05
Speed: 8
Tackling: 13.75
Agility: 44.75
Throwing: 9
Jumping: 8
Catching: 8
Stamina: 26.75
Carrying: 8
Vision: 13.75
Kicking: 8
Confidence: 24.05
Punting: 9

Setting the constraints to sum to 1, I've never done such a problem, and can't imagine why it would give better results. It would basically be reducing a 3 variable problem into a 2 variable problem...

AX + BY = C we're using now, otherwise its

AX + B(1-X) = C
AX + B-BX = C
(A-B)X = C-B

So I can rearrange the data to take the major average minus the minor average, and plot it against the scouting bar minus minor average, lets see...


DIFF 0.833478393
So I'll try fitting it with .83 and .17...

50.47755 51 good
30.689 30 good
36.758 36 good
52.23344 54 eh..
16.478 15 a bit better than before..
16.478 15 same
50.52392 54 was iffy before, now awful
46.87644 45 worse
45.22438 45 good
45.48134 48 bad
41.63292 42 good
42.28761 42 good
49.06256 51 eh...
45.04492 45 good
46.87644 45 bad


This gives a worse fit than before. It just takes a degree of freedom out, which will always make a model fit worse. The usual question with dropping degrees of freedom is correlations or overanalyzing and whatnot. I don't think there is an argument for that.



I'll PM you, I'm doing this all in excel, so I'm not even trying error diagnostics (checking normality) until I come up with a good model.


I still think we need about 3x as much data to do much, so keep pouring it in!

Originally posted by Kayoh
http://goallineblitz.com/game/player.pl?player_id=949201 <-- 63.2 natural AGI
http://goallineblitz.com/game/player.pl?player_id=148920 <-- 50.78 natural AGI

Great data point on the top. The low strength came out well, the formula undershot a little bit, AND the current formula favors blocking. If anything, this makes a very strong case that blocking matters a lot more than strength in the blocking bar, and that possibly agility is underrepresented at .075

new confidence ranges...


0.167403646 0.245153836 Strength
0.399552946 0.495903317 Blocking
0.054901222 0.1210146 Agility
0.005719398 0.103135428 Confidence
0.05328407 0.135461253 Vision
0.032128652 0.218528734 Tackling


Getting much more accurate on the strength and blocking.
My current proposed, though slightly off, model is

0.2 STR
0.45 BLK
0.1 VIS
0.1 AGI
0.1 TKL
0.05 CONF


I hate that tackling in there. Speed and Stamina are still coming up with nothing.


Running it without tackling gets..

0.160420562 0.242535461 Strength
0.408570938 0.509230743 Blocking
0.073427593 0.13778537 Agility
0.026332255 0.124267152 Confidence
0.064375982 0.149289557 Vision


What I like most is that strength and blocking are CONSISTENTLY coming up as 0.2 and 0.45, and agility and vision are consistenly the same. Confidence seems significant, but without it..


0.247751747 Strength
0.513456816 Blocking
0.157044566 Agility
0.182407674 Confidence


Obviously starts to shade up the str/blk.


What I like most about including the tackling is that the coefficients sum to 1, whereas taking it out, they sum to .95. What still gets me is why stamina isn't there.

What this leads to, though, is the confirmation that stamina is ONLY an attribute modifier, through energy. Why wouldn't bort include stamina in your bars? Because it will change how you play on the field, not how good your player is on play 1. Just a thought...

heres my level 21 guard http://goallineblitz.com/game/player.pl?player_id=860481

Strength: 56.37
Blocking: 72.37
Speed: 9
Tackling: 15
Agility: 38
Throwing: 10
Jumping: 9
Catching: 9
Stamina: 31.68
Carrying: 8
Vision: 18
Kicking: 9
Confidence: 28.37
Punting: 9

overall bar is 39
blocking bar is 54

when i did the original equation, i got 52.37 for the blocking bar; thought maybe i could save you a bit of time

That last one is a nice one to have as it has str much higher than blk. Edit: I mean Blk > Str!

FYI Octowned I created a player, threw 15 into stamina and the overall bar went up.

http://goallineblitz.com/game/player.pl?player_id=1131735

Looking at the data, it is more complex than a simple weighted major/minor attributes, so a regression analysis is necessary - over to you!

As I keep adding data, the str/blk/agi/vis all stay the same. Now all of a sudden, tackling is out and speed is in it's place. I think the biggest problem is the lack of variation with tackling.

Finding an o-linemen with over 50 tackling will be hard to come by though.


The str/blk/agi/vis relationships are all but confirmed, though. I have many outliers where the formulas continue to hold, such as all agility OTs, no strength OTs, etc.

Originally posted by Octowned
As I keep adding data, the str/blk/agi/vis all stay the same. Now all of a sudden, tackling is out and speed is in it's place. I think the biggest problem is the lack of variation with tackling.

Finding an o-linemen with over 50 tackling will be hard to come by though.


The str/blk/agi/vis relationships are all but confirmed, though. I have many outliers where the formulas continue to hold, such as all agility OTs, no strength OTs, etc.

When I retire my DE I'll build an OT and throughout the D-Leagues I'll only pump up tackling.
Just make sure this thread is still bumped then so i remember

.


BTW, I think I had the labels on vision/confidence switched. I've been moving a lot of columns around. I double checked my data, and its all fine (all the 0.25 decimals for the Gs are on vision, the .0x, .6x, .3x on confidence), and all my players line up.


That doesn't change the accuracy of my results, but it appears confidence is the one with agility, better than vision.


Doing some "overall" stuff, they are very different by position, but a few common themes pop up..


strength, agility, stamina and speed have an impact. I'm a bit iffy about that speed.

blocking does less than strength, agility, and stamina, for the overall bar. vision doesn't do much at all, and tackling does NOTHING.


Also, when I took off the intercept=0 constraint, it came out to a very even 3.0, which I thought was interesting, and worth investigating. If you had all 0s, would it read as a 0 or 3? Maybe bort needed an image width, so 3 is actually where you start. This slightly curves the data down.


Tackling is getting less and less of an impact as my data set grows. Strength and blocking still haven't strayed from .2 and .45, and agi/conf are still equal, usually right at or below 0.1. What keeps changing is the stamina, tackling, vision, floating between 0 and 0.07, which does enough to throw things off (figure if you've got 40 sta/vis, you're talking about a contribution of between 0 and 4 from these, so these coefficients are hard to estimate and can cause outliers).



Most discrepancies I'm finding are due to high strength. However I have some no strength builds who are doing fine under the model.

My last step once I get about 150 data points (at 64 now), will be to take this to a real statistical software. What they do is take out one variable at a time, whichever has the least impact, until taking one out drastically changes the results. It would give much more accurate results than my current bouncing around, and it helps to pick up the cross-dependencies between the predictors.

If I'd have paid attention in my regression class, I could help you out with this in SPSS...

I could ANOVA the hell out of it, though. Too bad it'd say nothing.