Unit Cost Equation

[sushiboat]

The coefficients are derived from the data, so you can't just change the
positive/negative sign without hurting the equation's performance. I could
take out all but one of a set of highly correlated predictors. However, all
predictors in the equation are making a highly significant contribution.
Perhaps a better solution would be to create a composite variable. A new
variable called Bigness could be the sum of the standardized scores of Size,
HP, and Strength. A little information would be lost by not keeping them
separate, but not much.

Edit: EvilDave, I don't have the output in front of me. (It's a Mac Classic app,
and I don't want to fire up the Classic environment just now.) I do remember
the predictors I kept having p-values that were very low, less than .0001. I
removed the predictors with p-values of .10 and higher from the three most
recent regressions. It just happened that there was a huge gap between the
predictors that weren't statistically significant and the ones that were.

[BigDaddy]

I DO realize that forcing the numbers will affect performance. But, building small strong units is trouble right now. . . SO, I would see if a Str and size composite (excluding HP) would be effective, it is a very good idea. I'd like to see HP remain alone, if possible. Then, little information would be lost.

Note to modders (who like this sort of thing), YES you can easily use this equation to min/max like never before and claim fairness. . .

[Scott Hebert]

Claim, yes. Hide behind for long, probably not.

[Evil Dave]

Scott Hebert said:

Unfortunately, I don't very much at all about regression testing, so it's hard for me to grasp much of the jargon.

No problem. Say you wanted to model the damage the units do in Dom2. You'd probably guess it was some function of strength and weapon damage.

So, you'd make a model like:

damage = A * strength + B * weapon damage

"strength" and "weapon damage" are the regressors (or contributors). A and B are their coefficients. The regressors also each have p-values, which is the likelihood they're due to chance -- 1%, 5%, and 10% are typical cut-offs for scientific work. The whole model has an R-squared, which is the fraction of the variation of the data it explains. R-squareds range from 0 to 1.

So, in this toy example, strength and weapon damage would probably have very small p-values (probably less than 0.01), but the R-squared would probably be pretty low, because the 2d6oe tends to swamp out the effects of the regressors.

Hope that helps.

[sushiboat]

Scott, I would be interested in seeing your figures for magic paths and cost.
Unless you are using the Unit Database spreadsheet and have numerical
entries for all mage units, I wouldn't be able to plug your numbers in directly
to my approach. As I said, I do plan to add magic paths to the analysis, but
no time for that today.

[Scott Hebert]

Dave: So... the p-values are like that 95% value for confidence testing, and the R-squared value is how much of the sample that the equation explains.

It sounds like math modeling and differential equations to me.

Sushi: If I had the Unit Database, I could supply that for you. It would be easiest if it were in Excel. It uses the Illwinter cost method.

[BigDaddy]

Scott,

Its here (and in Excel Format)

http://www.dominions-2.org/files.htm#docs

Under "Unit DB - by Edi"

[Scott Hebert]

Downloaded. I hope to have the magic path costs entered sometime today... if my New Era Pangaea game doesn't suck any more time away...

mmmm... Dryad Hoplites....

[Scott Hebert]

Oh, question. Where do you want me to put the Magic path cost, Sushi?

[BigDaddy]

Hmmm. . . I think Ill work on a raw data Unit spreadsheet. . .