but any planes attacking Rome still attack. Which sucks for them.
Statistical Model for Predicting the Winner
-
I’m guessing you could do a decent job of predicting who is going to win based on the total unit value of each side, and/or the IPC income.
This is what I did in 2007 with Revised. Has anyone tried this? Or want to work on it with me? I like SPSS.
-
I read the post, pretty cool… id be interested in seeing it done for G40, although id think it would be significantly harder due to the many game mechanics, many different powers that can surrender (Italy, France, China,), and strategy that could be a long term investment such as sea lion.
-
It might surprise a lot of people, but a lot of the so-called strategic positions of countries are mostly irrelevant. Or perhaps more accurately they can be controlled for as the relative strategic position of the Axis vs Allies is approximately the same in most games.
Also in almost all cases, if you have a better than average position, it will be seen very soon with a reward in terms of territory or unit losses from the other side. So the total unit value, and total income catches this - albeit with a small lag.
-
Yes, I agree in the long run. But say you took are poised to take London, but haven’t taken it, and you run a calculation, it wouldn’t account for the future allied loss of capital, IPCs, and whatnot. The calculator also would think Germany was at a disadvantage because they haven’t expanded towards the east and gotten those IPCs, but in reality they have a large unrealized investment towards the west.
I think it would be cool to add certain things into the mix, like 1 or 0 values that the person can input, such as is London at significant risk of falling, same with all other capitals and important strategic territories. This reduces the simplicity but adds accuracy if you’re running the calculation towards the begging of the game.
-
I don’t think Sea Lion happens that much at the pro level. At the less experienced player level, the model is likely to be less accurate as new players will be making large mistakes.
Even with a data set of over 100 games (which will be a pain to collect unless we can automate the collection of both the files and the variables), you’d still not want to have that many variables as you risk over-fitting.
-
We might be able to learn some good ideas from anyone who has developed an AI for the game.
-
@akreider2 yeah, that’s a good point.
Separately I think it might be cool to get a place for people to report their win/losses with certain common strategies vs other common strategies. Might be cool to have a model for what to use versus what, optimize the game-play for those who play competitively online, although I suspect they’d already have a good idea of what works versus what.
-
It’d be fun to train an AI using pro games. I wrote an AI once, but it was for Connect Four (a very simple game) and many years ago.
-
@akreider2 yeah… that’s pretty out of my league. Id simply like to analyse some game data from many different players of different levels to develop a little cheat sheet that would tell you how to play against each strategy and skill level.
-
How does the model handle force projection?
Example: USA has 50 tanks in Brazil and no transports
Germany has 50 tanks in Germany and no transportsWhich nation will kill more units and take more IPC and take more VC?
-
Hhhmmmmm…
-
the model doesn’t, i don’t think. What i was proposing was to make a variant model for G40 in which the user would manually account for things not represented by troops or IPCs.
-
@Imperious-Leader The proposed model wouldn’t be that effective at handling stupid moves, and mistakes.
You could partially control for mistakes with a variable that is based on the player’s league ranking (if they are in the league).
-
As a baseline test, let me know if you can find a game where the Axis player has equal or greater total unit value to the Allies and isn’t winning.
Alternatively, can you find one where the Allies have a total unit value of 400 more than the Axis and they aren’t winning (this latter might be possible if the total unit value is very high)?
-
One of the big issues for pulling this off - does anyone know if you can extract data (notably total unit value and other statistics) from the Triple A save files?
Is it in XML, CSV, JSON or some other easy to use format?
-
Barnee know any of that ?
-
@SS-GEN
yea you would want to ask @redrum I know tuv in tripela isn’t represented 100% accurate although it’s pretty close. I believe they’ve made some changes recently that make it better.
In global it seems the allies are overvalued a bit from what i remember
Edit
Actually I think it counts it right but it includes infrastructure. So for combat unit vs combat unit it’d b a bit different because of all the allied infrastructure. -
I remember wondering whether I should include the infrastructure and forget what decision I made. The infrastructure is often useful and the model would be controlling for the discrepancy at the startup. So I think it’d mostly count new infrastructure which is typically important.
-
@akreider2
Hi akyea I was just thinking you need to allow for it somehow but allies have kind of an excess I guess you would say that doesn’t really get used. At start anyway
idk redrum probably reply in a day or so. He’s usually pretty good at that
-
yea counting new might be a good way to go. Probably better than total. Or somehow how PU totals vs how much could be used for factories. Planes and ships for AB and NB.
Idk that seems to be a lot of work for maybe a small reward. Even if u get close with just new it should be of some value I would think
Suggested Topics
