Jump to content

MaxL_1023

Mathematics Contributor
  • Content Count

    1,494
  • Joined

  • Last visited

  • Days Won

    5

Everything posted by MaxL_1023

  1. We are discarding the previous assumptions and directly modelling from the statistics we have - if cap points turn out to be important we will include them, if not we won't. I suspect if there is a term there it would be polynomial in nature - an increase in WR up to a certain number of Cap/Game then a decrease higher up. Whatever we end up with for WN9 will have a rigorous statistical base.
  2. Cool People are Fucking Models at Midnight. I am Fucking Modelling at Midnight. A small but statistically significant difference.

  3. Ok, I ran your script. This is what I eventually got for the model: > summary(modelStep) Call: glm(formula = victory ~ nKILLS + nCAP + nDAR + nDAT + nDAMAGE + nDEF + nSPOT, family = binomial(link = "logit"), data = CTFResults.e) Deviance Residuals: Min 1Q Median 3Q Max -8.4904 -1.0555 -0.9544 1.1952 2.0787 Coefficients: Estimate Std. Error z value Pr(>|z|) (Intercept) -0.5001565 0.0064363 -77.709 <2e-16 *** nKILLS 0.2084454 0.0036110 57.725 <2e-16 *** nCAP 0.1374640 0.0036268 37.902 <2e-1
  4. Never try to run one of Gryphon's R scripts. Need a second Hard Drive for the required R packages.

    1. TheMarine0341

      TheMarine0341

      F

      I hated using R.

  5. Can you list off the packages you have? I keep needing to install more to run your script.
  6. I think R can fit gammas organically - it should be part of a toolbox somewhere.
  7. They don't have to be normally distributed - fitting to a gamma works. You just have to analyze the residuals differently and make sure you take the unique mathematics of the fitted distribution into account.
  8. Zoom in and make the bins smaller - what you probably have are left-skewed gamma distributions. K is probably near (but a bit higher than) one for most of them. The Gamma tends to Gaussian as a limit when K is large. One more thing - try normalizing dmg taken by vehicle HP. It will give you a more sensible number.
  9. Waterwar explained it well. You have to take into account many factors to determine the best outcome - having a preset play will usually result in a suboptimal strategy. You need to know the limits of your tank, the limits of your team and really need to have at least a general idea of what the enemy team is likely to do. I tend towards aggression when possible - I believe that I have a skill edge over individual enemy tanks and therefore expect to be able to outplay them under most circumstances. I spot when I can, attack when I see an avenue and only defend/camp when neither of those op
  10. If the bush is bugged, you get spotted anyways.
  11. Right now my idea for WN9 would go something like this: Base Modelling: 1. For each tank model solo win rate as a PDF with a mean, variance and skewness characterized by tracked stats from solo games 2. The model predicting the mean win rate would be used in WN9, the other parameters for developer analysis. Note that the model will be polynomial if possible, if not we may need to linearize about different performance regions (use a different model for very high or very low stat players, analysis pending) Per Player: 1. For each tank determine their predicted mean win rate
  12. I set WR cutoffs at 45% and 65%, but really it only works between 48% and 55% or so.
  13. I am beginning to wonder if it is even possible to directly model win rate vs statistics. Need more R skillz.

  14. Ok I ran some analysis on E-100 data. There is an obvious platoon effect present - the fitted models always overpredict win rates <50% (getting worse the lower you go) and underpredict win rates above 50% (again getting worse the farther from the mean you get). The reason is because the majority of data points lie near the 50%WR range, meaning that this effect is not accurately modeled. Basically, the average players are predicted very well but the outliers in WR are shit. Is there a database of only solo games?
  15. I got the previous version for free from my prof in university last year. I think they are at 9 - I got 8.
  16. It should be possible to take those non-orthogonal statistics and separate them into orthogonal and parallel components. Many experimental designs (such as definitive screening and min-run resolution) end up with correlated effect terms - generally you navigate this space by using stepwise regression analysis to develop the most likely model(s), then fit them to validation data. WN9 will not be a pure multidimensional vector with orthogonal components unless the statistical analysis produces a linear set of single-variable polynomial terms. In reality you will get interaction effects at m
  17. We are working on WN9 and already attempting to take all forms of padding into account. No matter what metric we make there will always be ways to artificially inflate it, no matter how obvious and pointless they turn out to be. This is essentially a non-issue and does not need to be further reviewed.
  18. I can finally play decent tanks - finished the 183 on-track. Tortoise is not what it used to be - what little armor it had is negated by APCR/HEAT.

    1. no_name_cro
    2. KenadianCSJ

      KenadianCSJ

      Doesn't have the deeps though

    3. Shifty_101st

      Shifty_101st

      Conq with food is decent. You need BIA and Vents to make it really good DPM wise. But I will still take a conqueror over a conway in a mid range, to close range 2nd line fight. The conway is just to papery

    4. Show next comments  12 more
  19. When driving a Tortoise Win Rate is luck. 6 straight losses where my team was dead before I could get to the battle.

    1. Flametz

      Flametz

      Stop yelling at your platoonmates :P

    2. Cunicularius

      Cunicularius

      I'll save you, senpai!

      In a few hours. :3

    3. HELIONIST

      HELIONIST

      Had you in one game on Prok Encounter team was pretty shit.

    4. Show next comments  12 more
  20. Have you tried using a tool like Design-Expert to do a full multi-factor analysis? It can only work with 32000 rows at a time but that would be enough for any particular tank.
  21. I am nowhere near driven enough in real life to apply anything close to what I think I could be capable of. I know I could probably have a PHD by now if I spent the time I spent playing tanks studying and doing research, since even during the crunch time before deadlines I still barely spent half my time doing actual work. Even assuming a logarithmic style diminishing returns (result = log(effort)) I would still probably have accomplished much more. However, I don't have the motivation to do anything else besides playing games. I just don't care enough - some neural connections or chemica
  22. I try to act as a damage multiplier and rarely play tanks fast enough to steal damage. Even in mediums I am usually too busy making half the team show their turret sides to Kewei's autoloader, so he ends up with 10002003004039403049928749374583 damage while I was hull down/sidescraping/baiting or what have you. He could drive backwards in front of me and half the time they hold their shots for me.
  23. have you tried dropping the Rstats and just using the raw data?
  24. Try setting the minimum battle count higher. It takes 100+ games for RNG to even out as opposed to 50 for tanks.
  25. I told an enemy JPE he had the situational awareness of Schrodinger's Cat. I flanked him in my Tortoise.

    1. Gashtag

      Gashtag

      maybe I'm prone to laughter right now but that made me almost piss myself hahahaha

    2. Medjed

      Medjed

      kekekeke that was funny indeed xDDD

    3. FreddBoy

      FreddBoy

      Literally clash of the titans. Like, how the FUCK do you not notice 80+ tonnes of steel trundling around your tank?!

×
×
  • Create New...