Jump to content

MaxL_1023

Mathematics Contributor
  • Content Count

    1,494
  • Joined

  • Last visited

  • Days Won

    5

Everything posted by MaxL_1023

  1. Potential Solution to the Russian DaddyTanker Problem: Use WR as an account filter - only use players in the x-y% range to get expected values. Worth a look?

    1. peregrine

      peregrine

      imho, yes.

      The whole population bias thing has to be analyzed going forward.

    2. spencer

      spencer

      How bout the german wehraboo problem or the MURCA problem?

    3. spencer

      spencer

      Also the regressions use data from VBAddict which has very few russian users.

  2. No, with the special tanks the expected value calculation is itself flawed in that it is assuming a representative dataset where none exists. Basically, it is treating the top 25% (mean x variance y) as a whole sample (mean x-a due to player sampling, variance y-b due to central limit theorem) and then taking the average of the top x% of the incomplete sample, which due to the higher variance is higher than it should be to start with. It is messing up in 2 places, resulting in expected values consistently 10-15% too high.
  3. So for the STB there was basically no change. What about the 140 and 430? I am curious about recency and how large of an effect it has. If recency seems to not be an issue it would be easier to adjust the special tank values by normalizing the variance and then shifting the mean.
  4. The mean of the top half of a Gaussian is farther from the overall mean the larger the variance is. The small sample tanks will nearly always have a larger variance, automatically producing a high bias. How large of an effect does recency have on a normal tank? When the STB-I was released how high were the first expected values compared to today?
  5. Weight the player stats by Z score. That will remove the effect of variance. You can set it up so that the expected value will always be a fixed percentage above or below the mean regardless of what the mean or variance is. You would decide how to normalize the variance based on the variance you get with comparable vehicles - basically the 907 would have high mean and variance now - you would normalize the data points to have the same variance as the 430 (or an average of all 3 normal mediums). You can then directly compare the mean win rate and Rstats for each tank. You can then adj
  6. Regarding WN8 Expected Values: Trust me, it's Oompa-Loompas.

  7. Honestly XVM focusing is killing it for me. What is the point in playing when the game itself breaks just because you are there? Besides that there is the fact that map rotations can often be terrible and it is impossible to carry dead weight solo because of all the skill equalizers (SPREM, Arty, Map Design, HD Armor, Accuracy still better than it used to be) which means that any random 49%er can damage you from almost all angles just by being smart enough to memorize a few weakspots or use a mod which tells them where to fire. Basically, the game is just not as much fun as it was be
  8. On a more serious note you can't do something like this without a better source of data than a few pounds of organic jelly.
  9. Before we do anything we should see if there are any players currently playing both the reward tanks and something in the same class/tier. If we have no data then there is no point in even a manual adjustment - we might as well just set each tank to the class/tier mean or remove them altogether and be done with it. I don't think anyone (myself included) could play a reward tank then regular tanks tanks and figure out their relative expected values in such a way as to make them applicable to the entire playerbase. There is just too much subjectivity and variation to make a good decision.
  10. I would honestly be more interested in figuring out why the formula spit out absurd expected values to begin with. If you want to hotfix the values to a placeholder it wouldn't really make a difference - I think it is more important to make the system itself more robust so we don't have to make these subjective adjustments.
  11. 5% is about the variation you see between most vehicles in the same class/tier - it is still significant. Even if it is easy to improve above what we have now the utility stems more from the current values being terrible than the method being fully developed. I agree that something needs to be done but we need to flesh out a procedure with at least some objective, statistical support before we start changing numbers. At the moment we have not gotten that far.
  12. The issue I see is that the adjustments may not make sense to everyone, even people eminently qualified. When comparing similar vehicles you will end up with most of the performance difference stemming from the playstyle of any particular driver, acting as a large noise source overlaying the actual differences between the two vehicles. An example is the 907 vs 430. They are very similar in principle. What if (for example) Garbad thinks the 907 is better, I think they are equal, you think the 430 is better and the entire PP population is split so that neither "camp" has a majority? Who
  13. The idea is sound - reward tanks and the like are difficult to model considering we are basically working with 1% of a dataset skewed heavily to the right. It makes it difficult to compare them to "regular" vehicles. However, we need to remember that an adjustment such as this is at its heart subjective. Someone (or a group) will have to make a decision regarding how to adjust the expected values. There will be disagreement and a lot of players may resent the departure from pure statistics. Before we do anything, is it possible to get some recent data from players who drive both a re
  14. Taking a break from Tanks to play HW:R. When I dreamed I was a Taiidan HC shooting ion beams at an E-100 on Himmelsdorf I knew I couldn't handle both games at once.

  15. You can use partial fractions to get it all in the numerator.
  16. Somebody on your account tried to platoon with me - I asked to see his real account stats and he quit.
  17. I think that I will still consult on WN9, but my rigorous modelling approach would best be suited to potentially forming WN10. Basically, have WN9 as an update of WN8 with WN10 being the major revamp. If WN9 performs better than expected it might not be needed, but the workload I have with university means that for a few weeks I won't have time to do a lot with it. However, WN9 should be good for 6 months+ so we would have plenty of time to rebuild the WN system if there is still performance improvement in sight.
  18. It was a 183 not a 215B - that can't fight against tier 10 meds there when your teammates are basically bots. I thought my platoon mate would set up to shoot anything attacking but he didn't engage effectively.
  19. Logit is a linear model with a log transform. It is basically Log(WR) = X*A + Y*B + ... The issue is that you are only capturing the average gradient vector across the entire response surface - you miss peaks, valleys and any second order effects. The point of the model is not to match the outcome itself but to match the probability of each outcome. That required building PDFs out of the binomial data and distribution fitting, and I doubt a normal distribution would fit well. Everything on the old WoT-News site was Gamma-style, with an extended right tail and a mode lower than the mea
  20. I don't think Logit will end up with a good model since it only includes linear effects. Additionally, damage blocked is the same as damage taken. You need the ratio between the two to account for armor effectiveness.
  21. We could try modelling in three regions (<47%, 47-51%, 51%+) and see if we get different behavior.
  22. I disagree with the be-all and end-all nature of the logit analysis, but I do agree that we can't discount any statistic this early in the process. We HAVE to include capture points in the analysis for no other reason to see if they have an effect. If a higher capture/game score makes you more likely to win then it makes you more likely to win, subjective arguments aside. It doesn't even matter how we design the formula - it will be paddable as long as it remains a linear formula. The only way we can make it unpaddable is to make it a multidimensional response surface model with a clearly
×
×
  • Create New...