Jump to content


Mathematics Contributor
  • Content Count

  • Joined

  • Last visited

  • Days Won


RichardNixon last won the day on May 23 2016

RichardNixon had the most liked content!


About RichardNixon

  • Rank
    Not Purple Enough

Profile Information

  • Gender
    Not Telling
  • Server

Recent Profile Visitors

8,918 profile views

  2. It doesn't scale well. For damage it's about +8% for each +10% you'd gain in an average tank. That's actually worse skill scaling than tier 6 arties, although better than tier 8+ arties. I don't like the term "skill ceiling", as it implies a hard limit beyond which skill doesn't matter, which doesn't fit the evidence in WoT.
  3. Nah, WN9 is a two value system. Sorta like xTE except with the values expressed differently and the lower one not being pathetically wrong.
  4. Yes, but that was never actually true. WN8 was probably intended to be contribution-linear, but the platooning and tier-scaling problems in rWin wrecked that idea. Consider the effects of the 0.22 rDmg cap and the frags*dmg term. 1.34x expected damage for 2450 WN8 is about average in practice.
  5. 1. I don't know where targetdamage.com gets 1450 from, but it's probably making a bad assumption. WN8 isn't anywhere near linear, so if your rDmg, rFrag, rSpotC and rWinC scale similarly (which they usually do) then you only need 1.34x the expected values for unicum, or <1300 damage. 2. leggasiini is correct on the skill-scaling point. WN8 assumes that all tanks have equal skill scaling, which they obviously don't. The centre point for the expected values is around the average tier 10 skill level. Slow tanks generally pad for players below that skill level, and anti-pad for players above that skill level. 3. The O-I just had a big nerf (roughly 8% damage output), so the WN8 value will be harsh even for players around the expected value level.
  6. I don't collect 2 & 3-man data separately, and the data's thin because the matchmaker was rewritten in 5.8. Typical advantage of platooning appears to be 2-4%, with no clear dependency on solo winrate, which is odd. Either players don't platoon with players of similar skill in WoWS, or there's some partial skill balancing applied to platoons.
  7. Well, looking at the whole signature, it does appear to be picking the value from the right that's numerically closest to 250 battles. The odd thing in your case is that it's not showing anything between 1d and 90d.
  8. The standard overall & recent WN8 methods don't use the tanks/stats API, so they're unaffected. Per-tank WN8 is affected, but most sites implement a workaround where the account/tanks results are used to filter tanks/stats. This can have a significant performance impact. I'd rather not build support for an obvious bug directly into the metric, although I suppose I should document the workaround. From my perspective it's actually more useful to see the old tanks. Cuts down on the cross-server reroll bias.
  9. Ok, here are distribution tables for two-month (min 100 battle) recent WN9 on RU, EU and NA. This uses the internal scale with 1.0 = meets expected values. NA EU RU 10.00% 0.278 0.279 0.317 20.00% 0.370 0.368 0.423 30.00% 0.454 0.452 0.508 40.00% 0.535 0.532 0.581 50.00% 0.612 0.608 0.649 60.00% 0.688 0.685 0.716 70.00% 0.770 0.767 0.787 80.00% 0.868 0.862 0.869 90.00% 1.007 0.992 0.982 96.00% 1.157 1.131 1.106 99.00% 1.350 1.304 1.266 99.40% 1.413 1.361 1.320 99.70% 1.494 1.431 1.392 99.90% 1.606 1.536 1.504 99.96% 1.685 1.619 1.593 99.99% 1.813 1.726 1.711 NA vs EU: The player distribution on these servers tracks pretty closely until the 90th centile, where NA increasingly races ahead. Three possible explanations: The good players on NA are actually more skilled. The lesser players on NA are uniformly worse. All the freebie tournament gold on NA causes the better players to stretch ahead. RU vs EU: Below the 80th centile, RU players get a lot more WN9. This is probably because the average tier is higher, although as they're also more experienced they may also be more skilled. The top players do somewhat worse. Possible explanations: The good players on RU are actually worse than on EU. The lesser players on RU are better. The metagame and/or higher arty counts on RU don't suit better players. If we assume that the top ~1% of players on each server are similarly skilled, then NA players should be given less WN9 and RU players more WN9. I suspect this wouldn't be very popular.
  10. The tanks/stats API doesn't clear the old tanks when you reset an account, so account WN9 is working with 26000 battles rather than 3500. Shouldn't be a new issue though. You get 1823 tanks/stats WN8 on the same data, for comparison.
  11. Well, they're in the queue because they're not in a game, so queue numbers are a bit misleading. The light tank battle count did rise sharply since the missions (especially on EU), but it seems to make surprisingly little difference to the class & skill balance. They don't have much practical impact on available HP because they don't deal much themselves, and unlike arty, they probably don't have a disproportionate impact on better players. You would expect lights to take spots from other tanks (which would reduce the available WN9), but the parallel increase in arty counts seems to cancel that out. The idea is that the expected values method just minimises numerical variance over common tier 6-10 tanks. However, the fixed scale coefficient (currently 666) would be replaced with a table of per-server per-metric coefficients which would be included in the JSON file. I'll have full recent 2-month recent data for RU, EU and NA shortly. I'll try unicum = top 2% and see what it looks like. It's possible that server variance only exists on overall data due to rerolls. Comparing distributions for overall data is tricky due to the reroll variance. I'd probably use a single server (EU) and adjust the others by comparing long-recent distributions.
  12. Ok, changed my mind. There's a relatively straightforward way to implement per-server and per-metric distribution scaling, mostly with data that I'm already collecting. I'll throw something together and see what it looks like.
  13. Tier 10 armour doesn't work as well against tier 10 guns as tier 8 armour works against tier 8 guns, but that's not what the matchmaker gives you, so the bounce percentages are quite similar. The relative strength of mediums at tiers 9 and 10 is probably due to the alpha jump, which has already been flattened out in sandbox. It comes down to the decision that tier 9-10 TDs and MTs had to be equally valuable as heavies. The 704 would have been a decent tier 10 TD if they'd stuck to the tier 5-7 balancing rules.
  14. Winrates usually change more, because they shifts drastically every time WG tweak the matchmaker. There's a limited point in measuring overall stats in a game that's changing frequently. Otherwise, the proportion of solo winrate in the metric should depend on battle count. If the playstyle variation (damage+frags vs winrate) is similar to WoT, you'd use mostly winrate beyond 1000 battles but mostly damage+frags at 100 battles. It's all terrible unless the expected values are good though.
  15. The 50/51 doesn't have the popularity (yet) or the damage output for mediocre players. That effect was absorbable, even together with the M48 & E5 overbuffs, partly because gradual changes are more acceptable than sudden ones, but also because the Grille had a much larger effect. The Grille-spam's more like a 100-150 damage penalty at that level, assuming that you don't play it yourself. I suspect there are stronger historical effects, notably an increase in arty counts since personal missions, map changes, and player-base stagnation leading to higher average skill & tier distribution. There is an option of setting the scales according to actual player distributions, but this only makes any sense if you set them separately for each server, would need a huge amount of additional update work (the current WN9 method is bad enough), and may lurch around with seasonal changes. There's also no guarantee that the top 1% of players on each server are equally skilled. The alternative is to minimise numerical jumps on updates, and accept that the unicum mark varies by server, and may become more or less difficult over time. I'm not sure anything else is practical. I can publish distribution tables occasionally, but I don't think they can be part of the update process. My data doesn't go very far back. The best I can do is older EU accounts from late 2014. More general samples from mid 2015 at the earliest.
  • Create New...