Jump to content
Korean Random

Praetor77

User
  • Content Count

    48
  • Joined

  • Last visited

Community Reputation

20

Contacts

  • Nick
    Praetor77
  1. Previous 1200 number was wrong. It has always been 1500, it was a mistake in my maths. :) Fixing E-25... Fixed: http://www.mediafire.com/view/jrfa2d3qt12lqy9/expected12.csv
  2. The reason for using a ratio scale is that it decompresses the distribution from the middle. What is the sense of a scale if every single player you meet in game has 900-1100 rating? It loses usefulness in being able to predict how well that player normally plays. http://en.wikipedia.org/wiki/Level_of_measurement The reason for changing to a ratio scale is that, once again, it decompresses the values from the middle, making a wider bell curve, which makes WN8 more effective and useful at predicting player competence/skill. With interval scale (WN7), 0 rating meant 0 stats, 0 kills, 0 damage, 0 spots. The problem is that even a bot gets some spots, and even some kills and damage if it is a more complicated one. With ratio scale, 0 rating means 0 contribution to winning. You will notice no players had WN7 scores below 300 (unless they had tier penalty applied to them). In WN8 the worst players in the game get 0-100 rating, regardless of the fact that they do not have 0 stats, they do contribute nothing to wins, for example a 38% win player. And I don´t know why you find it hard to believe that a player can contribute 4 times more to victory than an average player. This is true. This is why an average player "wins" 7% of matches (49% winrate, 7% above the winrate for doing absolutely nothing), while the best players in game can get winrates of around 70% while playing solo (no platoon), which is 28% above the winrate for doing nothing. So yes these players do 4 times more than the average player to help his team win. Of course, this is not linear for each stat, they don´t do 4 times more damage and kills than an average player, which is what makes the ratio scale in WN8 clever and useful. Again, the interval scale used in WN7/Efficiency/etc. is not really good at comparing one player to another. There is a HUGE difference in skill between a 500 WN7 player and a 900 WN7 player, which will now be better represented in WN8, as maybe they would have 200 and 1000 WN8. The same goes for the unicum players, with a huge difference in skill between a 1750 WN7 player and a 2000 WN7 player, which will now be better represented in the WN8 ratings, with maybe 2200 and 3000 WN8. About the 1-100 scale, I was simply suggesting something which I feel would be more useful for me than the standard 1-100, 50 median scale. Doing what I suggested, dividing WN8 by 32 for the 1-100 scale (with all players above 3200 WN8 having 100), maintains the ratio scale properties of WN8. If, however, of course you are more than free to use the same process of bringing the rating to a 0-100 scale that has been used before, this removes the interval scale properties of WN8, and turns it back into an interval scale, which is less useful to me, IMHO. PS:Oh yeah, about the per-tank values they would need to be updated every patch. PS2: The coefficients in step three were determined using Eureqa, a program which uses iterative genetic algorithms to find mathematical relationships between sets of data. http://www.nutonian.com/products/eureqa/ Eureqa determined that these coefficients are the best way in which we can approximate/estimate how much a player should win, given his stats. It is explained better and in more detail in the Word document. PS3: Another big change from WN7 to WN8, is that we now have per-tank normalized damage. So now we can fairly compare the average damage of two players who played absolutely different tanks. Using tier-based damage normalization in WN7 made average damage less useful, since we all know that if one player played a lot of games in AMX40 and another in T40, their average damages will not be comparable. In WN8, using per-tank damage comparisons allows this stat to be much more useful, even more useful than average kills. Therefore, WN8 is more heavily weighed on damage than on kills, which prevents WN8 farming, as some people did with WN7, farming kills in low tier games, and damage in high tier games.
  3. If you guys feel it would be best, I could write a smaller summary. My intention was to get the formula to the XVM developers so they could start thinking about implementation in XVM. The full text of the Word article is more for people who really want to know in depth how WN8 works. :)
  4. WN8 is finished. I hope Seriych can explain it better to you guys, since Google translate is pretty bad. Basically, a per-tank "expected" stat table is used. The stats for each tank are multiplied by the number of games played on that tank, and then added all together. Then the total stats for the player are compared with these total "expected" stats. Then we use a couple formulas to calculate the final value of WN8 like this: The Steps of WN8 – Step 1 rDAMAGE = avgDmg / expDmg rSPOT = avgSpot / expSpot rFRAG = avgFrag / expFrag rDEF = avgDef / expDef rWIN = avgWinRate / expWinRate Step 1 takes the counts of tanks played on account, and multiplies them by the expected stats to get the account total expected values. Then the actual account totals (your total dmg, frags, spots, def, win-rate) are divided by the expected values to give the ratios. Step 2 rWINc = max(0, (rWIN - 0.71) / (1- 0.71)) rDAMAGEc= max(0, (rDAMAGE-0.22) / (1-0.22)) rFRAGc = min(rDAMAGEc+0.2 , max(0, (rFRAG-0.12) / (1-0.12))) rSPOTc = min (rDAMAGEc+0.1 , max(0, (rSPOT-0.38) / (1-0.38))) rDEFc = min (rDAMAGEc+0.1 , max(0, (rDEF-0.10) / (1-0.10))) Step 2 sets the zero point for the ratios. See the assumptions section for more info on why this happen. min and max are functions to ensure the ratios stay within bounds. The constants are in the format of (rSTAT – constant ) / (1 – constant) To normalize that a player with all rSTATc of 1 would receive 1500 WN8. Step 3 WN8 = 980*rDAMAGEc + 210*rDAMAGEc*rFRAGc + 155*rFRAGc*rSPOTc + 75*rDEFc*rFRAGc + 145*MIN(1.8,rWINc) Step 3 takes the weighted (in Step 1) and normalized (in step 2) performance ratios and processes them through the coefficients determined for the final formula, reported above. This puts the scale on the more meaningful 0-5000, gives the relative weights of damage and reflects the interactions between frags*spots, def*frags and dmg*frags. Here is a document explaining everything in detail, it is in English, but I hope Seryich can translate it into Russian soon: http://www.mediafire.com/?23qbgbc9ydjcyff WN8 has many advantages over WN7. First of all, it can measure player skill regardless of what tanks he chooses to play. So WN8 cannot be "farmed" by playing a large number of games on OP tanks. Also, no more farming by playing tier 1 and 10 as was possible for WN7 (see document for more details). Secondly, the fact that it is a ratio scale, allows people to directly compare scores. A player with 4000 WN8 contributes 4 times as much to winning a game as a player with 1000 WN8. Thirdly, WN8 should be easy to translate into a 0-100 scale. Simply divide WN8 by 32. Approximately 99.99% of players have less than 3200 WN8. Those players above 3200 WN8 will have 100 on the 0-100 scale. An average player will have a score of around 25. I would be happy to answer questions and concerns, and I would be thrilled if WN8 could be implemented into XVM. There is no doubt in my mind it is a much better metric than WN7. Oh yeah, this is the updated per-tank expected stats table: http://www.mediafire.com/?lp574zkezhht1ge
  5. shnog anwered some things... but here I go... http://www.vbaddict.net/wot.php Average was tested first and does not work well. WE tried the average between average and top and that did not work well either. Exactly.
  6. Using google translate which is pretty terrible, but the idea is very simple. Use player-tank data to calculate top100 damage and kills like what appears in the service record when using XVM. This will allow honest comparison of each tank´s "power". For WN7, if you have the same player play an A-20 and a Su122-54, his WN7 would surely be lower than if he played Tetrarch and Object 704. The idea is to use per-tank top100 stats to multiply times the number of games a player has on each tank, then divide that by number of battles. Then you divide his overall account average by this number. We call this rSTATS, for ratio. So rFRAGS would be the players average frags divided by what the top100 would have gotten for the amount of games he has played in each tank. Same for rDAMAGE, rSPOT, rCAP, rDEFENSE, etc. So, for example, for M4 Sherman top100 numbers are 1000 damage and 2.5 frags per game. For T1 3 frags and 400 damage. For Je100 1.5 frags and 3000 damage. Lets suppose a player opens 4 accounts and plays different tanks on each of them. Lets say player A plays 10000 games on T1 averaging 3 frags and 500 damage. Average tier 1. WN7 (including low tier penalty) is about 1700. Player B plays 5000 games on T1 and 5000 on JE100, averaging 2.25 frags and .1700 damage. Average tier 5.5. WN7 is 2750. Player C plays 5000 games on M4 and 5000 on JE100,average 2 frags and 2000 damage. Average tier 7.5. WN7 is 2430. Played D plays 10000 games on Je100. WN7 is 1840. For WN8, his score would be around 2150 for all four cases.
  7. Hi guys, just wanted to let you guys know that I am working on something similar to the Teff system for WN8. It should increase the WN accuracy inmensely, also becoming inmune to farming stats with "OP" tanks, like many people are doing with T49, Hellcat, F155, etc.
  8. Oh I noticed it is probably because of the flat penalty to low avg tier players for WN6. With WN7, the penalty is only applied once a big number of games has been played.
  9. That is VERY weird. Don´t understand the huge difference then... but I definitely don´t like the categories you have for WN6 scores. Could you tell me how it would look filtering by 2500 games?
  10. I have recently had access to the wotlabs.net database, and I analyzed th server population for scores. I have pretty different numbers to this: I think this is using the ENTIRE server population which is silly because many, MANY accounts have very few battles on them. I put in a filter of at least 1000 battles and got these results: 1000 Battle filter WN7 better than 300 5% 450 10% 600 20% 700 30% 800 40% 900 50% 1000 60% 1100 70% 1150 75% 1250 80% 1400 90% 1550 95% 1850 99% 2050 99,9% Efficiency better than 550 5% 650 10% 750 20% 800 30% 900 40% 950 50% 1050 60% 1100 70% 1150 75% 1250 80% 1400 90% 1500 95% 1750 99% 1950 99,9% Based on this, your color table ends up having WAY too many of the active players in the yellow and green areas. I suggest something like the following: Or, you can run your analysis again on your own data setting a lower battle limit as 1000 or 2000. I would also suggest adding a category of 90%sup, since a 1450 WN7 players is MUCH MUCH better than a 1150 one but they are currently colored the same in XVM. Even i you don´t want to use the extra category, you take it out and you would have a very close, but better version of your color table, much more close to reality.
  11. For calculating avg tier played, it was proposed arties count as two tiers higher. No formula... However, it was decided not to do that since we needed API to do that. When WG reclassifies arties into tier 1-10 the problem with WN7 overestimating arty player skill will be fixed.
  12. Any plans to calculate per-server E scores? Due to russian server size, for most tanks, NA server will manage 10 X players at most, instead of 100...
  13. First of all I think we are not understanding each other, I am not complaining about my E score. I am just saying E value should be calculated with stats for each server. I may be a 7 on RU server, but on NA server I would surely have a higher score. Using RU server data for E calculation, for some tanks, there may not be ANY players in NA server who get E scores of X, for example. Am I clear? What I mean is, E values should be calculated with top 100 and average stats for each server, or they kind of lose their usefulness.
  14. I think these stats should be server dependant. I can assure you I am not a 7 on M48 in NA server... For the top damage and kills, do you take the very best player on server with at least 100 battles? Or do you take the 100th player? Taking the very top player should be tricky... I would take the stats of the 100th player...
  15. Also, this seems weird to me... With M48 I have 2527 average damage and 1.71 average kills, but I get an E value of 7? PS: Also, is it possible to place an E column on the left side of the service record, beside number of fights?
×
×
  • Create New...