Gryphon_

WN8 - Change to Averaged Expected Values

29 posts in this topic

As discussed in other threads, in the near future we will transition WN8 away from per tank expected values to averaged expected values for tanks of same type and tier. So, for example, all tier 8 mediums will have same expected values. This is necessary as the constant stream of new tanks requires a lot of work under the current system, and interest in doing all these updates is fading along with interest in WN8. 

The biggest question everyone will have is what change to WN8 will you see when this happens? Also, behind the scenes, the math guys are asking whether the averages should be simple averages of the expected values as published, or should they be weighted to reflect tank popularity? Well, using data on 20,000 accounts from vBaddict I can now answer both questions. The histogram in the spoiler shows the change in WN8 overall using both 'simple' (red histogram) or 'weighted' (blue histogram) strategies. 

As you can see, the simple averages will result in the vast majority of accounts seeing just a +/- 50 point change, with all but a few seeing +/- 100 points. A handful see up to +/- 400 (lol - play the right tanks guys). The weighted averages produce a similar result but with a distinct negative bias. As a hit to most WN8 scores is undesirable, we will plan to move to SIMPLE weighted averages. 

Note: the change was measured against normal expected values (with the v29 joke values for E50, M6, etc, corrected in advance.)

 

WN8%20Change%20-%20Average%20Values.jpeg

Share this post


Link to post
Share on other sites
7 minutes ago, Luna said:

So basically we are throwing it out the window at this point? I don't see how this is any different than just attributing a color based on DPG and calling that wn8.

(Can we also get a list of all the value changes?)

WG threw it out the window with their record keeping.

Share this post


Link to post
Share on other sites
44 minutes ago, Luna said:

So basically we are throwing it out the window at this point? I don't see how this is any different than just attributing a color based on DPG and calling that wn8.

(Can we also get a list of all the value changes?)

Its different to a 'color based on DPG' because the full WN8 formula, or which dmg is one of 5 factors, will still be used.

The value changes always appear on wnefficiency.net to show old to new deltas for everything.

Share this post


Link to post
Share on other sites
12 hours ago, Gryphon_ said:

Its different to a 'color based on DPG' because the full WN8 formula, or which dmg is one of 5 factors, will still be used.

The value changes always appear on wnefficiency.net to show old to new deltas for everything.

But since dmg is like 70% or however much of the value, it really kind of devalues the metric as being superior in any way compared to dpg.

Share this post


Link to post
Share on other sites

Looks like the team is on board with this, so lets move onto the details: the expected values. In the spoiler below you can see the complete table of all values by type and tier. Values in black are the ones calculated using averaging I explained above. However, to future proof the system in case WG adds low tier arty, TDs, or heavies, I dropped some placeholder values in there (red text). Those should be both non-controversial and of zero interest.

That leaves us with the only significant decision needed: the expected damage values for the lights at tier 9 and 10.

The spoiler shows a graph of expected damage for all type (1 =  lights, 2 = mediums, 3 = heavies, 4 = TDs, 5 = SPGs). Tier 9 lights value as calculated from the pre-9.18 data is of course based on those tier 8 lights that went up a tier, and the result was 919. (The highest value currently of any of them individually is the Lwt, at 999). As shown in the table in red text, I intend to replace the 919 calculated value with 1000 as the Tier 9 lights expected damage. This rise will impact owners of any tier 8 light that went up a tier, but I think a modest rise now is needed to avoid that value being seen as too low in the future - we don't want to have to update this again. 

For the tier 10 lights, I eyeballed the graph, compared light values with medium values, and picked 1250. Thats somewhere between a tier 8 and tier 9 medium in damage terms, which I think is about right - unless anyone thinks the new tier 10 lights can do as much damage as tier 9 mediums. 

So please go ahead and comment on the damage values for tier 9 and 10 lights (red dots on the graph). Once we have a consensus on that, we should be done.

 

Averaged%20Expected%20Values.jpg

 

eDAMAGE%20v30edit.jpg

Share this post


Link to post
Share on other sites

Old tier 8 LT ranged 1600-2000. New 9s seems to be 2200-2800. This is my personal performances, which are pretty stable. 10 LTs are holding 2500-3300 (less games, less sessions, more variable). Tier 9 meds sit 2700-3100. Perhaps the ratio of these to the midpoints can inform. I propose a general bump to all 6-8 LTs, to address that they're not shit right now.

We might want to commit to a 6 month check in too, as the new MM may change values? I know I'm seeing 10% higher dpgs across the board. 

Oh and fuck arty.

Share this post


Link to post
Share on other sites

I would like to put up a few cautionary arguments.

1) I would vote against calling this wn8 as what is being proposed is not wn8, it's deceptive.

2) The massive changes to wot are not done with, there may be substantial alterations to the changes this patch. We also have yet to fully realize the effects this patch will have on the game.

3) Before adopting a new system that will have to be sold to others, we should know how this new system compares to WG's PR rating system.   If it has less efficacy then what is the point?

 

We should measure twice before cutting once. :)

Is anyone willing to do one more update to wn8 so we could have time for the dust to settle?

 

Share this post


Link to post
Share on other sites

Is there some reason tank values aren't just calculated by a program? With the API's 10 call per second limit (without special permission) you can skim "last played" data of all players in around a day and a half. Assuming not all 110 million players that WG claims to have are active, you could actively skim tank data from players accounts from /every/ active account over a period of 15 to 30 days. Then just use a system of averaging and line fitting to automatically update values. Said system would then also have individual tank data 30 (or fewer) day history for every active player in the game. The system could even run an API that could feed to XVM or an XVM like mod that would then show players recent rating (or even recent rating for the tank they're driving).

In its most basic form though (which would save a lot of server storage space) it could just be used to update values on a 30 day rotation.

Since individual accounts would be skimmed, finding new tank id's would be done very quickly and the system could be setup to actively add new tanks as it finds them. Is server money and someone to write the initial program the /only/ limitations to this, or am I missing something big?

Share this post


Link to post
Share on other sites
3 hours ago, Android25 said:

Is there some reason tank values aren't just calculated by a program? With the API's 10 call per second limit (without special permission) you can skim "last played" data of all players in around a day and a half. Assuming not all 110 million players that WG claims to have are active, you could actively skim tank data from players accounts from /every/ active account over a period of 15 to 30 days. Then just use a system of averaging and line fitting to automatically update values. Said system would then also have individual tank data 30 (or fewer) day history for every active player in the game. The system could even run an API that could feed to XVM or an XVM like mod that would then show players recent rating (or even recent rating for the tank they're driving).

In its most basic form though (which would save a lot of server storage space) it could just be used to update values on a 30 day rotation.

Since individual accounts would be skimmed, finding new tank id's would be done very quickly and the system could be setup to actively add new tanks as it finds them. Is server money and someone to write the initial program the /only/ limitations to this, or am I missing something big?

Well, they are calculated by a program - R. Years ago I published an R script to process data provided by Phalynx of VBAddict, and when I 'retired' from the constant grind of producing new values every patch or so, Kitten took over. However he is also now finding it hard to find the time to run the script, deal with all the issues caused by API changes, etc, and publish updates. As part of research into various candidate WN9 systems we discovered that averaged expected values (by type and tier) perform just as well at account level as per-tank values, unless players play disproportionately large percent of their battles in a handful of tanks. The eventual WN9 was developed by RichardNixon using a pull from the API similar to what you describe, and Kitten was looking at it to figure that out. However, RN left, and WN9 was never adopted by any website. Others have mentioned WG PR, and now that is a pretty good rating (it has unparalleled data access!) but it doesnt do recent values. There are many possibilities for new ratings, limited only by the API data, but the problem with WN8 and all the rest is the same - how do you sustain it when the originators have gone away? 

Thats what this thread is about. Just need consensus on a couple of expected damage values for tier 9 and tier 10 lights, then we can leave WN8 to sail on with no-one at the helm, knowing that it wont collapse a bunch of websites and mods due to missing data. Ideas of other systems and other ratings are welcome and can be started in their own thread. 

Share this post


Link to post
Share on other sites

Tier 9 light tanks feel power-wise like a tier 8 MT, but with possibly 1-1.2 spots per battle, probably closer to something like a Mutz for expected (about 1250 per). Tier 10s, Id be looking about 1550 expected (tier 9 MT levels), and about the same spots per. Could argue a bit less, as their carry potential is kicked in the nutz

 

Share this post


Link to post
Share on other sites

tier 9 lights are still performing somewhat better than they used to as tier 8s, only problem is that some of them are new and some aren't so it's hard to compare them for what they are, the ru251 and lightweight are for sure better this patch, HP buffs, ltwt got HD and RU got an HP buff it needed

however, im still worried you'll have a metashape due to the new metric like wn8 did with the tier 9 meds in general, some exceptions are kind of needed. the entire tier 2 padding scene pretty much died with WN7 so i think it's fair to say that the metric has made its footprint on the game and will keep doing so

the 907, arguably the stupidest tier 10 medium around right now, will be by far the best medium padder of the metric in the game. you can play it like a braindead monkey and overperform for your skill level, not to mention it's very popular as it is already, and throw a padding incentive into it and then you have the stupidest tier 10 tank in the game being by far the best padder

the tank is OP, very stupidfriendly and hard to do poorly in. 

 

 

u have the same issue with the maus, it's wr is up by a lot just this previous patch and currently it looks it'll keep these buffs. the tank is stupid easy to play and downright op and will be the best padder for its class, it's literally the T18 syndrome all over again

i think some tanks need a bracket of their own at least, they're just too strong for the tier they play in and giving them paddable values doesn't really make any sense to me even if it'd arguably not change anything of current player stats, it would in the future and then you'd have to start manually adjusting it, just like wn8 required. to make even more light of the issue you have the fact that the Strv103B has over 5k dpg potential (and 3k does itself) while other TDs struggle to just reach those numbers completely. A 268 and a 103B do not belong in the same category, they're just too far apart imo to be put together

 

an example of what i mean: Leo1, AMX 30B, CAX, Obj 430, E 50M, BC25t and some others might work together, some stats (avg spots) will be very different but their raw dpg/wr numbers shouldn't be too far apart anyway

but throw 907, TVP and T-22 in there and you have 3 massive overperformers in general, especially among the mid-high end. Id consider giving these their own bracket until something changes or something similar to it

Share this post


Link to post
Share on other sites

Well coming right back at you: for years, WN8 has encouraged players to play bad tanks because the system compensated you for it. Now if you cant play as well in a Leo 1 as in a 907, Strv, etc, your WN8 will suffer and a lot of people might stop playing tanks they consider bad - for WN8. Then, maybe WG will notice and start balancing their tanks. We arent doing ourselves or the game any favors by papering over that crack. 

Also, the reward tanks arent too much of a concern as not many have them among the mass player population.

Share this post


Link to post
Share on other sites

Ive had a few suggestions by PM to put some tanks in a higher category, ie treat a Defender as a tier 8 TD rather than heavy. Im not going to go there, as where does it end? How do we decide whether a tank should be moved either way, and to what category? Also, it would greatly complicate the coding the websites have to do which, otherwise, this change will make much easier: WN8 can now run off one table with 7 columns and 50 rows...

Still looking for more opinions on the damage values for tier 9 and 10 lights....

Share this post


Link to post
Share on other sites
On 30/04/2017 at 8:16 PM, TheMarine0341 said:

Tier 9 light tanks feel power-wise like a tier 8 MT, but with possibly 1-1.2 spots per battle, probably closer to something like a Mutz for expected (about 1250 per). Tier 10s, Id be looking about 1550 expected (tier 9 MT levels), and about the same spots per. Could argue a bit less, as their carry potential is kicked in the nutz

 

How about a compromise between the values I suggested: tier 9 1000, tier 10 1250, and yours:1250, 1550? We need to get this done.

Are you ok with 1100 for tier 9, and 1400 for tier 10? I'm reluctant to go higher on the 9s as 1100 is now a 20% increase over what they were when those tanks were tier 8s (919), and that will hit everyone's account pretty hard.

 

Share this post


Link to post
Share on other sites
1 hour ago, TheMarine0341 said:

@Gryphon_ is WG going to continue to read the new tier 9 LTs as tier 8s, with the exception of the new CH LTs?

If thats the case, yes. 

Not quite sure what you mean, but if you are asking whether the entire stats history of the Lwt, RU, and T49 is still there despite the change from 8 to 9, then yes. Any changes we make to expected values that accounts for them getting more powerful / higher tier is going to also apply to all those millions of battles they fought while they were tier 8s. 

The LWT, for example: I have 379 battles of which only about 10 are post-9.18. Expected value is currently 1000. When we change to using 1100 for all the tier 9 lights, that will drive the WN8 down for everyones LWT, and in my case all 379 of my battles will be recalculated using 1100 as expected rather than 1000.

But I think we can deal with that in the interest of future-proofing.

Share this post


Link to post
Share on other sites

Here are the results. The v30 file is configured like the usual updates, so it can be posted on wnefficiency.net and the site will pull from it (wotlabs, xvm, etc). This is the last version in this format.

For use going forward, v31 is the table with 50 rows of values by type and tier. The sites will have to recode to accept that format, but it shouldnt be hard, and there is plenty of time until WG releases more tanks, right?

average_expected_values_v31.csv

average_expected_values_v30.csv

Share this post


Link to post
Share on other sites

In a little over a month I'll be able to give you very accurate multi-server wide recent values for light tanks, assuming all goes well on my end, which so far is all green lights.

Share this post


Link to post
Share on other sites

If you are experiencing difficulties in calculating the expected values, maybe we can automate calculation on the XVM server? If you provide an algorithm.

This will allow to update the values daily and quickly add new tanks.

PS. Found a topic: 

 

Share this post


Link to post
Share on other sites
3 hours ago, seriych said:

If you are experiencing difficulties in calculating the expected values, maybe we can automate calculation on the XVM server? If you provide an algorithm.

This will allow to update the values daily and quickly add new tanks.

PS. Found a topic: 

 

Maybe I'm wrong, but I was under the impression that the XVM servers did not store account data, just the calculated results of that data (ie. stats). Without account data you wouldn't be able to do any expected values calculations.

I don't want to hijack Gryphon's thread again so I'll just point you to where I was talking about the possibility of XVM implementing a WN9 like stat except for recent values.

 

Share this post


Link to post
Share on other sites
1 minute ago, Android25 said:

Maybe I'm wrong, but I was under the impression that the XVM servers did not store account data

No, we store data on all active players in our own database.

Share this post


Link to post
Share on other sites
2 minutes ago, seriych said:

No, we store data on all active players in our own database.

I feel like some sort of recent stat has been asked for within XVM for a long time. I guess if you wanted to calculate change values you'd have to almost double, if not almost triple (depending on whether it was kept separate or used as part of what you already collect) your database, which could be a deterring factor, but is it the only reason XVM hasn't developed their own recent type rating? For anybody who uses stats on XVM, a recent stat rather than an overall account stat would make the application immensely more valuable.

Share this post


Link to post
Share on other sites

  • Recently Browsing   0 members

    No registered users viewing this page.