Login

russian armor

Heavy Tank AI Performance - A Comparison

MMX
12 Feb 2020, 17:29 PM
#1
avatar of MMX

Posts: 999 | Subs: 1

alright,

as all of you know, heavy and superheavy tanks received a pretty decent boost to their anti-infantry performance with the most recent changes to their respective scatter values and AOE profiles. and while the impact of these changes is quite heavily debated among the community, comparisons between old and new AI performance, as well as the AI capability of the individual tanks seems to be mostly based on anecdotal evidence due to the lack of somewhat reliable data.

hence, i've tried to generate some actual numbers that would hopefully give a bit of a clearer picture of the overall performance of heavies pre- and post patch, and might also give some indication where some adjustments could be made.

however, since in-game testing to generate a sufficiently large sample size would literally take centuries and precise calculations taking scatter and AOE values into account turned out to be quite a bit more complex than what i am capable of doing (without any programming knowledge and thus relying on excel to do the heavy lifting), i've decided to conduct a numerical simulation to approximate the actual in-game behavior of a tank firing at a 6-man, full-health squad until all entities are killed.

with that out of the way, here's what i did so far:




probability of killing 6 80-hp entities as function of fired shots, d = 40 m


probability of killing 6 80-hp entities as function of time, d = 40 m


average number of shots to kill 6 80-hp entities at range 10, 20 and 40


average time to kill 6 80-hp entities at range 10, 20 and 40

as you can see, the average number of shots it takes to kill a full-health 6 man squad at a distance of 40 m decreased significantly, mainly due to the substantial reduction of scatter. also, the tiger seems to have benefited the most from these changes, cutting the number of required shots / time by roughly 1/3 (~1/4 for KT and ~1/5 for IS2 and Pershing, respectively) and now being almost as lethal as the Pershing. together with the more reliable damage delivery this leaves heavy tanks in a very good (too good perhaps?) spot. what do you guys think?

--UPDATE--

as pointed out earlier in this thread by blvckdream, some of the AOE profiles currently in the live game seem to differ more or less drastically from the intended values given in the patchnotes (SEPTEMBER 2019 BALANCE UPDATE and onwards). a lot of people have thus suggested these discrepancies to be responsible for the surprisingly high lethality of certain heavies, in particular the tiger. so let's see what all the fuss is about.

assuming the stats extracted from the attributes.xml represent the live values it seems that:

1) Tiger I AOE (n/m/f) is 1/1.5/3 instead of the intended 0.25/1.5/3
2) Pershing AOE damage (n/m/f) is 1/0.4/0.175 instead of 0.75/0.375/0.175
3) IS-2 AOE (n/m/f) is currently 0.25/1.35/3 instead of 0.25/1/3.

now the question is, will fixing these presumably bugged stats change anything in terms of AI performance and, if yes, by how much? here's what the numbers indicate:


probability vs fired shots for live and patchnote AOE profiles, d = 40 m


probability vs time for live and patchnote AOE profiles, d = 40 m

while a marginal decrease in performance for all three tanks would indeed be expected by changing the AOE profiles to their intended values, the actual reduction in S2K or T2K would likely only be about 2-3%... in other words: barely noticeable. so unless there are some other bugs hiding in the code that are about to get ironed out, i'd rather doubt hotfixing the numbers above will do enough to please the crowd hoping for a significant performance decrease of heavy tank AI capabilities.

12 Feb 2020, 17:48 PM
#2
avatar of elchino7
Senior Moderator Badge

Posts: 8154 | Subs: 2

Give this man a medal :hansGG:
12 Feb 2020, 17:51 PM
#3
avatar of Vipper

Posts: 13476 | Subs: 1

I wonder if the scatter are is calculated correctly (because the shape is a bit wierd, the intersection of 2 parallels lines and 2 circles) and if you checked for
Distance_scatter_offset.

I would just compare AOE damage circles, dived by scatter area and calculate probability of damage...

But great work. Well done.
12 Feb 2020, 17:56 PM
#4
avatar of elchino7
Senior Moderator Badge

Posts: 8154 | Subs: 2

Suggestion: start the post with a TL;DR explained in layman terms as the way you present it might be a bit intimidating.

Basically, i made a simulacrum of how Heavies tanks AoE work pre and post patch. It ran 10K times and here are the results.
12 Feb 2020, 17:59 PM
#5
avatar of T.R. Stormjäger

Posts: 3588 | Subs: 3

Please add labels to each axis for every graph with measuring units. I had a hard time deciphering.
MMX
12 Feb 2020, 18:11 PM
#6
avatar of MMX

Posts: 999 | Subs: 1

jump backJump back to quoted post12 Feb 2020, 17:51 PMVipper
I wonder if the scatter are is calculated correctly (because the shape is a bit wierd, the intersection of 2 parallels lines and 2 circles) and if you checked for
Distance_scatter_offset.

I would just compare AOE damage circles, dived by scatter area and calculate probability of damage...

But great work. Well done.


thanks for the heads up. I've used the values for scatter angle, distance scatter max, distance scatter ratio and distance scatter offset. from the vague description found in the attribute editor i think distance scatter offset basically just shifts the whole scatter area forwards or backwards (with 1 making all shots overshoot and -1 making all undershoot), is that correct?

Suggestion: start the post with a TL;DR explained in layman terms as the way you present it might be a bit intimidating.

Basically, i made a simulacrum of how Heavies tanks AoE work pre and post patch. It ran 10K times and here are the results.


Please add labels to each axis for every graph with measuring units. I had a hard time deciphering.


all good points, i'll update the OP if i have some spare time tomorrow.

12 Feb 2020, 18:17 PM
#7
avatar of Vipper

Posts: 13476 | Subs: 1

jump backJump back to quoted post12 Feb 2020, 18:11 PMMMX


thanks for the heads up. I've used the values for scatter angle, distance scatter max, distance scatter ratio and distance scatter offset. from the vague description found in the attribute editor i think distance scatter offset basically just shifts the whole scatter area forwards or backwards (with 1 making all shots overshoot and -1 making all undershoot), is that correct?


Think this will explain it easier.

https://www.coh2.org/file/10620/scatter-model2.jpg

but yes as far as I know distance scatter offset work as you describe.
12 Feb 2020, 18:21 PM
#8
avatar of blvckdream

Posts: 2458 | Subs: 1

Are you using the in-game (bugged) values or the theoretical(intended) values published in the patch notes?

I am also pretty sure these stats don't say much about the actual in-game performance because it doesn't take into account cover, elevation, world objects etc.

The KT, for example, is very unreliable in the live game when there is even a hint of elevation or possible world objects that can block the shot. Its shots are often way off-target because of it.

I guess it's because the balance team (forgot) to adjust the distance_scatter_offset when they changed overall scatter values? I don't know what this value does but it could be an explanation for how the KT gun works.
12 Feb 2020, 18:24 PM
#9
avatar of Vipper

Posts: 13476 | Subs: 1

Are you using the in-game (bugged) values or the theoretical(intended) values published in the patch notes?

Value are, if I read correctly, from the editor thus the ones that in force in Live.
12 Feb 2020, 18:30 PM
#10
avatar of blvckdream

Posts: 2458 | Subs: 1

jump backJump back to quoted post12 Feb 2020, 17:51 PMVipper

Distance_scatter_offset.


What does this value do?
12 Feb 2020, 19:03 PM
#11
avatar of distrofio

Posts: 2358

Now this is some real stats, GJ!

I have not enough time now but i will come back later and analyze all the data.
12 Feb 2020, 19:18 PM
#12
avatar of JibberJabberJobber

Posts: 1614 | Subs: 3



What does this value do?


It just moves the whole scatter area forwards a bit, which tends to help with hitting scatter shots against tanks and hitting models behind the model your tank is targeting.


For instance, the Sexton (and the Priest) have a 0.3 scatter_offset modifier. This is one of the most extreme modifiers for artillery in the game. A 0.3 offset modifier means that the entire scatter area is displaced by 15% (or 0.3 = 30% of the "radius" of the scatter).
MMX
13 Feb 2020, 01:25 AM
#13
avatar of MMX

Posts: 999 | Subs: 1

jump backJump back to quoted post12 Feb 2020, 18:17 PMVipper

Think this will explain it easier.

https://www.coh2.org/file/10620/scatter-model2.jpg

but yes as far as I know distance scatter offset work as you describe.


thanks a lot, this looks pretty much like the scatter area i calculate (basically a pie slice with the outer bounds defined by angle_scatter, twice the distance_scatter_ratio times distance or distance_scatter_max, whichever is smaller and finally offset from center by distance_scatter_offset).

Are you using the in-game (bugged) values or the theoretical(intended) values published in the patch notes?

I am also pretty sure these stats don't say much about the actual in-game performance because it doesn't take into account cover, elevation, world objects etc.

The KT, for example, is very unreliable in the live game when there is even a hint of elevation or possible world objects that can block the shot. Its shots are often way off-target because of it.

I guess it's because the balance team (forgot) to adjust the distance_scatter_offset when they changed overall scatter values? I don't know what this value does but it could be an explanation for how the KT gun works.


yes the results use the values found in the attributes.xml, that is the (bugged) live AOE profiles, together with those reconstructed from the changelog for the old values. i plan to do a comparison with the intended AOE profiles vs. the live ones and see how that would shuffle things around.

regarding the correlation of the test results with the actual in-game behavior you are of course right to point out that a lot of other variables (elevation probably being the worst contender here) factor into the overall outcome that are not modeled by this simulation. still, i believe there's at least some value to be found in the results since these tests provide a comparably large dataset obtained under defined and equal conditions.

lastly, i don't think distance_scatter_offset is the reason for the KT's alleged tendency to overshoot. while it is different from the usual value for most tanks (0.185 vs 0.25), the IS-2 has the same offset and seems to work fine.
13 Feb 2020, 02:39 AM
#14
avatar of Jae For Jett
Senior Strategist Badge

Posts: 1002 | Subs: 2

jump backJump back to quoted post12 Feb 2020, 17:29 PMMMX


Were the TTK calculations made using the assumption that a shot is instantly fired?
MMX
13 Feb 2020, 03:15 AM
#15
avatar of MMX

Posts: 999 | Subs: 1


Were the TTK calculations made using the assumption that a shot is instantly fired?


yes, first shot is fired at t=0, thus excluding aim time and other delays that may apply for sake of simplicity. all consecutive shots factor in reload + other delays however.
13 Feb 2020, 04:05 AM
#16
avatar of KiwiBirb

Posts: 789

thank you so much for this! really cool & informative

great work
13 Feb 2020, 05:56 AM
#17
avatar of mrgame2

Posts: 1793

Thanks for effort.

Imo this is part of story. Any one can do a TTK a heavy tank? I feel it is easier to kill a heavy now than old days. Especially if you have allies 60td

As such their AI performance is increased to compensate their lower survivability.

Iirc heavy = heavy lumbering armor. Hard to kill.
MMX
13 Feb 2020, 15:37 PM
#18
avatar of MMX

Posts: 999 | Subs: 1

thanks for the kind words everyone.
i've updated the OP with some clarifications and a comparison between some presumably buggy in-game AOE profiles for the Tiger, Pershing and IS-2, and those announced in the patchnotes from sept 2019 onwards.
TLDR, if you are hoping for a substantial decrease of lethality from fixing these stats, don't expect too much...
13 Feb 2020, 18:38 PM
#19
avatar of elchino7
Senior Moderator Badge

Posts: 8154 | Subs: 2

jump backJump back to quoted post13 Feb 2020, 15:37 PMMMX
thanks for the kind words everyone.
i've updated the OP with some clarifications and a comparison between some presumably buggy in-game AOE profiles for the Tiger, Pershing and IS-2, and those announced in the patchnotes from sept 2019 onwards.
TLDR, if you are hoping for a substantial decrease of lethality from fixing these stats, don't expect too much...


Pushing them back to old CP values + small adjustments "might" be enough as a psychological change. More so if other changes are in the pipeline to units related to them.

MMX
14 Feb 2020, 00:50 AM
#20
avatar of MMX

Posts: 999 | Subs: 1



Pushing them back to old CP values + small adjustments "might" be enough as a psychological change. More so if other changes are in the pipeline to units related to them.



fair point, i guess we'll have to wait and see how things turn out in the grand scheme of things.

on a slightly different yet related note, does anyone know what exactly the scatter reduction most heavy tanks get with vet does or does not affect, i.e. angle_scatter, distance_scatter_max, distance_scatter_ratio (my guess would be all of these, but i'd like to know for sure)? it would be interesting to see how powerful these vet bonuses are, but i couldn't find any info on that so far.

1 user is browsing this thread: 1 guest

Ladders Top 10

  • #
    Steam Alias
    W
    L
    %
    Streak
Data provided by Relic Relic Entertainment

Replay highlight

VS
  • U.S. Forces flag cblanco ★
  • The British Forces flag 보드카 중대
  • Oberkommando West flag VonManteuffel
  • Ostheer flag Heartless Jäger
uploaded by XXxxHeartlessxxXX

Board Info

442 users are online: 2 members and 440 guests
DIRTY_FINISHER, SneakEye
7 posts in the last 24h
33 posts in the last week
87 posts in the last month
Registered members: 44639
Welcome our newest member, trickproblem
Most online: 2043 users on 29 Oct 2023, 01:04 AM