Gun accuracy calculation broken

[Firestorm01X2]

Players reporting of massive reducing in firearm accuracy.
Forum thread:
http://smf.cataclysmdda.com/index.php?topic=14523.0

Looks like we have massive regression after that:
#20941

This issue should be set as milestone to 0.D and probably marked as "Priority" because shooting is one of the core mechanics of the game.

[kevingranade]

Need expected behavior vs actual behavior, i.e. how accurate do you expect it to be?

[Leland]

Was #20941 expected to change accuracy? And if so, do we know to what degree it was changed?

I ask, as leading the discussion with a breakdown of how accurate it was prior to that PR vs. how accurate it is now should allow for a reasonable approach to how accurate firearms should be.

[kevingranade]

I did not expect #20941 to strictly maintain the previous level of accuracy, no.

There was no standard for accuracy in the first place, so while there might have been a change, you can't really call it a regression.

Again, give me a scenario, an expected result, and an actual result, and we can get this party started. Otherwise it's just going to go in circles indefinitely.

[John-Candlebury]

Just for reference #18943 (EDIT:mb I linked the wrong one) has the pre-revert unit tests. Which would serve as the previous standards players are comparing current dispersion against.

Someone would need to recreate them for the actual state of the game and make the comparison however.

[kevingranade]

Yea that's one if the incredibly frustrating things about that code. That unit test doesn't actually do anything. It added a "determine engagement range" method and tested that, not the actual accuracy of the weapons.

You may notice the test doesn't make any assertions about effective range or accuracy, because it isn't testing that, just that the function itself has certain properties.

[AlecWhite]

Again, give me a scenario, an expected result, and an actual result, and we can get this party started.

What do you mean by a scenario? Something like player skill/stats/modifier, range, gun stats, target skills/stats/modifier and whatever variables are needed to get the a shooting result?

I'm sorry, I don't understand what is the current problem with guns, only that they work "wrong". And if the shooting target formula is decide, and the code works, it would only need the tweaking of how much each variable weights for a successful shot.

[kevingranade]

scenario

yep, what you said.

only that they work "wrong".

People keep saying it's "wrong", but won't or can't say what's "right". If we don't have a way to check whether it works correctly, how do you expect to fix it?

[Coolthulhu]

Yea that's one if the incredibly frustrating things about that code. That unit test doesn't actually do anything. It added a "determine engagement range" method and tested that, not the actual accuracy of the weapons.

That method was a great measure of actual accuracy - simple, reliable, consistent, meaningful.
It probably can't be accurately recreated for the much less natural sum of uniformly distributed variables, but if it could, it would be good enough.

The problem with the current system is not the numbers, it's the distributions and the way they're combined.
Summing non-negative, uniformly distributed variables must weight the system towards expert, naturally gifted marksmen, creating a situation where guns require a lot of training to be useful and reward going from "good" to "a bit better than good" more than going from "average" to "competent".
In most cases, Joe Average will be better off practicing with a knife than using a gun, even if the gun was fully silent.

But for the scenarios: I can't think of a better measure than just recreating the old tests, except by adding some extra upper caps to prevent the system from blowing up for skilled marksmen.
Those tests already exist, are balanced, are consistent with what players were used to, can be understood without even learning what dispersion is in the current revision. The only thing they don't do is testing the implementation details - but I don't think they should even test that.

[kevingranade]

Those tests already exist

I'd like you to point them out, because I couldn't find them.
I believe this is where they were supposed to go: https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L190

To break down the rest of that file:

This test: https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L147
merely asserts that for a given aim level and player/weapon/target, projectile_attack_roll() agrees with projectile_attack_chance().

This one: https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L34 asserts that for all guns at selected aim level/ player skill / range, gun_current_range() agrees with projectile_attack_chance()

This one https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L64
and this one https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L72
assert that a special overload of gun_engagement_range() reports a longer effectve range for higher aim levels.

This one https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L80
asserts that gun_effective_range() reports a longer range when dispersion is lower.

This one https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L93
and this one https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L108
assert that gun_current_range() scales range based on requested accuracy level.

This one https://github.com/CleverRaven/Cataclysm-DDA/pull/20941/files#diff-ae9fb308eca169e424e32c565235db95L108
asserts that all else being equal, smaller targets are harder to hit.

Except for the last test, they are all testing the methods themselves, not anything about game data. For the most part they're checking for consistency between two methods, which doesn't demonstrate correctness, just that if errors exist, they exist in both methods.
None of these construct a scenario and assert an outcome.

[Coolthulhu]

OK, looks like I remembered wrong then - my bad.
The dump.cpp data was compared manually, there was no complete regression test.

Still, the values themselves are good:
#18020 - very good analysis of the math involved, with good sample numbers for a relatively broad set of weapons, at 3 different character skill levels
Just approximating those numbers would be a really good starting point.

[PlasmaChroma]

From my current testing/play of the new system, I'd say what feels "wrong" to me is that skill levels seem to not have as big an impact as one would hope. Both gun-specific and general Marksmanship slightly move the actual performance in the players favor. Probably somewhere between 30% to 80% more from that would feel better.

Although in terms of balance, getting the actual skill levels is unnatural as well. If you're trying to train weapons via practice it seems to take excessive amount of time to skill up anywhere high. Whereas a lucky character can just read a few books or train via NPC and "instantly" be as good a shot as a player who has killed countless hordes over months and months of time to gain the exact same skill. The actual impact of skills in the case where they are gained "cheaply" seems like it would have about the improvement that it does now. But a hardened veteran with real combat experience is under appreciated.

Perception also helps a bit here, although I imagine a lot of players are likely running low-ish perception characters, at least I always have until now. Typically I wouldn't push it past 10, but on my latest run I'm experimenting with a natural 14 to see how big the impact is.

To just ballpark it, I'd think at 25 yards with a decent rifle and iron sights you would expect to hit fairly often even with low-ish skills and some steadying against normal sized target.

[Coolthulhu]

From my current testing/play of the new system, I'd say what feels "wrong" to me is that skill levels seem to not have as big an impact as one would hope.

They do have an effect on accuracy, it's just that now it's only really visible near level 10, with great equipment and high perception.

[JacobKessler]

I posted these in the forum thread and, while I didn't get any comment on the proposed scenarios, they seem reasonable to me. They obviously aren't exhaustive, but are at least a start toward some tests:

• guns 0 without steadying effectively cannot hit. Wild, unskilled fire should be ineffective
• guns 10 without steadying shouldn't be able to hit often outside of 5 tiles, regardless of weapon
• guns 5 should make shots 'effective', meaning a 75% chance of hitting at mid-steadiness, at around the following ranges (modulo gun differences):
• pistols: 5-7 tiles
• SMGs and shotguns: 10-15 tiles (shotguns have raw damage, SMGs have automatic fire and common ammunition)
• rifles: 20-25 tiles
• guns 10 with a single '.' of aiming should be able to reliably hit within 'close range', 5-7 tiles
• guns 10 with mid-steadiness should give effective ranges of ~50% more than guns 5
• guns 10, with scope, a decent rifle, and max steadiness, should have a 75% hit range of at least 60 tiles.

[kevingranade]

#18020 - There are literally over a hundred comments in that PR, "do what that novel over there says" is not a valid bug report, especialy when various parts of the comments almost certainly contradict each other.

@JacobKessler that's the kind of thing I'm looking for, thank you.

guns 0 without steadying effectively cannot hit.

That's certainly testable, but I'm not totally sure it's accurate, what about a point-blank sawed-off shotgun blast? I could see just excepting e.g. spread weapons and calling it a day though. Actually lets just assume spread weapons aren't covered in general.

guns 10 without steadying shouldn't be able to hit often outside of 5 tiles, regardless of weapon

Needs to be a bit more specific, what's "often"? (1 time out of 10 would be a testable example).
This scenario roughly matches effective quick-draw range for competitive shooters, so something like https://www.cowboyfastdraw.com/images/stories/pdf/Forms/GunslingersRules&Guidelines_9th_Final_11-28-16.pdf
tl;dr page 19: 24" targets at 21', 17-3/16" target at 15'
Could use a little more figuring, but after scaling that sounds reasonable-ish.

guns 5 should make shots 'effective', meaning a 75% chance of hitting at mid-steadiness

This middle part is SUPER arbitrary, for one thing, why mid-steadiness? Especially for the first batch of tests, its seems more productive to lock down the limits rather than test scenarios in the middle.

guns 10 with a single '.' of aiming should be able to reliably hit within 'close range', 5-7 tiles

It doesn't make much sense for this to be uniform across all guns. In the extreme case you're saying that the worst gun in the game can hit reliably at a range of 7 (~25') with a single aim action (which is itself fairly arbitrary). It makes sense to put limits on the best guns can be, but limiting how bad they can be doesn't make much sense.

guns 10 with mid-steadiness should give effective ranges of ~50% more than guns 5

Let's not get into scaling and distributions until we have limits locked down.

guns 10, with scope, a decent rifle, and max steadiness, should have a 75% hit range of at least 60 tiles.

Define "decent" :/ Also "at least" is pointless here since you can't shoot further than that.
Otherwise though, sure, it makes sense to have a test around some whitelist of endgame-tier rifles being able to shoot things at the edge of the map.

[Coolthulhu]

OK then, breakdown:

• Each gun+skillset pair should have medium range, at which it is accurate enough to actually hit ("good hit" rating) at 50+% chance
• Each gun+skill should have a maximum usable range, beyond which the chance for a good hit is below 10%
• Point blank hits may be stupidly accurate for now, as the most proper way to balance it is include it is to cause target dodge to disrupt aim, which is beyond the scope now
• Low skill characters should not be penalized extremely hard, as they are now
• Getting last few points in skill/stat should not grant disproportionately good rewards, as it is the case now
• Getting a slightly better scope/gun should not increase all ranges heavily, as it is the case now

Let's define 3 characters:

• Novice: 8/8/8/8, 0 in all skills, 10 encumbrance on all parts, 300 dispersion (gun+ammo), 90% confidence (novices have a big time penalty for aim)
• Competent: 10/10/10/10, 5 in all skills, 5 encumbrance on all parts, 150 dispersion, 95% confidence
• God: 14/14/14/14, 20 in all skills (just to check the caps, can be 10), 0 encumbrance, 0 dispersion, fully aimed (0 recoil)

Novice character checks the lower bounds. Competent character and novice together ensure that the lower levels have sensible numbers.
God character is only there to make sure the numbers don't blow up, meaning adequate formulas are used. For example, "5 dispersion for every point below cap" is a very bad formula for uniform variable summation, and the god character ranges should catch that.

In case you want to balance it with per-gun settings (you proposed something like this a ~year ago), let's also check different guns:

• Pistol and buckshot-loaded shotgun
• SMG and pneumatics
• Rifle
• Slug shotgun could be either at rifle, or halfway between SMG and rifle

If not, just use the rifle numbers for everything. This will buff everything, but guns are currently very bad, so it's better to overbuff (except for free ammo ones, like pneumatics, and for endgame characters) than to nerf.

Character	>50% good	<10% good
Novice pistol	3	5
Novice SMG	5	8
Novice rifle	8	15
Competent pistol	7	10
Competent SMG	10	15
Competent rifle	15	45
God pistol	10	15
God SMG	20	30
God rifle	30	60

The numbers I picked were selected so that:

• Pistols are short ranged at all times. Their strengths should come from quick aiming and high damage output per time unit. Volume and ability to pull them out quickly should not be considered a sensible advantage, because they're nearly irrelevant in actual game. Maybe wielding guns with 1-handed melee will become an advantage in the future, but that's the future
• Rifles always allow some sniping, even at low levels. Their disadvantages should come from slow aiming and somewhat from harder to find ammo. Other disadvantages are too situational to matter - slow reloads are already handled by slow aiming, large volume is nearly irrelevant, slow draw time is a minuscule disadvantage.
• SMGs, being in the middle of the road, have the highest damage output at middle range and - when aimed - allow devastating short range bursts. Single shots should be slower to aim than for pistols, though.
• Novices can actually use guns. Slowly - note that novice will take a long time to reach 90% confidence - but can still score good hits against targets that are assumed to be immobile.
• Competent shooters can effectively use guns at "city" ranges with some time spent aiming
• Numbers don't become stupid after that

[Firestorm01X2]

This issue should be set as milestone to 0.D. It will be bad to forget to fix this.

[Firestorm01X2]

Temporary fix (fork containing non bugged aim) by Coolthulhu:
http://smf.cataclysmdda.com/index.php?topic=14767.0
https://github.com/Coolthulhu/Cataclysm-DDA/releases/
Use that until it will be fixed in this repository.

[kevingranade]

That's a fork, not a fix.
A fix is in progress at #21468

[ghost]

It's a fork containing a fix.

[Coolthulhu]

Note: the actually implemented tests and balance standards are significantly less strict than the ones I proposed.
The guns should be useful now, but this doesn't mean they are balanced against other options or against each other.

Notable balance problems are: rifles are still just plain better than SMGs and pistols, "endgame" characters can easily reliably snipe much farther than they can see, perception got less useful despite already being much worse than dexterity.

[TeMPOraL]

Are those numbers based on gut feelings and back-of-the-napkin math, or did you have a mathematical model for them? It would be great if there was a well-defined one, which could later be enshrined in unit tests as well to safeguard against future balance-breaking changes.

[Firestorm01X2]

Likely it is "game feeling by personal experience".
But digits looks nice and there are no other balance outline so that is it. It was accepted ofically.

And unit tests already were impelemented here: #21468
Tests just were implemented not completely corectly based on that design.

[kevingranade]

As Coolthulu stated, the numbers are based on a number of game balance goals.

The only input from reality is the quickdraw based tests, whose upper bounds are based on the records established by actual quickdraw competitions.

Please open or comment on current issues for future discussion.