Suggestions for AI Improvements

[Pursuivant]

Originally Posted by majorfailure

If estimates are correct, then almost half of planes shot down in real life were unaware or aware too late of their attackers. If Il-2 reached close to 10% I'd be pleased.

In some cases, it was more than 50%. So, there has to be a way of "dumbing down" AI, especially for single-seater planes, and especially for attacks from the rear and from below (and out of the sun, but that's possibly modeled).

In some cases, even multi-crew bombers can be surprised if the crew isn't keeping a proper lookout. But its much harder because each bomber has 1+ guys whose job description includes looking out for fighters.

Another big deal is lack of Situational Awareness. Flying in close formation means that you lose most of your SA. Attacking means you (probably) lose SA. Trying to navigate means you lose your SA. Dealing with damage or pulling a high-G maneuver means that you might lose SA. So, in addition to reducing sighting chances when you're doing those things, there also has to be a chance that you "lose acquisition" of a previously "acquired" target. That is, you lose sight of it and you can't figure out where it is.

Originally Posted by majorfailure

I'd say if you additionally allow for radio comms (e. g. better AI in the region/same flight calling out the bandits/bogeys), and also do not make spotting skill rigidly connected to general AI skill (there were pilots with keen eyesight even as rookies, and some of those had nerves of steel - future aces to be)

It's currently possible to call out bogeys, and AI will occasionally call out bogeys. It will also occasionally call out that you've got a bandit on your tail.
Generally, though, "In combat" AI communication is a weak point, which I forgot to address in my initial wishlist barrage.

It's probably too much to hope for for realistic bomber crew behavior (i.e., individual crew skills and traits) and communications. But, fighter squadron behavior might be fixable.

* The lag problem for in combat commo needs to be fixed. Sometimes it might be several seconds until the AI calls out a particular command, leading to odd situations where the dead pilot of an exploded plane says, "This is 2 (calm voice), I'm bailing out (panicked voice)."

* AI needs to distinguish between "bogeys" (i.e., unidentified aircraft) and "bandits" (i.e., hostile aircraft).

* Wingmen should always call out bandits attacking the lead.

* Friendly planes in your flight/squadron should call out enemies they happen to see attacking another friendly within the squadron.

* Wingmen should tell the lead how to break when a bandit is incoming to set up a "drag and bag" attack. (i.e., break left/right, climb, dive)

* If maneuvering by sections or flights, flight/section leaders should tell the other section/flight how to maneuver to avoid/attack incoming bogeys.

* There should be individual squadron call signs - at least for the air forces that used them.

* Anything other than rookie and unqualified AI should identify themselves by squadron, flight and number. If they know it, they should also call out squadron, flight and number for a plane they're addressing. E.g., "Tare Red 2, break right!, or "Tare Leader to Tare Green 3, rejoin formation!" (This is actually a bit unrealistic and clunky, since typically pilots used personal names or nicknames in combat, e.g., "Mickey, bandit on your 6! Break right!" But, it would be easier to program.)

* The should be "calm" and "excited" voices for squadron, flight and number identifiers. It's sometimes a bit confusing to have "This is 2. (calm voice) Help me! (panicked), especially when there are multiple flights or squadrons in the air.

* Unqualified and rookie pilots might "step on" commo channels, making it impossible for other pilots in the squadron to use the radio.

* No radio means no radio. Currently, AI planes which don't have radios or which had historically unreliable radios, can maneuver and respond to commands just as effectively as those without (as far as I can tell). Realistically, if you're flying without radio, or your radio is dead, commands like "rejoin" or "cover me" are impossible except at very short ranges.

Originally Posted by majorfailure

It is realistic to try. And it is realistic to hit when target approaches on a low angle low deflection path, dead six for rear gunner. It is not realistic to make hits dead center when target approaches fast, off angle and from above - in a situation where even hitting the attacking fighter at all would be lucky or extremly skilled. And I have had this more than once - either real bad luck or too high chance to hit.
(On a sidenote, I have been shot three times by now by rifle caliber guns, me flying an nearly undamaged IL-2, technically impossible IMHO)

I spent a lot of time flying QMB missions in arcade mode against ace bomber formations to try to figure out how IL2 engine damage models work, especially against small caliber bullets. My experience is that, while there are plenty of problems with AI flexible gunnery, it's not unrealistic that they aim at center of mass.

And, if you are moving at high speed and high angle off, particularly if you are maneuvering in 3 dimensions (like a diving pursuit curve) you can throw off the AI gunners and you'll "only" get hits near the edge of your airplane. Likewise, fighters making high deflection shots often hit you in the wing or tail rather than center of mass.

Your point about rifle caliber bullets being excessively lethal is spot on, though. I've documented this extensively in previous posts. Even .30 caliber AP bullets shouldn't be able to penetrate armor plate at anything but short ranges.

AI gunnery problems

Doesn't factor in plane vibration
Doesn't factor in air turbulence
Doesn't factor in wind buffeting on exposed gunners
Doesn't factor in slipstream buffeting on guns
Doesn't factor in gunner reaction time (in some cases. In other cases the AI seems right).
For hand-turned guns, it doesn't factor in the limits at which the gunner can move his body to accurately sight the guns as opposed to maximum cone of fire. Realistically, this should make ventral guns in planes like the He-111 or A-20 pretty well useless.
Gunners have unrealistically good Situational Awareness which allows them to perfectly track attacking aircraft.
Gunners start shooting at historically inaccurate ranges - this is inappropriate for anything other than rookie/unqualified gunners
Gunners have unrealistically good ranging ability.
Gunners have perfect coordination with the pilot, such that they know exactly how the plane will move and can instantly adjust their aim accordingly.
There is no stoppage of fire when the gunner would realistically need to reload ammo cases/drums.
Aircrew aren't always injured by nearby cannon shell hits (realistically, a 20 mm explosive shell going off in an enclosed space like a gun turret is at least going to stun you).
Flexible guns never jam or freeze
Gunners never panic
Gunners never get fixated on a target
.30 caliber/7.62 mm bullets are unrealistically dangerous, especially at longer ranges.

Originally Posted by majorfailure

Equal numbers, first plane I hit usually gets labeled "lesser threat" - at least when I have the feeling I hit something that might be vital

This would actually be fairly easy for AI to distinguish if some of the damage modeling info were available to it. When a particular aircraft shows a "heavy damage" skin for some vital part (e.g., wing, engine, cockpit) then AI set for "defensive" level aggression or less, or coping with superior numbers of enemies, will ignore that plane.

Something that I didn't really touch on was how to get AI to recognize that the odds are/aren't in their favor. This would actually be easy to do.

Give each plane in the game a combat effectiveness/threat rating of 1-10. Ratings go down for damaged or seriously damaged planes. Multiply that by the number of planes. Divide the number if a plane is far away - more than 1,000 meters. Multiply or divide that by positional advantages (altitude advantage, frontal or stern attack). Divide or multiply the number if you're over hostile/friendly territory.

If the total number for your side is less than that for the total number for the enemy side, you're disadvantaged and react accordingly (unless AI is set to be Aggressive/Suicidal). If the total number for your side is greater than the total number, you've got an advantage and react accordingly.

The same simple calculation could be used to determine which threat an AI plane reacts to first. Close threats to the rear and/or above are generally going to be a priority compared to more distant threats or threats from below and to the front, but there are exceptions.

If you want to get a bit more complex, you could also assign each plane in the game a "maneuver rating," an "attack rating", a "speed rating" and a "defense rating"

If your maneuver rating is better than the enemy's, you go for turn fighting. If your maneuver rating is poorer, but your speed rating is better, you go for BnZ.
If your maneuver and speed rating are poorer, you go defensive.
If your attack rating and defensive rating is better than the enemy's, you go for head-on attacks. Otherwise, you avoid them.
If your defense rating is lower than the enemy's you maneuver more defensively to avoid taking hits.
The lower your defense rating, the more cautious you will be about making attacks against large bomber formations or ground attacks against targets defended by flak.
If your speed rating is lower than the enemy's, you know that it's futile to go for shots from stern or to chase a fleeing enemy.

These calculations could be made once every 15 seconds or so (longer intervals for less experienced AI levels), so the total effect on fps would be minor.