Taking this out of the SE5a thread ...

What AI issues do you think need to be resolved? What kind of improvements would you like to see?

Before any criticism, I want to say that the AI in RoF is some of the finest that I have ever seen. It is a very difficult task and probably a subject of endless debate.

Now, my #1 request: I feel that AI accuracy is the #1 AI issue right now.

1. In some aircraft the AI uses a non-historical, unrealistic maneuver to consistently kill the player.
2. The effectiveness of the pull up into a stall and snap off a well aimed shot scenario is an AI issue, not an FM issue. The airplane is behaving just fine, it's the AI that is not.
3. The general accuracy of the AI, front and rear, is an AI issue that can be solved.

The solution is to impart simulated human limitations to the AI. In the case of accuracy, the idea that I proposed would simulate human limitations by requiring the AI to take a period of time to acquire a firing solution. This would prevent the AI from achieving consistent success under unrealistic conditions.

In the SE5a thread I threw out a solution that - if I were coding - I would explore. Since I'm not coding, my opinion means about as much as any other opinion posted on the internet . However it may be done, I do hope that 777 takes a look at many aspects of the AI with an eye towards simulating more human behavior.

I support your overall contention that AI accuracy is the number 1 issue in an overall high quality AI implementation. It is by far the one thing that could improve my enjoyment of this sim more than any other (and I do enjoy this sim from so many perspectives). Especially item 3 on your list. I spent hours the other day reading over every post, past and present, that I could find on this issue trying to figure out what I could do better to deal with AI accuracy (especially of rear gunners). I subsequently went into the sim and tried these techniques, many times, and none really moved me forward.

Yes, it's the non-humanity of the AI that annoys me most of all.

The diving for the deck.
The notion that my body is actually 1/2 the size of the plane, meaning 50% of every bullet that hits is hitting ME in the cockpit. That might be pilot hitbox size or the AI's ability to aim for the pilot directly.
The ability for gunners to compensate for any move I make while his pilot is also making moves to perfectly blow my nose off (not the plane's nose, see above).
Snapshots from historically unlikely angles as mentioned.
An unconfirmed notion I have that enemy AI will seek you out over your wingmen/allies whilst those AI allies seem to make little if any effort to follow you. Hitting the F2 key I'll often see me vs 3 enemies and my 3 wingmen vs 1 enemy. In a flight of 8, it's just worse.



The Jedi Master

They need to make it so the AI will take shots at planes from non-perfect firing positions. The problem with the AI now is that when they fire they hit you because they wont fire unless they are lined up. A real human will make guesses and push their plane in a general direction in hopes of getting a shot. The AI does not do this and it is evident.

They also need to destabilize the AI guns. A random value that gets added to the bullet arc would fix it. The idea here is that the AI lines up to fire on you and it controls every aspect of the plane to draw a perfect bead and keep the plane aligned then it fires directly at your pilot. If they cant fix it so it takes non-perfect shots then they need to add a random value to the bullet trajectory (on a per burst) so that the burst moves away from the direct pilot shot in a random direction at a random angle (all very minor but just enough that a pilot hit is not achieved with every pull of the AI trigger.

I haven't seen the RoF AI code, but most combat flight sims do introduce some random element to when the AI take a shot, whether it is an 'angle or arc of fire' that the target must be in for the AI to take a shot. It amounts to a cone of uncertainty that gets larger with less AI experience (ace=narrow cone, green=broad cone). RoF's cone may be too narrow, thus sniping.

I've said this before in other threads, for other sims as well, but there is a limitation in modeling the AI, that is the AI has one objective: to get a guns solution on you. It does not work to improve its position.

What ends up happening is the "pull up and snipe" maneuver, rather than the AI trying to seek energy parity.
What also ends up happening is that the AI will fly like you do because both of you are trying to get guns on target. You circle and the AI will circle, you go vertical and the AI will go vertical, stall fight... almost like your moves are being mirrored. Try it. Is there any resolution? I don't see a way out, but I'm not a programmer.

There was a terrific board war game called Squad Leader decades ago. In that "sim" platoons of men are in combat, and under certain conditions (such as their squad leader dies...etc) they have to check morale. Imagine a morale check being implemented in RoF: Each AI aircraft has to make a certain percent 'check for morale' to engage or not engage, OR in worse case scenario, to split-S and run like he!! when:
1) flight leader goes down (morale check with a heavy penalty)
2) damage to aircraft (depends, if gas tank hit and leaking, then almost 100% run because the fumes are super dangerous to ignite)
3) out of bullets (again...)
4) being out numbered
5) being at lower altitude
6) etc...

The various factors can be cumulative penalties to the check. This might yield a more human AI, but one that may be reluctant to engage, which may piss some people off. Playing late war German Air Service could be frustrating with the SPADs and SE5s running for it.

Again, the AI does not have to run-like-a-little-girl, but if the player has not already engaged and they fail morale check they can choose not to engage and fly away to seek more altitude. If this isn't clear, or you think it sh!te, say so.

Boelcke Dicta

Yes, but how do we implement it?

It's certainly not easy. For a start I would like to see the accuracy issue addressed. That I really do believe can be done.

Making the AI understand combat moves would be really, really difficult. It can definitely be done, but doing it is the equivalent of teaching it to play chess.

Here's an example of how it could start to work = ... maybe :):
One would have to code a decision tree -
Pick mode:
- Track and shoot.
- Flee.
- Setup.
- Evade.

Mode selection would be based on AI position, enemy position, odds, immediate circumstances (being shot at), damage state, etc. The programmer has to be the AI pilot - what are the factors and what is the final decision?

Once the mode is chosen the specific action must come next.
Flee:
Turn towards home and dive as fast as possible.

Evade:
In immediate danger - turn, spin,some other violent evasive action. Speed, altitude, aircraft capabilities all come into play.

Track and shoot:
Similar to current AI behavior. Try to make corrections for a shot.

Setup:
This is where the chess analogy comes into play. No short term solution presents itself - what to do? A list of potential maneuvers is available. The list has to be weighted for best choice depending on current position. Once this is done the maneuver should be carried out to completion. This is how to escape the "mirror image" AI behavior.


How might this play out?
AI sees player. Mode = setup. Maneuver = merge.
AI approached player head on. Mode = Track and shoot. Maneuver = slight correction for solution.
AI has passed player. Mode = setup. Maneuver = wingover.
AI completes maneuver. AI is above and to the right of player. Mode = setup. Maneuver = diving turn.
AI completes maneuver. AI is behind player at 4:00. Mode = Track and shoot. Maneuver = turn right for solution.

and on it goes. It is actually more difficult than the chess analogy because the always does the best thing. For AI you have to randomize it and sometimes make a bad decision based on AI level.

All of the above probably barely touched on the difficulties. Definitely not easy. However, I do think that a chess like weighted algorithm with an element of randomness would be interesting (something that I have always wanted to code too )

The thing that pi$$es me off more than anything is my worthless, piece of cXXp wingmen who do NOTHING TO ENGAGE THE ENEMY THAT'S ATTACKING MY A$$ OFF!!!!!!!!!!!!!!!!!!!!

I've been tempted to shoot their sorry a$$ down out of spite!!!!!!!!!!!

I would just like it so that the AI (both friendly and enemy) would understand when to disengage and not chase a target half way across France.

I'd like to see the AI break off when they are out of ammunition. Some amount of inherent survival instinct would be good too "...Oh, look Ma, no wings...hmmm time to head back to the aerodrome".

My bloody arms were dropping off this morning as I tried and tried again to get my attendant 'sightseers' to earn their corn and follow instructions but they steadfastly ignored the balloon and armed convoy. With this bunch of loafers I didn't fancy taking on approaching E/A so beat it back to base. Result: 1 ballon. Poor return for a five plane sortie.

OK, calmed down and had a nights sleep.

I'm wondering if it's my fault and there's something I'm not doing correctly when I try to "manage" my flight-but I don't think so.

I have the default "k" key as well as a button on my joystick programmed to tell my wingmen to attack. I've gone to the external plane view and I can see "me" waving and pointing at the enemy targets but they don't do diddly. I've tried slowing WAAAY down so my wingmen are in my back pocket and are close enough to "see" me-doesn't seem to matter...

Has anyone had success getting wingmen to follow commands?

1. Able to turn horizontally without losing altitude
2. Able to turn near maximum turn rate (current AI don't).
3. Able to dive at a target
4. Better at shooting from the rear hemisphere than from the front hemisphere
5. Natural deflection shooting - no more banking the wings 110 degrees and pushing negative g's for a shot
6. Actually respond to orders to engage, etc.
7. Correct tactics for different aircraft, e.g. hit and run with Spad, etc.
8. Self-preservation, i.e. will not engage 4vs1 if there's an option to run for it.
9. Greater variety of defensive tactics besides spiraling earthward
10. Cooperative tactics

Honestly, over the years I've noticed in many sims, including RoF but PARTICULARLY in Il-2, you have useless wingmen. You get attacked? They don't care. They'll go off on their own and attack whatever.
HOWEVER, once you get on someone's six and are laying into them? Guess who shows up, almost colliding with YOU as they steal your kill from under your nose???
I would have 2x as many kills in the average Il-2 mission if I didn't have wingmen stealing them from me. Of course, they always show up AFTER I've put dozens of rounds in. So I waste 2/3 of my ammo load shooting up planes I don't get to take down.



The Jedi Master

Less precision in coordinating turns.

A lot of problems with the AI in gunnery, flying, etc. would be solved if they weren't such perfectionists at keeping the ball centered.

If you haven't seen it yet HERE ...

So all our hypothesizing, and theoretical solutions, are likely for naught and we will be left with the warts and syphilis of our current opponents and enemy. This is disappointing given the high fidelity of most other aspects of the sim.

I think I recognize people from this community who used to fly (may still do...) First Eagles (FE) by Thirdwire. While the flight models and physics engine of that 'simlight' do not approach RoF, I have to say that FE's AI (both your wing men and hostiles) at least give you a very realistic fight. The AI does as commanded, the opponents try to energy manage, etc.

I think it all comes down to the programming foundation that was established and whether the soul and determination is there to make a better AI in RoF.
Is it too late to make a silk purse out of the sow's ear?

If funding is the issue, why not work towards RoF II, which could correct the AI foundation and could include a channel map for $xx.xx. Given the vocal support for funding continued development, I do not see why this wouldn't be successful.

Il-2 had terrible AI until the 4.11 patch. Now it's tons of fun to fight the AI. If a few guys working for free could turn something so bad into something so good, I don't see why the same couldn't happen for RoF. The first thing is to incorporate Criquet's changes into the default game.

I fly with Criquet's mod, while there is a slight improvement, it isn't much at this time. Just an opinion, but it looks like the problem lies a bit deeper than the files we have access to.

I'll try to find the thread that explains how the AI are tasked within a mission, and this explains a lot about why they behave so strangely in many contexts.
I'm sure Pat can chime in to verify given his knowledge of mission generation.

I fly with Criquet's mod, while there is a slight improvement, it isn't much at this time.

Ditto. If you're into flying Mods On, this one's a must. While it's far from perfect, it does make the rear gunner AI a little easier to deal with and the scout AI flies a good bit better. I was actually able to down a DFW CV in my new 8 RNAS Career in a Sopwith Pup, while taking no damage from the rear gunner! I dove in from above, making my profile move constantly from the gunner's point of view, leveled out about 150 yards behind the enemy aircraft, lead, and shot, and wound up flaming the engine. The gunner's shots were close, but it was far more realistic than the previous sniping-the-pilot-no-matter-how-perfect-the-attack-is AI.

The main thing that bothers me about the AI isn't the enemy. It's when I lead a flight, and my wingman(men) will go running off after the first enemy aircraft they see regardless of whether or not I gave the order to attack. I have to fire a flare telling them to maintain formation and cover me to get them to return. And this doesn't happen just once per sortie. They'll take it upon themselves to attack anything they see as many times per sortie as they spot something. It's unrealistic and very annoying.