AI vs AI Kill ratio

[taly001]

I think the problem with AI in ALL computer games is that they can track the player once spotted with 3D X-ray vision. In real life when sight is lost the player needs to re-find the enemy each time, and then to re-analyse the maneouver the enemy is making. The AI doesn't seem to need this and can track and follow your moves all the time.

This may partly be why AI vs AI is such a kill-fest. They never loose track of each other so will fight to the death. Forcing AI to reacquire the target after loosing sight would require CPU cycles, drawing predicted target flight path from time sight lost and have AI look there first, then scan around to find before been allowed to manoeuver to attack.

[Storebror]

I think your idea bears some useful points to start thinking on.
Already now, IL-2 has the ability the check whether actually AI can see an enemy plane at all, taking own plane obstacles into account as well as buildings, mountains etc.
But you're right that AI will instantly reacquire the target once it comes in it's sight again.
The latter could be changed closer to reality by taking a few parameters into account, e.g.:

• The larger the relative movement between own aircraft and target (when it moves out of sight), the longer you need to scan (because it will become hard to predict where the enemy will be).
• The smaller the relative movement between own aircraft and target (when it comes back in sight), the longer you need to scan (because it's harder to spot non-moving targets).
• The more G's are pulled by enemy aircraft, the longer it takes to scan (the future position will become hard to predict under heavy maneouvering).
• The more G's AI pulls themselves, the longer it takes to scan.
• The further the enemy distance is, the longer it takes to scan (dots becoming smaller).
• If the target is below the horizon when it comes back in sight, add a scan penalty.

With these factors a scan time could be calculated where AI would not immediately track a target when it comes back in sight, but just after this scan time has elapsed.

Best regards - Mike

[Pursuivant]

Originally Posted by Storebror

But you're right that AI will instantly reacquire the target once it comes in it's sight again.

That removes several important elements from the OODA Loop.

When you reacquire a target, your brain spends a fraction of a second on the Observe, Orient, and Decide portions of the loop before you Act. AI should do the same, if only by briefly hesitating before acting, and with a small percentage chance that AI won't automatically reacquire (because the pilot briefly had his head down in the cockpit, or his vision was briefly obscured by "Fog of War" - literal or otherwise.)

My experience playing the game is that Storebror's list of factors to take into account is very true once you lose Line of Sight (LoS), but sometimes makes maintaining a "lock" on an airplane you can see easier.

For example, a plane which is fast-moving, or committed to a high G maneuver is in some ways easier to predict because a fast moving plane is more or less committed to a particular speed and vector. Likewise, a plane pulling serious Gs has "no place to go" but to a lower G state. Same thing for a plane in a very poor energy state, or a plane with certain types of damage.

But, I'm not quite sure how to get AI to recognize all those things, even if it's possible.

[Storebror]

One thing we shouldn't forget is that AI is comparably "stupid" when it comes to draw conclusions based on the facts they know.
Human players can utilize their experience from previous sorties and can adopt promising options even in uncertain or unknown situations.
AI can't do any of this.
AI has a simple tree/branch decision scheme, sometimes leading to rather stupid moves.
In that regards, letting AI "cheat" in another regime to compensate this lack of experience might be a valid decision to some degree.

Best regards - Mike

[sniperton]

Originally Posted by Storebror

One thing we shouldn't forget is that AI is comparably "stupid" when it comes to draw conclusions based on the facts they know.

It would be intriguing to know what are the ‘facts’ they know.

1. There seems to be a 'visibility' range for contacts. Whether it’s general or depending on AI skill is unclear to me. It’s also unclear under what conditions combat AI kicks in and takes command over from the mission script (or gives the command back).
2. Once a contact is within this range, the AI has no problems to identify it as friend or foe. In this respect the AI always flies ‘with icons on’.
3. How many contacts the AI tracks simultaneously is unclear. It’s also unclear whether and how this depends on skill and/or other factors (apart from blocked LoS). Anyway, the data of all contacts are theoretically available to the AI, even if some of them are disregarded.
4. The AI has exact knowledge of contact position (vector and distance), perhaps with some inaccuracy margin to reflect skill.
5. The AI has a very good knowledge of target travel (direction and speed), probably with an inaccuracy margin to reflect skill. (This ‘knowledge’ is a prerequisite for gunnery, particularly for deflection shots.)
6. The AI seems to be aware of the general type of the target aircraft. There are different attack schemes against fighters and bombers. The differentiation is probably based, in some simplified way, on the types listed in post #46.
7. The AI is aware of the terrain (but not the objects!) it is flying over.

All this is just guessing, of course, and probably far not complete, but brainstorming could be made more effective if we knew the cruel 'facts' about the AI.

[majorfailure]

Originally Posted by Storebror

One thing we shouldn't forget is that AI is comparably "stupid" when it comes to draw conclusions based on the facts they know.
Human players can utilize their experience from previous sorties and can adopt promising options even in uncertain or unknown situations.
AI can't do any of this.
AI has a simple tree/branch decision scheme, sometimes leading to rather stupid moves.
In that regards, letting AI "cheat" in another regime to compensate this lack of experience might be a valid decision to some degree.

Best regards - Mike

Experience is not needed to have an option to disengage - besides the obvious out of ammo/fuel low option. A script can decide if numerical superiority has gone, a script can decide if positional advantage has gone, and a script can track losses, a script can count friendly planes in sight and so on.
Experience is neither needed to obey your commander. Which current AI does not and most times not at all. I have the experience, and if I tell them we do the mission my way they should not question it -if all goes south, I will have to face the music, not them.

Experience is not needed to shoot semi-accurately at nearly non moving targets, at least if you have been through flight school (possibly with the exception where pilots were needed fast and anyone half capable of making a landing stick was allowed to fly(e. g. Germany end 44).

And you don't need experience to set a limit for maximum contacts to be able to track. Script can do that. Trow in some parameters, as distance to enemy, current LoS, movement across or away/towards, time in LoS, etc and it will be better than all-seeing, and if done decent maybe will let us forget for a few moments we are battling ones and zeroes.

[dimlee]

Ones and zeroes they are but they never stop to surprise us mighty humans.
I like to meet AI online where their skill levels are unknown to me. Sometimes when I relax too much and expect another easy kill... I'm punished by that mysterious super sharp sniper gunner who sends his bullet right into my cockpit from 1000m. I consider that as a "compensation" mentioned by Storebror and continue to enjoy the game.

[Storebror]

Right dimlee.
We've lately adopted the 4.13 AI into the good old Ultrapack 3 which we run on our SAS Gameserver (where we're sporting COOP missions mostly, with lots of AI) and it turned the game into something completely new, where we yet have to figure out all the new moves AI has learned.

majorfailure has a lot of good points there.
What I'm trying to say is that regardless of it's superduper abilities in terms of situational awareness, AI already now sucks bad in many other elements of the fight.
Let me give you a few examples:

• When AI is outnumbering human players and can keep the fight going, they do a really great job (with 4.13 AI code) to help each other, distribute their planes across human opponents etc., all fine, but...
• When AI is outnumbering human players but a single human player has the fastest plane in the set and can run away from the fight, if he's managing the distance to AI planes (keep it at slightly above 1km, never separate further than 1.5km from the closest opponent), he can group up all AI planes behind him in a big swarm and drag them away to a point of choice on the map, then run away and leave them there, completely confused.
• When human players are outnumbering AI and AI planes have no option to climb and run away, AI sucks big times in defending their lifes. They will do a couple of weird maneouvres but this will stop quite soon, so all a human player has to do is stay on an AI's six for half a minute and wait for it to go straight - it will keep going straight, even if you start shooting parts off the plane, until you finally kill it.
• Outnumbered AI will immediately stop helping each other, instead they'll all fight a fight on their own.
• AI bombers don't try to stay in formation when being attacked - when you hit one, he will bounce out, even if he could stay in as well.

This is why I'd suggest not only to think about how and where AI uses superior powers at the moment and how to tone them down, but also how and where AI currently lacks desireable abilities and how to improve them.
Because if we'd just tone down AI's situational awareness, this would have to be compensated elsewhere, and with AI's current abilities the only compensation available would be to give the superpowered AI flightmodel even more super powers of to make the snipers even more sniper like.

I don't think anyone would want AI to become more stupid and more deadly by surprise.

Best regards - Mike

[taly001]

And you don't need experience to set a limit for maximum contacts to be able to track. Script can do that

I just saw a show about a concentration and memory test done with computer graphics object tracking, tracking 4 objects was the "peak point" where high skilled gamers outscored non or low skilled gamers. At less than 3 or more than 5 objects player skill level mattered less.

Storebror AI items listed I agree with, I just add the easy kills of planes flying straight for home that don't evade when hit is the most annoying one, I noticed it worst with recon/patrol fighters.

Il2 AI is now pretty good for dogfighting, and most things.

[Pursuivant]

Originally Posted by Storebror

One thing we shouldn't forget is that AI is comparably "stupid" when it comes to draw conclusions based on the facts they know.

It doesn't have to be. For relatively simple "pure" air combat (i.e., no bomber defense, no ground attack, no strategic objectives, just a dogfight) the "decision tree" can make AI fight fairly realistically. In fact, I would guess that the real trick is to keep it from being too good!

I'd bet that "deep learning" AI programming - were it to be implemented into an air combat sim - would result in some frighteningly effective and realistic AI behavior after a short period of time, to the point that the AI is legitimately unbeatable by all but the most talented human players. Of course, that's a pipe dream given the current time and cost required for machine learning (unless you work for DARPA or General Atomics).

Originally Posted by Storebror

In that regards, letting AI "cheat" in another regime to compensate this lack of experience might be a valid decision to some degree.

I agree, but any AI cheats have to seem fair to the players. In particular, AI "reflexes" can't be any better than a good human player's - no laser-guided gunnery, 360-degree radar vision, or instant, perfect control inputs.

I think that the best way to prevent cheats from being obvious is to have a simple percentage chance based on skill that the AI will screw up and do something random and/or stupid, rather than acting with killer robot efficiency.