4.13 development update discussion and feedback

[IceFire]

Originally Posted by majorfailure

I don't know if this necessarily is the case. Wouldn't that be similar to fighting a cautious but good human opponent? If not ALL ace AI would behave the same, a very challenging opponent for occasional duels. And maybe he would even be easier to escape from, as his cautiousness would forbid him from following you at all costs, he would rather keep his superior position. And some cautiousness wouldn't hurt the AI all across the board - and at least the better AI could once in a while try to avoid a fight at bad terms. Or to just get away when they see the fight is not going their way.

And another thing AI usually is bad - using a speed advantage. Planes like Bf109 in AI hands are not employed well against nimble but slow opponents - and I think at least regular to ace AI should know a little about what their plane is good at - veteran and ace to some degree what the enemy planes are bad at. They should of course not be omniscient. Maybe too complicated, but say an ace AI has virtually brought down 10 P-40 in his Bf109, then he should at least know he can outclimb them, and maybe know he is usually faster - and has a little disadvantage turning and rolling.

From what research I've done in the literature surrounding AI and AI programming, you're describing something that is still kind of a dream. We can program in specific behaviours and I'm sure an abstracted system might be developed to boil it down for the AI so they can make some artificially generated judgment calls - but to truly have the ability to size up their opponent and suss out a tactical plan based on that is still something only a human can do.

This is especially true if we try and factor in a fog of war scenario where you know your opponent's aircraft generally but maybe not all of his capabilities or weaknesses. Developing that knowledge organically is the stuff of neural network research and they have made great strides in that area but it's still just in its infancy I think.

A long time before we'll see a game AI with the abilities that we'd all like them to have. And by then... AI might be a little scary to behold. Just ask Elon Musk about what he thinks of that

[Pursuivant]

Originally Posted by IceFire

From what research I've done in the literature surrounding AI and AI programming, you're describing something that is still kind of a dream.

Currently "Artificial Intelligence" is still an oxymoron.

The good news is that air combat is a very limited sphere of activity, closely bound to the laws of physics, and further bound by historical doctrines and the limits of human physiology. With those limits in mind, AI can often be abstracted into decision trees and flow charts.

For example, currently damaged enemy bombers often behave "stupidly" when choosing whether the crew bails out or crash lands/ditches. A simple decision tree or flow chart could be used to make them behave in a much more realistic fashion. For example:

Can I hold altitude? N/Y > Am I over friendly territory? Y/N > Can I get to an airfield long enough to land the plane? Y/N > Is there open ground where I can crash land? Y/N > Can I reach any water within 300 m of land? Y/N? > Fly towards land > When within 300 m of land, turn parallel to the wind and ditch.

[IceFire]

Originally Posted by Pursuivant

Currently "Artificial Intelligence" is still an oxymoron.

The good news is that air combat is a very limited sphere of activity, closely bound to the laws of physics, and further bound by historical doctrines and the limits of human physiology. With those limits in mind, AI can often be abstracted into decision trees and flow charts.

For example, currently damaged enemy bombers often behave "stupidly" when choosing whether the crew bails out or crash lands/ditches. A simple decision tree or flow chart could be used to make them behave in a much more realistic fashion. For example:

Can I hold altitude? N/Y > Am I over friendly territory? Y/N > Can I get to an airfield long enough to land the plane? Y/N > Is there open ground where I can crash land? Y/N > Can I reach any water within 300 m of land? Y/N? > Fly towards land > When within 300 m of land, turn parallel to the wind and ditch.

Completely agree. AI is still kind of an oxymoron. It's more of a decision tree that it follows and it's really only as good as the assumptions that the programmer has made.

More assumptions can be programmed like what you have up there. That's fairly "easy" to check for I would imagine... although I'm not really sure if the AI would know if it's in friendly or enemy territory or if that kind of thing is passed to the AI at all. Would be interesting!

[Pursuivant]

Originally Posted by IceFire

More assumptions can be programmed like what you have up there. That's fairly "easy" to check for I would imagine... although I'm not really sure if the AI would know if it's in friendly or enemy territory or if that kind of thing is passed to the AI at all. Would be interesting!

I don't know if front markers are passed to the AI. But, given as they're points marked on a map and the game engine already keeps track of aircraft position and vector, it doesn't seem like that big a problem to add them in.

For missions where no front lines are marked, just assume that all territory is friendly, or all territory that isn't within X meters of a hostile ground unit is friendly.

Even so, my original partial decision tree for bailout decisions shows the sort of work that is necessary to make aircraft behave in a "smart" fashion for just one small aspect of flight.

Humans have plenty of experience with "don't do this, it's probably dangerous," so we understand the ideas that friendly territory is better than enemy territory, landing is (usually) better than bailing, and it's (usually) better to crash land or bail out over land than water. We also have the ability to extrapolate from basic principles.

Computer AI is like programming a baby. The computer doesn't automatically "know" anything, and has to be "taught" that certain things or behaviors are bad. Even worse, it has no ability to extrapolate and it's typically really poor at certain types of visual pattern recognition that humans take for granted.

[majorfailure]

Originally Posted by IceFire

From what research I've done in the literature surrounding AI and AI programming, you're describing something that is still kind of a dream. We can program in specific behaviours and I'm sure an abstracted system might be developed to boil it down for the AI so they can make some artificially generated judgment calls - but to truly have the ability to size up their opponent and suss out a tactical plan based on that is still something only a human can do.

This is especially true if we try and factor in a fog of war scenario where you know your opponent's aircraft generally but maybe not all of his capabilities or weaknesses. Developing that knowledge organically is the stuff of neural network research and they have made great strides in that area but it's still just in its infancy I think.

A long time before we'll see a game AI with the abilities that we'd all like them to have. And by then... AI might be a little scary to behold. Just ask Elon Musk about what he thinks of that

Correct me if I'm wrong, but AI does not have to learn, it does just have to simulate learning. Ever played an RPG(pen 'n paper)? Then -did your virtual character ever learn anything? No, you just simulated that.

Now do the same for AI - and even a basic system would work IMHO. But that may be too complicated - I do not know if it is possible to program an AI with traits without having to write a new code for every small change in AI. Only then independent skill sets were possible.
As for the other thing, AI not beeing able to use speed advantage - either classify the planes as fast/slow - then a Bf109 will try to run from P-51 - which is okay for rookie/regular. Or give them a faster/equal/slower table, and let only ace know all, veteran and lower will get a table with more and randomly selected wrong entries. Make that table for turn and roll and climb too - and use the result to give the AI rules to follow.

[sniperton]

BTW, does anyone know how radio communication is simulated in case of AI flights? I mean, do they inform and warn each other (>added up situational awarenes), and do they issue and obey orders (>combat cooperation)? Do they effectively use the same range of radio orders as we do?

[Pursuivant]

Originally Posted by sniperton

BTW, does anyone know how radio communication is simulated in case of AI flights? I mean, do they inform and warn each other (>added up situational awarenes), and do they issue and obey orders (>combat cooperation)? Do they effectively use the same range of radio orders as we do?

I don't think that radio communication is modeled at all. In any case, it wouldn't make much difference, since AI doesn't seem to cooperate beyond the leader/wingman.

[sniperton]

If the AI is fighting his war alone, then the AI is very much handicapped...

Perhaps the easiest way to improve AI performance would be to teach them to communicate.

Historically, radio made a huge a difference in RL. It would be great to have this difference in game.

[Pursuivant]

Originally Posted by sniperton

If the AI is fighting his war alone, then the AI is very much handicapped...

Yep.

To be fair to IL2 and DT, I don't think that there's a combat flight sim out there where AI uses intelligent section, flight or squadron tactics.

There's absolutely nothing like one flight going high to cover another flight going low to attack a foe, or two sections in a flight splitting up to "box" a single opponent.

Originally Posted by sniperton

Perhaps the easiest way to improve AI performance would be to teach them to communicate.

Easier said than done. Without radio, signals can only be transmitted if you can see and understand gestures. That might be 50 meters, and only if you're at 2-4 o'clock or 8-10 o'clock level with respect to the signaling plane, and only if you're in clear visibility and are both traveling basically straight and level.

[sniperton]

Originally Posted by Pursuivant

There's absolutely nothing like one flight going high to cover another flight going low to attack a foe, or two sections in a flight splitting up to "box" a single opponent.

AI programming is a black box for me, but I guess there are basic tactic roles assigned to individual planes: lone wolf (fighting alone for his life), wingman (copying the wing leader and making opportunity attacks), wing leader (copying the flight leader, but acting independently when attacked), flight leader (responsible for flight attack tactics). But no matter how many flights are there, all flight leaders are on the same tactics. In other words, there are no distinctive tactics for 2 flights together, or 3 flights together. Or at least it seems so.

Originally Posted by Pursuivant

Easier said than done.

Sure, and I have more questions than answers.

AI used to have mystic vision; now it cannot see through clouds and obstacles. It has been made more 'realistic'. Dunno whether this has been done to communication as well.
For AI does communicate. I hear him cry 'help me' and the like. I can mute him by unchecking the 'radio' checkbox in the FMB. Does it mute him only for me, or does it completely stop communication? If I switch off the radio for a whole squad, does it affect its combat performance? Or do they still rely on some mystic thought transmission?