|
Post by Dutch Owen on Feb 23, 2020 14:30:16 GMT -5
All testers,
Now that 0.10.1 is out we have a basis for developing AI gunfighting skills. Gunfighting is core to AI combat abilities - everything else is just added on to those skills. We have to get this as right as we can. Any potential encounter can end up in a close-in fight even if it starts with missiles.
So here I'd like to start a discussion thread on the ways to avoid mistakes and optimize the AI pilot skills, mostly using Tacview to analyze particular encounters.
We've already made progress, and if you record a fight with 0.9.13 versus 0.10.1 you'll see a dramatic difference in what the AI pilot does. Before he just flew around as if trying to find you, going up occasionally, and it was very easy to get offensive on him and just wait till the top of his inevitable climb and take him out. Now you will see he homes in on you wherever you are and tries to get opportunities and if he does he will hit you.
But he still makes lots of mistakes, and misses some opportunities too. That's what I want to figure out why and correct those.
I'll probably make a reply shortly analyzing the last test I did, there's lots to talk about there.
Dutch
|
|
|
Post by xpel on Feb 24, 2020 2:14:02 GMT -5
I watched the Tacview file & improvement is obvious. Good job Dutch. I'll test this the next few days, since weekend is over. But I'll share my thoughts about the whole (AI) Air Combat concept based on the basic - essential rules of it :
1. Decision making at start of engagement I (turning dogfight or boom & zoom?) - Energy Advantage : Energy of each plane is the sum of Kinetic (speed) + Dynamic (height) energy of a plane at a given time. In a 1 vs 1 fight, defines who has the upper hand at start of engagement. For programming reasons, can be computed with something like this :
- WWII : 50Knots speed = 1 energy unit, 500meters height=1 energy unit, thus a plane flying at 3000m with 250 knots gets a sum of 11 energy units - Jets : 100knots = 1 energy unit, 1000m height = 1 energy unit, so plane above has 5,5 energy units
So, if there is a lot of difference in energy, the attacker (skill above average maybe) won't decide to go into a turning combat, but instead will perform boom & zoom attacks (high speed passes for opportunity shots, by diving on defender and then zoom to height again to retain his energy advantage & repeat)
2. Decision making at start of engagement II, when AI doesn't have the upper hand : Skill dependant, AI may decide to either : A. Avoid imediate engagement and fly straight (after merge) to gain speed + slight pich down (increase speed energy = extend) and re-engage after say 15-20 secs with higher energy, or B. Totally avoid engagement (wise decision when in disadvantage), by performing the above with a steeper pitch down, trying to escape from area in max speed.
3. Escape decision making (when AI under pursuit) : A. For long distance between the planes, it is already implemented in FSCAI B. For close - guns distance : There are three basic manouvers for AI to avoid being hit : B1 : Break turn. Applies for incoming missiles also. If attacker overshoots, then fine. If not, has to go to one of next two choices : B2. Scissors : AI should change turn direction from left to right & vise versa, every few seconds, giving a hardtime to his opponent for a firing resolution. This can be an auto procedure, even done randomly (it will work & it'll be effective, depending on attackers skill) B3. Barrel Roll : Since FSCAI controlled AIs cant do complete 360 rolls, an imitation of it : Bank right (or left) and go up (right ascending turn) - pitch down for 2-3 secs - left descending turn - level up & repeat if enough energy.
4. General : If speed is less than = , then fly straight in shallow pitch down for few secs, until speed is higher than = (retain energy)
|
|
|
Post by Dutch Owen on Feb 24, 2020 11:12:44 GMT -5
Thanks Chris! I feel like we're on the right track and although there's a lot of work ahead it will be possible to make the AI a feared opponent. Although his abilities will be a little limited by the sim's limitations he has certain advantages that humans lack - he knows where you are and what you're doing exactly at all times and he can make literally millisecond decisions which a human can't make because our brains aren't fast enough. The bane of most AI is that it becomes predictable. If we can give him a large enough bag of tricks and randomize it enough he won't be entirely predictable.
Your advice is very logical and will surely help and I'll implement it as we move forward.
Right now we simply have to improve his turn rate. I can turn (in some fighters) at the end of blackout, he can't. Yet.
I have an idea that's very promising and I'm working on that today. I think we're going to get him to be able to turn at max human G - and then, watch out!
I'm also pretty sure I can have him do a barrel roll, I have already forced him inverted and holding it (albeit a bit wobbly) in preparation for implementing a split-s and Immelmann. After I get the turns tighter that'll probably be the next priority - defense needs a little love too after all this improvement on offense.
What about a spiral dive? I have that working fairly well in testing.
Dutch
|
|
|
Post by Dutch Owen on Feb 24, 2020 12:47:41 GMT -5
Here's exactly why turn rate is crucial:
Even with the improvements, because of turn rate I can beat the AI every single time (unless I am in a poor turning fighter which some are.) Here's how: At the merge I do a slight climb usually at about a 10-20 degree angle versus his approach vector. I do not want to let him get a shot at me at close range in the merge- he is deadly if allowed to do that. When he gets close but still outside gun range I make a sharp dive. He's committed to his climb at that point and won't pitch down fast enough to keep a firing solution.
Once we pass I simply make a max-rate 180 while climbing back up to his altitude. At the end of that turn I'll see him out there also making his hardest turn to get pointed at me. But I am pointed at him before he can get more than about 3/4 around towards me. I then just angle off a bit and he will pass by me without being able to turn enough to get a firing solution. I then slow down and just max-rate around again and I'm on his six and if I can close up he's a dead pilot. He may go into defensive at that point and try to climb or dive but I can almost always stay on him and shoot him down.
The game would be completely different if he turned at the same rate (or better) than I do. Then after that first 180 he would be pointing at me and we'd be in a scissors situation and I'd have to evade again because of his deadly aim on the merge. Eventually I'd likely make some mistake and he'd be on me from the rear or side. Otherwise we'd just keep on scissoring.
At this point his defense is poor but with max-G turning I'm sure he could stay on your six in Pitbull mode and would be very hard to shake.
Dutch
|
|
|
Post by xpel on Feb 24, 2020 14:27:50 GMT -5
All that sound very promising indeed !
Actually a vertical barrel roll with the advantage that speed energy is retained. The best defensive card in AI's sleeve currently. I remember in such a case I was pursuing an AI about 2 FSCAI editions ago. It was very exciting (& spectacular) indeed.
|
|
|
Post by Dutch Owen on Feb 24, 2020 14:36:09 GMT -5
I have it! I have him turning at max-G. Take a look at this: Test 109 vs 109 max rate turn method 4In this test I have my own aircraft invulnerable and I'm not firing at him. You can see that twice in the max turn fight where I am turning at max G before blackout he gets his guns on me and in both cases the log shows he fired a burst that would have mostly likely hit and unless I was lucky put me out of the fight. Eventually, since he couldn't hurt me, he made a mistake and I got on him and his dive couldn't shake me. But if I weren't invulnerable I wouldn't have got to that point... He is very tough in a turn fight now. Now we need to get his defensive moves better, then move on to energy mode fighting. IN this test, after his dive failed to shake me he tried a zoom climb, but not zoom enough, and then at the top with me still in firing range he dropped his landing gear! I am never telling him do that so that's some FS sim thing. I'm sure I can make him raise the gear if he ever starts lowering them. Dutch
|
|
|
Post by edakridge on Feb 24, 2020 15:41:57 GMT -5
I flew a test mission this morning and the Aggressor AI are much smarter now. They no longer wait and attack the bomber formation from the rear. I had to turn their skill level down so that the bombers will be able to get through.
|
|
|
Post by Dutch Owen on Feb 24, 2020 15:52:51 GMT -5
I flew a test mission this morning and the Aggressor AI are much smarter now. They no longer wait and attack the bomber formation from the rear. I had to turn their skill level down so that the bombers will be able to get through. Good news! Glad to hear they are more formidable. Did you have your gunners firing at them, assuming you were flying a bomber? It's possible that if you have a custom version of a bomber defined in Config_USER it won't have the weapons configured correctly. The weapons config for a self-defense user piloted bomber is different - it needs 4 new gun definitions, you for each quadrant - front, rear, left and right. Check out the way a B-17 or B-24 is defined in the standard edition to see how it's done. If done right you'll hear your gunners firing at anything within range and the sound is calibrated to the quadrant. The forward guns will be loud, the side guns will be softer and panned left or right, and the rear guns will be even quieter; you may not be able to hear them over the engines. But this means they are firing and they can hit and damage the enemy aircraft. That's really why I had to change the attack mode for the OPFOR pilots - they were slowly coming in behind and getting raked by the rear gunner. Dutch
|
|
|
Post by xpel on Feb 24, 2020 16:29:14 GMT -5
I just finished a test with Ed's Tacpacked P-40 vs my AI Normandy BF-109. The improvement is about 80% better experience since previous version! I certainly found it interesting & didn't manage to bring him down (out of ammo). Rookies and starting people with WWII dogfight, will have a great fun & time with this, as is...
But the Tacview sample, is really the thing ! Greaaat job Dutch.
Obvious bug fixes also in ASL missions. I can spawn simobjects now & get them on route. Thank's.
|
|
|
Post by Dutch Owen on Feb 24, 2020 16:55:02 GMT -5
I just finished a test with Ed's Tacpacked P-40 vs my AI Normandy BF-109. The improvement is about 80% better experience since previous version! I certainly found it interesting & didn't manage to bring him down (out of ammo). Rookies and starting people with WWII dogfight, will have a great fun & time with this, as is... But the Tacview sample, is really the thing ! Greaaat job Dutch. Obvious bug fixes also in ASL missions. I can spawn simobjects now & get them on route. Thank's. Take a look at this. I'm in the Aeroplane Heaven Hurricane (which previously wasn't a problem defeating the 109) and this time it's a totally different experience. I never got my guns on him, he stalked me ruthlessly and hit me three times, the last sent me down in flames (where the recording stops) - you can see me desperately trying to shake him and I can't do it. Scary good. I'm going to have to scale the turn boost factor by pilot skill to give non-expert folks a chance. This is level 9 "ace". Hurricane vs BF109 TestI will post up a version with the changes as 0.10.2 soon. I want to get a little better defense in for him. He'll need it against other AI (but not against me any more!) Dutch PS. This is very much the feeling of my experience in DCS. All my efforts just go into postponing the inevitable when I will hear the bullets whacking into my airplane.
|
|
|
Post by xpel on Feb 24, 2020 22:32:22 GMT -5
This is dramatic improvement. The efficiency in level plane turning, draws over the efficiency in use of the vertical also. Exciting times & perspectives ahead. I feel we're almost there, at the goal. It's 05:00 here and I'm on my pc. You have officialy ruined my sleep for tonight. But I'm not complaining...
|
|
|
Post by xpel on Feb 25, 2020 1:57:23 GMT -5
Ok. Last half hour I flew 2 x test flights :
A. 2 x F-14s vs 2 x Mig23s. I got shots in head on and one engine of mine flamed out. Fortunatelly, before that, my wingman had fired an aim-9 that damaged one of Migs. Underpowered, I turned for the other. My wingman was more efficient. Shot 2 more missiles and last Mig splashed on sea.
B. My first ASL mission in Free Flight. My F-5 took off from Nea Aghialos to intercept one F-102 of THK. Lot's happened during engagement to describe. I shot him down with guns.
TESTs Results : 1. There is a huge gap between AI offensive vs AI defensive mode efficiensy in current version. It's more profound with jets & high speeds, compared to WWII. (Something you already know Dutch and improve for next version) 2. Is there an AI defensive mode for incoming missiles also (attempt to avoid somewhat like with user at six)? 3. Scenario creations are limitless !
|
|
|
Post by Dutch Owen on Feb 25, 2020 11:59:57 GMT -5
Yes defense vs missiles is lacking, defenders need to jink to avoid incoming missiles and use flares/chaff. Right now they just fly & pray.
|
|
|
Post by Dutch Owen on Feb 25, 2020 12:59:17 GMT -5
Let's talk about G force for a bit. We now have the capability to get the AI pilots to turn at any rate we want, including beyond human or aircraft capability to survive the turn. Obviously we don't want to exceed reality with our AI so we have to limit it in some way.
Right now the limit is a 5 G turn. According to the info I've researched the average human can withstand 5 G for about 10 seconds before beginning to black out. The designers of FSX/P3D must agree because if you turn on G effects you will start to black out at about 5 G. That's why we chose 5G - to be fair and put the AI ace at the same turn limit as a human with G effects turned on.
Of course in the sim you can get an immense advantage by turning those effects off. Then you can safely turn at any rate your airplane will allow. And since G suits exist it might be realistic in some cases. Certainly in the Korean war American pilots had an advantage over their MiG-15 adversaries when they had a suit. But of course a G suit is not magic and there's still a limit.
Human ability to withstand G force varies, with the extremes being that some few people will conk out at 3G and some few can withstand 6G for long periods of time. Also, withstanding G can be trained both by physical techniques and by practice in centrifuges.
So it seems that leaving our AI limit at 5G always isn't accurate.
But what rules can we have that are more accurate?
FSCAI can check to see if the user has his G-effects turned off. If so, should we then jack up the AI limit too, to even the fight?
How can we simulate G suits? And "G-monsters" as those that can withstand high-G are called?
Should an AI Ace be able to make a 5.5 G turn? Some WWII aces could withstand higher G's and it was one of the things that made them an ace.
I'm not sure what the limits are on G suits but should jet pilots automatically have an increase in their turn rate limit because of this?
I'm reluctant to add another parameter about G suit to individual aircraft models, but it may come to that.
Your thoughts welcome.
Dutch
|
|
|
Post by StuB on Feb 25, 2020 22:23:54 GMT -5
It also depends on the airplane. From what I have read, the pilots of the BF-109 were able to sustain higher G forces because the seating position in the plane was so low that their legs were almost straight out in front of them.
There are two things that a G suit gives...... higher G resistance and longer G tolerance.
Where a pilot could handle 6 G's for a second and 5 G's for 10 seconds, with a G Suit they could perhaps handle 7 G's for a second and 6 G's for 20 seconds.
During the Korean war, the US pilots had G suits and the Russian pilots did not. The better G tolerance helped the F-86 pilots negate the MiG-15's slightly better turn performance.
|
|