2.7.14.23966 BVR AI First Impressions

[Exorcet]

The AI is definitely better at BVR combat. In previous patches they would easily die to ~30 nmi AMRAAM shots reliably. Now it seems like high Pk shots have been pushed back to be at 20-25 nmi.

While the AI is more effective, tactics don't seem to vary much. Thankfully they no longer use airbrakes while defending and bleed speed unnecessarily, but they overwhelmingly prefer to give up altitude when faced with approaching missiles. Ultimately this ends up with them being on the deck, always. This predictability can be exploited and it also puts the AI in a bad position for follow up shots. The AI will also rarely crash into the ground when pressured by a missile (not necessarily a bad thing, real pilots do this). Diving seems to be aircraft independent as well. As likely to happen with a Su-27 as it is with a dedicated interceptor like the MiG-31.

Something else I noticed watching the AI is that it seems indecisive. It will roll a lot and also make small sudden pitch movements, only to counter than a second later. I'm guessing this is the AI continually assessing the situation, which is true is great to see, but it reacts a bit overzealously maybe and makes unneeded control inputs.

Overall, this is definitely an improvement, but I think more variation in tactics are required. High altitude fighting and extending away from WVR for example would be huge additions. I know this is only the first step, but I think the earlier feedback is provided the more helpful it can be.

 

EDIT Summary of ideas

High speed/altitude profiles for intercepts

Low altitude/ground clutter approach

Imperfect AI SA, ie add ambiguity to the AI's knowledge of incoming missile range/angle and enemy aircraft position, speed, etc.

Incorporate timelines into AI tactics

Give the AI some self preservation, like the ability to run away if wingmen are destroyed

Edited October 21, 2022 by Exorcet

[TheBigTatanka]

The real question is, can tactics be set to match the role of the AI flight? If on an escort or OCA role, perhaps a banzai flow mentality. If on a DCA, a preference to turn cold and drag shots. If the AI could use red air pre-planned maneuvers, that would be really cool too (to overcome their fox-1 only limitations and shorter Fpole limitations).

Sent from my Pixel 3 XL using Tapatalk

[Dragon1-1]

11 hours ago, Exorcet said:

Something else I noticed watching the AI is that it seems indecisive. It will roll a lot and also make small sudden pitch movements, only to counter than a second later. I'm guessing this is the AI continually assessing the situation, which is true is great to see, but it reacts a bit overzealously maybe and makes unneeded control inputs.

This is a valid tactic to make the incoming missile bleed its energy. If they think you might have launched at them, this is a proper response. AFAIK, the current tactics overhaul is focused on 1:1 fights, where going to deck is usually a good choice. The more complex tactics require multiple ship flights to execute successfully, and as such, I think they'll come when they get to group tactics.

[BIGNEWY]

Work on flights of more than one AI will come in the future, this first change was focused on the 1 v 1 scenario. 

thanks

[Exorcet]

4 hours ago, TheBigTatanka said:

The real question is, can tactics be set to match the role of the AI flight? If on an escort or OCA role, perhaps a banzai flow mentality. If on a DCA, a preference to turn cold and drag shots. If the AI could use red air pre-planned maneuvers, that would be really cool too (to overcome their fox-1 only limitations and shorter Fpole limitations).

Sent from my Pixel 3 XL using Tapatalk


 

It would be nice to not only adjust tactics for the mission, but to have strategic settings as well. For example if the enemy generally has better aircraft the AI might be set to be more defensive and favor withdrawing rather than pressing until they take losses. In the opposite situation, they could be set to be more aggressive. In a campaign a setting like this might continually change depending on how well each side is doing.

2 hours ago, Dragon1-1 said:

This is a valid tactic to make the incoming missile bleed its energy. If they think you might have launched at them, this is a proper response. AFAIK, the current tactics overhaul is focused on 1:1 fights, where going to deck is usually a good choice. The more complex tactics require multiple ship flights to execute successfully, and as such, I think they'll come when they get to group tactics.

The pitch maneuvers, if you could call them that, are far too small to effect incoming missiles. The plane's attitude barely changes.

Trading altitude for energy is generally a good idea, but so far 100% of the time the AI ends up skimming the ground and doesn't really have any options if another missile comes along. It also hurts their ability to fight back, depending on their own missiles. The Phoenix and R-33 for example have terrible acceleration at low altitudes. I'm always at least 10000 ft above AI and my missiles enjoy the slight range advantage it brings.

[Exorcet]

I moved on from testing against peer adversaries to older aircraft. It looks like the new AI has more believable sensor limitations as planes like the MiG-21 won't attack unless they get very close (Although the Su-27 seems a bit shortsighted). I need to test how this interacts with AWACS/EWR later. One problem with this lack of response though is that it applies to STT lock. The MiG-21 will happily fly straight and level despite being locked on. Of course it still won't know the exact range of the enemy targeting it from well beyond its own radar range, but it should probably start to do something to mitigate the threat. Perhaps it should put the enemy RWR at radar gimbal limits, start snaking, attempt to break lock by beaming, or dive in an attempt to hide in ground clutter. Some kind of reaction to indicate that the pilot is trying to survive would make it much more believable in this case.

[Exorcet]

So, the new AI has been out for a while. I'm surprised there has not been more discussion given how many have wanted this for so long.

Anyway right now I'm curious about further development. The AI only has one real BVR tactic at the moment, which is just to dive. It's certainly more effective than before, but it's also easy to counter. It also doesn't fit some mission types, like high altitude intercept.

Will the AI receive more varied behavior in the future?

[M1Combat]

I believe they've said it will and is a WIP at the moment.

 

Regarding the community volume on the subject...  I think many people are awaiting the dogfight AI release as well as the likely BVR updates before they "go loud"...  Probably an improvement over people who, historically, have just outright attacked any non-perfect first implementation of any new feature IMO :).

[BIGNEWY]

3 hours ago, Exorcet said:

So, the new AI has been out for a while. I'm surprised there has not been more discussion given how many have wanted this for so long.

Anyway right now I'm curious about further development. The AI only has one real BVR tactic at the moment, which is just to dive. It's certainly more effective than before, but it's also easy to counter. It also doesn't fit some mission types, like high altitude intercept.

Will the AI receive more varied behavior in the future?

Not true. Different AI skill levels have different tactics and even within the chosen level there are variants. 
The team continue to work on the AI and we will see other improvements in the future.

thanks

[Exorcet]

6 hours ago, BIGNEWY said:

Not true. Different AI skill levels have different tactics and even within the chosen level there are variants. 
The team continue to work on the AI and we will see other improvements in the future.

thanks

Thanks for the reply BIGNEWY, and I'm glad to hear that more work will be done in the future. Can you provide more detail on the AI tactic variations? I have done a sweep of BVR tests against different difficulties, but they seemed to behave similar to each other. It might be because I tried to make the test repeatable which constrained the AI response. Having more direct control over AI tactics in the ME might also be a nice feature to have. Would that be something the dev team might consider? I understand if these things are not yet concrete.

[Dragon1-1]

Note that single ship BVR tactics are not particularly varied even IRL, and a repeatable test means one tactic is a way to go for the AI in this situation. The really interesting stuff will be multi-ship BVR tactics, where you could have a lot of variations, unlike in a dogfight, when you'd be hard pressed to do much cooperation beyond the first turn, unless numbers are on your side.

[BIGNEWY]

Based on skill level, the AI will or will not use tactics such as adjusting altitude, adjusting airspeed, cranking, cranking and forcing look down, deciding to skate or banzi, and deciding to maximize A-Pole or E-Pole. All of these are in now if you understand BVR tactics and know what to look for.

[Mad_Shell]

5 hours ago, BIGNEWY said:

Based on skill level, the AI will or will not use tactics such as adjusting altitude, adjusting airspeed, cranking, cranking and forcing look down, deciding to skate or banzi, and deciding to maximize A-Pole or E-Pole. All of these are in now if you understand BVR tactics and know what to look for.

Hi BN.

My main grip with the current BVR AI is that when it launches a missile, it will always immediately crank and dive. While cranking makes sense, diving should only be a thing if the AI thinks there is a possibility of incoming missile. If the enemy aircraft is identified (RWR) and has no BVR missile, or ones with smaller range, diving is a bad decision, as you give up you altitude for nothing.

If it hasn't fired a missile yet, the current BVR AI also never defends before an enemy missile goes pitbull, which is too late most of the time.

Edited August 13, 2022 by Mad_Shell

[Exorcet]

On 8/13/2022 at 4:22 PM, Mad_Shell said:

If it hasn't fired a missile yet, the current BVR AI also never defends before an enemy missile goes pitbull, which is too late most of the time.

Yeah, they push really hard until they fire their own missiles, but on Ace difficulty is seems to work against Rmax shots, at least against AIM-54 when chaff is involved.

After some more testing I think at least Ace has one of the problems carried over from previous AI, which is that it's too perfect. It knows exactly where a missile is coming from, how fast it's going, and in many cases knows whether it can outrun the missile or not.

I know for simplicity sake the AI might need to "cheat" at certain things, but can the information fed to the AI have some randomized error that increases with lower difficult settings? This way Ace will no longer be able to perfectly notch or outrun a missile. It will either need to use safety margin in calculating maneuvers or it will occasionally make misjudge the situation and be more likely to get hit or take emergency action.

This should also be incorporated into the dogfight AI when it comes in the future. Current AI knows your airspeed down to the knot and will instantly change tactics depending on it. Instead the AI should only get your airspeed +/- X% so that it won't have perfect knowledge (the exception would be a modern aircraft with a radar lock, though even in this case, the pilot won't be staring at the radar target airspeed for the whole dogfight, so some error in this case would also make sense).

I hope the developers can at least consider these changes, thanks for the hard work.

[MARLAN_]

I'd love to see the AI incorporate timelines into their tactics so they aren't just always reacting and can literally just act on a defined timeline. Lower skill can "miss" their timeline when they shouldn't, and higher skill can sometimes elect to press the timeline when they are able to.

e.g. Shoot at NLT SR, Skate at MOR, etc.

Edited August 22, 2022 by MARLAN_

[Exorcet]

Along with the need to have interceptors maintain speed and altitudes, we probably also need a low altitude attack option, especially for less capable aircraft. They could use a setting where they attempt to sneak up on other planes by flying low in ground clutter. They should probably reduce radar usage as well or try to rely on external radars like AWACS and EWR.