AI awareness and tactics - and the ability for AI to conduct hit-and-fade attacks over Korea

[Avimimus]

AI is the hardest area to develop, and is the weakest area of all flight simulators. This is saying a lot - given that flight simulators deal with the complexities of fluid dynamics.

 

However, there are some potential areas for relatively 'simple' improvements. I'm very curious if the developers are planning enhancements to AI tactics.

 

Anyway, I've been thinking a lot about the following two areas:

 

1. More realistic spotting and vision/awareness limitations for aircraft:

- The visual awareness of pilots is too predictable currently. Implementing a system similar to what DCS is developing (i.e. Petrovitch and now ground vehicles) would allow tuning AI awareness and introducing more diversity.

- AI spots opponents too quickly (it is always looking around, too often and too quickly). It also re-acquires opponents a bit too reliably after losing sight of them. More realistic AI vision might make players complain less about their difficulty 'spotting' opponents (if the AI has difficulty spotting opponents too, people might complain less about their own limited abilities).

- It would also be good if AI had a bit more trouble seeing through clouds (it currently re-acquires opponents too quickly and even when they are partly obscured).

- Tweaking situational awareness would be the quickest way to significantly improve AI. If the system could be implemented for existing titles (e.g. Tank Crew) it would also lead to significant improvements in realism.

 

2. An ability to disengage:

- Disengaging after the first pass: Aircraft with a significant altitude or energy advantage should consider making a single pass, then climbing away or diving for the deck to disengage (depending on whether they have an advantage in climb or an advantage in dive speeds over the opponent). A lot of historical accounts report this tactic. It was the standard tactic used by Germans towards the end of WWI and towards the end of WWII, as well as a standard tactic for Mig-17 interceptors in North Korea.

- An option for 'realistic morale' levels: Aircraft which are outnumbered (or separated from allies) should have a probability of trying to disengaging (and regroup or return to base if they are unable to regroup). Flight simulators have been known for unrealistically high attrition levels in campaigns because pilots lack self-preservation... so having a realism setting which would allow AI to disengage once isolated would be very welcome.

- Of course disengaging aircraft which have failed to gain enough distance from climbing or diving should attempt to turn and fight (especially if they lack a speed advantage).

 

P.S. Ideally developing a modular structure for the AI, so that different tactics could be implemented for different aircraft (e.g. props vs. jets) and different eras would be a good idea to support long term development. It would be great if a new manoeuvre could be coded in for one aircraft without it risking changing the behaviours of all other aircraft!

 

Note: Some people might object that this is less fun than dogfighting, but one can combine the two in the same mission - with some enemy flights 'jumping' the player flight and then disengaging, while other flights of fighters choose to dogfight. So, I think adding an ability for disengagement wouldn't negatively impact gameplay. Realistic morale would have much more of an effect - but it could be an option that could be turned on or off.

[Aapje]

The DCS solution is still far too predictable in that they seem to just start a timer based on the skill of the enemy, so it doesn't have real variation in pilot skill, consider luck, keep in mind how visible you are, etc.

 

I'd like to see the system calculate a chance to be spotted. Something like so:

 

If in the blind spot, chance of getting spotted (Csp) is 0.

 

If in front of the pilot, keep the base chance of getting spotted. If to the side, reduce the Csp by a certain percentage, based in part on the AI pilot spotting skill (Ssp). If to the rear, reduce it way more still, also based on Ssp. Ssp is how well the pilot looks around him. When creating the AI pilot, it gets a semi-random Ssp based on the pilot skill level. So an Ace AI will tend to have a very high Ssp, but it would still vary per pilot. And for an easy AI pilot, the average Ssp should be way lower, but also have more variation. Pilots with a low enough Ssp never check behind them, so you would forever be invisible to them as long as you stay behind them (Csp becomes 0 in this situation).

 

The Csp should be lowered based on the distance and size of the plane.

 

The Csp should be lowered when there is reduced visibility between the two planes. So thick clouds should reduce the chance to zero, and partial obstructions should just lower the chance.

 

The Csp should be lowered at lot when the pilot is looking into the sun.

 

The Csp should be lowered based on the eye sight of the pilot (Seye). So again, each pilot would get a semi-random Seye, with aces getting way higher values on average than easy pilots.

 

The Csp should be increased when the AI is aware of there being an enemy. For example, when the AI has lost track of an enemy and is trying to reacquire them, or when they get told over radio that there are enemies.

 

Quote

An ability to disengage

 

Yes, in general, the AI should value their energy, with higher levels of AI being less willing to bleed off their energy. So a spitfire can lure a fast-flying/high, lower level AI, 109 into a turn fight, but not an Ace AI.

 

Ideally, the AI should also give up after a while if boom and zooming isn't working out well for them.

 

And the AI should have a chance to ignore planes that would cost too much energy to engage.

 

Quote

Some people might object that this is less fun than dogfighting

 

It opens up a new dimension in that you would be able to convince the enemy to let you go if you dive for the deck and evade effectively, so they would need to bleed their energy and thus become a sitting duck for others, if they keep going after you.

 

And you would still be able to change the AI level, to make the AI more prone to dangerous turn fighting.

[MajorMagee]

Here's an old programming trick. If you add two random number together the sum will result in a normal distribution of values. You can then use this distribution with appropriate parameters to create realistic variations centered around a nominal value. Most of the time it will result in something near the expected behavior, but occasionally you'll see something different enough to seem lucky, or unlucky because it's relatively rare. 

 

[Avimimus]

Very interesting conversation!

  

On 7/4/2024 at 1:14 PM, Aapje said:

The DCS solution is still far too predictable in that they seem to just start a timer based on the skill of the enemy, so it doesn't have real variation in pilot skill, consider luck, keep in mind how visible you are, etc.

 

Well - see the last patch that just dropped. They moved the AI detection to a separate thread and they now have a system that uses probabilities based on range, and probabilities to simulate scanning limited areas of the sky at any given moment (with an adjustment for vehicles which are buttoned-up vs. ones which aren't). This was explained in the 28 June 2024 newsletter.... so that might be the new standard.

 

On 7/4/2024 at 6:31 PM, MajorMagee said:

Here's an old programming trick. If you add two random number together the sum will result in a normal distribution of values. You can then use this distribution with appropriate parameters to create realistic variations centered around a nominal value. Most of the time it will result in something near the expected behavior, but occasionally you'll see something different enough to seem lucky, or unlucky because it's relatively rare.

 

It is amazing what can sometimes be produced with simple calculations. I recall Notrium (a free indie game) had a system whereby aggression could be set as a function of distance to player... but it could be set for multiple 'rings' of distance (instead of a simple curve). So, I modded it to make the predatory aliens have alternating bands where their aggression would rapidly increase or rapidly decrease (depending on the distance to the player). This created the illusion of stalking, and also created a kind-of complex reaction to player movement - where small changes in player motion/distance could lead to passive or hostile reactions in a semi-unpredictable way!

 

I think the big challenge with AI isn't producing a behaviour, but making it predictable and consistent... something that can be debugged and lead to reliable gameplay.

 

As an additional though - it might really help the gunner AI to separate estimation errors (i.e. ranging, velocity) from firing errors (e.g. aiming the gun slightly incorrectly). By having two layers (estimation error and firing error) it could probably fix the issue with the maths which has traditionally led AI gunners to become excessively accurate at one specific range (when the sources of error cancel out)... I've long suspect that this was the reason why gunners are seemingly excessively accurate in this engine.

[Stonehouse]

One AI facet missing from the current series, that I would hope to see in the Korea module is flak avoidance maneuvers by the AI. Especially considering heavy AAA was much more likely to be aimed using radar or other such height finding devices by Korea's time frame. Even a simple fluctuation of say 10% of altitude around the planned height for the current leg of the flight plan would make a big difference. 

 

Ground attacks too should be better too. By the end of WW2 Fleet air arm tactics for ground attacks specified aircraft attacking simultaneously from different directions etc to reduce effectiveness of AAA. I believe other air forces and air arms of the UK forces had similar doctrines by the end of the war gained by hard practical experience.

 

Likewise, formations need to improve. Pretty much all air forces used some variant of the German finger four formation and fought in pairs by the time of Korea. eg fluid four. So, a group of say 12 aircraft would need to act as 3 separate spaced 4 aircraft formations which broke down into co-ordinated pairs in combat. Separation between aircraft also tended to be generally more than we have now. Whereas in the current IL2 you would see 1 huge 12 aircraft formation usually quite close together which generally breaks down into singletons in combat. Formation turns need to be better handled going forward too.