FC3 Wishlist - Air Combat AI

Close dogfight (gunzo) is too easy - AI "swallows the bait".

One player vs two AI:

- AI isn't trying to do maneuvers that will help them achieve superiority (example: page 205 figure 5-3)

...

Hi Jim,

That book is great, but despite its volume it's missing a vast number of things when it comes to team-work.

Further, it tells you nothing of the AI. Right now the AI is very reactive (or so it seems to me) and has no ability to analyze and predict the bandit's moves. At the same time, this analysis is useless if it cannot formulate plans for each choice the bandit may make.

This isn't easy to do with AI, but also not impossible - and this really is a deep AI science issue, not just a dog-fighting issue.

The AI need to be able to make decisions based on sensor input - ie. if it is 1 AI vs. 2 enemy planes, he should be trying to keep tabs on both bandits and apply some 1v2 tactics. But this also implies that the AI must build an internal 'track file' on its bandits, so that loss of sight is properly simulated - ie. the AI would 'look' at where the bandit is supposed to be after checking up on another bandit, etc.

After the above are resolved (1v1 and 1vx), -then- communication between flight members could be worked on to achieve realistic team-work.

All of this is practical for BVR A2A and A2G as well, although the tactics change.

In short, the AI currently:

- Does not act proactively

- Does not take good snapshots (it pretty much takes no snapshots, instead waiting to align wings to fire)

- Gives up when you're on its six

- Gives up when damaged

- Does not understand BFM (this goes back to giving up when you're on its six)

- Does not understand DACT

- Does not understand teamwork

- Appears to cheat when it comes to pulling g's, but I might be wrong

etc. etc. etc.

It is a very very long subject, but here's a reasonable start perhaps:

- AI should recognize if it is at least doing one or two circle fighting

- AI should adjust its plan when it recognizes who is winning that circle fight

- AI should aim to take snapshots (but subject to some human-like error induced into it) if it cannot achieve a tracking solution (an F-15 vs. a flanker for example should be biased towards quick defeat using high energy and snapshots, where the flanker might be more inclined to turn and should be more capable of saddling up)

- AI should understand how to cause BFM problems when it is defensive

Each of these statements requires fairly lengthy and detailed explanation, which I won't do here.

I have write a long post about it http://forums.eagle.ru/showthread.php?t=81295

But simple :

For me the best, faster and simplest way to obtain a good AI is to give to every people the possibility to totally rewrite with an SDK the AI script, cause ED can't write it alone, its too long and anyway the community will have AI with 2 or 3 good maneuver and no more without possibility to do it by ourself...

Another/additional possibility can be something like that i say here : http://forums.eagle.ru/showthread.php?t=80658

The possibility to save into a track what our aircraft do, for, after with another aircraft fly with, example, do a hard flying inside a canyon the first time and after training for extreme formation fly, or make really advanced mission sp/mp...

Ok you will say : what is the rapport against the first idea/AI improvement and the second...

Its simple, a quick script editor, just we do a maneuver, we record it, we convert it into a script and we complete it with informations/condition, and like that we can really faster make a good and complex maneuver for AI script, that can open a lot of possibility and give to AI extreme good skill in dogfight/AG strike, SAM evasion and other maneuver, can be use for thing like : breaking formation, advanced strike like that :

GG summed up the actual combat AI issues pretty nicely. So instead I will comment on the setup of A2A AI in the mission editor and how they ought to react in the game-world.

1. We need a functioning Escort role where an aircraft or ground group is assigned to be escorted. The aircraft with the escort role does not need waypoints as they will fly based off of the escorted group and will be reacting to threats to the escorted group. However waypoints should be allowed to exist but are simply ignored once the aircraft is tasked a group to escort. (Note) I mentioned a ground group simply because I noticed if you assign a Predator or Kiowa to FAC on a moving group they will orbit the targeted group, lets have that orbiting functionality for friendlies but as a CAS!

2. A multiship group of AI should use different formations en-route to an engagement that are not restrictive to a setting in the mission editor. Formation should be determined by the tactical situation.

3. Aircraft weapon status should be a RTB condition and RTB needs to be a trigger condition.

And then there is this one...

4. Handling of multigroup CAPs need to be improved. We need something that acts as an AI command and control over an airspace CAP. Thus removing this aspect of the mission from the rigid trigger system. The mission editor is a powerful tool, but to command AI at a macro level... well it really doesn't have that functionality.

What I am suggesting is an automated system that keeps track of fuel load, weapons status, flight status (number alive), known enemy threats, and the current combat situation to command AI air forces on a larger scale. Essentially it should balance maintaining the CAP with tasking AI flights to engage enemy groups, friendly re-enforcements, airborne refueling, friendly groups RTB. Lets look at a practical example of what currently happens, and what I propose should happen...

Lets say we have 4 flights of 2 F-16Cs performing a CAP in roughly the same area.

Situation: Setting up the Cap

Current Sim: Builder must adjust position and timing of groups to ensure CAP area is will covered. This goes out the window once flights go engage or refuel.

CAP Plan: Mission builder draws an area with a spline tool to define the CAP zone. Mission builder can assign weighted values to splines of zone to define the priority on certain areas of the CAP zone to ensure higher density of CAP flights. In mission, CAP groups are ordered to flight plans distributed with the area they must protect and where enemies are likely to come from.

Situation: A single enemy aircraft enters the area

Current Sim: All 4 flights will likely press and engage the target until it is destroyed

CAP Plan: The nearest flight with capable fuel and weapons is ordered to intercept the contact. All remaining groups maintain their approx CAP areas.

Situation: Flights running low on fuel

Current Sim: Builder can guesstimate when flights will need fuel and stagger "refuel" tasks to flights. Alternatively nothing could be set and the AI wingman will likely need fuel first, he will leave his flight lead to get fuel, go full burners back to the flight lead. Shortly after the wingman returns the flight lead will need fuel, leading both of them to the tanker for gas. If no reaction to threats exist, the other 3 groups will react the same way, likely at the same time...

CAP Plan: Fuel status of all flights, distance to tanker, and approx time to refuel the flight is taken into account when giving refuel orders to CAP flights. Active CAP flights adjust their flight plan to counter act the void created when one flight leaves to refuel.

Situation: Reinforcements are needed

Current Sim: If the CAP flights have unique waypoints, each flight could use a replacement group with identical waypoints that are activated once the first lands or is shot down.

CAP Plan: Flights are placed into the cap plan and are activated as needed until no more flights are needed. CAP plan activates flights based on needs of the CAP.

Situation: Flight lands undamaged

Current Sim: Flight is useless

CAP Plan: After a set rearm/refuel time flight is added back into available flights for the CAP plan to use in the future.

IFF systems and behavior should be implemented. IFF/enemy recognition is the main pivot point of tactics. Right now it feels that the entire airspace is a killbox.

What I am suggesting is an automated system that keeps track of fuel load, weapons status, flight status (number alive), known enemy threats, and the current combat situation to command AI air forces on a larger scale. Essentially it should balance maintaining the CAP with tasking AI flights to engage enemy groups, friendly re-enforcements, airborne refueling, friendly groups RTB. Lets look at a practical example of what currently happens, and what I propose should happen...

Yeah, there should be some AI system for handling the flights tasking. Currently it's more like air quake in a way.

BTW, don't know if it was modified in FC2, but I remember that interceptor AI (e.g. MiG-31) was rather useless - it would get into close combat when instead it should try to separate (using it's high speed) and attack from afar.

IFF systems and behavior should be implemented. IFF/enemy recognition is the main pivot point of tactics. Right now it feels that the entire airspace is a killbox.

Definitely, but if implemented, I think it should be configurable per side (e.g. several levels of reacting to unknown contacts, e.g. from close to visual identification to shooting them down regardless).

IFF is something you can probably forget about. It is far too complex an issue when there are plenty of other things that the AI still does not/cannot do or understand. Having perfect knowledge of friend or foe isn't as huge a deal (in simulation) as you might think - just pretend they already executed all their ID matrix stuff and got a perfect result.

The 'killbox' thing isn't a huge deal either, and either way, with the A-10C's introduction of new ME parameters, you can modify the reaction to enemy contacts.