Jump to content

Better AI needed all round


RuskyV

Recommended Posts

I've just been looking through the forums and noticed the issues people have been having with the ai and how little anyone has commented or even engaged in conversation with any of the threads.

 

This sort thing i would have thought would be a little higher up in importance seeing the paid content for campaigns are based around AI assets that seem to be exhibiting really bizarre behaviour, specifically in this case the AI aircraft, especially the AI enemy.

 

-My initial interest got peaked the other day while i was doing a mission i generated and i noticed that there were some really funny things going on, one of the things i noticed in tacview was that when i fired an aim120c at the same time a Su27 fired an R27 my aim 120 guided on his launched missile for about 2 seconds and that was enough for the 120 to then have bled enough energy that when re-acquiring the su27 it had lost all its speed to not be effective once the su27 went defensive, I've not seen this behaviour from the enemy missiles ever being spoofed friendly missiles in this way.

 

After seeing this somewhat strange behaviour i started to look through more tacview files that showed even more bizarre behaviour when fighting enemy AI air assets on our server. One of the first things i noticed was the massive spamming of countermeasures coupled with this death spiral they do some times near stall speed to spoof incoming missiles.This weird behaviour is quite easily repeatable which will often see them just smash into the ground.

 

If the enemy escapes the BVR onslaught with their death spiral completed and without hitting the ground, once out of counter measures they often just bug out and leave.

 

On our server the other day (under slightly different circumstances) i chased a su27 from 5000ft in a constant climb up to 30k while slowly and rhythmically dumping all its flares after firing just one AIM9X while down low, i actually closed within gun range and shot it down without it even attempting to evade or go defensive while in its climb.

 

The only part of the AI logic that seem to be better, (not perfect) is within the visual area where they present a good challenge, the capabilities of the individual aircraft could be brought into question here but this thread is already far too long for that....

 

 

-I could go on about other issues i have seen over time with workaround logic needed in the mission editor just to get AI to land on aircraft carriers. How people make such awesome campaigns has me perplexed.

 

I can say for sure the AI is not up to scratch and does not exhibit any behaviour that could have me tricked into thinking im fighting or maneuvering with another opponent that has any intelligence at all.

It's just a shame AI exploits seem to now work better as tactics than the capabilities and strengths of the given aircraft fighting for or against.

 

Given MAC is supposed to be the new chapter in DCS world i would have thought that this would be given some serious love....

 

Peace:thumbup:

Link to comment
Share on other sites

People, including the devs are well aware of the AI shortcomings. There has been talk of improving it. However, AI is one of the more complicated things to do on a higher level, it also tends to be very resource intensive, especially as unit numbers ramp up. There is also the aspect that ''one AI'' won't work very well, it either needs to be tuned for individual aircraft or it needs to be able to work abstractly with different aircraft. Both are time consuming and the second is extremely resource intensive.

 

It's possible with enough time to make highly competent AI, but that takes a lot of time and is demanding on the computer resource side. So, it's never going to be particularly great, maybe ''competent'', but it'll never be on par with humans. Same basic reasons the AI doesn't use the same FM as we do. Too much resources required for a home PC.

Де вороги, знайдуться козаки їх перемогти.

5800x3d * 3090 * 64gb * Reverb G2

Link to comment
Share on other sites

You can exercise some pretty specific control over AI behavior; how they react when attacked, where & what types of targets they will attack, whether they will jettison AG stores, and more. Most people just don't know or care to make the effort to do so. You can't just slap crap down without putting in any time to learn how to achieve the desired result....complex missions require complex design.

Could behavior be improved? Absolutely, but it's not nearly as bad as people make it out to be.

Link to comment
Share on other sites

yes there is quite a bit of different totally useless control you have over the AI. but most of them are totally useless and 9 times out of 10 you use the same setup for every plane in the mission, just so that they will not get a stupid idea of doing something that will break the mission.

 

In fact, I have been using the mission editors since back in the first iteration of the flanker series, and the state of the AI is today at an all time low. in my opinion.

I turn of every part of the AI that is possible because if you give them just a tiny bit of self control, it will most likely destroy your mission.

Instead of setting up a nice SEAD flight to clear enemy air defences, I tend to use standoff pinpoint strikes to make the AI kill AD assets. Because that way you don't risk them just deciding to overfly 2-3 active sam-sites on the way to the sam site the AI decided was the most important one.

And forget about using the search in area or attack group, or ROE controls, most of the time the AI either just decides there is no targets to destroy and goes back home without attacking, or attacks a different target that it deems worthy.

 

Ohh, and the standoff strike can not be directed at the vehicle or group itself, that would be to easy. It has to be directed at the actual ground under the vehicle, because the AI with all the fancy FLIR and pods cant be relied on to find a target in the open.. it is just impossible sometimes.

 

I have stopped creating content for DCS because it is to much of a fight against broken AI and the mission editor. The only saving grace I can com think of is the MOOSE framework that at least makes enemy air a little bit less troublesome to add to missions. Even if that to has a long way to go before it will make mission editing fun and rewarding.

 

If people don't see a reason to make missions anymore, of find that the AI behaviour break their immersion and don't feel like it is a simulation anymore DCS will be a very boring, to the point that i want care how many cool modules there is for the game.

Link to comment
Share on other sites

In fact, I have been using the mission editors since back in the first iteration of the flanker series, and the state of the AI is today at an all time low. in my opinion.

The AI absolutely needs improvement (and ED has acknowledged this) but I disagree that it's in a dire state. Earlier on in DCS it was much worse, and it's now actually usable. It might require trial and error to get the AI to behave as you intend, but I think this might be sort of inevitable with advanced AI.

 

 

 

Instead of setting up a nice SEAD flight to clear enemy air defences, I tend to use standoff pinpoint strikes to make the AI kill AD assets. Because that way you don't risk them just deciding to overfly 2-3 active sam-sites on the way to the sam site the AI decided was the most important one.

This is a current AI weakpoint, and I try not to make SAM defenses overly complex if the AI needs to do SEAD. However, I can still get them to work with a little management. Never use the default tasks, they're too general. The specific tasks that dictate engagement range, weapon type, etc are required for any realistic mission. In addition I find that SEAD works better with many individual flights instead of one large flight with many planes. Large flights just don't fire enough missiles (but it depends on what you're trying to do).

 

 

 

And forget about using the search in area or attack group, or ROE controls, most of the time the AI either just decides there is no targets to destroy and goes back home without attacking, or attacks a different target that it deems worthy.

I don't like giving the AI omniscience when it comes to finding targets, so I almost always use Search and Engage, it seems to work well, but you need to understand each unit's sensors. The sim could provide more information on this internally which could help in mission planning.

 

 

Ohh, and the standoff strike can not be directed at the vehicle or group itself, that would be to easy. It has to be directed at the actual ground under the vehicle, because the AI with all the fancy FLIR and pods cant be relied on to find a target in the open.. it is just impossible sometimes.

CAS is the task for attacking ground units. Always use that, it should also let you force the AI to find the target if you want to. The mission task system isn't that great and I think that's the bigger problem here. Scripting lets you assign any task to any group and I think the ME should do the same, along with expanding some of the other options (like attack group lets you specify with what weapons, while Search and Engage doesn't).

 

 

If people don't see a reason to make missions anymore, of find that the AI behaviour break their immersion and don't feel like it is a simulation anymore DCS will be a very boring, to the point that i want care how many cool modules there is for the game.

While I don't think the ME is even close to unusable, I agree with the sentiment. The modules do need a strong environment to fly in to really be of value.

Awaiting: DCS F-15C

Win 10 i5-9600KF 4.6 GHz 64 GB RAM RTX2080Ti 11GB -- Win 7 64 i5-6600K 3.6 GHz 32 GB RAM GTX970 4GB -- A-10C, F-5E, Su-27, F-15C, F-14B, F-16C missions in User Files

 

Link to comment
Share on other sites

totally useless control you have over the AI

 

Except that said control isn't useless at all, because without it you'd be even more up the creek. Which, OTOH, isn't to say that the AI isn't pretty bad when jumped on in a one-to-one situation for example, because it is.

 

So you're both right and wrong IMO ;)

The DCS Mi-8MTV2. The best aviational BBW experience you could ever dream of.

Link to comment
Share on other sites

Same basic reasons the AI doesn't use the same FM as we do. Too much resources required for a home PC.

 

I can see the logic behind what your saying but i don't see this as a valid reason when you take into account the multiplayer AFM (used by others) is the same as what you would be flying and yet this gets broadcast across a network not only to yourself but other connected clients at the same time.

 

This should be far more resource intensive when you take into account some of the servers like blue flag that not only deal with all the AFMs being flown but also having to track all the ground units.

 

How the AI uses any advanced flight models shouldn't need to be any more cpu intensive than transmitting this across any normal multiplayer network.

Link to comment
Share on other sites

How the AI uses any advanced flight models shouldn't need to be any more cpu intensive than transmitting this across any normal multiplayer network.

This isn't comparable. What happens in MP is the server takes positions calculated by clients and then shares those positions with all the other clients. Each individual client runs the physics for its own aircraft, so the load for all the planes is shared among many computers instead of one. It's sort of like multithreading. AFM for AI planes would be many many times more resource intensive than AFM in multiplayer.

Awaiting: DCS F-15C

Win 10 i5-9600KF 4.6 GHz 64 GB RAM RTX2080Ti 11GB -- Win 7 64 i5-6600K 3.6 GHz 32 GB RAM GTX970 4GB -- A-10C, F-5E, Su-27, F-15C, F-14B, F-16C missions in User Files

 

Link to comment
Share on other sites

Each individual client runs the physics for its own aircraft, so the load for all the planes is shared among many computers instead of one. It's sort of like multithreading. AFM for AI planes would be many many times more resource intensive than AFM in multiplayer.

 

I agree with the part about only the positional information being shared across a network, but my question still stands about the AFMS and why the AI using them would be no more CPU intensive than in a multiplayer environment when as you say each local client runs its own AFM.

 

Take for example blue flag or any server that would be running many aircraft, i think its fair to say a large percentage of that figure would be AFMs? i don't know the exact count of flying AFMs but take into consideration any of the paid single player campaigns and tell me the AI air unit count would be wildly different.

 

My point is if AI is just a complicated set of rules used to govern any aircraft (AFM or not) why should it be any different in CPU use than individual clients running AFMs in a multiplayer environment?

Link to comment
Share on other sites

Every player-operated aircraft in DCS at this point is either EFM (3rd party "PFM" i.e. F-14B), AFM+(some systems not modeled but airflow across surfaces modeled i.e. Su-25T/A-10A) or PFM (F/A-18C, A-10C etc.) the old Flaming Cliffs flight models have all been updated.

 

DCS is essentially a wind tunnel. Performing the calculations for airflow across multiple control surfaces and how it affects the aircraft's behavior is extremely CPU-intensive and not something a home PC could handle for several aircraft at once. In multiplayer, each user's own computer is managing those calculations for their own aircraft, then sharing the results. This allows multiple aircraft with complex flight model calculations to exist in the same space simultaneously.

 

Think of it this way: if I tell you the square root of 750 times 5 is approximately 137, that's a whole lot faster and less labor intensive for you than you grinding out the calculations by hand yourself.


Edited by feefifofum
Link to comment
Share on other sites

Every player-operated aircraft in DCS at this point is either EFM (3rd party "PFM" i.e. F-14B), AFM+(some systems not modeled but airflow across surfaces modeled i.e. Su-25T/A-10A) or PFM (F/A-18C, A-10C etc.) the old Flaming Cliffs flight models have all been updated.

 

DCS is essentially a wind tunnel. Performing the calculations for airflow across multiple control surfaces and how it affects the aircraft's behavior is extremely CPU-intensive and not something a home PC could handle for several aircraft at once. In multiplayer, each user's own computer is managing those calculations for their own aircraft, then sharing the results. This allows multiple aircraft with complex flight model calculations to exist in the same space simultaneously.

 

Think of it this way: if I tell you the square root of 750 times 5 is approximately 137, that's a whole lot faster and less labor intensive for you than you grinding out the calculations by hand yourself.

 

That makes sense thanks for explaining:thumbup:, so we should have plenty of overhead for AI if the only complex calculations done are for the plane you are flying and the rest just being positional information.

Link to comment
Share on other sites

I recently got back into the Ka-50 Blackshark after a couple of years absence and was trying the Instant Action - Nevada - Shooting Range mission. We both sat in a Hover about 6 clicks from the target site and I asked him to attack ground targets. His response was to fly around me in circles at low level, firing of chaff as if it was confetti and then crashed into the ground :huh:

 

This didn't give me a lot of confidence that the wingman AI has improved much during those two years. Just to confirm I am using the latest 'Stable version' of DCS World.

 

Glen

GAJ52

 

Intel i7-7700K @ 4.60GHz | 32 GB Ram | Win 10 Pro 64 | GTX 1080i 11.00MB | Saitek X52 Pro

Link to comment
Share on other sites

That makes sense thanks for explaining:thumbup:, so we should have plenty of overhead for AI if the only complex calculations done are for the plane you are flying and the rest just being positional information.

 

That only applies to client aircraft in multiplayer, which makes AI a moot point because everything is being driven by a human.

 

In a single player environment, your PC has to handle rendering the environment, calculating your own flight model, PLUS dictating all the behavior of other units in the map. That means basic flight models for every AI-driven aircraft, plus the pilots' behavior, plus any units on the ground, etc. Things can get bogged down very quickly.

Link to comment
Share on other sites

Things can get bogged down very quickly.

 

Just out of interest what sort of unit count would you class as an upper limit in your opinion. Our dedicated server runs upwards of 750+ units that can swell quite alot with the Moose stuff.

 

would you have any input on what you might think would be a good compromise? A bit off topic i know:book:

Link to comment
Share on other sites

Yeah things like AI units following a waypoint breaking off and going back home because they were attacked en-route.

 

EDIT: actually it only happened with fly-over waypoint, with the normal turning-point it continues.

 

 

Every player-operated aircraft in DCS at this point is either EFM (3rd party "PFM" i.e. F-14B), AFM+(some systems not modeled but airflow across surfaces modeled i.e. Su-25T/A-10A) or PFM (F/A-18C, A-10C etc.) the old Flaming Cliffs flight models have all been updated.

 

DCS is essentially a wind tunnel. Performing the calculations for airflow across multiple control surfaces and how it affects the aircraft's behavior is extremely CPU-intensive and not something a home PC could handle for several aircraft at once. In multiplayer, each user's own computer is managing those calculations for their own aircraft, then sharing the results. This allows multiple aircraft with complex flight model calculations to exist in the same space simultaneously.

 

Think of it this way: if I tell you the square root of 750 times 5 is approximately 137, that's a whole lot faster and less labor intensive for you than you grinding out the calculations by hand yourself.

 

 

Except atency is the biggest bottleneck in such cases, once in a while it would be cool some community got together to have a good quality LAN match and streamed to Twitch ofcourse.

Modules: A-10C I/II, F/A-18C, Mig-21Bis, M-2000C, AJS-37, Spitfire LF Mk. IX, P-47, FC3, SC, CA, WW2AP, CE2. Terrains: NTTR, Normandy, Persian Gulf, Syria

 

Link to comment
Share on other sites

Yeah things like AI units following a waypoint breaking off and going back home because they were attacked en-route.

you can determine this behaviour with an advanced waypoint action, i believe its called reaction to threat

CPU: AMD Ryzen 5 1600X

GPU: AMD RX 580

Link to comment
Share on other sites

Just out of interest what sort of unit count would you class as an upper limit in your opinion. Our dedicated server runs upwards of 750+ units that can swell quite alot with the Moose stuff.

 

would you have any input on what you might think would be a good compromise? A bit off topic i know:book:

 

DCS is basically geared around single sortie, small flights/squadrons and appropriate relevant unit scales. Normally this would be on the order of 'a few dozen'. That people CAN create missions with hundreds of units and dozens of players is beside the point, it's outside the scale of what's intended.

 

I mention that because, the only reason these really large scenarios are possible is because of the shortcuts via simplified AI modeling and intelligence and the relative power of modern PCs compared to when DCS family started. You can have lots of items simply rendered, or a few items rendered to a high level or detail, but when you start trying to do lots of units to a high level... then you are in trouble.

 

Translation : If and when they start upping the AI complexity, unit counts will be forced down. No two ways about it and no amount of ''optimisation'' will alter that. A gallon jug, no matter how you turn it, ultimately only contains a gallon.

 

Personally, I think fewer but better is ideal, but many server owners and mission designers think player and unit count automatically = better. Can't have both, though, quality + quantity are to an extent mutually exclusive.

Де вороги, знайдуться козаки їх перемогти.

5800x3d * 3090 * 64gb * Reverb G2

Link to comment
Share on other sites

DCS is basically geared around single sortie, small flights/squadrons and appropriate relevant unit scales. Normally this would be on the order of 'a few dozen'. That people CAN create missions with hundreds of units and dozens of players is beside the point, it's outside the scale of what's intended.

 

I mention that because, the only reason these really large scenarios are possible is because of the shortcuts via simplified AI modeling and intelligence and the relative power of modern PCs compared to when DCS family started. You can have lots of items simply rendered, or a few items rendered to a high level or detail, but when you start trying to do lots of units to a high level... then you are in trouble.

 

Translation : If and when they start upping the AI complexity, unit counts will be forced down. No two ways about it and no amount of ''optimisation'' will alter that. A gallon jug, no matter how you turn it, ultimately only contains a gallon.

 

Personally, I think fewer but better is ideal, but many server owners and mission designers think player and unit count automatically = better. Can't have both, though, quality + quantity are to an extent mutually exclusive.

 

 

Thanks for the insight, i guess when you start adding external scripts in with high unit counts in order to add in some resemblance to some sort of intelligence or tactics then you are forcing the "system" to work beyond what it is designed to do currently.

 

On our dedicated server we have to have a system where we can leave it and it be semi dynamic where you can jump in and out without having and end goal, i guess this is where you start running into issues as you have to rely on external scripts to counter for the (sometimes) dumb AI when you already have a unit count that is proberby 3 times what it should be. Guess i cant blame the "system" when its already overworked.

 

That being said i would love to have some of those externally made scripts built into the editor with some sort of total unit count with maybe a built in warning when you are pushing the limit. Being more informed on how things would run would save hours of play testing (which often end up being spoilers for those involved). The editor which is really so much better than it used to be could do with more power over AI units that is a more user friendly for its users, especially for people like myself who have little time to even learn how to script AI outside of DCS.

 

Some of the external programs being used to ease mission generation are really nice and i've often wondered why these are not features in the editor..:doh:

Link to comment
Share on other sites

  • 8 months later...

I’ve been playing almost exclusively in SP for a little over a decade at this point. While I feel the AI is nowhere near unusable, I also feel it is in great need of improvement. The air to ground is ok in my experience. I do use advanced settings, and search and engage in zone a lot. But that’s what advanced settings are for. Hell I wish there were even more advanced settings. In air to air the old Ai, circa LOMAC, gave a much better gun fight, and a totally shitty BVR fight. Over the years I’ve noted a ton of little tweaks from ED, and the air to air Ai has improved greatly, especially in BVR. The biggest issues I notice ATM are, the extremely predictable dogfight behavior, the lack of randomization in a dogfight, the lack of ability to fine tune AI skill. In a WVR fight the AI ATM will always try to defend by going vertical, in many cases this is a horrible decision, and makes for way to easy of a kill. This is not the way the older simpler AI worked. It never really went “defensive” it just constantly tried to turn into you to get on your six. But it made for a much more realistic feeling gun fight. I’ve seen the AI do some pretty interesting things in between patches. Everything from turning cold and lighting the burners, to pulling a “maverick” and slamming on the air brakes at the bottom of a split S. What I feel needs to happen is we need all on that, but with a randomizer. I also feel that attributes such as aggressiveness Ned to be scalable. That way one “good” AI doesn’t necessarily act the same as another “good” Ai. I do not feel that the Ai using the SFM is a bad thing. As long as the performance lines up pretty closely one the performance graphs, and I think mostly it dose. ED has made some pretty good progress this year, and are saying that they will be making more in the coming year, so I’m hopeful. Wile I do not think it’s all doom and glum, I definitely feel that a not very interesting Ai is one of the main limitations on my enjoyment of this awesome sim.

Link to comment
Share on other sites

  • Recently Browsing   0 members

    • No registered users viewing this page.
×
×
  • Create New...