Suggestions for AI Improvements in Future Updates

[ben_wh]

It seems that a recent AI discussion thread took a less than ideal turn, however I believe that the AI improvement is a worthwhile subject to be explored further due to impact on offline gameplay.

I understand that it is difficult to code AI player to be both realistic (like real human) and fair, but since DT started on the AI improvement path I think there may be the willingness to make incremental refinement in this area.

IMHO, these are at least two areas that may warrant further further investigation by the team if AI improvement is planned for 4.13 and beyond:

(below are personal observation only and would like others' input)

1) AI gunnery of rookie and regular pilots - from both personal experience in game and reading others' posts, lower level AI player (rookie and regular) still seems to make too good a shot (high speed, difficult angle) at times.

2) AI wingman warning - it is great to see that enemy AI act more as a team, and enemy AI wingman will come to rescue to its leader if the leader is targeted by the human player. Given this AI behavioral improvement, it would be very beneficial for the human player to receive some warning from the wingman, if the human player is being targeted. For example, the wingman can proactively issue a warning if enemy is within a short distance around (especially behind) the player, provided that the wingman is within range of course. Although this may get excessive in a crowded fur-ball environment. Another option is , similar to the implementation in another sim (Battle of Britain 2 Wings of Victory), have a command that enable the player to query his team whether the player's six is clear. (For example, 'Check my six!') The wingman can then respond 'Clear!' or 'Bandit at seven o'clock!' (for example) accordingly, again if the wingman is within range.

Not sure if either of these is feasible but I hope these suggestions at least stimulate some thoughts. Would love to hear what DT and other players think.

Cheers,

[IceFire]

Some good suggestions Ben!

I had added in that other thread that some AI flown aircraft still have VERY high roll rates that simply aren't possible at the speeds the AI are performing them at. Something is obviously simplified for them. Oddly it doesn't affect all aircraft... the Ki-61 will roll VERY quickly in the hands of the AI while any Bf109 rolls much more slowly.

[ben_wh]

Thanks, IceFire.

I think the 'AI roll' has been around since the earlier days of IL-2. I was glad to see that removed when I was flying some of the mod packs but it is very noticeable back in 4.12 now.

I do not have any documented proof of whether this - and the very high-G maneuvers pulled by some of the AI crafts - are realistic or not. But it does make for some eyebrow-raising moments when one sees an AI plane rolling endlessly while diving or pulling up vertically and very sharply all of a sudden during a dogfight.

Cheers,

[IceFire]

Yeah, definitely not a new problem at all. I'm hopeful there are ways to mitigate it in the AI programming without it being a huge task.

[majorfailure]

1) AI gunnery of rookie and regular pilots - from both personal experience in game and reading others' posts, lower level AI player (rookie and regular) still seems to make too good a shot (high speed, difficult angle) at times.
For a while I thought that too -I felt I got hit and shot down by rookies/regular at ridiculous distances too often - but then it didn't happen that often any more, so I think that was just bad luck. Part of that is also because aces AI does usually not shoot at great distances (700m+) like rookies do. And have you ever watched rookie/average AI shoot at a nearly non evading target? - You just want to facepalm continously.
The only thing that I think rookies are too good at is torpedo bombing - while average/veteran AI usually only gets 2 out of 4 torps on a large moving ship (battleship, carrier) rookies most of the time get 3 on target.

2) AI wingman warning - it is great to see that enemy AI act more as a team, and enemy AI wingman will come to rescue to its leader if the leader is targeted by the human player. Given this AI behavioral improvement, it would be very beneficial for the human player to receive some warning from the wingman, if the human player is being targeted. For example, the wingman can proactively issue a warning if enemy is within a short distance around (especially behind) the player, provided that the wingman is within range of course. Although this may get excessive in a crowded fur-ball environment. Another option is , similar to the implementation in another sim (Battle of Britain 2 Wings of Victory), have a command that enable the player to query his team whether the player's six is clear. (For example, 'Check my six!') The wingman can then respond 'Clear!' or 'Bandit at seven o'clock!' (for example) accordingly, again if the wingman is within range.

+1
I had this happen maybe 3 times in 20 missions, that my wingman actually told me there were bandits near me, and one time he told me they were at 12o'clock -duh, really....
But I think either my perception of my wingmen is different or they have changed, they often stay with you in a fight and if told at the right moment actually cover you on command - and even reform on your wing when told so (most times). Not everything is fine, sometimes thy still wander off aimlessly. And when your mission is to escort something, you can just forget about the AI -they either try to fulfil their escort mission, if you command them anything, or they follow you if told to do so, but will not engage, even if attacked.

[ben_wh]

For the AI warning, the idea is to have the communication with AI more meaningful and useful in making tactical decision. So having the AI wingman to warn you of immediate danger, even if passively on your request, is along that direction.

Other aspects I'd like to suggest (_if_ DT plans to tweak the AI more in the future) are:

1) More self-preservation instinct among AI planes - on when to break off an attack/bomb run, when to jettison bombs, break formation and head for home when its formation is decimated

2) More variations in behavior among AIs - some are more disciplined and press home the attack no matter what, some break formation easily when attacked.

The following is not a request but reminiscing of earlier flight sim experience ...

If I recall correctly, in Rowan's MiG Alley, not only are there training and behavioral differences among AI pilots in Russian, Chinese and North Korean MiG squadrons, but recent successes or defeats in the dynamic campaign also affected AI behavior. For example, AI squadrons that had been successful in recent missions exhibited more aggressive behavior, while those that suffered attrition sometimes hanged back or disengaged more readily. In some missions you would see some brightly painted MiGs from an expert squadron and you knew you'd be in for a tough fight. In such case the AI behavior was very successful in raising the immersion factor in that you, the human player, suspended your disbelief and almost attributed human quality to your computer-controlled opponents.

Cheers,

[sniperton]

If I recall correctly, in Rowan's MiG Alley, not only are there training and behavioral differences among AI pilots in Russian, Chinese and North Korean MiG squadrons, but recent successes or defeats in the dynamic campaign also affected AI behavior. For example, AI squadrons that had been successful in recent missions exhibited more aggressive behavior, while those that suffered attrition sometimes hanged back or disengaged more readily. In some missions you would see some brightly painted MiGs from an expert squadron and you knew you'd be in for a tough fight. In such case the AI behavior was very successful in raising the immersion factor in that you, the human player, suspended your disbelief and almost attributed human quality to your computer-controlled opponents.

Yeah, a dynamic AI would be great, but it's only possible in campaign games which the Il-2 engine is not designed for (campaigns are generated as a series of static missions by 3rd party applications like DGEN and DCG). IF the AI were composed of various openly accessible skills (as major_kudo suggested on the pattern of CLoD, and many others opposed with reasonable arguments), THEN it would also be possible to dynamically affect AI behaviour via 3rd party campaign generators. I mean we could create "brave" AI squads without corresponding good shooting skills, and the like. Just brainstorming, don't shoot me. :-P

[major.kudo]

These are the pictures which I drew before.

First, I think this is just a problem.
http://gachopin.no-ip.info/kudo/Fighter_2.jpg
I think the cause of this problem is a thing by too exact shooting.
I want to play more historical air combat.

I thought that it would be solvable by this.
http://gachopin.no-ip.info/kudo/AI_Level.jpg
However, this may change greatly the victory or defeat of a battle for dynamic campaign etc.
I think that the result which the maker of campaign does not mean may be brought about.

Then, I thought that it was a thing that what is necessary is just this.
http://gachopin.no-ip.info/kudo/AI_skill_edit.jpg
It seems that however, it is not so reputable now.
:
:
:
So, I added change.
http://gachopin.no-ip.info/kudo/AI_skill_edit2.jpg
Everyone, what do you think?

-

I point out another one.
http://gachopin.no-ip.info/kudo/Unreal_gunshooting.jpg
At least, I think that impossible to the pilot of rookie and average.

Sorry, my poor English.

[ben_wh]

To sniperton,

I understand the limitation of the IL-2 engine and the MiG Alley reference was not a request but more of a reminiscence of past flight sim experience.

Who knows, may be useful as creative brainstorming when DT is thinking about AI update for 5.00? :)

[sniperton]

I understand the limitation of the IL-2 engine and the MiG Alley reference was not a request but more of a reminiscence of past flight sim experience.

It was clear to me, but I liked the idea, and upon some conditions it could be even feasible IMO. Campaign generators do improve the skills of the AI over time, but they do it according to how AI skill can be set in the game -- in a 'monolithic' way (rookie>average>veteran>ace), where one particular skill (say, shooting accuracy) develops hand in hand with others (say, agressiveness). I'm pretty sure that an 'average' Japanese pilot in RL had a different 'skills composition' than his American adversary, even if they both were on the same 'average' level in general. What I suggest is not a full breakdown of all particular skills via sliders (like in CloD), but to allow different 'attitudes' within the same skill level. Just think of the difference in habits between Manfred and Lothar von Richthofen. :grin:

[majorfailure]

These are the pictures which I drew before.

First, I think this is just a problem.
...
I think the cause of this problem is a thing by too exact shooting.
I want to play more historical air combat.

I thought that it would be solvable by this.
...
However, this may change greatly the victory or defeat of a battle for dynamic campaign etc.
I think that the result which the maker of campaign does not mean may be brought about.
Part of the problem may be too good gunnery by the AI, but greater part is IMHO the fight to the death attitude - AI usually doesn't retrat - real life pilots in a faster plane wouldn't wait till all of their wing is wiped by superior opposition.


Then, I thought that it was a thing that what is necessary is just this.
http://gachopin.no-ip.info/kudo/AI_skill_edit.jpg
It seems that however, it is not so reputable now.
:
:
:
So, I added change.
http://gachopin.no-ip.info/kudo/AI_skill_edit2.jpg
Everyone, what do you think?

Better, use only single gunnery skill maybe (bad marksmen necessarily are bad deflection shooters, and good marksmen probably are good deflection shooters) and add maybe bombing/torpedo ability, add character (agressive/timid) and maybe some leadership skill.
And make much less grades of each talent, three to five (inept/(less than average)/average/(good)/excellent)



I point out another one.
http://gachopin.no-ip.info/kudo/Unreal_gunshooting.jpg
At least, I think that impossible to the pilot of rookie and average.

Sorry, my poor English.
It may be unrealistic, but as a player I do it a lot, too, you have nearly endless ammo in a 190 so you can try to get the enemy even with low probability snap shots. And you get a feel where the enemy is and where he will be - even if he is temporarily out of sight you can make the shot, and even as a not so great pilot you can IMHO - difference should be ace makes maybe 1 in 5 no-look-shots and rookie gets 1 in 100 or the like

[Notorious M.i.G.]

It may be unrealistic, but as a player I do it a lot, too, you have nearly endless ammo in a 190 so you can try to get the enemy even with low probability snap shots. And you get a feel where the enemy is and where he will be - even if he is temporarily out of sight you can make the shot, and even as a not so great pilot you can IMHO - difference should be ace makes maybe 1 in 5 no-look-shots and rookie gets 1 in 100 or the like

The difference between AI and player guesswork tends to be evident, though - try rolling into a different angled turn while you're in their blindspot and the AI still seems to visibly track your movement :-|

Personally, since TD made it clear that AI have subset skills a la CloD, I think it would be nice to see a popup window like CloD's FMB that gives the mission builder an option to nudge around their strengths and weaknesses, as per major.kudo's example. Having a good maneuver fight is fun, but sometimes the single .303 to the head from 500m in a sharp turn dampens the excitement a bit. I know these things happen, but I seem to get a disproportionate amount of PK's from the AI as compared to human players (maybe because the player is generally near the "center" of the 3D model, and the AI will always use the central point of the model for targeting as opposed to any firing solution that will connect)

[Pursuivant]

I know these things happen, but I seem to get a disproportionate amount of PK's from the AI as compared to human players (maybe because the player is generally near the "center" of the 3D model, and the AI will always use the central point of the model for targeting as opposed to any firing solution that will connect)

Same here. While it's reasonable for AI to aim at the aircraft's center, less experienced human pilots will give too little deflection, while IMO more experienced human pilots will give a little bit too much lead in an attempt to take out the engine on single-engined planes, since it's a larger target.

A quick fix for AI gunnery accuracy would be for Rookies and Average pilots to not lead their targets sufficiently, Veterans to give their target too much lead, and for Aces to get it "just right" aiming more or less at the target's CG or vulnerable parts of multi-engined planes.

[IceFire]

I point out another one.
http://gachopin.no-ip.info/kudo/Unreal_gunshooting.jpg
At least, I think that impossible to the pilot of rookie and average.
Even a rookie pilot will make a guess and fire a bunch of rounds to see if they can score a hit in those kinds of situations.

Prior to the AI upgrades the AI could see 360 degrees and would make these shots regardless. With the upgrade the AI could no longer see through their aircraft but that meant adding some routines to change how the AI did aiming. With the upgrade they will now estimate the target angle and make a deflection shot guess not unlike how human pilots behave.

You are probably on to something that the Rookie pilots are perhaps too successful in this area. The effect can be toned down surely. But its not unrealistic for an average to veteran to ace pilot to be able to make this shot.

With the FW190 in particular I got to a point where I was firing blind a lot of the time and destroying my target. I'd follow the target for a while and then once I was into firing position I'd pull the stick back and although I couldn't actually see the target I'd have the whole trajectory worked out in my head and I'd score a killing shot say 7 or 8 times out of 10. No reason the AI at perhaps the veteran level to be able to do that kind of shot semi-reliably. The rookie AI... maybe not as much as it does.

[Jami]

Same here. While it's reasonable for AI to aim at the aircraft's center, less experienced human pilots will give too little deflection, while IMO more experienced human pilots will give a little bit too much lead in an attempt to take out the engine on single-engined planes, since it's a larger target.

That's just what happens with human pilots in virtual skies. Rookies that I know shoot the enemies always when possible and hope to get some hits - and waste their bullets. More skilled guys try to hit the fuselage and most skilled aim at engines and even bomber pilots when attacking head on. That's how it goes in my sqn with pilots who have 1 to 10 years of experience in virtual combat flying.

A quick fix for AI gunnery accuracy would be for Rookies and Average pilots to not lead their targets sufficiently, Veterans to give their target too much lead, and for Aces to get it "just right" aiming more or less at the target's CG or vulnerable parts of multi-engined planes.

If we got this change, my sqn would be veeery happy...

[major.kudo]

Probably, many people thinking about me "this guy is adhering to AI's deflection shooting".
It's right. But I have a reason.

I prepared the server of the COOP only and was playing together some players.
Since A.I comes to have too exact shot, it became impossible to pleasant play.
Because "human's rookie player" is also contained in "some players".
They are merely shot down, before doing any action.
This is no good. So, we can not play air combat mission now.

I hope DT observes this thread.

[major.kudo]

Part of the problem may be too good gunnery by the AI, but greater part is IMHO the fight to the death attitude - AI usually doesn't retrat - real life pilots in a faster plane wouldn't wait till all of their wing is wiped by superior opposition.


retrat = retreat?
I think about retreat.
http://gachopin.no-ip.info/kudo/retreat1.jpg
and
http://gachopin.no-ip.info/kudo/retreat2.jpg
I want to transmit my idea to you well.
Is this unclear?

[Pursuivant]

You are probably on to something that the Rookie pilots are perhaps too successful in this area. The effect can be toned down surely.

But remember, there's some dispute over what "rookie" skill actually means. Does it mean "straight from training" with about 350 hours flying time and about 50 hours flying time "in type", plus an adequate gunnery and air combat maneuver training, or does it mean "straight from the farm" with less than 50 hours total flying time and no gunnery and combat maneuver training?

But its not unrealistic for an average to veteran to ace pilot to be able to make this shot.

Agreed. Some real life pilots described doing exactly this sort of thing.

It's also the reason that Veteran or Ace pilots will also break in an unexpected direction when they're "under the nose" of an enemy. They will anticipate the firing solution their opponent is trying to achieve on them and move to counter it.

No reason the AI at perhaps the veteran level to be able to do that kind of shot semi-reliably. The rookie AI... maybe not as much as it does.

I can't imagine even a well-trained rookie being able to pull off this shot.

The most realistic fixed-air gunnery training of the period was against towed targets pulled by a target tug. But, target tugs were (mostly) slow-moving obsolete planes to begin with, even ignoring the drag produced by towing a giant wind sock or banner.

That means that (most) target tugs just didn't have the speed required for trainees to make shots that required extreme amounts of lead.

I also can't see any responsible training officer allowing novice pilots to run the risk of accidentally shooting the target tug by making a badly miscalculated high deflection turning shot. I haven't found anything really detailed about exactly how fighter pilots attacked towed targets, but my guess is that attacks mostly consisted of "pursuit curves" which ended up with "high side attacks" against the target drogue, while keeping the target and the target tug in sight at all times.

In such situations, when a "trained rookie" can mentally plot his intended shot in advance and can keep his target in view at all times, accuracy from a high-side attack at 30-60 degrees "angle off" from the target should probably be about 2-3% at 200-300 meters, maybe a bit less for 60-60 degrees "angle off. (3% was the expected accuracy standard for UK and US gunnery schools, 5% was considered to be very good.)

For an "under the nose" attack where a "trained rookie" can't see his target, accuracy should be less than 1% at best, and might result in the rookie ending up colliding with his target, getting ahead of it, or losing awareness of its location.

For pilots with no gunnery training (e.g., 1941-43 Soviet rookie pilots, post 1943 German and Japanese rookie pilots, and many pre-1944 Chinese rookie pilots) any sort of deflection shooting at much more than 20 degrees "angle off" from the target's front or rear should be a waste of ammo.

[Treetop64]

I agree that Rookie fighter pilots are still too good in the game, especially in I-16s, while Aces now often do things in certain aircraft that they definitely shouldn't be doing. For example, play any early Barbarossa scenario with AI Ace pilots flying Bf109F-2s and AI Rookies flying I-16-18s and 24s, and watch as the Aces constantly get themselves into trouble by trying to turn fight with the I-16s - and even I-153s - and get shot up in the process.

Indeed, you see this in any theater aircraft; Veteran and Ace AI pilots in BnZ type aircraft are constantly trying to turn-fight against dedicated turn-fighters, and get stomped as a result, even when their foe is Rookie or Novice. This happens far too often to be attributed to the "occasional chance of a less skilled pilot beating a more skillful one" explanation.

With all the great improvements done to the AI recently, this is the one area where things got messed up in the process.

[ben_wh]

A few observations that I hope DT would take into consideration when they decide to further refine the AI:

1) Reading the discussion again there seems to be some consensus about the gunnery accuracy of rookie and regular AI pilot. Would love to hear more thoughts from more sim pilots.

2) (With respect to Treetop 64's comment) It is also my experience that the AI for BnZ planes tend to go into turn fights with slower but more maneuverable opponents. Furthermore, sometimes the high turn rate makes it easier for an AI pilot to to move the nose around for a shot, and this makes for some very interesting (some may say unhistorical results) in many occasions. e.g. Ace Bf-109F Vs novice/regular I-16/LaG-3 (as pointed out above) and Ace late US fighters Vs novice/regular Zeroes.

3) Self-preservation instinct - it was pointed out before that AI may not necessarily want to only retreat when its own plane suffers significant damage. Highly unfavorable tactical situation may also prompt this reaction (as graphically illustrated by major.kudo). This may help address issues of high loss rate in battles - e.g. whole flight wiped out in a fur ball - which is not very often historically since at some point the rest of the flight would retreat. This may apply for bombers as well - if 2/3 of their flight went down and there is no fighter escort - would they still press on to the target?

4) Linked to (2) above is whether AI can be made sophisticated enough to use team tactics. It is admittedly a very difficult task, but if, for example, AI can be made adopting 'Hit and Run' tactics, forming 'Lufbery Circle' or initiating 'Thach Weave' then it would truly take offline battles to a new level.

Of course whether or how DT may adopt this is up to the team, but I hope that discussion here may help stimulate ideas for future patches.

Cheers,

[Jami]

Would love to hear more thoughts from more sim pilots.


I agree 100%. This thread has got over 750 views but only few replies.

It'is understandable that for those who fly dogfight against human pilots on line the issue of AI skills is not so interesting but for off line players it is fundamental. And for those who want to fly with their friends coop mission with AIs involved. And it comes more and more important when you try to create historically accurate missions.

[Furio]

I think that giving “more power to the player” could lessen some of the AI issues. As I see it, there should be the option to pause the game, or briefly hand your plane to autopilot, to take some decision that no AI can take.

Some examples.

In the landing pattern, Number 4 is damaged and should land first, but continues to go around leaving unnecessarily precedence to others until it crashes.
Player should be able to take over control tower and order Number 4 to be the first to land.

During an escort mission, an enemy fighter appears. All your squadron’s pilots chase it, leaving alone the escorted bombers.
Player should be able to order: “Number 5 and 6 go ahead with the chase, all other return to escort position”, or click on the selected planes to obtain the same result.

Player has damaged and enemy bomber, and is turning for the coup de grace. Three more squadron’s AI mates run to finish it, allowing other enemy planes to bomb the target undisturbed.
Player should be able to select a target to be ignored by AI pilots.

The list could go on, but the concept is clear, I think. You pay a little price in immersion, because you need to stop the action and momentarily assume the rank of god, but you avoid that immersion be killed by AI pilots unrealistic actions.

[majorfailure]

2) (With respect to Treetop 64's comment) It is also my experience that the AI for BnZ planes tend to go into turn fights with slower but more maneuverable opponents. Furthermore, sometimes the high turn rate makes it easier for an AI pilot to to move the nose around for a shot, and this makes for some very interesting (some may say unhistorical results) in many occasions. e.g. Ace Bf-109F Vs novice/regular I-16/LaG-3 (as pointed out above) and Ace late US fighters Vs novice/regular Zeroes.

Yes, AI seems to get in trouble when fighting in a much faster less agile plane - and tries to TnB -with usually disappointing results. Though in some situations they seem to get it right, F4F(good) vs. A6M often turns out okay.

3) Self-preservation instinct - it was pointed out before that AI may not necessarily want to only retreat when its own plane suffers significant damage. Highly unfavorable tactical situation may also prompt this reaction (as graphically illustrated by major.kudo). This may help address issues of high loss rate in battles - e.g. whole flight wiped out in a fur ball - which is not very often historically since at some point the rest of the flight would retreat. This may apply for bombers as well - if 2/3 of their flight went down and there is no fighter escort - would they still press on to the target?
Also AI should consider avoid picking fights in unfavourable positions -if they still can. Real life pilots couldn't tangle endlessly with the enemy on most missions, limited by mission objective and by time (or fuel). And AI should sometimes give up attacking well defended targets(ground&air). There should be a difference in willingness to retreat depending on situation and airforce - while a USAAF rookie flying P-400 in the beginning of 1942 would probably retreat in next to any situation where there is no numerical superiority, an ace piloting an A6M in the IJN in 1942 would even try to attack in an unfavorable position in numerical inferiority.

4) Linked to (2) above is whether AI can be made sophisticated enough to use team tactics. It is admittedly a very difficult task, but if, for example, AI can be made adopting 'Hit and Run' tactics, forming 'Lufbery Circle' or initiating 'Thatch Weave' then it would truly take offline battles to a new level.

It's "Thach Weave". And they kinda do use that. Ever noticed how the AI tends to do their evasive turns towards their wingmans paths - so if you follow them you'll end up in front of AI's his guns? And when you are locally outnumbered, the AI that doesn't get engaged often climbs and positions itself to swoop in after you made a mistake or got dragged down by its wingmen.

[majorfailure]

In the landing pattern, Number 4 is damaged and should land first, but continues to go around leaving unnecessarily precedence to others until it crashes.
Player should be able to take over control tower and order Number 4 to be the first to land.

This should work automatically - with no involvement of the player. Damaged planes lost due to having no priority in the landing pattern are as bad for AI only flights - and player can't watch landing approach of 10+ fights.

During an escort mission, an enemy fighter appears. All your squadron’s pilots chase it, leaving alone the escorted bombers.
Player should be able to order: “Number 5 and 6 go ahead with the chase, all other return to escort position”, or click on the selected planes to obtain the same result.

Bomber escort missions are among the most disappointing as it is in this game. Either you don't get your flight to engage the enemy even if the bombers are under attack(!) - or the engage an unimportant enemy plane and don't want to let go.
And they cannot be commanded usfully when in "escort mode" You spot a group of enemy bombers heading for your carriers, and you decide to shoot them down - because a lost carrier is worse than a few lost attack planes -you are on your own. Your wingmen will not engage, they either follow you or go back to escorting the bombers.

Player has damaged and enemy bomber, and is turning for the coup de grace. Three more squadron’s AI mates run to finish it, allowing other enemy planes to bomb the target undisturbed.
Player should be able to select a target to be ignored by AI pilots.

This should be automatic, too. Planes with visible crippling damage should be less important targets. And while we are at that: AI (and especially FlaK) should try to shoot at planes with only dead pilots/shot out controls, at least if they are not experienced.

[ben_wh]

Yes, AI seems to get in trouble when fighting in a much faster less agile plane - and tries to TnB -with usually disappointing results. Though in some situations they seem to get it right, F4F(good) vs. A6M often turns out okay.

I often find some planes are particular problematic - for example, AI P-38 and P-47 do not exploit the strengths of these planes and often try to turn with opponents or follow opponents to the deck. F4F is actually a very agile planes in the sim and in reality - though during training pilots were told not to engage the Zeroes unless the Wildcats have clear numerical or tactical advantages

It's "Thach Weave". And they kinda do use that. Ever noticed how the AI tends to do their evasive turns towards their wingmans paths - so if you follow them you'll end up in front of AI's his guns? And when you are locally outnumbered, the AI that doesn't get engaged often climbs and positions itself to swoop in after you made a mistake or got dragged down by its wingmen.

I don't notice the Thach Weave as much (perhaps I should try more plane combo) although I do observe that the AI does use team tactics much better after 4.11, and human pilots will learn not to fixate on a target and watch one's six often. There has certainly been improvements here.

Regarding national difference in AI behavior - I thought about that and often wondered whether 'doctrinal behavior' can be implemented. However I can imagine the debate people will have regarding how their national AI should behavior vis-a-vis that of another country's ...

[Pursuivant]

retrat = retreat?
I think about retreat.

This is a good idea, but could be taken further.

Leaders of formations which have gotten badly scattered might well retreat, even if the planes in their formation aren't actually damaged or destroyed.

Squadron and flight leaders who are outnumbered by enemy aircraft - even if they haven't engaged - are also likely to retreat or avoid contact unless they have a clear advantage. Even then, they are likely to make a single "hit and run" attack in such as way that they can disengage before the enemy can respond.

The exception would be that badly outnumbered fighters will still aggressively take on large bomber or attack plane formations and won't be intimidated by them.

Single pilots will also be very cautious about attacking anything other than another single plane unless they have a clear advantage.

Rookies and average pilots will usually avoid contact unless they have an overwhelming advantage. Veterans and Aces might attack two or more planes when they have less of an advantage, but are more likely to try to maneuver to gain an overwhelming advantage, using sun and clouds to their advantage. Even then, they're likely to make just a single hit and run attack rather than staying around to fight.

For example, an ace pilot sees a pair of hostile fighters at 12 o'clock level and about 5 km out. Rather than flying straight in as Ace AI often does, he might try to duck into a cloud at avoid being seen. Then, he'll gain altitude and maneuver based on the sun's location and the anticipated course of his opponents, so that he's above and behind his targets and can dive out of the sun to attack them.

He will then make one fast attack to cripple or kill the trailing aircraft then disengage and analyze the situation before choosing to run away or attack the surviving enemy.

[Pursuivant]

Regarding national difference in AI behavior - I thought about that and often wondered whether 'doctrinal behavior' can be implemented. However I can imagine the debate people will have regarding how their national AI should behavior vis-a-vis that of another country's ...

There are a few places where national differences could be implemented without too much controversy. For example, in 1939-40 the RAF insisted on tight three plane fighter formations, USN doctrine from 1942 on emphasized high side attacks and lots of teamwork at the section level, and for a number of reasons Japanese pilots preferred maneuvering to BnZ tactics and were less likely to fight as a team.

On a more controversial level, RAF pilots reported that Italian fighter pilots performed more aerobatic maneuvers than Luftwaffe pilots did, and Luftwaffe pilots noted that the Soviets were much more willing to use aerial collisions as a tactic (the "Taran") - at least early in the war. Late war Japanese and German pilots noticed that some American pilots were undisciplined and aggressive - in that some were willing to break formation or otherwise take risks to "rack up a score."

Then there are well-known situations where pilots of a particular nation had good reason to behave in a certain fashion. For example, Kamikazes were known for being not very good at maneuvering, but willing to hold formation and take massive losses when any other pilot would have maneuvered defensively.

As another example, some U.S. fighter pilots reported that in 1944-45, German fighter pilots would occasionally bail out as soon as they got into a hopeless tactical situation or took damage. (This makes sense - Germany had airplanes to spare at that point, but not enough pilots to fly them, and any German pilot the USAAF encountered was probably bailing out over friendly territory.)

[ben_wh]

OK, summarizing ideas and implications so far:

1) Gunnery accuracy refinement (toning down) of rookie and regular pilots

2) More detailed engagement/ disengagement / retreat logic based not only on plane status (damage, ammo and fuel level) but also on tactical situation, for example

- Number of opponent Vs friendly
- Whether flight/section leader is lost
- Relative height to opponents
- Skill level of the AI, among others

3) Better command/communication - ability to ask wingman to check your six, for example

4) Potentially better BnZ behavior among AIs (this may need to be considered more since this is relative to the plane match-up: one plane is an energy fighter in a match-up but may be a turn fighter in another)

5) Doctrinal/national behavior by time frame - ideas: Vic formation for Commonwealth planes in 39-40; random (infrequent, occasional) kamikaze behavior for damaged Japanese planes in 44-45 - this one will needed to be teased out more as well; not sure whether AI behavior by nation is feasible / desired by players

Would love to see this refined / expanded further by others.

Cheers,

[Furio]

This is an interesting thread, with many good ideas.

In my opinion, however, we should pose ourselves a simple question: why all these good ideas aren’t already implemented? As far as I know is because they’re anything but easy, and the most difficult issue is about decisions.

An “AI” reacts according to a string of possibilities, strictly predefined. If there is any deviation from what is predefined, AI will not take any decision, or will take the wrong one. For this reason, I suggested some sort of time out and some simple tools to allow the one and only human mind in offline missions – the player – to take decisions. I understand that is not a perfect solution, but it represents a progress, perhaps in the only viable direction.

Of course, I appreciate the effort of all other people here. This thread is a sort of brainstorming session, and something useful should come out of it.

[ben_wh]

In my opinion, however, we should pose ourselves a simple question: why all these good ideas aren’t already implemented? As far as I know is because they’re anything but easy, and the most difficult issue is about decisions.


AI coding can be difficult no doubt. The challenge is not only to have competent AI but also believable AIs with behavior that feels human for the player. Still, the objective here is to stimulate new ideas - the hope is that (i) perhaps DT can find a smart way to code some of the suggested behavior mentioned, or (ii) some other current / future sim developer will come across this and influence their future work.

... For this reason, I suggested some sort of time out and some simple tools to allow the one and only human mind in offline missions – the player – to take decisions. I understand that is not a perfect solution, but it represents a progress, perhaps in the only viable direction.


Interesting - can you elaborate what you mean by this and how this can be accomplished in-game? Via the command menu? (e.g. 'Attack my Target!', 'Section -> Tactics -> Hit and Run', 'Section -> Tactics -> 'Form Lufbery Circle!'?) Would like to learn more.

Regardless, I personally liked the expanded 'Drop Bombs on my Command' and related options from 4.11. It provides the human player with more options and control without breaking immersion. More control to the flight lead on flight behavior/tactics would be welcomed.

Cheers,

[Furio]

Interesting - can you elaborate what you mean by this and how this can be accomplished in-game?

First of all, I’m not an expert in AI coding, so my suggestions can be only generic. That said I would start from the simplest problem (or from the one that looks simplest to me): the damaged plane in the landing pattern. The rationale of the solution is simple: when the situation is so complex that AI pilots and control tower are unable to take the right decision, the player should take over. To say it with more details:
The player sees a plane smoking being sent around by control tower, or the player’s plane is smoking, or leaking fuel, and is sent around.
The player hits a key meaning: “Taking Over Mode”.
He hits a second key meaning: “Controller”. He assumes the controller role, the same way as shifting through crew positions.
He hits a third key meaning: “This Plane Must Land First.”
He hits the call number of the plane.
He hits again the first key, returning to his pilot’s position.
The controller’s voice imparts the correct orders, and AI planes shift positions allowing the correct landing sequence.
It’s perhaps possible to make the whole thing even simpler, but the only real issue I see is… numbering of the planes, when there are several planes with the same call number. A smart and realistic numbering would be of help in many other situations, but is outside the scope of this post.
Another situation is absurd concentration of AI planes on the same target. Again, we can think of an appropriate list of commands.
First key: “Taking Over Mode”.
Second key: AI pilots behaviour.
Call numbers of AI controlled planes’ affected by order.
Third key: order: “Ignore your present target”.
First key again: exiting “taking over mode” and returning to your pilot’s position.

I say it again: I’m not expert and I’m using just common sense, but the whole thing looks reasonable to me, and can be reduced to just one concept: player takes decisions.

[sniperton]

As I see the main problem now is that issuing orders in real-time in any other way than voice communication is handicapped due to the fact that we have only two hands (already busy with flying and doing all the 'engeneering work'). :rolleyes:

[Furio]

As I see the main problem now is that issuing orders in real-time in any other way than voice communication is handicapped due to the fact that we have only two hands (already busy with flying and doing all the 'engeneering work'). :rolleyes:

In my opinion, that’s a minor problem: it can work the same way as shifting crew position. Surely the player can briefly leave his plane to autopilot. Or, the whole thing can be done in a timeout condition (sort of pause, but with the game still running).

I understand that some aspects of my idea are detrimental to immersion, but some AI behaviour can be even worse. The choice is up to the player, as always, and I would surely prefer the ability to take decisions to the frustration of helplessly watching a good mission going nuts.

[sniperton]

Well, I frequently fly autopilot when off-combat, but always manually when enemies are nearby. I frequently pause the game and take a coffee break when off-combat, but never when I'm actually fighting. You propose just the opposite, dont't you?

[Furio]

If I understand correctly the question: yes, but just briefly, while there’s usually not fighting during landings. As for me, I’d be glad to have the option described above.

[sniperton]

Oh, I see. It could be implemented for landings, but all this wouldn't resolve the other issue: how to effectively command AI teammates during fight. And as for landings, it might generate new problems: when you switch to autopilot, it would not necessarily stick to the route/altitude/direction you were flying in, that is, there's a chance that YOUR AI would change direction etc. while you're busy with your orders.

[Furio]

I don’t want to defend my idea at all cost. It’s just that, an idea for just an option.
In my opinion, it would need some sort of time-out with the game still but not paused.

Of course, I understand other can have different tastes. The core of my proposal is simple: I don’t believe we can expect true decision-making ability for AI, at least with present day technology. I think the only viable alternative is to entrust decision-making to human beings – offline players.

In the end, I don’t want to steal this thread, and I’m stopping here, for the moment.

[sniperton]

Taking control over from the tower during landings is a great idea IMO, and can be done real-time. I was just arguing against automatically switching to autopilot. :grin: Completely pausing the game while giving orders is a bit against my taste (it's a sim, anyway, not a Total War battle), but the option is still there (I have a difficulty option in mind) to enable commands while the game is paused (like in a Total War battle). ;)

[majorfailure]

The idea has its merits, but I'm not convinced because:
-How will you handle multiple flights (10+) approaching their respective different bases -you can't be everywhere?
-Online it will be difficult, making a break is akward at best.
-I believe such a task as setting landing priorities for (damaged) planes can very well be done by AI -there are not to many variables there IMHO -and it could be done with a decision table like that:
-Highest priority: Plane on fire, Pilot bleeding.
-Second: Plane already out of fuel or engine dead.
-Third: Plane with heavily damaged engine or running out of fuel in very short time(e.g.<120sec).
-Fourth: Plane with any other engine damage, plane with injured pilot, plane with fuel low, but enough for say 5 minutes.
-Fifth: Plane with any other damage.
-Sixth: Plane in undamaged condition.

[ben_wh]

The 'Take-over Mode' is an interesting idea, although we may have to think about implementation.

On one hand, the Take-over Mode would enable much more precise tactical control at individual plane level.

One the other hand, this may necessitate going to the external view (or map view), jumping to different planes in your flight and assess their status regularly. For example, using the landing priority example, the human player would have to know which plane in the flight is heading close to base, ready for landing, and which is not. If the flight is long distance apart, for example, result of a plane disengaging and returning home early - the human player may have to manage its landing priority while shepherding the rest of the flight during combat. But it is an intriguing idea that has a lot potential ...

Any additional thought on potential AI improvement?

How about ship AI? Simple, predefined capital ship evasive maneuver when under attack for example? Would it be feasible/desirable?

http://www.nasflmuseum.com/uploads/4/9/5/8/4958573/7185130_orig.jpg

Cheers,

[Pursuivant]

-How will you handle multiple flights (10+) approaching their respective different bases -you can't be everywhere?

Here's one idea:

A) Map makers and mission builders could define fixed camera views at each airfield as "Ground Control" positions. One Ground Control view per airfield.

B) Pressing some key allows you to cycle through the different ground control views.

c) Pressing Ctrl-C (or some other bound key) has the same effect as switching to a different crew position, in that the previous crew position you occupied goes to "autopilot", but rather than switching to a different position within your plane, you instead switch to the ground control "crew" you're currently monitoring.

Alternately, there can be keys bound which allow you to instantly move from an air crew position to a ground control position.


-Online it will be difficult, making a break is akward at best.

Yes, but in that case you can have the option of assigning a player to just play ground control for some side. This could also include options like allowing a player to be a Forward Air Controller who can shoot off flares, lay marker panels and pop smoke grenades to mark targets, or a operations officer in charge of vectoring flights of friendly aircraft towards unidentified radar contacts.


-I believe such a task as setting landing priorities for (damaged) planes can very well be done by AI -there are not to many variables there IMHO -and it could be done with a decision table like that:
-Highest priority: Plane on fire, Pilot bleeding.
-Second: Plane already out of fuel or engine dead.
-Third: Plane with heavily damaged engine or running out of fuel in very short time(e.g.<120sec).
-Fourth: Plane with any other engine damage, plane with injured pilot, plane with fuel low, but enough for say 5 minutes.
-Fifth: Plane with any other damage.
-Sixth: Plane in undamaged condition.

Additionally, planes which fall into the first four categories should divert to the nearest airfield and shouldn't bother circling prior to final approach, but should go right in to land.

Realistically, planes with damage that doesn't require them to land immediately, but which does make it more likely that they'll crash should either be diverted to a different runway or should land last so they don't risk delaying landing for the other planes using the airfield.

Badly shot up carrier aircraft shouldn't even try to land. Instead, planes which are on fire or are very damaged should attempt to ditch close to a friendly ship.

Finally, crew aboard planes which are very likely to crash on landing (e.g., landing gear inoperable) should attempt to bail out at a safe altitude over a friendly airfield or next to a friendly ship leaving just the pilot and co-pilot to try to land the plane.

[majorfailure]

How about ship AI? Simple, predefined capital ship evasive maneuver when under attack for example? Would it be feasible/desirable?
Cheers,
Desireable, Yes, highly.
Feasible, maybe. It could make matters worse, if then regularly ships crash into another when evading.

Additionally, planes which fall into the first four categories should divert to the nearest airfield and shouldn't bother circling prior to final approach, but should go right in to land.

Realistically, planes with damage that doesn't require them to land immediately, but which does make it more likely that they'll crash should either be diverted to a different runway or should land last so they don't risk delaying landing for the other planes using the airfield.

Badly shot up carrier aircraft shouldn't even try to land. Instead, planes which are on fire or are very damaged should attempt to ditch close to a friendly ship.

Finally, crew aboard planes which are very likely to crash on landing (e.g., landing gear inoperable) should attempt to bail out at a safe altitude over a friendly airfield or next to a friendly ship leaving just the pilot and co-pilot to try to land the plane.
I just made a short list as a general outline, if used it could and should be refined and tuned to cover next to any possibility. And additionaly there could be a random element added, something like stubborn-not-listening-to-you aircontroller, if desired.

And even giving any damaged plane a higher priority than an undamaged plane would be a -though small- improvment over status quo.

[sniperton]

... Additionally, planes should ... divert to the nearest airfield and shouldn't bother circling prior to final approach, but should go right in to land.

BTW, does the AI positively 'know' where (friendly or enemy) airfields are located (other than those predefined by its own mission waypoints)?

[Aviar]

BTW, does the AI positively 'know' where (friendly or enemy) airfields are located (other than those predefined by its own mission waypoints)?

Simply put, AI planes are not 'aware' of airfields, be they enemy, friendly or neutral.

When taking off or landing at an airfield, they only do so because it is a pre-defined waypoint. For instance, a damaged AI plane cannot 'seek out' a nearby friendly airfield at which to land. The AI is not programmed for that kind of 'thinking'.


Aviar

[Pursuivant]

Simply put, AI planes are not 'aware' of airfields, be they enemy, friendly or neutral.

That explains a lot about AI landing and critical damage behavior.

Would it be possible for the AI to "check in" with the appropriate map file ever minute or so to determine the closest airbases to the plane's current position?

Then, if the plane needs to land in a hurry, it can be programmed to fly in a direct line to the nearest airbase.

Additionally, runways for stock airfields could be given some sort of coordinates as to exactly where they start and end.

That way, when a plane needs to make an emergency landing, rather than following programmed waypoints, it would just maneuver so that it's on a heading and altitude to land at one end of the nearest airfield.

Obviously, there would need to be a bit more complexity, with things like collision avoidance routines and determining whether an airfield is friendly or hostile, to make fully automated landings work realistically, but it might be possible given the way that IL2 works.

[ben_wh]

AI emergency landing has been a problem fro day one.

1) For the landing back at a friendly field issue, perhaps, as Pursuivant mentioned, having airfield marker one can assign side to (Red Vs Blue) would allow a plane to look for friendly airfield close by to land. May still need substantial coding to achieve such AI improvement though.

2) Another aspect of AI landing is even more complex and potentially challenging to code: when a plane is in trouble and there is no airfield nearby, the AI has a choice to make - where to crash land. Right now AI plane often chooses to fly into mountain or landing into a forest - i.e. ensuring total destruction of its plane (and pilot) - rather than chancing it by picking a relatively flat field to land. This is something very natural for a human pilot but I imagine can be very hard to code into AI behavior.

Cheers,

[sniperton]

A more simple (and feasible) solution would be an 'RTB' command combined with an "Ignore Waypoints' command. Then the AI could head directly for the pre-defined home base coordinates. In addition, the AI could even check whether there are other friendly airfields in the vicinity (I mean the take-off coordinates for other friendly flights). Remember, the AI can only be aware of locations which are pre-defined by the mission designer. If the mission designer decides to leave a 'physical' airfield to remain inactive, then it would be strange if the AI could land on it. Similarly, if the mission designer created a new airfield (not on the physical map), then it would be strange if the AI ignored it.

[Pursuivant]

A more simple (and feasible) solution would be an 'RTB' command combined with an "Ignore Waypoints' command.

This would be a nice option for the FMB, either on its own or included with my idea.

If added to my idea, it would complement the idea of having the AI being able to "see" airfields, and would allow mission builders to create missions where friendly aircraft would divert around otherwise friendly airfields (to simulate things like the airfield being fogged in, bombed out, or otherwise unusable). All they'd need to do is mark a particular airfield as being "invisible" to friendly planes.

Likewise, the mission builder could make certain airfields "secret" by flagging them as being "invisible to hostile aircraft, or flag certain airfields as being unavailable to either side by making them "invisible" to all planes.

Or, if you wanted to keep it simple, any mission could automatically have AI planes which must RTP immediately try to go to their landing waypoint, or their take-off waypoint if it is different and closer. That way, all the AI would have to do is refer back to one of two waypoints, rather than having to check the map and do calculations to determine what the closest airfield is, or requiring the mission builder to determine what the plane's "home airfield" is.

[Pursuivant]

2) Another aspect of AI landing is even more complex and potentially challenging to code: when a plane is in trouble and there is no airfield nearby, the AI has a choice to make - where to crash land.

Yes. Also, AI has a very tough time making realistic emergency decisions that would come naturally for a human pilot due to its limited terrain recognition ability. For example, AI planes might have the bail out just inside hostile territory rather than attempting to hold on a bit longer to reach friendly territory. Or, AI planes might have the crew bail out over the ocean when attempting to ditch is preferable.

"Intelligent" emergency decisions would require AI planes to be able to "see" and recognize areas of ocean and open flat ground which are suitable for crash landing/ditching, as well as being able to "see" front markers.

Once that's done, the flow chart as to what to do is fairly simple.

[Pfeil]

All the AI would have to do is refer back to one of two waypoints, rather than having to check the map and do calculations to determine what the closest airfield is, or requiring the mission builder to determine what the plane's "home airfield" is.

From the mission builders perspective, sure.

Though the AI always knows its exact position. Taking the coordinates for a list of airfields(which wouldn't be that long, not over 100 at least) and comparing them to see which is the shortest distance from any given AI's location(or even a whole mission full of AI) is hardly a CPU intensive calculation.

In fact, such a calculation already exists: Runway lighting.
This already determines the nearest friendly base(Try it, it only works in close proximity to airfields, otherwise they're permanently lit).
There is obviously a way to "detect" airfields, so it's quite possible it could be adapted for use by the AI.

[6BL Bird-Dog]

Would it be possible to alter the default behaviour of formations of aircraft when a mission is started to match that as set in the mission builder .For example I am working on a missions for the Solomon islands and have a Squad of B-17 air start in 2 v formations .By Default the game always spawn in echelon right formation so the two second flights from each v have to drift across into echelon left .The Ai seem to do it ok after trial and error in the FMB but for a large group of players this leads to all kinds of confusion unless a detailed explanation is given in the brief and those joining actually read it :) . It would be far easier to write the brief also .
eg .Mission Air start at 6000ft: Heading 270deg :indicated airspeed 160mph :
Starting positions follow...
B-17 5th Bombardment Group 70% Fuel 20x250lb Bombs
flight 1 echelon right in starboard section of lead v:
fligh2 echelon left in port section of lead v:
flight3 echelon right in starboard section of v :
fligh4 echelon left in port section of trail v.
Target etc........ Bird

[major.kudo]

AI emergency landing has been a problem fro day one.
For the landing back at a friendly field issue, perhaps, as Pursuivant mentioned, having airfield marker one can assign side to (Red Vs Blue) would allow a plane to look for friendly airfield close by to land. May still need substantial coding to achieve such AI improvement though.

If AI Decision for landing airfield is range within enemy artilleries, what does AI do?

-

another one.
http://gachopin.no-ip.info/kudo/Bombers_formation.jpg
I have continued thinking that "this is a little strange".

[Furio]

If AI Decision for landing airfield is range within enemy artilleries, what does AI do?


This is just one of the possible questions. For what I know (I can be mistaken), AIs don’t really take decisions. They react to a list of conditions with pre-set actions.

If the condition is 1, action is A, if condition is 2, action is B, and so on.

Some of the proposals require a rather more complex listing. Something like: if condition here is 1, condition of the plane is 7, condition of the pilot is 14, condition of the nearest friendly field is 2, condition of enemy planes is 16, condition of artillery is 5, then the AI’s action is C, followed by E, followed by B. Change any of the conditions (numbers) and you’ll change the actions (letters).

This is just an example, of course, and I would be glad to be corrected by someone more knowledgeable than me.

If I’m more or less right, this thread is precious, because it shows us how much complexity is required in AIs “decisions” and subsequent actions to obtain a realistic behaviour.

[Pursuivant]

If AI Decision for landing airfield is range within enemy artilleries, what does AI do?

The same thing applies if an airfield is under attack by enemy planes.

Historically, if they had to, planes would try to land in spite of the bombardment. Otherwise, they'd divert to another airfield or wait until the bombardment was over.

AI planes could be warned to divert or delay landing if the game registers damage to objects on or near the airfield where they were going to land.

Damaged AI planes would still go straight in for landing, regardless of whether the airfield was under attack.


http://gachopin.no-ip.info/kudo/Bombers_formation.jpg

This is another excellent idea. Currrently, if a lead bomber in formation is damaged, the rest of the planes in the formation will follow it until it is destroyed or the crew bails out. In a few cases, I've seen formations of bombers follow a mortally wounded leader almost down to the ground.

Realistically, if the lead bomber can't hold speed or altitude, it should drop out of formation and leadership of the remaining planes in the formation should pass to the #2 plane in the formation.

Players should also have the option of passing off command of a formation to an AI plane, or taking over command of a formation if they have sufficient rank.

[yak9utpro]

that bomers formation change would be very great. And i think easy or at least not so hard.

[ben_wh]

Hi all,

This is another attempt to summarize the ideas so far (in simplified form):

Combat AI Behavior

1) Gunnery accuracy refinement (toning down) of rookie and regular pilots

2) More detailed engagement/ disengagement / retreat logic based not only on plane status (damage, ammo and fuel level) but also on tactical situation, for example

- Number of opponent Vs friendly
- Whether flight/section leader is lost
- Relative height to opponents
- Skill level of the AI, among others

3) Potentially better BnZ behavior among AIs (This issue can sometimes be seen in set-up such as Ace P-38 AI Vs. Regular A6M Zero AI. This may need to be considered more since this is relative to the plane match-up: one plane is an energy fighter in a match-up but may be a turn fighter in another.)

In-flight Behavior

4) Emergency Landing: potential for AI plane to automatically (or on command of the human player) divert to airfield marked as 'friendly' (e.g. same colour - red or blue - as the AI plane)

5) Potentially to implement routine to let AI recognize distance to different marked landing sites (e.g. take-off way point and landing way point) and make routing decision based on distance to site and its own condition (e.g. damage, fuel status)

6) AI flight leader can yield command if severely damaged so that the next-plane-in-command (AI wingman or section lead) can take over, to avoid the whole AI flight getting 'drag down' by a damaged leader


Communication with AI

7) Better command/communication - ability to ask wingman to check your six, for example

8 ) 'Return to Base' command to individual plane which will ignore any preceding way point to get home

9) Potentially for human player to 'take control' over from the airfield tower to issue command and assign landing priority to different AI planes


Other AI Suggestions:

10) Doctrinal/national behavior by time frame - ideas: Vic formation for Commonwealth planes in 39-40; random (infrequent, occasional) kamikaze behavior for damaged Japanese planes in 44-45 - this one will needed to be teased out more as well; not sure whether AI behavior by nation is feasible / desired by players

11) Potential simple evasive maneuver (with land avoidance) routine for capital ships when under attack

Please feel free to fill in any gap you see and more suggestion welcomed.

Cheers,

[Pursuivant]

Good summary of the discussion.


4) Emergency Landing: potential for AI plane to automatically (or on command of the human player) divert to airfield marked as 'friendly' (e.g. same colour - red or blue - as the AI plane

4.1) "Smarter" AI behavior for mortally-wounded planes, such that planes that are on fire or which can't maintain altitude after jettisoning ordinance will immediately try to land, crash land or ditch - in friendly territory if possible. If it isn't possible to "land" the plane (i.e., flying over mountains or woods), crew should try to bail out at 500-1,000 meters above ground level. For maximum historical realism, planes that ditch should attempt to do so near friendly ships or land held by friendly forces.



7) Better command/communication - ability to ask wingman to check your six, for example

7.1. Ability for player to call out the presence of enemy aircraft to AI without having to use the padlock enemy command. Also the ability for a player to warn specific AI planes about enemy aircraft coming at them from a particular vector (e.g., "Red 3, Check your 4 o'clock low).


10) Doctrinal/national behavior by time frame - ideas: Vic formation for Commonwealth planes in 39-40; random (infrequent, occasional) kamikaze behavior for damaged Japanese planes in 44-45 - this one will needed to be teased out more as well; not sure whether AI behavior by nation is feasible / desired by players

"Weaver/Tail-end Charlie" vic or echelon formations for Western allies 1941-42.

"Combat line" formation for USAAF planes in 44-45, and "section" tactics for USN/USMC planes from 42-45.

Lack of formation for Soviet attack aircraft 1941-43 (they usually flew in irregular "gaggles").

Soviet fighters in 41-42 should occasionally use "taran" attacks (i.e., controlled collisions) if they're out of ammo.

Japanese planes in 1944-45 will very occasionally deliberately collide with heavy bombers.

Mortally wounded Japanese planes in 1944-45 WILL attempt kamikaze attacks against enemy ships.

Pilots of any nationality flying mortally wounded aircraft will occasionally make kamikaze attacks against enemy capital ships.


11) Potential simple evasive maneuver (with land avoidance) routine for capital ships when under attack

Option for mission builders to make ships automatically travel in formation, keeping station with each other. Option for mission builders for ships (on their own or in formation) randomly zig-zag to simulate historical anti-submarine tactics. Carriers will always turn into the wind and travel in a straight line when retrieving planes and will always travel at top speed in a straight line and in line with the wind when launching planes.

[majorfailure]

1) Gunnery accuracy refinement (toning down) of rookie and regular pilots

Yes for rookie. No for regular. They are pretty bad shots already, and are usually unable to get more than two kills/mission even in a target-rich environment.
-Different solution: generate a level below rookie, and make them bad at everything, real bad-representing cannon fodder thrown at the enemy with zero experience and next to no training.

Consider tuning down torpedo hitting ability of rookie and regulars - it may be only due to my limited statistics, but I don't see any evidence that they hit much worse than veteran and ace - even on the contrary.

2) More detailed engagement/ disengagement / retreat logic based not only on plane status (damage, ammo and fuel level) but also on tactical situation, for example

- Number of opponent Vs friendly
- Whether flight/section leader is lost
- Relative height to opponents
- Skill level of the AI, among others

I would be glad if AI with smoking engines, that are not dead yet, or shot off or out controls or ... would try to get home ASAP for now. Even AI that is hit, without any apparent damage should quit sometimes, real life pilots did that for fear of invisible damage sometimes.

3) Potentially better BnZ behavior among AIs (This issue can sometimes be seen in set-up such as Ace P-38 AI Vs. Regular A6M Zero AI. This may need to be considered more since this is relative to the plane match-up: one plane is an energy fighter in a match-up but may be a turn fighter in another.)
I don't know if possible, but data like speed of their and enemy's plane, turn time could be made available, with a guessing factor, the worse AI, the more off -representing intimate knowledge of their and enemys plane of Ace.


Communication with AI

Make it possible to specify WHERE the target is - using pilots "o'clock" system, where directly ahead is 12 and so on. Or simpler, use four quadrants. AI already must know where the enemy is relative to others of your flight, as they do warn you and others with "Bandits on your XX o'clock.

11) Potential simple evasive maneuver (with land avoidance) routine for capital ships when under attack

Yes, if possible -but if the cost is losing ships due to collisions, then at least make AI's torpedoing abilities versus groups of ships worse.

[Jami]

I would be glad if AI with smoking engines, that are not dead yet, or shot off or out controls or ... would try to get home ASAP for now. Even AI that is hit, without any apparent damage should quit sometimes, real life pilots did that for fear of invisible damage sometimes.

I couldn't agree more. Sometimes those smoking AIs fly and fight like nothing had happened with their planes. It is understandable that when you chase a smoking enemy it fights back, but if you try to leave it and head to home it may start chasing you. I think that most of the human pilots wouldn't do that.

[sniperton]

As to ships, why not to implement what 3rd party mission generators like DCG already do: to combine them into 'shippacks' or 'convoys'? The concept is here:

A shippack is an abstract naval unit, a fixed formation with slots for individual ships. Each slot represents a relative position to the formation centre (a vector with distance). Once a ship is assigned to a slot, it becomes a subordinated element of the formation, and from then on, the mission designer can adjust the travelling speed and direction of the whole pack all at once. Direction changes can be twofold:

1) individual elements turn, but the formation's heading remains the same (good for zigzagging or for turning into the wind while starting aircraft);
2) elements turn together with the formation (allows only minor direction changes as the 'outer' ships has to travel more/faster, while 'inner' ships less/slower).

Max speed is of course equal to the slowest element's speed, max turning rate ditto.

Ship formation coding on this basic level is pure mathematics, and it seems to be more easy than to 'teach' individual ships to behave more prudently when in company ;)

[yak9utpro]

ok AI require heavy fixes.especialy when some (friendly or enemy) AI had began the landing procces they will not pay attention to anything even if their oponent is directed at their 6 in low distance.This causes another error when you ask them to help they respond ( here we go attacking bandits ) but nothing.I died so many times in this way. (I think one day i will begin hunting those deserters). another nice idea is if TD creates charachters for AI and create deserters is that other friendly AI will begin hunting them.

[ben_wh]

... it seems to be more easy than to 'teach' individual ... to behave more prudently when in company ;)

Probably true - I'm trying to learn that myself still ... :)

Agree that ship AI doesn't need to be very complicated; even an illusion of evasive maneuver (designed with collision avoidance in mind) would be welcomed.

... especialy when some (friendly or enemy) AI had began the landing procces they will not pay attention to anything even if their oponent is directed at their 6 in low distance

Experienced this as well - actually if the mission designer does not built in a lot of AAA to 'discourage' intruders you can follow a disengaged enemy flight back to their base in some relatively compact maps (short distance between enemy base and yours) and shoot them down relatively easily if you wait till they get into landing pattern.

Also tried some more match up recently - AI certainly do not fly to the advantage of the BnZ planes, making some rookie and regular AI turn fighters, the Zero for example, more dangerous than they should be.

Cheers,

[majorfailure]

ok AI require heavy fixes.especialy when some (friendly or enemy) AI had began the landing procces they will not pay attention to anything even if their oponent is directed at their 6 in low distance.This causes another error when you ask them to help they respond ( here we go attacking bandits ) but nothing.I died so many times in this way. (I think one day i will begin hunting those deserters). another nice idea is if TD creates charachters for AI and create deserters is that other friendly AI will begin hunting them.
This is not entirely true. Sometimes they behave like that, but there were some changes, I have see it by now a lot of times, that an attacked landing plane does floor the throttle and pulls up gear and figths back -or tries. And in former days the AI stuck to the landing circle once they entered it, and no earthquake or hurricane could change that. It was a funny day in the AIs home when I approached their field with their landing planes after one of the last patches and was ready for some easy prey, and got in turn my ass handed to me. Sometimes they still revert to that former catatonic state, and maybe that is realistic, a pilot fully concentrated on his approach and not aware of enemy planes present may have reacted the same.

[Pursuivant]

Agree that ship AI doesn't need to be very complicated; even an illusion of evasive maneuver (designed with collision avoidance in mind) would be welcomed.

One way to do this would be to allow mission builders to "copy and offset" waypoints in the FMB, and to assign them to different vehicles.

For example, say that you want to get a line of ships to zigzag in line abreast formation. You'd set the speed and zig-zagging waypoints for the first ship. Then you'd tell the FMB to copy that series of waypoints a certain number of times, offsetting each waypoint by a set direction and distance.

For example, for a formation of 5 ships, you'd tell the FMB to offset waypoints vertically (say north and south if the central ship is traveling east or west) by 300 m, centered on the first ship, with 2 iterations. That would give you a formation of 5 ships - one in the center, one offset by 300 m north, one at 600 m north, one at 300 m south, and one at 600 m south.

You could then edit each ship to make it different.

Alternately, or additionally, the FMB could allow the mission designer to include "station keeping" where certain vehicles are automatically linked in formation and are programmed to move identically to the "central" vehicle in the formation as best they can.

If the central vehicle is destroyed, or if a following vehicle cannot keep station for some reason, it automatically acquires the waypoints of the central vehicle, and attempts to follow those waypoints on its own.

This is already automatic, to some degree, for aircraft, but could be applied to any vehicle in the game. And, since simple "station keeping" routines in the FMB aren't true AI (with collision avoidance and enemy detection) they might not be that hard to program.

As a third idea, mission builders could be provided with "stock" sets of waypoints, such as zig-zagging, weaving, circling or altitude change which would modify the usual straight line path between waypoints.

For example, when the mission builder lays out waypoints, the FMB could give him the option for "zig-zag" with a set or "random within a given range" distance between zig-zags, and with a set or "random within a range" distance for each "leg" of the vehicle's path.

The mission builder would place his first waypoint normally, then "draw a line" to the second waypoint, but the FMB would interpret that as a whole series of zig-zagging waypoints and place them appropriately. That way, the mission builder just needs to place two waypoints and the FMB does the rest.

Pre-programmed movement using this option could be:

Zig-zag/weave - alternating 5 to 85 degree left and right turns with equidistant "legs" to each "zig" or "zag" and an offsetting turn of the same angle so that the vehicle consistently crosses its baseline path at the same interval of time or distance traveled.

Random zig-zag/weave - random 5 to 90 degree left and right turns that eventually get the vehicle to "point B".

Rectangular search pattern/overlapping squares/rectangles - The vehicle travels 2x meters, then turns 90 degrees left or right and travels y meters, it then turns in the 90 degrees in the same direction as it did before and travels for x meters in the opposite direction from its baseline course, it then turns again and travels back y meters, until it reaches its baseline course and turns 90 degrees again, at which the cycle repeats.

Circular search pattern/far escort/loiter - The vehicle travels 2n meters, then makes a circular turn to the left or right with a diameter of n meters.

Oval search/fare escort/loiter/"Race track" - As above, but the vehicle flies a series of overlapping ovals.

Corkscrew - aircraft makes a diving turn 5-15 degrees right or left and loses n/2 meters of altitude moving left or right until it until it reaches a point equal to (altitude to be lost/2). At that point, it reverses its turn and loses another n/2 meters of altitude until it reaches its baseline course and bottom baseline altitude. It then climbs and diverts from its baseline course until it regains (altitude lost/2) and reaches and equivalent point from its baseline course. It then reverses its turn and regains its baseline altitude and course at which point the cycle repeats.

Regular altitude change - As for zig-zag, but aircraft regularly gains and loses altitude within a set range.

Random altitude change - As for random zig-zag, but aircraft randomly gains and loses altitude within a set range. Always keeping a baseline distance above the ground.

Terrain hugging - Aircraft always adjusts altitude to remain X number of feet above the ground.

[ben_wh]

Question - does the ship AI today capable of path-finding - i.e. avoiding obstacles/hazards automatically?

If so then there are more options to be explored - if not, the third option:

"As a third idea, mission builders could be provided with "stock" sets of waypoints, such as zig-zagging, weaving, circling or altitude change which would modify the usual straight line path between waypoints."

is probably the most feasible. The first option seems to be a lot of work for the mission builder if there is a fair number of ships in the formation.

Cheers,

[Aviar]

Question - does the ship AI today capable of path-finding - i.e. avoiding obstacles/hazards automatically?

AI ships have no 'path-finding' abilities. They are only programmed to follow their waypoints. They will collide with anything that gets in their way.

Aviar

[sniperton]

AI ships have no 'path-finding' abilities. They are only programmed to follow their waypoints. They will collide with anything that gets in their way.

Aviar

My humble proposition is to take into consideration my humble proposition as in #60.

[Pursuivant]

My humble proposition is to take into consideration my humble proposition as in #60.

The idea of having "ship packs" is a good one, and would save mission builders a lot of time since the FMB could be loaded with preset historical packs like "1942 Western Allied Convoy" or "Midway, June 4, Japanese Carrier Task Force".

The same idea could be applied to aircraft and ground vehicles as well, with formations such as "1944 U.S. Heavy Bomber Group Box Formation" or "1942 Soviet T-34 Tank Formation"

If this is implemented, another nice feature might be for the FMB to allow mission builders to create and save custom formations. That would allow TD to just include a few generic formations (e.g., "German destroyers line abreast") and let the fans create actual historical formations.

[majorfailure]

The idea of having "ship packs" is a good one, and would save mission builders a lot of time since the FMB could be loaded with preset historical packs like "1942 Western Allied Convoy" or "Midway, June 4, Japanese Carrier Task Force".

The same idea could be applied to aircraft and ground vehicles as well, with formations such as "1944 U.S. Heavy Bomber Group Box Formation" or "1942 Soviet T-34 Tank Formation"

If this is implemented, another nice feature might be for the FMB to allow mission builders to create and save custom formations. That would allow TD to just include a few generic formations (e.g., "German destroyers line abreast") and let the fans create actual historical formations.
The "fun" of this starts when your formations need to turn. Without collision avoidance you will need a preprogrammed turning pattern for each formation. For player-created formations, player would need to create a pattern, too. There would be more than one mission designer willing to create and test such formation patterns I assume, so if created make them user-accessible, and let us experiment with it.

[sniperton]

The "fun" of this starts when your formations need to turn. Without collision avoidance you will need a preprogrammed turning pattern for each formation. For player-created formations, player would need to create a pattern, too. There would be more than one mission designer willing to create and test such formation patterns I assume, so if created make them user-accessible, and let us experiment with it.

The first step would be to create pre-defined formations with fixed positions and pre-defined turning patterns. Hard-coded. If they work flawlessly, then the customization can begin, but not before. Say, you create a hard-coded formation like

_1_2
3_4_5
_6_7

Numbers here represent slots. The user can decide whether he/she fills a slot with whatever ship. If the turning pattern is defined in a parametric way (according to the slowest and worst turning ship in the pack), then it's pure mathematics to calculate each element's route (avoiding collisions).

[Pursuivant]

The "fun" of this starts when your formations need to turn.

True. But the rules for collision avoidance during a turn are simple - follow the leader and maintain the same relative distance and "angle off" with respect to all the vehicles around you as you turn.

Certainly the calculations required for such changes of direction already exist since those algorithms are commonly used to simulate the behavior of groups of birds, fish and herd animals.

I'd also consider it to be realistic if a sharp turn at high speeds results in collisions, or vehicles skidding, stalling, tipping over or temporarily losing formation.

If IL2 can't handle the sort of math required for perfect station-keeping and collision avoidance, parameters could be built into the formation to limit the maximum speed and direction at which the formation can travel to make collisions or unrealistic behavior impossible. Again, this is realistic, since groups of disparate vehicles in convoys have to travel at a bit less than the highest sustainable speed of the slowest vehicle in the formation and formation leaders need to take the limitations of the least maneuverable vehicle in the group into account when making maneuvers.