[qwadtep]

Make the AI run away when under heavy fire.

Thats it.

One of the single biggest pains in the ass in BPmulti is trying to chase down a Kent or Atalanta that doesn't want to be hit and is actively focused on getting away. Eventually the strategy was formed that if you want to kill it, you have to leave it alone for a while, turn your back, and wait for it to come to you. The fact that most AI attempt to out turn you and attack you instead of getting away is one of the reasosn fights are so quick.

That would mean: Low courage, high get away chance.

Ironically, the most cowardly enemy is the one which is most difficult to kill...

And then you realize you're 2km away from the ship you're escorting and a bomber wing just jumped in on the other side. Clever bastards.

[Nuke]

when you consider that a fighter that is chasing another fighter is no threat to your warships or bombers, then its a damn valid and effective tactic. it was one employed a lot by sf dragons in fs1, iirc.

[BengalTiger]

Create AI profiles not only for broad definitions of the ships they pilot (fighter, bomber, cruiser, etc.), but make variations for each type.

Interceptors should use burners to get away from enemy fire as often as possible- a Loki needs to move around a lot to survive.
A heavy fighter like an Erynies may be more successful trying to shoot the enemy down in the first pass before the engagement changes into a turning fight, even though this requires getting hit a few times.
Light bombers should disengage their capship target and assist their escorts in dealing with enemy fighters, but an Ursa might be better off launching its bombs under heavy fire first- it won't succeed in any form of dogfighting, unless the enemy is a cruiser.

Have fun in the next couple weeks, staging dogfights in FRED and watching the AI duke it out to fine tune the settings... 
Adding and balancing new species is even more work if each one has a different ship building philosophy (and sometimes technological level)...

[Fury]

AI profiles are on per mission basis, you cannot have multiple AI profiles in any one mission. AI classes on the other hand can be used in significant numbers, but IIRC there was a limit in number of AI classes as well. Doing AI classes for different ship types is doomed in insanity because number of AI classes grows so large it'll be next to impossible to maintain.

Creating 5 AI classes for light, medium and heavy fighters and bombers would already grow to 30 AI classes, plus 5 for capships. That's just stupid.

[BengalTiger]

AI profiles are on per mission basis, you cannot have multiple AI profiles in any one mission.

I was thinking of AI class, sorry.

Also, from ai.tbl:
; We have a virtually unlimited number of AI classes.  If you add a new AI class and
; you get an execution error about too many classes, increase MAX_AI_CLASSES, defined
; in ai.h.  On Jan. 30, 1997, MAX_AI_CLASSES = 50.

Gives enough room for the 35 required, and creating them will take just a few minutes. More time would be required to tune the values, but this is optional...

More in a few hours when I get this stuff done, I'll post a report on how custom AI for each ship class works and how it compares to the stock classes.

[Fury]

More in a few hours when I get this stuff done, I'll post a report on how custom AI for each ship class works and how it compares to the stock classes.

[BengalTiger]

So after getting light fighters working and tested...

There is no significant difference between stock AI and fighters with higher evasion and really high or really low courage. While the pimped fighters did maneuver a bit more, it didn't have any effect on the outcome of dogfights after having 6-8 battles of AI vs AI fighters.

All tests were performed with GTF Ulysses fighters on both sides, stock weapons, at first 6 vs 6 and 5 waves for each team; then 12 vs 12 and 5 waves for each team, difficulty on insane and with a slightly modified ai_profiles from BP 2. The AI class used as the 'stock AI' was captain.

What may be significant is the accuracy setting and boostsing afterburner use during defensive maneuvers (once I figure out how...), but that's for another time...

[BengalTiger]

After running a few more tests using stock AI it seems that the enemy team gets some hidden bonuses on insane, even though I use
$Autoscale by AI Class Index:    NO
and the same accuracy for both AI classes (captain and general) the AI running the enemy team scored 3:0 victories no matter what the pilot skill was set to.

I ran another test, like the one yesterday with one side having custom AI and the other one stock, both with the same accuracy. This time the custom AI was running the hostile team...

They wiped the floor with the friendlies.
Therefore the fact they were able to draw when on friendly side means that custom class is far superior and the stock class was holding it's own when hostile thanks to some hidden bonuses.

Where are the parameters for giving friendlies/enemies nerfs/bonuses so I could shut them off to make designing the other AI classes easier?

[yomi]

only
$Friendly AI Fire Delay Scale
and
$Hostile AI Fire Delay Scale
comes to my mind. But its default favoring friendlies, and its Even on insanity, and on most difficulties in BP2

[Commander Zane]

I could be wrong, but with BP AI in particular the rates are 1.0 for both sides on Medium and up.

[BengalTiger]

That is correct, which makes the results even more mysterious, because getting a 4:3 when custom AI controlled the friendlies and 3:0 when it controlled the hostiles is rather strange...

Now after more testing:

Neutral vs hostile- custom AI wins 3:0 when neutral and 3:0 when hostile, so guess this is the way I'll be testing from now on.

Friendlies vs neutral was pretty much like friendlies vs hostiles, with more/less equal results when custom AI flew friendly and custom AI victory when it flew neutral..

It seems that friendlies could be weaker for some reason...
Either that or I just consistently got improbable results and in an infinite sample custom AI would win when flying friendlies just as often as when flying others.

[General Battuta]

You know, I've had one player complaining that friendly AI has turned really ****ty since 3.6.12. Assuming that you've got a 3.6.14 build right now, could you maybe test again under 3.6.12 and see if things are any different?

[Cyborg17]

That would explain a lot.

[Fury]

Also, you do realize that there is no such thing as "stock AI" when you use custom AI profile? If you use custom AI classes, then you also have to use custom AI profile to get get full advantage from AI improvements. But doing so also alters retail AI classes to a point where they are no longer retail AI classes at all.

Which effectively means that it is IMPOSSIBLE to test custom AI vs. retail/stock AI in any one mission. All you can do, is compare how well AI performs but comparison must be done separately. Setting up a dogfight mission does not tell you anything because retail AI classes will then use default values from AI profiles where appropriate. Sure they lack some AI flags, but that does not mean they perform worse, actually in might just be on the contrary as "better" is very much subjective.

[BengalTiger]

Well the point is to check how changing courage, and other parameters (except accuracy) influences the performance of AI pilots.

The AI profile is custom, but the AI class on one side (Captain) is unchanged compared to FS 2, which I call "stock". It challenges an edited Light Fighter Captain AI class, which I call the "custom" class.
These could be directly tested against each other, as the profile is the same for both and doesn't give any side any advantages or disadvantages.

I could kick out the ai_profiles.tbl and use only the characteristics found in ai.tbl, but I have a feeling the original profile will fade away pretty soon and only custom ones will be in use.