Sorry for the mahuusive post! 
I was under the impression LFS dropped the 'learning AI' part a few versions ago? Certainly the last time I played with them (z25?) they seemed pretty unremarkable, still. I don't know if even the biggest LFS fanatics would suggest it's a great offline game due to the AI, or at least, I've never seen anyone do it
I did say they were idiots!
Anyway, I'd forgotten about them dropping the adaptive part (I have a several versions insatalled, for various reasons). It was quite annoying, because they were too impressionable, really, and couldn't
unlearn very easily. They're not very fast, either, and a bit blind to the player - hence the observed reticence to praise them. However, the devs
should be praised for trying a different approach!
I think it's important to keep track of what a lot of these AIs are designed to do. When people say something has "poor AI" I can't help but think they are kind of assessing things ass-backwards, in terms of writing it as part of an actual commercial or releasable product, I think a lot of racing games do pretty well.
Regarding your checklists, the whole point of a good game is that the technology and innovative development that goes into it should be transparent, and not interfere with the experience (argh, horrible word, but you understand) - so, actually, the end-user checklist is, in a way, more important than the developer checklist - and should feature on said checklist. Without such an attitude, it's difficult to make a truly great game - simply because, it's judged by the way it plays, not by how cleverly the devs have managed to control asset usage to remove loading screens. Of course, the knock on effect of this particular "tech" is quite strong (positively) on the game experience, so it's not a pointless exercise to do such things, but it has to work
for the game - think about how Kaz has alluded to having avoided true weather and damage effects until recently, simply because he couldn't guarantee they would work in a game-environment, irrespective of the fact that it works well in prototypes!
Of course, credit should always be given where credit is due.
The development of AI for racing games has a couple of unique challenges too - everything is in a fixed frame of reference, and the speeds are immense. In other cases, say, flightsims and space games, you have some 'outs' to get around the speed issues. Flightsims (at least IME from modding the Il-2/CFS series) get around this by using relative frames of reference for plane-to-plane combat and manuvering, and intense scripting for anything using fixed reference points on the ground, (again IME from Freelancer/X3) space games typically have very loose frames of reference and again can get around any close interaction with them through scripting.
Correct me if I'm wrong, but a relative frame of reference is surely the
only way to "do" human AI? We all see the world from our own frame of reference, which is
always relative. Maybe I'm mis-understanding the terminology, and there's an issue with a particular method of implementing "frames of reference" in this manner, for AI, but if a relative frame of reference is used for dogfighting aircraft, why not for jostling racing cars?
Presumably the fixed referencing is for the circuit (which hopefully doesn't move), but it does move
relative to the cars...? Could a hybrid approach be used for racing games, as it is for flight-sims?
Racing games though - not only do you have to make the AI fast enough to drive sensibly, it has to react to things at 300+ KMH, react extremely quickly to them, deal with minute obstacles, on a totally fixed frame of reference, times 15 cars, and cope with networked reactions to human input. And it has to look convincing, and do nothing much that isn't acheivable by a human player.
But all of this is applicable to dogfighting, too - and they travel much faster than 300 km/h (closing speeds of two aircraft can be well in excess of 2000 km/h)
I know it's tempting to knock things like AIW because "it's just a line", but it's pretty robust, and it works, with very few drawbacks.
I wouldn't knock it for "just being a line" (which it isn't, in any case - it's more of a "corridor") unless that's how it
played. It does "work" (depending on your checklist) but is that enough? It has drawbacks (in every application of this particular "format") - namely breaking of the suspension of disbelief. Not game breaking, I suppose, but hey - most people play those games online, too.
The thing is, [half-life's grunts] are absolutely, purely scripted, with just about every possible action the player can take pre-calculated with a fallback script ready to take it into account. They spent ages working on it. It looks and plays awesome when you use it, but it's not reacting on any "real" level, any more than a choose-your-own-adventure novel.
This is what I mean about the implementation vs. the end result. I was utterly convinced when playing this game for the first time, that I was fighting something that was trying to "out-wit" me. I even rummaged around the SDK trying to find the bit that controlled their grenade throwing abilities, because I was gob-smacked at how deft they were with it, even when I was in a doorway beyond a balcony, and they were down below, I'd often see the grenade land
behind me, flushing me out onto the balcony. Harrowing stuff when the adrenaline is flowing! Of course, they were far from perfect. Pausing for cover in the open, for example, or the grenade missing and landing well out of harms way - but these moments just made you think you now had the upper hand, and were actually, oddly enough, great for gameplay, even though they're obviously errors in the implementation!
Yep, absolutely. Again it's just a matter of looking at from maybe a slightly different perspective than "this doesn't do everything I want". If it does most things right and doesn't fail egregiously then that's
really good AI
Of course, you're right - but I'd rather not have to convince myself to consider the development of such a feature unless it's because I'm gob-smacked (see above) at how effective (as part of a game mechanic) it is.
and disembodied collision sounds 
. Theoretically we could run as many AI controlled cars as we wanted, but they each took time out of eachothers learning cycle, so not much point in that. We mostly tested with 4-8 cars.
)
This makes sense, why generate more data when you can adapt what you "have" already?
