Hi all

Went online today and there was SO Classic servers so i started a race with 10 AI cras. They slowly got better to laps of 45 secs. Still to slow to race with. If i keep them running will they get towards 42 secs?

It's amazing that your Ai actually can complete a lap of SO1. Mine just seem to hit the wall after bus stop and race over they can't get any better because they go nowhere. I think the Ai have a limited ability, they'll get quicker upto a point. Then they might be too fast for their own good and not know how to drive on cold tyres or a different set.

Quote from Davo :

or a different set.

different set - different .knw file.
thats not a problem.
but yea, they arent good for serious racing yet, i highly doubt that u will get them good enough to have some competition.

just try and limit yourself to the AI skill, use their brakeing zones etc

Although it is well known that Live For Speed AI is below par, I'm yet to play any racing sim where the AI are genuinely competetive and human like.
GPL was the best I've experienced, yet the AI still braked at wierd spots for some corners, had slow corner speeds and erred when trying to overtake.

I think realistic AI is just a mother hard thing to achieve.

lfs still does a pretty good job when it comes to ais ... they learn the tracks they can learn how to overtake without crashing etc

the only thing they never got the hang of ever since s1 though is driving fast on a car with wide tyres in the front

From what I understand the AI learns by trying different lines and throttle/brake usages and when this results in better times they keep it, otherwise they discard it. How exactly they do this I don't know but I figure its kind of hard to do that by laptime so they prolly cut the track in pieces and use the time/speed accomplished over that sector.
I've been tinkering quite a bit with the AI in S1 and noticed that this causes some problems. First of all the experimenting with different lines doesn't seem to work at all. At least not with hotlapping. When overtaking it does work. I've used track objects to force them to change their line but they keep crashing into the objects. Even when I throw away the AI files and start all over again they end up doing the same thing. I tried both large and small changes but neither of it worked.
The other problem is that a corner is simply a part of the track. When there is a decent straight after it. we tend to enter a bit slower so we can exit faster and that is something my AI never figured out. Thats what becomes fatal at the busstop chicane at SO-Classic. They probably exit the first sector at the little straightaway and keep that method because they messure a good time/speed. They can never exit anymore because they are way too fast but they never get that. So it becomes an endless loop. They will not settle for a slower time on the first sector because they don't see it as part of the track.
Since the AI is only in an early stage I don't have any complaints though. It won't take rocket science to allow them to draw conclusions about where they crash and how that affects the sector before that so there is a lot of potential. Interaction with other cars however seems to me much harder. I noticed they cant really understand an attack. When you get under them somewhere around the apex they will often either crash you or themselfes. They don't see the consequences of not being able to take the ideal line. At least not on time.

Pretty much everything artificial intelligence is based on search, the computer tries a lot of things and uses some heuristic function to estimate whether the direction its going in will lead to the goal state, i.e the fastest lap.

I think the reason the AI are limited at the moment is to keep cpu requirements reasonable. They could quite probably become ludicrously fast within their own representation of the world, by feeding them more data, and building in penalties added to the estimation for contact with other cars etc.

It would even be technically possible to produce some very human like behavior, with a few very simple additional rules to manipulate the search direction, especially when interaction with other cars is involved. This all costs time and cpu power to implement though, so I wouldn't hold your breath. Just look at the way ants work together, with just a set of very few simple rules "find food", "hord", "don't stray further than x from the next ant", "run like crap if you see an anteater", throw them all in together, you get complex behavior out the other end.

Does anyone know if the developers have considered using a neural net to run the AIs?

I would be interested to know whether the computer knows how well it's doing that by guaging that lap time, or the time to complete a section.

I'm a Computer Science & Artificial Intelligence student at Sussex university by the way, in case anyone wondered why I know anything about it.