Hello,
After driving on the Snetterton track, I also saw very soon that something
was terribly wrong with the AI cars. Exactly as the creators of the track
wrote in the readme file and in some postings on this newsgroup.
But now, I'm wondering:
- How does this work? Does the behaviour of the AI-cars has to be programmed
for each track? I thought that the AI-algorithm made them drive accurate and
fast on each track?
- What do you have to program? And can you make mistakes? I mean, is it
possible that the (human) input of data makes the AI cars too slow for some
corners or make them brake too soon for certain corners?
- Has this work also been done for Brands Hatch and the Ostereichring?
- If the behaviour of the AI-cars is based on human input, how can we be
sure that they drive on their maximum on the original GPL tracks?
regards,
danny
---------
[email hidden]
[email hidden]
URL: http://perswww.kuleuven.ac.be/~pb655180/
ICQ 50937950
"Danny Uten" <[email hidden]> wrote
> But now, I'm wondering:
> - How does this work? Does the behaviour of the AI-cars has to beprogrammed
> for each track? I thought that the AI-algorithm made them drive accurateand
> fast on each track?
Well somehow, no. Even if it's true that the AI algorithm in the new
Papyrus game engine used in GPL is much better than in their old game
engine, they still have to rely on a track-per-track model for each AI in
each situation. But it's a step forward towards an universal AI that
understand any kind of track.
If you look inside any track folder you will notice *.LP files. Each of
them control an overall AI situation. Example, there is pass1 and pass2 who
handle how the AI will perform under a passing situation, or pit.lp, which
obviously controls the AI behavior when pitting. To resume what's actually
modeled inside the LP files, it's XY calculations from a line that goes
around the middle of the track. Wow it's been quite a while since I worked
on that (my failed attempt to create a GPL AI for an Icr2 road track), but
for the old game engine it's not exactly in the middle. The overall
important LP file is the race.lp, which basically tells the AI the perfect
line around each track.
> - What do you have to program?
If you check out ThePits, you'll find an LP Editor.
> And can you make mistakes?
Well, that's where the new game engine is a step forward the old one. Check
out your main GPL root directory and open the "driver.ini" and "gpl_ai.ini".
Enjoy at all the settings for each independant driver. That's why Surtees
is notorious in r.a.s for his well hmm... behaviors around other cars ;)
> I mean, is it
> possible that the (human) input of data makes the AI cars too slow forsome
> corners or make them brake too soon for certain corners?
IIRC the Global Hype is calculated towards an overall calculation of a lap,
not how fast you are for each portion of the track. So theorically, no.
> - Has this work also been done for Brands Hatch and the Ostereichring?
Dunno, should be ;) But somehow, GPL is clearly not an offline AI racing
simulator, even if it features one of the best AI in any title.
> - If the behaviour of the AI-cars is based on human input, how can we be
> sure that they drive on their maximum on the original GPL tracks?
Hmm.. turn off the Global Scalling?
--
-- Fran�ois M�nard <ymenard>
-- May the Downforce be with you...
-- http://www.WeRace.net
-- People think it must be fun to be a super genius, but they don't realise
how hard it is to put up with all the idiots in the world.
I've been working on the AI files for Talladega and I will tell you one
thing; it's pretty danged hard to get them to drive the right line. The LP
files contain all the data they need to get around the track. I have no
clue how the guys at Papyrus did it, but I sure wish I did. It appears the
data gives the AI the best line around the track and the individual drivers'
settings in drivers.ini and gpl-ai.ini files alter how well each driver can
follow that line. It's taking a long time to figure out how all the data
interacts, but I'm close to getting it to work right. It's tough getting
them to want to pass properly.
Tom
"Danny Uten" <[email hidden]> wrote in message
news:thaC4.15125$[email hidden]...
programmed
> for each track? I thought that the AI-algorithm made them drive accurateand
> fast on each track?
> - What do you have to program? And can you make mistakes? I mean, is it
> possible that the (human) input of data makes the AI cars too slow forsome
Quoted text (8 lines)
> Hello, > > After driving on the Snetterton track, I also saw very soon that something > was terribly wrong with the AI cars. Exactly as the creators of the track > wrote in the readme file and in some postings on this newsgroup. > > But now, I'm wondering: > - How does this work? Does the behaviour of the AI-cars has to beQuoted text (16 lines)
> corners or make them brake too soon for certain corners? > - Has this work also been done for Brands Hatch and the Ostereichring? > - If the behaviour of the AI-cars is based on human input, how can we be > sure that they drive on their maximum on the original GPL tracks? > > regards, > danny > > --------- > [email hidden] > [email hidden] > URL: http://perswww.kuleuven.ac.be/~pb655180/ > ICQ 50937950 > > >Sorry guys, but I think the award for the best effort by programmers towards
an AI model must go to Ratbag for Powerslide and DTR. An interview with the
programmers (on their website) explained how each driver in the game was
trained using a neural network - maybe "bred" is a better expression. The AI
that performed best were picked for the higher difficulty settings. And the
effort shows: I'm still having problems trying to beat the AI on the third
level of difficulty (whatever they call it in the game). The neural network
approach is IMHO a much better way of programming AI than the human coded
racing lines, as the AI are far more likely to obtain optimal performance on
ANY given track, regardless of the track designer's skills. I think the
training only happened offline (i.e. there is no neural network code running
in real time - all training was completed before the title shipped), but the
next step will clearly be to integrate such neural networks into the real
time code. Neural network models are also more likely to be capable of
making human-like mistakes, and would be far more challenging over the
course of a game's lifetime (imagine a game where the AI "learn" from the
human player!). So there might be a use for those 1 GHz machines after all
;-)
I would have included a link to the interview with the Ratbag programmers,
but my browser is currently crashing and I'm in the middle of programming a
fairly delicate piece of embedded DSP code, so restarting is not currently
an option . . . maybe someone else can post a link.
Ryan
Does this mean, that the DTR AI is actually playing with the same "rules"
as human driver? I have not played DTR, but I think in other games AI is
cheating to save CPU cycles and programming effort.
Jari Jokinen
Ryan Mitchley wrote in message <8bcp32$4dp$[email hidden]>...
>Sorry guys, but I think the award for the best effort by programmerstowards
>an AI model must go to Ratbag for Powerslide and DTR. An interview with the
>programmers (on their website) explained how each driver in the game was
>trained using a neural network - maybe "bred" is a better expression. TheAI
>that performed best were picked for the higher difficulty settings. And the
>effort shows: I'm still having problems trying to beat the AI on the third
>level of difficulty (whatever they call it in the game). The neural network
>approach is IMHO a much better way of programming AI than the human coded
>racing lines, as the AI are far more likely to obtain optimal performanceon
>ANY given track, regardless of the track designer's skills. I think the
>training only happened offline (i.e. there is no neural network coderunning
>in real time - all training was completed before the title shipped), butthe
"Jari Jokinen" <[email hidden]> wrote in message
news:8bevp9$pmj$[email hidden]...
> Does this mean, that the DTR AI is actually playing with the same "rules"
> as human driver?
Short answer: YES.
Ryan
Short answer: YES.
Ryan
On Fri, 24 Mar 2000 11:51:59 +0200, "Ryan Mitchley" <[email hidden]>
wrote:
Long answer: "Yes, but with a few billion neurons less".
Just a detail, though :-)
JoH
Quoted text (9 lines)
> >"Jari Jokinen" <[email hidden]> wrote in message >news:8bevp9$pmj$[email hidden]... >> Does this mean, that the DTR AI is actually playing with the same "rules" >> as human driver? > >Short answer: YES. > >Ryan>>>>> "Jo" == Jo Helsen <[email hidden]> writes:
Jo> Long answer: "Yes, but with a few billion neurons less".
Jo> Just a detail, though :-)
I don't know...looking at drivers in rush hour I have a feeling the DTR
AI have more neurons than some humans :-)
-peter
