AI - the future...(theory)

[ZAGGIN]

the real proof that the ai in the next gran turismo will be 'better' than previous games, will not be how well it races, or how realistic it is, but how quick and well the cpu learns from its racing experiences. making the ai drive realistically, is nowhere near as hard as trying to get it to interact with us. many times ive raced gran turismo, and felt like i was part of a spearate race to the cpu. sure we were both gunning for the finishing line, but our motives were different. i wanted to win, the cpu was told it had to win. whether it be through the use of raw power, rubber banding, racing lines or minimum lap time limits, the cpu, doesnt really race for itself, its following orders. basically, the ai in gran turismo lacks passion, it has no will of its own, it has nothing to strive for. its not surprising, because it can be hard to replicate human emotions. until then, racing ai will always be inferior to our own thinking because we have free though, we can do what we want, the cpu cant.
one of the main problems with gran turismos ai, is that it is sandboxed - it only applies to the race that it is in. ive seen no evidence except with b-spec mode, that any learning takes place in gt4 at all. learning is the key to sucess, and it would eradicate all of the silly cpu carcraft that we see on a daily basis. learning would also allow the ai to evolve into something greater than the sum of its original programming. whilst the cpu does what is efficient, we humans rarely race at anything near 100% efficiency, though we all strive to reach that goal. we shouldnt talk about more realistic ai, we should be thinking about learning - real time learning. ive lost count how many times ive suckered the ai with a basic racing manouevre, with learning, that wouldnt happen.
now imagine gt5 online with a cetral ai database. now imagine if that database was updated with ai uploads from every gt5 game. whenever we raced online. that database could then form a central core, that could be downloaded, intergrated and updated into our own system. the input from thousands of users could go to form part of an ultimate race ai. as long as the ai is 'fenced in', the problems we are experienceing now will still persist. sure the programming will get better, but the ai wont be truly artificial or intelegent until some form of learning is implimented.

 

[G25RJ]

I couldn’t agree more with you but the part about the central database is not a great idea since the AI that is learning from you it’s only suited for you and not every body is racing by the same rules you are racing, this central database will only create a frankstain that nobody will like.

Not only learn from you but have variables the we could modify to suit our taste like aggression, mistakes, etc.

 

[qwazy|06]

the computer plays to win.
the people play to not lose.

 

[ferrari_chris]

Everybody enjoy's bagging the Ai, which is fine. Maybe Artifical Stupidity would be a better name that Artifical Intelligence. Does anyone know of any games that come with good or great AI?...

There must be some, surely...

 

[ZAGGIN]

I couldn’t agree more with you but the part about the central database is not a great idea since the AI that is learning from you it’s only suited for you and not every body is racing by the same rules you are racing, this central database will only create a frankstain that nobody will like.

your missing the point. everyone knows some of the best software on the net is open source? why? because everyone has a part in making it great. firefox is a good example of this.

ive read many articles on this forum about the ai in gran turismo, its either too hard, too easy, not aggressive enough, too agreessive, rarely makes mistakes, always makes mistakes, is very realistic, not very realistic and so on and so on and so on...
PD can never get the it right no matter how they impliment the next generation ai. as a base though, it should be tough, realistic, but not human like. why? because imo i think trying to replicate our thoughts and feelings during a race is impossible, the ai has to learn and adapt just like we do. with ai, its fallibility is measured on percentages or something similar, we humans on the other hand are never that precise. there are other factors at play when we race, and these factors are often fuzzy.
you will never see the cpu try and get a couple of extra laps out of those red raw tyres like we do, you will never see the cpu determine its pit stop strategy based on mood, or see the ai, impressed with its own new found skills apply them in an abstact manner like we can. when was the last time the you saw the cpu challenge us? we often attempt and win races in cars that are inferior to the competition we might face, the cpu never does that.
PD need to seriously look at the way it programs the race ai. until it can get the cpu to beat us humans in a race with an identical car, then we will always be in the same situation that we are in now, the only difference is that the game will look better. ai, will always be artificial, but it can be intellegent, but that intellegence has to be learned not granted.

pd also needs to stop using the power-to-weight ratio of a car as a major determining factor when it comes to furnishing the race with silicon competitors. on paper it is the truest way of comparing cars. in reality though, it is not ver realistic. there are other factors like capacity. for instance. a 1.4ltr kei car can be tuned to have a better pwr than a stock 10ltr v10 dodge viper, but it would never be able to out accellerate it to such a degree that the viper was left xx seconds behind after a couple of laps. capacity plays a big part in how much raw pawer a car has. a 1.4ltr car, can never muster the same power, as a 10ltr car. for a start just looking at the engine is enough to prove this. the kei car has 4 cylinders, the viper has 10. thats 2.5 times more cylinders, and thats before you even consider cylinder capacity. capacity was first implimented as far back as gt2, but it seemed to of been ditched when gt3 arrived, giving rise to ridiculos mismatches like the one ive just mentioned. im still running tests to see if capacity plays a big part in gt4.
kodiak posted a thread 'last laugh at seattle', based on the seattle endurance race on gt3. in it he seemingly won an impossible race with a kei car. i challenged his claim, and went on to prove that the race was a forgone conclusion. you can read the article i wrote on the second page of that thread.

https://www.gtplanet.net/forum/showthread.php?t=25304&page=2&pp=20

the purpose of my reply was not to rubbish his claims, but to prove the weakness of the ai which (on gt3) was modelled solely on the power to weight ratio. im still investigating the depth and scope of the ai on gt4, but once i finish my experiments, i will post my analysis in the gt4 forum.

 

[northflt]

again, the central database has its problems. the idea of having a consolidated database of knowledge is great. the problem is that not all players have the same skill... that's the same of getting all the knowledge of all chess players in the world together. even Kasparov would have trouble beating it.

if the GT AI knows everything at any time it would get to the point that it could always win and then I will throw my ps2 through the window because I bought the game to defeat it to 100%.

i think a real AI is the one that learns with the player its playing against. for instance, there is a limited number of tracks in the GT game and you have to play them in lots of circunstances. each of these races AI would learn with the player. to get to the point that by the end of the game (100%) it would be extremelly hard to beat it. and that works for everyone... if my 5 years old sister plays and gets out of the track lots of times, the AI won't learn anything, will keep on a dump mode until she starts beating it.

that could require an initial AI skill... that could easily be built during a session of tests (licenses).

what do you think?

 

[MasterDriver]

Right now i highly doubt GT4 used decision trees to do their a.i. to me it looks very state based. However I would like to see it done through "learning" as so pleasently put, but their are a great many things that need to be fixed before A.I. should even be looked at, and you guys at "PD" u know what i'm talking about, quite a few bugs that fell through the cracks hah.

 

[kikie]

What you mean is that when we race we should be part of the AI cars ro rekognized as another AI cars by the cpu??? And the AI cars must "think for themselfs"??

learning is the key to sucess

That is what I have been saying all along in other threads about AI. We need AI cars that learns from their mistakes and try to learnd how to drive properly.

but not human like. why?

I think you are wrong about that!!!!! Human like reactions is the key to realistic AI, just my humble opinion.

because imo i think trying to replicate our thoughts and feelings during a race is impossible, the ai has to learn and adapt just like we do

you're contradicting yourself. Learning like and adapting like us, are human reactions, we are no artificial computers!!! Our thought and feelings is what makes us human. If the AI must learn and adapt like us (humans, not cpu's), then you have to replicate our thoughts!!. A cpu with feelings is still science fiction.
KY said it himself, last year, the AI must have human reactions!!!!

 

[Arawn]

Well , if "learning cars" are the key to a more realistic overall driving experience then you will find this rather interesting. it's called NERO. The team behind the project have truly developed a new form of AI. You can even download their software and try it yourself and i must say it does work though in a primitive way.

i'll give an example , this is what i did : I instructed my army to capture a flag and placed some enemy's in front. left of the enemy's i placed a large wall. i instructed my men to get the flag and not get shot. at first they would run straight forward towards the flag and get shot again and again but as time passed they learned that it is better to take a longer run and go left from the wall not get shot and reach the flag.

imagine this new AI being perfected , you would have opponents who would try not hitting you and thus braking accordingly remembering the braking points adjusting their driving style and overtaking by exploiting that weakness.

this is the URL for those interested -> http://nn.cs.utexas.edu/NERO/

 

[SchuberT]

you're contradicting yourself. Learning like and adapting like us, are human reactions, we are no artificial computers!!! Our thought and feelings is what makes us human. If the AI must learn and adapt like us (humans, not cpu's), then you have to replicate our thoughts!!.

He's not contradicting himself, he's just setting boundaries. The real problem here is processing power. Sure the PS3 is supposed to be a computing powerhouse, but there's no computer as complex as the human mind. I think it'd be nearly impossible to replicate our thoughts. Instead, there must be someway to capture human movement in various situations.

BTW, the thing that makes us human is rationality (judging right from wrong) or free will. Along those lines is reason which is also something that scientists believe is unique to the human mind. (not trying to disprove what you said, but this is the more accepted definition of the human mind)

I find the idea of a learning AI very intruiging. Fighting games in the past have showcased this like Virtua Fighter 4 and the likes. Although this may be even more basic than NERO, it's still there. If the technology already exists, I don't see why KY couldn't find a way to port that idea to GT5 on the PS3.

 

[kikie]

Nope, sorry but he is!!!!!!!!!!!!!!!!!!!!!!!!!!!
He's even contradicting KY!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! KY said, that he wants an AI with human like reactions. Can't remember in which interview I read this.
Who drives race cars?? Humans, animals or computer??
Humans!!!! So we drive, we react, so the reaction we make are human. PD wants to make a real driving simulator. If he wats it to be real, he has to mimick the human reactions into the program.
The pitcrew and drivers for GT4 are based on human movements and behaviour, not based on computer simulations. AI will be exactly the same thing.
Or he programs human like reaction or he programs cpu reactions and that were he's contradicting.