A new idea for sims - ai constantly learning from humans

There’s a revolutionary idea for racing sim AI learning from humans that I’d love to see explored – where AI learn not from a pre-set algorithm or programmed racing line, but from mimicking the behaviour of thousands and thousands of race laps done by humans. An example of this kind of learning in natural language processing is the chatbot CleverBot (http://en.wikipedia.org/wiki/Cleverbot), which learns its responses from humans’ answers to conversational lines, rather than from a database of canned responses.

We have human racers world-wide churning out thousands of race laps both online and offline 24x7 every day. If AC, or any racing sim for that matter, could be constantly capturing metrics from human laps and consolidating them in a central database, it could learn how to behave in any given situation – where to brake, where to position the car, how to pass, how to chase an opponent. The sim would eventually learn very human-like and realistic behaviour, being constantly taught by the best – real, human drivers. A lot of the required information is quite possibly already being captured in replay files.

Just imagine how good AI developed and constantly learning in this way would be.

 

AI wreckers, eh? That would be a novelty

 

I believe Forza Motorsports has something similar implemented "Drivatar".
No idea how well does it work, but its tech that has been developed by Turn10.

Skynet/Cylons/Matrix... here we come!

 

You could pretty easily filter out data that differed from the average speeds or lead to crashes.

 

I can tell you how it works my son has Forza 5 and Horizon 2 with Drivatars. Image the rammers, pushers and all the nonsense you sometimes see with online races, now image most all the AI acting the same way that is what Drivatars are.

 

The Drivatar approach sounds like a step in the right direction, but I couldn't tell what kind of population the drivatars were learning from. If they are averaged over a big enough sample size, wreckers would just become noise that can be filtered out, assuming that wreckers are in the minority.

Edit: reading a bit more about Drivatars, that's not the same idea, if they only learn from one person. I'm talking about making use of the massive amount of data that could be captured from thousands of human players.

 

I also own Forza 5, but for me the drivatars are quite good, overall clean driving/minor rubbing but no dumb moves... it might have something to do with the fact that drivatars change for the different skilllevels the player has.

 

I think its a bit silly for the AI to learn from humans. In the end it would learn how to make mistakes.

There is no subjectivity in driving a car fast, its very straightforward for a computer to do that.

 

LFS has a learning algorithm.

 

"There is no subjectivity in driving a car fast."

I guess you must not be a very fast driver then if you believe that. You do know it's very difficult to get a computer to drive a car properly on the limit at a good human pace in dynamic conditions? Perhaps programming a single line with 100% scripted behavior makes the AI fast, but not so much with dynamic physics calculation and the AI computing what to do at all times.

Driving a car on the limit is extremely subjective and everyone does it somehow differently. Ways of getting on the limit and staying there are very similar but all unique.

As a sidenote, there must be mistakes in the AI's driving. Otherwise if it can be programmed to drive a car on the limit, no one can ever catch it and surpass it's laptimes, ever. Do you think you don't make tens, perhaps hundreds of small mistakes per lap? You do. Even pro drivers make a few small mistakes per lap all the time. A lap with no mistakes, on the limit of car potential, would be absolutely insane.

AI will never be as fast as humans, either. At least race computers can't churn out laptimes that can't be reached. Pretty much every time a "theoretic best" laptime is recorded, a human driver of the highest level passes it.


The whole idea is very interesting and perhaps, with a few million people who are good drivers sampling into it, we could get a very complete racing driver. It'd be extremely hard to get to work properly though.

 

If the AI drove like humans i'd just hot lap. I gave up racing humans 3 years ago. My blood pressure has gone down ever since.

 

there are quite a few ways to implement good AI. Neural Nets, machine learning, etc.

 

Says the guy who didn´t want to "copare dick sizes" in another thread...

Going fast isn´t rocket science, only three inputs, with the biggest constraint being traction available on the tires wich a computer can calculate or "feel" a lot better than any human, specially inside a sim where there are "absolutes" and all the AI has to do is use those numbers.

It would be difficult in real life, as it would rely on sensors and those could be faulty or not accurate enough, but inside a sim, its easy.

The hardest job in making an AI isn´t to make it go fast, is to make it more human, adding some minor mistakes here and there, some "personality".

 

Yes and no.

I've dabbled in machine and deep learning, and it is not exactly simple. Training a network to recognize and classify black and white images took me 3 days to train, and then all it could do is that. Building an ML or DL AI for driving and reacting to actual events on track is substantially harder. Harvesting laps from actual users could help, but even then the results are unclear, and how you would train it separately to try to avoid opponents.

Theoretically, yes. But there are some CPU (or even GPU) limits you'd be running up against, as the network would need to run on the driver's PC, rather than cloud, as cloud would be too slow.

 

If we have AI like that, I would deliberately ram them off the track and see how they will react and learn.

I think it would be awesome if they learn to throw dirty words at me, vote kick me, ban me, and ram me back.

 

Oh yes, very difficult and time consuming. Just wanted to throw the ideas out there, AI has no one solution.

 

You're talking about a pretty elaborate system that would require a lot of processing power. Programmers are only now getting supercomputers to comprehend and act like humans IE: IBMs Watson. AI in games now are very cut down to minimise CPU usage, I think most AI systems in games are just faking intelligence. If you want AI to act like humans their going to have to see the virtual world like humans, which means programming in constraints. Which would all require more processing power.

The AI in games could probably go around a track non stop setting perfect laps within milliseconds of each other. To make them seem more believable they have to be fallible. It might come with time but you're going to be programing AI to essentially make mistakes, which can be done easier than creating skynet.

 

Remember, computers are not intelligent. They merely mimic intelligence.

 

Computers are as inteligent as their programmers.

Like dogs

 

Rf2 has pretty good AI they have some parameters for the user to adjust in the plr.json

AI Mode#":"When in a test day with AI, AI will attempt to perfect his driving line, and save his knowledge for future use",
"AI Logic Override#":"use bitfields to disable certain AI behaviors. 1= blocking line",
AI Limiter#":"Range: 0.0 (no limiting) - 1.0 (limiting used to make racing closer but also make more driver differences on flat-out tracks)",

so they kinda learn as they go if you set it up ..