Better AI

[F_B]

Most AI in modern sims are hit & miss, personally I think Raceroom and Iracing (yes, this previously online only racing sim) have the best AI right now, especially the latter one offering many settings to customize AI grids (driver attributes such as aggression, smoothness, defending, age and much more) and assigning custom skins to each car.

This is where AC2 should improve (among many other things already mentioned here). The current AI is average at best and needs to be improved.

 

[Kingkoenig]

The AI has always been one of AC's weak points.

 

[MJQT]

I agree with the sentiment of "better AI" 100%. I'm heartened to see this has as many upvotes as it does!

But what in particular is "better"? For example:

Consistent difficulty level track-to-track? More awareness when racing side-by-side, 3-wide, etc? Better overtaking by slipstream? Blocking lines? Oval racing? Ability to handle faster classes lapping slower classes? F1-style blue flags? AI making in-car adjustments to turbo, rollbars, fuel map etc? Dynamic pit strategy calculations? AI making occasional mistakes?

( I could go on...)

And hence why we don't have markedly "better" AI in many recent games. Can it be done? Of course. Geoff Crammond and Papyrus did it. They implemented a lot of these features above. But it takes time, a LOT of it!

And given the popularity of online multiplayer racing and the (entirely reasonable!) demandingness of the community for high-quality visuals, hyper 'accurate' physics behaviour, immersive FFB, laser-scanned tracks, weather, day-night transition, and more, I can see why AI seems to have not been the focus of smaller dev teams. For the return on their time investment, it's just not worth it.

I think one question to think about is this: what feature(s) would you be prepared to not have in AC2 in order to have significantly better AI coded and calibrated by Kunos? Personally, I love racing good AI more than anything, so I'd be prepared to trade some features or additional content off against a really well-done AI experience. But I'm guessing many others in the community don't feel the same way.

 

[Dr.Dave]

Lets say we need multiple AI lines. Because AC has only one Line (Racing line) and they try with math formulas to guide the car, when problems (slower traffic) occurs. this is not a good approach. 2 diffrent ways would work better:

old school: Even Geoff Grammonds GP2 had better AI. iRacing does and RRE does and rfactor 2 does. But what exactly do they have in common? thei all have diffrent AI lines definded. They define lines for overtaking and it would also be possible to limp back to the pits or to go for a slow down lap etc. here the AI just needs to know which line to pick, which is easy to define.

new school: let them learn. it is possible nowadays to train them for the best results and make the AI really inteligent.

 

[MJQT]

Lets say we need multiple AI lines.

I think you're 150% right, Dr Dave.

Echoing your thoughts somewhat, what I think we need most of all is an innovative pathfinding system that dynamically plots and chooses lines based on the location of other cars around it and the location of track grip. Need to go two or more wide through a corner or on a straight? Devise an algorithm that takes the fast line and automagically plots a two-or-more wide path through the corner with sufficient room between cars (using car width in the calculation)and stays away from track boundaries -- and make sure it kicks in the moment the AI detects there is a car nearby, as if a spotter called it out to them (this is NASCAR Racing 2003's behaviour, interestingly). This would be useful for both side-by-side racing and overtaking manouevres of all kinds (including under multiclass and F1-style blue flags). Want the beautiful variability of lines that NR2003 AI run? Induce a little random variation in the radius of the line taken through the corner and braking points. And so on.

We have the statistical learning tools like neural networks to take some real-world training data from human sim racers from different hotlapping/racing scenarios and train the AI driving algorithm to drive the car and follow a specified line. This is a solved problem! And AC's AI shows this is possible even on player physics. The (large) challenge that remains for developers, then, is to thoughtfully design a versatile pathfinding algorithm that tells AI cars where they should go at any given time.

In the old days of Geoff Crammond's Grand Prix games and Papyrus Indycar and NASCAR, devs had to cheese the AI behaviour with quirky manual coding, in a kind of 'cop out' because computers couldn't possibly handle 43 cars of real-time physics. Makes for funny examples like this. But this made the devs disciplined, and they manually coded in a lot of useful behaviour that still shines if you play the best of those sims nowadays. We have to return to that spirit of hard-coding behaviour IMO! Back to the future. But we now have the advantage of multi-core CPUs that could run many AI cars' physics at once (unlike AC1's unfortunate choice to consign AI and player physics calculations to one thread) and advanced tools to train AI to drive with human-like inputs and results.

 

[Robert Marold]

As I said in another post on a simmilar thread. I'd rather play with myself than race bots online. All of that extra work for your PC, so that imaginary "people" can screw you around with their slow driving and stupid non reactions...... Imagine a bot, that could "see" an accident ahead, and drive over the grass to avoid damaging your car and theirs? And imagine a flag system that would reward you for doing that, instead of penalising you for doing the best thing for everyone in the game? That may be worth it, but while the bots follow a predestined "railway track", at a predetermined speed, then why would anyone want that.

 

[MJQT]

Imagine a bot, that could "see" an accident ahead, and drive over the grass to avoid damaging your car and theirs? [...] but while the bots follow a predestined "railway track", at a predetermined speed, then why would anyone want that.

Even better, what if the game coders didn't have to tell the bots how to "figure it out for themselves" with some unnecessarily complicated calculations? What if the game could look at the situation like an all-powerful deity, and dictate to AI cars where to go e.g. by telling them to follow one of several pre-determined "panic" lines that is not blocked? This is a simpler solution that wouldn't be that hard to implement, but would yield the same desirable behaviour you're describing. Devs need to start thinking like this again! Simpler, more deterministic, but yielding sophisticated results. AC2 would be a perfect time for innovation in AI behaviour like this!

As I said in another post on a simmilar thread. I'd rather play with myself than race bots online. All of that extra work for your PC, so that imaginary "people" can screw you around with their slow driving and stupid non reactions.

I think your comment points to a larger issue that's important. Part of the reason no-one races offline with AI is because the AI is often so half-baked in recent sims that it's not worthwhile. Provide good AI, and you might have more people using it!