AI Aggression - placebo setting at best

Something I wondered after watching the RandomCallsign video yesterday:

Do the AI get dumber if you run lots of them? Do the people who report bad results with the AI run lots of AI or is the behaviour exactly the same with fewer AI?

 

Please stop making excuses for bad AI. This has nothing to do with it. They are just completely brain dead when it comes to actual racing, racing lines , track limits etc. I haven't come across another series that actually has AI flat out racing you off the track around corners. The way they dart from offline to racing line is ridiculously unrealistic. I have not seen this behaviour in any other title.

 

Good morning to you too.

 

I agree with you albeit not entirely. Humans excel at creative problem solving but dont really do repetition well without error. Computers do repetition well without error but not problem solving. What does this mean for our conversation? During a race it is likely that humans make errors, some small some big, all of which will lead to pace variations, mistakes etc. I even find my mind wondering sometimes and lose half a second a lap for a while then am surprised when an Ai appears on my tail. Ai do not suffer from this because they are capable of maximum performance without error. There are examples of successful implementation of error in Ai in rfactor2 most notably. What is noticeable in AMS2 at the moment is that the Ai group together and also seldom if ever make errors in the race (I certainly dont remember an error that wasnt derived from close contact or similar in Ai in AMS2). This results in the characteristics above becoming obvious and the difference between Ai and human becoming obvious, therefore immersion lost.
On a positive note, its been stated elsewhere by those that know that the Ai is (and I loosely quote) 'not even close to its full potential' in AMS2. Various reasons why its taking time summarised as complexity and risk management. Im not going to offer solutions because Im not a software engineer, but the end result needs to provide that holy grail of motorsports- avoiding predictable racing with strung out cars always finishing in the same position with no overtaking and no mistakes. Should be simple..

 

That was certainly an issue with PC2. It also affected player physics if there were too many Ai, not sure if that issue has migrated too but I havent noticed it particularly. Its a reason why people in that title complained that setups developed in practice (on their own) reacted totally different once there were Ai everywhere. I personally run between 11-13 Ai with myself making the round number. Anything else is out of the points anyway so irrelevant in any championship that I run.

 

All this discussion has made me reinstall AMS1 again. That will be my Ai Championship race game, AMS2 will be my pretty looking, quick race blast with fun weather, and dare I say, RF2 will be my "want to race on ok looking track with mods"...until it pisses me off again with fiddling and loading times.

 

Dan I actually thought about purchasing AMS1 as I race offline. You seem to be in a similar place to me right now so would you recommend it?

 

Definitely. AI is decent and realistic to view. Quite often I found myself just watching the AI race instead of being arsed to race myself. Talent files can easily be edited, to create realistic performance during a season. No weather, and graphics starting to look dated.

 

I still play AMS1 quite a bit - I still prefer aspects of it over AMS2, including the handling a some of the cars (eg. Opala Stock Cars still feel better to me in AMS1).

In terms of AI, whilst I still rate the AI in AMS1 overall higher than in AMS2 (than whole talent file setup works very well, which includes likelihood of AI errors, etc), I now find that AMS1 is more annoying than AMS2 in terms of the AI taking me out from behind. It perhaps varies from class to class but I find the AI in AMS2 in the classes I race is very clean now.

As others have said, as long as you recognise the typical behaviour of the AMS2 AI, particularly that they'll try to stick their nose up the inside whenever they get the chance, and drive accordingly, it's perfectly possible to race side by side with them, through many corners, without incident. I'd go as far to say that I'd expect cleaner races with them now than I would with the typical online race.

 

Yeah, I would agree that AMS2 AI do race cleaner, but just totally unrealistically. Running 3 wide on corners and off track, maintaining speed etc.

 

Ok, thanks guys. Good to get your insight

 

This was my experience in AMS2 in the past, although recent versions seem to have fixed some of it. The FClassics for example no longer go nuts when I try to race more than 20 AI. But I wonder if a change was made, where instead of downgrading the player physics when there are "too many" AI for the CPU, now the AI logic gets dumber? Just speculating on what could explain the wide range of experiences.

Also, has it been confirmed if the "-pthreads 4" command line option does anything in AMS2? In PC1 and maybe PC2, the physics (and AI?) were running on two cores only, which could cause bottlenecking and weird bugs when racing the AI. I have this set just in case but could never confirm if it actually does something.

 

I'd forgotten about the pthreads command. Good idea! Also good to hear your observations about the ai. Once Ive finished doing the stuff Ive got to do, I'll try running a short versus long grid and see if I see any changes. Certainly if we can optimise the grid for our individual machinery then thats a good thing, and if an awareness of that is built on something solid then equally good.

Edit, some time later.... well I ran a full grid of 31 cars against a grid of one other ai and the good news is I certainly didnt observe any fundamental differences in either ai behaviour or my own setup behaviour. The only difference I observed was in performance which was to be expected on my machine, so if performance was assumed to be linked to ai or setup behaviour, I didnt see any evidence to support that hypothesis.

As regards aggression, with the one other ai test I varied it between 0 and 100 and observed that it was easy to keep up, pass and be passed at 0, but I was only able to keep up and have maybe one opportunity to pass in several laps at 100 aggression (110% difficulty, formula trainer, Brands Indy). I didnt see any other difference in behaviour, so for me at least the aggression slider is a fine tuning tool at best, pointless and a distraction at worst (I say worst because it doesnt adjust ai driving standards as such).

 

Thanks for engaging on this Renato.
I have to say I cant remember the last time I was rear ended by the Ai. If I judge that the car behind is carrying some excess speed I'll leave the door open and try and get the cutback on corner exit, or bide my time for an overtake.
For me the Ai in Rf2 is a good standard to reference. You can see the Ai cars sliding sometimes, they make mistakes sometimes (not often but enough) that rarely result in multiple pile ups. You shouldnt get collected if you drive to avoid and if you do.. no money lost, no lives lost, but its very immersive.
The Ai in AMS2 in really good fun to race against but its lacking that certain something right now that really drags the player in time and again. The problem is exemplified in my custom championship, second time around. Theres very little different from the first time around. Same cars in the same places racing in the same groups doing the same lap times to the same result with no surprises and no variation. Its this that I hope gets fleshed out, rather than little shuffles here and there. Hopefully it will happen, but Im sure that Im not alone in being slightly concerned at the legacy of having to deal with ai physics running alongside player physics to give an homogenous experience.

 

Thanks for explaining things from your side Renato. It makes a big difference to hear from developers and have things clarified.

Communication on some of the ways the AI works & what variables the AI settings affect changes perspective on things and hopefully helps with the varying opinions on AI, particularly that we now have a better idea on what to look for when changing settings or being in certain situations on track.

The primary concerns (AI being stuck in groups, too fast/slow on various turns, aggression behaviour - hitting walls, going off track, etc.) sound well heard and I'm looking forward to future AI updates and how it changes the racing experience.

 

I do wonder how far away we are from something like an A.I that uses the players time(s) or even lines as a base, then with variation in the field forwards and backwards speed wise. Is something like that plausible in the near future? Either way, things will keep improving.

 

As long as it doesn't use racing lines data from me, that might work!

 

Apologies if I came across a bit harsh. Just find AI frustration. No offence meant.

 

NP m8, we all have frustration from time to time.
And we all have an other definition on how a good AI should drive.