AI can't plan a pit strategy

[Javaniceday]

Well, on the other hand, arcade games can get away with using the same physics for player and AI, but the point of a sim is to simulate car behavior and today's computers cannot properly do that for 20+ cars simultaneously. Unfortunately we're still too far away from having AI really feeling like humans in sims or any other type of game really in our PCs/consoles.

All racing titles "fake" the AI to some extent. It's the way all racing titles work. Most developers aren't truthful as to what is really happening. Some titles make it less apparent. If the title has fewer car variants to support, they can make some convincing AI. The more variants to support, the more gaps you see in the AI behavior as it's not feasible to develop AI behavior per car type in a title that supports a high variation of car types.

Well, look. I don't think this is a controversial statement. The AI in sims of the past are generally better regarded than the AI in today's sims. It's not because the older ones were more sophisticated or "simmed" it better. Modern day AI surely is probably doing a lot more things than those old ones. But the end result of older AI - from games like Grand Prix 4, GTR2 and GPLegends, hell even super old school dos-based Grand Prix 2 - is that they were more successful in.. um, "immitating" or "faking" the behavior of actual people. They fooled us better than the current sims.

And the reason is probably because the stuff being simulated was more paired down, simpler, easier to code and manage. There was more parity between the player tire model and what the AI was doing. So now, in modern sims, the tire model are so complicated, they've outpaced the fidelity of what can be achieved by the AI. So my opinion is that, for the sake of gameplay, for the sake of replicating human behavior more successfully, things like the tire model need to be simplified. I don't know if GTR2 also didn't have any AI tire wear - but if it didn't, they faked it more successfully than pc2 did. And I think SMS knows this, which is probably part of the reason they nixed tire wear entirely for pc3. One less variable for the AI to fake, one less thing to break immersion and be a "tell" that you're racing against bots.

I understand all "AI" in these video games are all faked and not true AI, as your video demonstrates. I'm saying you can have better results faking the AI if you go back to the fundamental established in older sims.

[cpcdem]

So you are suggesting to make the driving less realistic/convincing, in order to make the AI easier to develop. I know a game or two that actually did that recently and weren't very successful...

Anyway, sharing the exact same physics with the player is not what makes a good AI, those differences can be emulated well enough. What is most important is a lot of effort and persistence to make the AI better balanced corner by corner, track by track, for every condition. See for example AMS2, even though it uses the same engine and limitations with PCARS2, its AI is already so much better than that of PCARS2 in every level and it will keep improving, give it a try (now, after the latest updates) if you haven't already.

[Javaniceday]

So you are suggesting to make the driving less realistic/convincing, in order to make the AI easier to develop. I know a game or two that actually did that recently and weren't very successful...

Anyway, sharing the exact same physics with the player is not what makes a good AI, those differences can be emulated well enough. What is most important is a lot of effort and persistence to make the AI better balanced corner by corner, track by track, for every condition. See for example AMS2, even though it uses the same engine and limitations with PCARS2, its AI is already so much better than that of PCARS2 in every level and it will keep improving, give it a try (now, after the latest updates) if you haven't already.

I don't know what the right balance between how much simpler the tire model should be in order for the AI to match the player. That would be up to the developer to determine that, but I do think that not having the AI experience tire wear - if indeed that is true which if true is incredible to me - probably is not the right balance - evidence by the fact that we all could tell the AI wasn't right.

Maybe the right call is simplifying the tire model. Maybe it's having fewer cars, maybe it's having fewer cars on track, maybe it's having fewer tracks. Maybe its having just one type of dry tire per car like Reiza did (no softs or hards option if you noticed on AMS2). Or maybe it's taking more care and effort in faking the AI better, as Reiza is doing. I enjoy their AI immensely - I have noticed their behavior is superior to that of pc2. Does anyone know if AMS2 AI experiences tire wear? I still can't believe pc2 doesn't. I thought it was just pc2 not pulling it off rather that it not existing.

To answer your question directly, though - look, compromises and simplifications are already being done. The technology exists to have complex fluid dynamics dictate downforce in a very real way. But it's too computationally heavy to include that in a video game like pc2. That's a reasonable compromise - you simplify the air dynamics model for the sake of gameplay. That balance with the tire model is perhaps also something that should be looked at, seeing as how the AI doesn't even experience tire wear.

If we set what the AI experience is like in sims from 15 years ago as the standard, which we should, then I don't think its unreasonable to say that the balance between sim accuracy and gameplay is off.

[Mahjik]

I understand all "AI" in these video games are all faked and not true AI, as your video demonstrates. I'm saying you can have better results faking the AI if you go back to the fundamental established in older sims.

I cannot agree with that statement. The AI of those older sims were awesome, for the time they were released. Playing them today, you will see some no so great behavior. You are also mentioning titles that had AI developed for specific car types. If you look at titles like the F1 Series and ACC, you'll notice the behavior is more complete. Why? Because the AI is only built to handle a specific type of car.

PC2 uses most of the same bones from GTR2 as to how the AI works. i.e. it still uses AIW's to give the AI a path its needs to follow. Sure, the actual AI logic has been rewritten for how it makes decisions, but it's same principles for how it works. If you ever tried to put non-GT cars into GTR2, you would have seen some really bad AI behavior. Why? Because the AI was tuned for the GT cars. Trust, the AI of back in the day isn't any better than what we have today. It's just that the titles developed back then were smaller focused titles which hid the issues we see today.

[Javaniceday]

I cannot agree with that statement. The AI of those older sims were awesome, for the time they were released. Playing them today, you will see some no so great behavior. You are also mentioning titles that had AI developed for specific car types. If you look at titles like the F1 Series and ACC, you'll notice the behavior is more complete. Why? Because the AI is only built to handle a specific type of car.

PC2 uses most of the same bones from GTR2 as to how the AI works. i.e. it still uses AIW's to give the AI a path its needs to follow. Sure, the actual AI logic has been rewritten for how it makes decisions, but it's same principles for how it works. If you ever tried to put non-GT cars into GTR2, you would have seen some really bad AI behavior. Why? Because the AI was tuned for the GT cars. Trust, the AI of back in the day isn't any better than what we have today. It's just that the titles developed back then were smaller focused titles which hid the issues we see today.

I can certainly agree that the AI in GTR2 and other older sim titles masked their short comings by focusing solely on one class of car. What I'm also saying is that to reach the effect those titles gave the player/driver, perhaps compromises need to be considered to reach the standard set by previous titles. What that compromise is is dependent on whatever factor one wants to consider.

But I think, the choice of not having the AI experience tire wear in a title where tire wear is fundamental part of the experience is not a good choice. Other choices/compromises should be made. And obligatory, nixing tire wear entirely for the next title is also the wrong choice. The better choice, as Reiza is demonstrating, is to simplify things by either a-having only one dry compound, b-smaller car selection, or c- whatever other compromise to be had - among them could be having a simpler tire model for the same reason SMS didn't use state-of-the-art fluid dynamics simulations to create their downforce model.

But the end result must be, in my opinion, the standard set by titles of 15 years ago. I don't know if GTR2 or Grand Prix 4 also didn't have AI experience tire wear. But we sure didn't notice.

[cpcdem]

I think that we just had so much less expectations in titles 15 years ago, it was just awesome having even some small degree of realism back then, so the AI felt good enough. Try driving against those AIs now, treating them and fighting them like real humans, do you still find their behavior realistic? I would really doubt that.

I do not know for sure if, how and which today's games have wear for the AI, but even those that don't have it realistically calculated at realtime, they can still simulate it. For example the devs can run a simulation with driving a car in given conditions for 20 laps with the player physics, see what the average wear was for every lap, and then hardcode this for the AI for those conditions. Even arcade games like Grid Autosport had tire wear, for endurance races...

Actually it may even not important to think in tire wear terms at all. AI only knows how fast it should go out of each corner and devs try to make the pace in every corner match that of a human player. If they run a simulation and see that the human player loses say 1% of pace in every lap for whatever reason for otherwise static conditions, then they can model this to the AIs as well. It is not important if the loss of pace is due to tire wear or tire overheating or both, as long as the end result is relatively close for both the AI and the player, then it's good enough.

It just takes a lot of time and effort to calibrate a lot of those factors to make them match the player in as many conditions as possible. Some devs are really devoted and are prepared to put that huge amount of effort in this aspect and many other parts of the sim, too, to make it really shine, while others are not really that committed to it, that's what makes the big difference at the end IMO.

[dault3883]

@CPCDEM i remember playing gran turismo 2 on ps1 and the ai spinning out constantly going through the cork screw at laguna seca

let me tell you i learned how to drive that turn real quick as it was a spot i could make up lots of time at

[cpcdem]

@CPCDEM i remember playing gran turismo 2 on ps1 and the ai spinning out constantly going through the cork screw at laguna seca

let me tell you i learned how to drive that turn real quick as it was a spot i could make up lots of time at

Yeah, and still now, in every sim/simcade etc, in every track, there are some corners were the AI is a bit or extremely slower or faster than the player. Putting a lot of effort in calibrating them improves things, but it never really feels "real". It's the same in every game, take also CS:GO for example, when you practice with bots you quickly figure out the patters in how they move/shoot and learn to deal with them. Then, when you play against human opponents, almost nothing you have learned will help you! Unfortunately we are still far away from having realistic AI in computer games.