Observation on AI in GT7

[Samus]

the AI is not falling back due to a script, you are literally outpacing them.

Yes they are. This has been tested and proven since GT6. The AI will run laps 10+ seconds faster or slower than it did in the early laps of the race depending on what it is planned to do, or how it adjusts to the player pace.

As people have said, the AI is not straight up programmed to go as fast as possible and try to win like other games. It's programmed to give you the illusion of an exciting race.

 

[Sven Jurgens]

I'm a very consistent 'lapper'. In the 'scripted' Autopolis race my lap times were like clockwork all race long. My FL was 2:27.9, but standard deviation from 2:29 was very low. The 2:27.9 outlier was when drafting several AI that gave up near the end, draft pass after draft pass. The rest were all 2:28.5 to 2:29.5. I also ended on 2:29s with the AI still further giving up and giving me the lead with 6 minutes left, then dropping another 12 sec back. The AI FL was 2:26 right from the beginning.

I've driven N24 thousands of times and also there I lap like clockwork, my lap times mostly affected by rain. Which is another area where the AI completely deviates from the player with different physics. Yet while driving consistent laps I see the AI speeding up and slowing down around me. Fast when they come out of the pit behind me, slow when they're ahead and it's time to move up to that position.

In dry races on the Nord, with very consistent laps, AI is fast at first, taking up position well ahead, then drops back. I hardly ever get the fastest lap in those races, usually it's smashed by 20 to 30 seconds! during one of their catch up boosts. In the end I either win far ahead or with the AI capitulating in the final lap. And sometimes when I'm in a slower car the last 1 or 2 cars teasingly stay just ahead. The rest of the field far back.

Sure it's anecdotal, but it has been consistent with everything I have observed so far in 163 hours of not grinding driving time. I'm not going faster, the AI has a speed profile to go fast first, fall back later. It's the same in many other games with AI although mostly in arcade games like DriveClub which did the same thing. However in GT7 the rubberbanding is more obvious than in anything I've raced before. Maybe because the races are much longer the rubber banding stands out more.

The intent is indeed the illusion of a close race. GT7 does it in such a wonky way that the illusion never holds up.

Maybe it will be better when the pit stops are incorporated in the AI planning. Atm it seems they simply pit when laps remaining goes below 1.0, or when tires or track wetness reaches a certain point. Problems here are that the laps remaining estimate is often too optimistic, which explains the AI sometimes running out of fuel not pitting on time and often having to pit again in the final lap. Failure to correctly calculate the fuel still needed to end the race.

They don't plan their pit stops and won't change tires for intermediates or wets until they get the sudden switch from all fine to got to slow down because of water on the track. Then they stop again for heavy wets when the next trigger is reached, but they never look at the radar, always pit too late. They only go directly to heavy wets if they got caught out on the track and the rain percentage has already reached the heavy wet trigger.

When the track dries it goes the other way. From Heavy wets to Intermediates first, then from Intermediates to Slicks while it's still too dangerous for yourself to drive on Slicks. The AI doesn't have much of a problem going over the still wet parts, it's like for them the whole track is dry already. But they can't figure out to stay on Heavy wets one lap longer and skip the Intermediates... It's a simple, IF current water% > X move up one tire compound. IF current water% < Y move down one tire compound. Not even some scripted intelligence, just a simple rule.

And since their pit stops aren't planned, merely reactionary, they are not incorporated into their time tables either. Hence their silly speed boosts on the outlaps to get back on schedule. But they other directive to stay x distance ahead of player until y will make them slow down again after you pit. Best is to pit when the AI pits to avoid the extra rubber banding. And always tank a little more than laps remaining shows!

 

[O604]

GT is a simulator but it's a Human in the Loop application. It's very different from a pure computer simulation because human behaviour is very unpredictable.

Also GT is very different from a lab environment where simulations are studied. Subjects will behave differently in a lab setting compared to unsupervised, on the couch at home.

So what we have is basically a Human in the Loop simulation experiment in an unknown setting. Imagine attempting a scientific experiment this way.

 

[Sven Jurgens]

GT is a simulator but it's a Human in the Loop application. It's very different from a pure computer simulation because human behaviour is very unpredictable.

Also GT is very different from a lab environment where simulations are studied. Subjects will behave differently in a lab setting compared to unsupervised, on the couch at home.

So what we have is basically a Human in the Loop simulation experiment in an unknown setting. Imagine attempting a scientific experiment this way.

True, but it's an experiment that has been going on for over 25 years now!
It seems the more data PD gets, the worse the AI gets

 

[Tenacious D]

People who think this A.I. is bad need to experience Grid Autosport or any of the Forzas.

 

[O604]

True, but it's an experiment that has been going on for over 25 years now!
It seems the more data PD gets, the worse the AI gets

I would argue that the AI has improved. But the average abilty of players is lower, due to increased popularity of games in general.

More data isn't necessarily better. For example more data used to make things easier for beginners would work against someone like you.

 

[fmq75]

People who think this A.I. is bad need to experience Grid Autosport or any of the Forzas.

well,while it‘s true there is one thing i still like more in Forza: it can be a very unpredictable mayhem.Somehow i like this more than this robots on rails.

And the rewind helps in worst situations.Here you are screwed if the AI is in carnage mode.

 

[AutumnalGlow]

People who think this A.I. is bad need to experience Grid Autosport or any of the Forzas.

Forza Motorsport's AI is not this bad. Horizon is a totally different game and is irrelevant to a track-based AI discussion.

And conveniently, they've already talked about their plans to remove rubber banding and cheating from the next game's AI. Whether that will 100% happen is still unknown, but Turn 10 isn't likely to throw all of the AI improvements away. Unlike Polyphony, they don't want to have 5 years of development with nothing to show for it.

 

[Samus]

People who think this A.I. is bad need to experience Grid Autosport or any of the Forzas.

All games have issues with their AI but most of them get the core requirement right - outright speed. Where the other games you mentioned have issues is with aggression and decision making, but you can adapt to that and still have good races.

GT on the other hand has the fundamental issue of slow speed, and we don't get to see many aggression issues because PD space them all apart.

In short, given the choice id take the fast but agressive/sometimes sloppy AI of those other games over an AI that is just very slow and programmed not to try and win.

 

[MrNukki]

If I remember right, only one person was responsible for coding the AI for GT7, feel free to correct me though. Considering how herculean of a task it is to create a competent artificial intelligence system, I'd say that that person did okay. It's a stupid AI, but it works for what its worth.

At least this AI doesn't drive on rails 100% of the time, employs basic pitting strategies and kinda compensates for oversteer? Leagues ahead of GTs of old, or even Forza Horizon, where the AI literally glues to the ground and has no physics off road. We've yet to reach the Skynet levels of AI that Sophy promises to deliver, if that'll even happen at all.

 

[Samus]

If I remember right, only one person was responsible for coding the AI for GT7, feel free to correct me though.

Yes, it's the same person who has been doing it since GT5. On GT5 and GT6 there was a second person credited, for GTS and GT7 it was him alone.

Nothing against him personally of course, but It doesn't seem he is up the job or at the very least, he needs help.

 

[Tenacious D]

you can adapt to that and still have good races.

I haven't touched any of my Forzas since I played Forza 6 for two months. I downloaded FW7 for PC but never installed it, and the bots are the biggest reason. So make of that what you will. And as far as I'm concerned Grid doesn't exist. I grew to absolutely hate that series.

But if you think this A.I. is slow, you've obviously missed all the posts complaining that this race or that is impossible to beat. Maybe you mean slow around turns, but you typed what you typed...

Plus, for all the talk of all the races scripted from start to finish, I still don't agree. Yes, events are clearly scripted, like the Huracan sliding off the track on LeSarth. But then, how does Kokubun end up in third place at the Tokyo WTC600? Or no, fourth? Wait, seventh? Or this time first? Did they not get the script? Am I dreaming all those bot scuffles in replays?

 

[Samus]

I haven't touched any of my Forzas since I played Forza 6 for two months. I downloaded FW7 for PC but never installed it, and the bots are the biggest reason. So make of that what you will. And as far as I'm concerned Grid doesn't exist. I grew to absolutely hate that series.

But if you think this A.I. is slow, you've obviously missed all the posts complaining that this race or that is impossible to beat. Maybe you mean slow around turns, but you typed what you typed...

Plus, for all the talk of all the races scripted from start to finish, I still don't agree. Yes, events are clearly scripted, like the Huracan sliding off the track on LeSarth. But then, how does Kokubun end up in third place at the Tokyo WTC600? Or no, fourth? Wait, seventh? Or this time first? Did they not get the script? Am I dreaming all those bot scuffles in replays?

Races are hard to beat because the AI start so far ahead, not because they're fast.