Why I won't be doing an adaptive-AI championship...

... in short: because it's too unpredictable.

After doing several adaptive AI races in the Hyundai at Bathurst, I could see the AI zeroing in on my pace.

For my next race I put myself at the back of the grid in 25th, ready for the challenge...

... lights out... by turn one I had passed about 10 cars, another 8 by Griffin's. And yes, 6 more across the top of the mountain. As I hit the brakes at skyline for the first time I was leading.

Last to first by skyline on the first lap!!!

I then proceeded to lap an average of 2.4 cars per lap over the rest of the race to finish first having lapped the car in second place in the final corner. So in ten laps I passed every car twice!

I've never lapped anyone before and I was also a tad disappointed that the AI ignore any blue flags they might be getting (at least in the corners - they seem to pull over on the straights).

My race result is below. I have the fastest lap by 18 seconds! And it wasn't that great a lap given the traffic I had to contend with!



This sort of unpredictability would kill a championship.

So... I return to my gentle plea to allow the user to select the AI level for each track when they create a championship. Please please please. (And I know it's one of a million things we all want...)

PS Still love the game... it's my one and only which is why I want it to be the best it can be...

I'd also like to be a beta tester... is there a form I can fill in?

 

AI initially used to work not well. Then something was changed in it, and I became a fan of it. Once one has understood how the process to "train" the AI works, per class and track, it was okay. However, having to do all that training was what finally nevertheless killed the concept for me again, an I now use fixed AI settings of 97 and 98 again.

 

Something's screwy here.
AI lap-times at around 2:38 means it was running at level 80.

Off the top of my oddly-shaped head, I can think of three possibilities.
1 ) You accidentally set it to 80 instead of Adaptive. They're right next to each other in the drop-down menu.
2 ) The index-file got corrupted somehow. If it's corrupted R3E will delete it and create a new one. One way to check is to see if it contains any other entries than the race you just ran. Alternatively, check the file-size, a fresh index-file takes up about 270kb.
3 ) Somehow there's a bug in the championship use of AAI. I've never run a championship, so I wouldn't know.
But I'll try and set up a mini-championship using the WTCR18 and see if I can replicate the issue some time this week-end.

 

Just ran a 3-round championship with AAI, and it works fine.
Bathurst, Imola and Laguna, and all races ran with the correct AI.

One more thing it might be.
You mentioned you'd run a few races first, to get the AI up to speed.
Did any of these races include a major off from you?
The reason I ask is that if you e.g. whack the wall, and spend 12-15 seconds getting the bloody thing turned the right way around again, those seconds are added to your average laptime.

 

Thanks Cheerfullyinsane - I really appreciate the effort you have gone to in having a look at this!

In response:

No - definitely not. Slightly insulted (although I know you had to ask!! )

Ummm… you might be on to something here. As a relentless data collector, I can tell you that my average lap time on the previous ten laps was 2:27.782. There was one 3-4 minute lap in there due to a massive AI pile-up at skyline (the AI drivers need to learn more about managing weight transfer!).

The ten laps before that averaged 2:21.814. So my average had just dropped by 6 seconds per lap prior to the race.

That is likely to be it I guess but as I'm sure everyone appreciates it's not ideal as a bad lap (in this case not my fault) means the following race will be a waste of time. It would seem to me to be a (relatively) easy fix to take an average of your 6 best laps in the last 10 or something like that. This would pick up the variability but not the outliers.

Thanks again for the response.

 

No problem, I happen to enjoy this sort of problem-solving.
(And yes, I'm a little weird. )

Couldn't agree more. I've gotten into the habit of manually deleting any extremely bad laps after each race, but it shouldn't be necessary. So as you say, remove the fastest and slowest lap, and take the average from the rest.
In fact, I had a bit of a think on it last night, and there's an easier way to do it.
Simply have the index-file list your 5-10 fastest laps throughout your 'career'.
As long as all your race-laps are slower than those in the index-file, nothing happens to the AI level.
Once you start improving, the AI then gets a notch or two upwards. But once at the right level, it'll never adjust downwards, meaning offs, spins and the like become irrelevant.

@Christian Göpfert Any chance you can pass some of this along to the AI guys?

 

Yes - I think something like that would be much better.
I just did another race at Bathurst and, in its current configuration, it looks like it takes 2-3 races to recover from a real shocker.
Editing the file - now there's a thought! (I mean seriously - just to clean out the %^%$).

 

It takes ten laps. How many races obviously depends on how many laps each race is.
So a major mistake on the Nordschleife is a bit of a bugger.

 

I tried the TCR cars at Bathurst just to have a look. I also tried the AI at various set levels (between 90%- 95%) rather than Adaptive. My lap times were similar to yours, but the AI really seemed to have difficulty through the Esses with regular crashes into the wall and holding everyone up.
I suspect this might be more of a Bathurst issue than how the Adaptive AI works.

 

Out of interest do the ai have an actual racing line to follow on each track?

 

It is. Like any AI there are some combos where it performs better than others.
Give them a bit of downforce, and they do quite well at Bathurst.
Equally it does fine with the TCRs at e.g. Hungaroring.
But the combo of tail-happy cars + downhill turns + walls isn't a good one.

Yep. Or rather, I'm 99% sure of it.
If you check the results.txt file after a race, it'll reference an .AIW file.
The files are encrypted, but I'm pretty sure they're there nonetheless.

 

Good question, does the AI use the same car setup as the player?

Also, looking at my aiadaption.xml file, it is veeery long. How do you know which track to find your latest times etc?

 

I just did and we're discussing possible improvements.

That's correct, AIW stands for AI Waypoints if I'm not mistaken and they're given a racing line plus some alternative corridors for certain situations. It's rather intricate and one of the reasons physics changes need lots of time, cause all those lines and corridors have to be adapted for all cars.

This is also correct.

 

I wouldn't. Slight misunderstanding there, I meant that if you ran Bathurst in a car with more downforce.
F4s at Bathurst is phenomenally fun for example.

You look for the track and car IDs.
Here ya go:

Spoiler: Track List

Anderstorp
Grand Prix: 5301
South: 6164

Bathurst: 1846

Brands Hatch: 2520

Chang International Circuit
Full Curcuit: 4253
D Curcuit: 4252

Circuit Park Zandvoort
Grand Prix: 1678
National: 1680
Club: 1679

Circuit Zolder: 1684

Eurospeedway Lausitz
Grand Prix: 2468
Automobilsport: 3291

Falkenberg: 6140

Gelleråsen
Grand Prix: 5925
Short: 6138

Hockenheim
Grand Prix: 1693
National: 1763
Short: 1764

Hungaroring: 1866

Imola: 1850

Indianapolis
Grand Prix: 1852
Moto: 2014

Knutstorp Ring: 6137

Lakeview Hillclimb
Full Run: 1682
Reverse: 2181

Macau: 2123

Mantorp Park
Long: 6010
Short: 6167

Mazda Laguna Seca: 1856

Mid Ohio:
Full: 1674
Chicane: 1676
Short: 1675

Monza
Grand Prix: 1671
Junior: 1672

Moscow Raceway
Sprint: 2473
Full: 3383
FIM: 3683

Motorsport Arena Oschersleben: 2384

Nordschleife
Nordschleife: 2813
VLN: 4975
Tourist: 5093
24 Hours: 5095

Norisring: 2518

Nürburgring
Grand Prix: 1691
Short: 2011
Sprint: 3377

Paul Ricard
Solution 3C: 2867
Solution 1A: 4246
Solution 1C-V2: 4247
Solution 2A Short: 4248

Portimao
Grand Prix: 1778
National: 1783
Club: 1784
Club Chicane: 1785

Raceroom Hillclimb
Full Run: 1705
Reverse: 2214

Raceroom raceway
Grand Prix: 263
Bridge: 266
Classic: 264
Classic Sprint: 265
National: 267

Red Bull Ring Speilberg
Grand Prix: 2556
National: 5794

Road America: 5276

Sachsenring: 3538

Salzburgring: 2026

Sepang
Grand Prix: 6341
North: 6578
South: 6579

Shanghai
Grand Prix: 2027
Intermediate (WTCC): 4041
West Long: 4042

Silverstone
Grand Prix: 4039
International: 5816
National: 5817
Historic GP: 5862

Slovakia Ring: 2064

Sonoma Raceway
WTCC: 1854
Sprint: 2016
Long: 3912
IRL: 3913

Suzuka
Grand Prix: 1841
East Course: 2012
West Course: 2013

Spa-Francorchamps
Grand Prix: 3870
Classic: 4542
Combined: 4543

Stowe
Long: 6055
Short: 6056

Twin Ring Motegi
Road Course: 6658

Zhuhai: 3464

Spoiler: Car List

FRJ (20): 253
Aquila (24): 255
HillClimb Icons: 1685
GTR1 (41): 1687
GTR3: 1703
GTR2: 1704
GTRX: 1710
Touring Classics (39): 1712
GTO CLassics (17): 1713
Group 5 (44): 1708
P1 (15): 1714
Silhouettes (24): 1717
DTM 13: 1921
WTCC 13: 1922
P2 (46): 1923
BL Cup: 2322
ProCar: 2378
ADAC 13 (24): 2922
DTM 14: 3086
ADAC 14: 3375
DTM 92 (25): 3499
WTCC 14: 3905
DTM 15 (24): 4260
ADAC 15 (19): 4516
WTCC 15 (23): 4517
FR2 (22): 4597
Audi TT 15: 4680
Audi TT 16 (20): 5726
NSU (21): 4813
F4 (20): 4867
F3 (26): 5652
Audi TT RS (20): 5234
DTM 16 (24): 5262
KTM (29): 5385
FR US (20): 5383
WTCC 17 (16): 6309
BMW 235i (23): 6344
FRX (20): 5824
Porsche GT4: 6648
Porsche GT3 Cup (32): 6345
Zonda Cup (25): 7110
WTCR 18 (25):7009

(Number in parentheses is the number of available liveries for that particular class)

Tah muchly.

 

Our single best result, or the average of our all-time top 5 or 10 best results?

For wildly inconsistent drivers like me being matched agaisnt my personal record sounds painful

 

How cool would it be to be able to choose how the calculation derives

Options could include

1. Average results of last 10 laps
2. Average all time best 5 laps
3. Average results of the last 15 laps excluding the bottom 5 results (effectively ignores those slow laps due to incidents)

Option 3 is my preference

 

I wouldn't worry too much about it. The AI lap-times that the player-times are compared to are still going to be based on the average speed of all the grid (or so I assume).
So even if you have variations, it shouldn't result in a blow-out.

 

I thought it was based on the highest placed AI driver, not the average of all AI drivers ...mmm

I wish there was a definitive explanation sticky in the forums from the developers on how it is calculate albeit I guess they don't want to give away their formula to their competitors

 

If it's proprietary, it's gotta be the worst kept secret in the world.
Here's how it works (or at least as far as I can figure):
After each race, you get a results.txt containing the fastest lap-times of each driver (you and AI both).
All the fastest lap-times of the AI is then averaged, which subsequently becomes the indexed lap-time for that particular level.

In order for the lap-times to not vary too much between each race, once you start racing the same AI level there's a further averaging when adding the new times to the index-file.
The current lap-time is weighted by the number of times you've run it, and then averaged with the newest lap-time.
Or in other words;
New indexed lap-time = ((Old indexed lap-time * counter) + new lap-time) / counter + 1

The whole thing is actually pretty clever. Since the AI fastest times are compared to your average times, not only do you end up somewhere mid-pack, you're also likely to move upwards assuming you're consistent.
Suffice it to say that the AAI is one of the major reasons I keep coming back to R3E.