AI : Editing the STRENGTH of a mod. need ADVICE

[cangrejin]

hello everyone... as the title says... I'm EDITING the TALENT files (.rcd) of a specific mod.... which in my opinion... are somewhat disorganized.

things I ALREADY DID:

- Values like "reputation", "courtesy", I left them as they were (they were not modified)....after much reading...they warned that they don't matter or are useless.

-values like "composure", "crash", "recovery", "completedlaps" were readjusted .... although many say that little affect the ai
(It may be that "Composure" does something...so if I made a SCALE from best to worst) all of this...I made a scale from best to worst...in a reasonable way.

And now the IMPORTANT:

AGGRESSION: I noticed that the fast ai and the slower ai had a very marked aggression gap (or difference). (100% to 65% from best to worst)

and to make it easier...here is the table of "MinRacingSkill" and "SPEED"
(minracingskill...every once in a while, the ai does a slow lap...you know, they're not always on their best lap, this value is the PERCENTAGE, it means HOW MUCH they can drop their lap pace...(they'll never drop more than that value)

HERE THE TABLE OF MINRACINGSKILL AND SPEED:

TO CONSIDER:

POINT 1)- As you can see the AI are in groups....
this list DOES NOT (NOT!) MEAN that THE DRIVERS will qualify this way (or finish the race) in this order
it means that THE GROUPS (not the drivers) WILL END this way....
...I mean, in each group there will be a fight for position...but rarely will they be able to reach the next group...

POINT 2)- some groups overlap a bit, (they get quite close to the next group, or even occasionally they manage to overtake a car from the upper group)
EXAMPLE, in real life, JARNO TRULLI sometimes managed to be ahead of FRENTZEN (frentzen's SPEED value is 90, and trulli's is 88, ....but if frenzen loses rhythm or speed in some lap (frenzen : minracingskill= 85), trulli could pass it.

Some groups are closer to or overlap with other groups than some others. (the group of frenzen and the group of trulli....or the group of herbert and heidfeld)

POINT 3)- I intentionally said that some cars can NEVER overtake others...example: mazzacane can never finish ahead of diniz. another example: trulli and his group will not be able to finish in front of barrichello

I FORGOT AT WHICH POINT I'M GOING)- so that there are interesting fights...I made some cars faster than their group...but slower at their slowest pace (minracingskill)
example: JEAN ALESI, he is faster in his group...so be slightly ahead...in he can even fight HEIDFELD...but slower laps from him (value = 80) will cause a fight and advances between both groups.

I forgot something VERY VERY IMPORTANT....
As I clarified in another thread, there is a problem in this mod (and in almost all rfactor mods):
-ai only does ONE STRONG lap in qualifying (makes the others very slow)...so if a fast ai was NOT able to do his best lap on his one strong attempt...there will be weird mixes of ai (like hakkinen in position 20...or mazzacane in position 5)

That's why I made the AI with a VERY MARKED DIFFERENCE... how did I do it? :

in ALL ".gdb" files (they are in the files of each circuit)... I UPLOADED the value and line "AIRANGE=" (airange) TO "0.8" (previously it was 0.1) (also if you want you can take it to "1.0")
what does this value mean?

is the OVERALL "compression" level of ALL ai's in that RACE....at 0.1 means the slowest ai will be at 90% of the fastest ai (!!!!!) (0.1 means 10%... at 90 and 100%)
(explained in another way, all the ai will be compressed to 90% ... of its real values in the talent rcd)
(this is VERY little difference between HAKKINEN AND MAZZACANE, in the mod....if hakkinen had a bad ranking...he could go down to 20th place)

(also with 0.1, having SO little difference between ia, the original values of the talent rcd files were too EXAGGERATED... to be able to make a difference)
(explained in another way: hakkinen had a SPEED value of 95....and mazzacane...50....but being so compressed (0.1 = 10% = between 90 and 100) you didn't notice the difference)

by taking the value to 0.8 (80%), (between 20 and 100) THE AI WILL ONLY BE COMPRESSED BY 20% (of its real values in the talent rcd).

if you set it to 1.0 ...there will be no compression at all...the circuit will take the REAL values from the talent rcd files)
(also with 1.0, when handling real values of rcd files, you can be more moderate with the differences...
(explained another way: hakkinen now has a SPEED value of 99....and mazzacane...81....this is closer than the original value, but since there is NO compression (1.0 = 100% = between 0 and 100) you notice the difference, see how they respect the values)

remember... the compression is towards the UPPER part... that is xx at 100%

 

[cangrejin]

THANK YOU FOR YOUR PATIENCE...AND SORRY FOR SO MUCH READING.

I just have one question for you:
any advice? some help? Is this table of values "okay"?

 

[cangrejin]

(by the way....if you think mazzacane was VERY bad...you should know that:

- was "consistent", although he is clearly the SLOWEST, finishing 12 of 17 races.
-had a reliable minardi that year, only 4 mechanical failures in HIS car)
-I didn't finish last in the general classification, the PROST had the WORST year of their lives. (even being very superior car)
-finished 8th at the nurburgring...in the rain, when others didn't survive.
-was 3rd place ahead of hakkinen in usa, by 4 laps, with better tires.
)

That's why I decided that the slowest ai has a SPEED value of 81 (no f1 driver, no matter how bad it is, will go "at half throttle", and even giving a bad lap it will be 80% of its value (80% of the 80% speed value)
That's why I modified the AIRANGE of the circuits... to 1.0
originally... with the original rcd (of mazzacane)... his SPEED value of it was 50. with the ai range at 1.0.... mazzacane crawled very slowly (exceedingly slow) very unrealistic.

 

[cangrejin]

another driver...luciano burti, only ran one race...so i thought he shouldn't have a good ai, like herbert or irvine (they were full time)...just an anecdote about burti that year .


UPDATE EDIT:
Silly me... I forgot to mention the most important thing of all...

that obviously modifies the values

-Auto Line Smooth=3
-AI to AI Collision Rate=40
- AI Brake_Power Usage=0.35000
-AI Brake_Grip Usage=0.97250
AI Mistakes=0.20000
- AI Limiter=0.00000 (very important, otherwise, none of the rcd will work !!!!!!!!!!!!!)

I just have one question for you:
any advice? some help? Is this table of values "okay"?

 

[jgf]

"composure", "crash", "recovery", "completedlaps"-
Used to sort AI if you skip or accelerate qualifying sessions

AGGRESSION-
Basically controls how close that AI will get to other cars (AI or player); high value means they will go fender-to-fender with you, lower values makes them back off. This does not make AI directly faster, but lower values make them spend more time avoiding each other and you. Too high and they will sideswipe you, and each other, often causing damage; too low and they bob and weave all the time trying to stay apart.

Skill-
Consider this as "experience", higher values will hold the line better and make fewer mistakes; this will not directly make them faster, but they will have better, and more consistent, lap times.

AIRANGE=
This is the spread of AI around the settings in the track files. In the .aiw file are (values vary):
WorstAdjust=(0.8600)
MidAdjust=(1.0800)
BestAdjust=(1.2200)
These correspond to the the AI menu settings of 80%, 100%, and 120%, and correspond to the speeds set at each waypoint on the track. So that midadjust setting of 1.08 means at 100% AI setting in the game menu the AI will run 108% of the speeds set in that track's aiw file. The airange sets how far the AI will vary from that; a setting of 0.1 is 10%, so the AI will be +/-5% of the adjusted (worst/mid/best) settings of the track's aiw speed. A setting of 0.8 means the AI can vary +/- 40% of the speed settings - some will be amazingly fast (possibly to the extent of flying off the track) and others will drive like they're stuck in first gear. This is then further modified by talent file settings.

Two other things can greatly improve AI performance:

Give the AI setups for each car and track; first set "Fixedsetups=1" and "FixedAIsetups=1" in your plr file. I use my own setups but with top gear extended a click or two (otherwise the AI are likely to blow the engine). For example, you create a setup for your Mustang at Goodwood, set the top gear a little higher and save with different name; copy that setup (svm file) to the root folder of that Mustang (where the hdv and pm files are located) and rename it Goodwood.svm, now any AI driving that car at that track will use that setup. Look in the track's gdb file for "AISetting" for the proper name for the setup, otherwise it will be ignored.

And train the AI at each track, this must be done for each type car at each track. There are numerous tutorials for this online. https://www.racedepartment.com/threads/how-to-make-ai-learn-the-racing-line-for-a-track.88801/

 

[MJQT]

Fantastic stuff @jgf – I always love to see more clear info about these parameters written up!

One small question: is the "track learning" or "training" of the AI at a given track is dependent on the AI corridors defined in the AIW file? in my experience, it seems that way. In other words, if the corridors are too narrow, the trained AI may not use the full extent of the track you want him to use, and if the corridors are too wide, the AI may take risky lines out into the grass/runoff in pursuit of faster times (which leads to faster hotlaps but poor racing). Is this correct?

@cangrejin one quick response to your proposed parameter values – you may want to experiment with lower values of "AI to AI Collision Rate" like 18 or 20. I sometimes see better AI behaviour using a value of 40, but also sometimes see the AI drive erratically. Try it out, and see which you prefer.

 

[jgf]

The training will make the AI adhere more closely to the "ideal" racing line as defined by the aiw file waypoints, so they will only be as good as that file allows (if you get an "improved" aiw file for the track you will have to train again). They will only deviate from this line to pass, avoid collisions, or from the occasional driving error, so the width of the path only comes into play in those situations. At the end of a training session you may get a notice like "80 yards left", this means the line the AI drives is still 80 yards longer than the "ideal" line, repeating the session should lower this; it may take several sessions but you can usually get that to zero or very close. If after several sessions you still have a notable discrepancy it is most likely the waypoints were designed for a quite different type of car and your car cannot follow that line exactly (a GT3 Viper and an F1 car would have quite different lines around the same track), but as long as it is using a reasonable line consistently the training is working (and "notable discrepancy" is relative, 50yards at Mid-Ohio is a major problem, at Nurburgring it is virtually irrelevant).