So I'm still trying to work out how to get the adaptive AI up to speed without having to run too many races. And I think I might have cracked it..... If I'm right (and right now that's a big if), I'll be able to supply a complete index of lap-times for every car for the major layout of every track. And for every AI level. No more training, assuming I can talk @pixeljetstream into re-coding the Primer Tool with the new data. But I do need data to check the algorithms against. So far it's looking good, with the predicted lap-times getting within 2% of the actual lap-times. But that's against the same data-set that the algorithms are based on in the first place, so not really conclusive. What I need from you are your aiadaption.xml files. The more data in them the better, but I'll take anything. It'll be in your Documents\My Games\SimBin\RaceRoom Racing Experience\UserData\Player1 folder. Simply zip/rar it, and attach it to your message, and I can compare it to the predicted results. Again, the more files I can get my hands on, the better, but I figure I need at least 5 or 6 before I feel confident enough to publish anything. There is however one massive caveat to this. I can only use XML files that only contains data from post-January '16. The AI update in the Jan '16 patch made the AI a LOT faster, so any XML file that contains data from before that will completely skew the results. So please check when your XML-file was created (right-click it then choose Properties, and check the date) before uploading anything. Thanks in advance.
The XML-files? There is no fast way. They're generated/updated automatically whenever you race against the AI, recording car, track, lap-times and AI level for each race you do. So the only way to get them is to do races. And if you're thinking "I'll just do a lot of 1-lap races, that'll be quicker", then stop thinking it. Yes it'll be quicker, but the lap-time generated for the AI is an average of the fastest lap for all the AI drivers. So in order for them to actually reach the fastest lap (and thus not skew the result for that level), they need some time to get spread out. Four laps is the minimum, in order to get a decent average.
A little preliminary excerpt. So far only at AI level 100, since I want to make sure the basis is sound before adding in more variables. And yes, there are a few hiccups with some of the combos, but since I plan to bracket each entry with +/- 4 levels anyway, the adaptive AI should take care of the rest. Still very early days yet, but I'm pretty happy so far.
This gave me a moment of pause, because i was thinking to myself, "Dang, January was only 4 months ago." And then I noticed the '16 part there and though, "Dang, Im getting old." Here's my AI file. Windows says it was made in January of 2017. Hopefully it can offer something useful. Dont have much else to say about this, as digging into those files is over my head, but pretty interesting project here.
Ah, thanks. Just what I...... Wait a minute, you've run the FR US at Salzburgring??? Why in gods name would...... Never mind, not my place to judge. Thank you. And just the XML files will do nicely. Assuming I can get the number-crunching right, all I really need is someone to code something to insert the times automatically. I mean, I've been mucking around with the AI for so long I can simply insert the data manually, but for those who aren't so inclined, an automation of some sort would be nice. A bit more time-checking, and still looking pretty good. There's still the ugly wart here and there, but an average variation of less than half a percent? Make that 1.7%. Slight flaw in the spreadsheet. Still pretty good. That's roughly 3 AI levels. Still need to collect data for the last 4-5 classes, and the shadow of the Nordschleife is still hovering over me. Once that's done I can get to work with adding the different AI levels. Oh, and one thing I've learnt in the process. If you do 4 laps of Hockenheim GP with the NSUs..... Bring a packed lunch.
More good news. I finished getting data for the last of the car-classes, and have begun introducing AI variations into the equations. There are still a few bad spikes. Anything over 3% variance makes me rather unhappy. On the other hand, an average variance of 1.6% should be good enough. Insert the predicted AI level, one +5, and one -5, and the adaptive AI should be able to take it from there. Since the baseline is 100% AI, it'll presumably be more accurate when the needed predictions stay around there. Say, between 90 and 110. Still think it'll be able to do the extremes, it'll just be more precise in the above interval. Ha, the bloody thing actually works. Don't get me wrong, there's still a ton to do. I still need more data to compare against. I still need to do some practical tests; i.e. try inserting some predicted AI levels on combos I haven't run, and see how it copes. Not to mention I still need to do four laps of the Nords to get the last bits of data. Plus the fact that right now it's just a look-up table with absolutely zero practical application for the average user. But still..... This has been a very good day.
To sunny here to spend indoors, but just a quick check as I don't want to waste your time, will my rather slow speeds skew the figures in any way?
Nope. I couldn't care less about the player-times. They don't get into the spreadsheet. All I care about is AI data. (And what's a 'sun'? Some sort of Scottish beverage? )
I will brush up on your training manual again and do some over the next week, except tomorrow because the sun's back again and I need that vitamin D
Nah, don't bother. I'm 95% sure I've got it right. Only done a few practical tests so far, but they're all looking good. In case anyone is interested, the complete list looks like this (You'll need a zoom-button. )
First draft finished. https://forum.sector3studios.com/index.php?threads/the-new-adaptive-ai.5013/page-17#post-150322