AI Learning / Training

All about making tracks for MX Simulator
Post Reply
Donut MX
Posts: 131
Joined: Wed Jul 06, 2016 2:21 am
Team: Privateer
Contact:

AI Learning / Training

Post by Donut MX »

Anybody know if JLV has shared the technicalities of how the AI --learns tracks? I couldn't find anything.

I've been toying with teaching AI more difficult custom tracks with updated tileinfo / erode settings so I can set the erode lower for motos instead of using a higher and more jagged erode setting by myself. They're not quite there yet lol.

I'm curious about how the --learn feature works. Will they try all the lines I rode with --learn on, and then ultimately choose the quickest? The most consistent? Is it all completely random what they choose? It seems like the AI have a pretty tame maximum lean angle, are there certain things I should do while riding that helps them go faster? etc..

I'm sure my hopes of a good erode moto are unlikely, but at this point I'm just curious about how to train the AI to be fast as possible. It'd make for an interesting stream lol. I'd assume I'm the only person whose used this feature in years. Please shed some knowledge JLV :lol: :lol:

this storken vid is all i could find - https://www.youtube.com/watch?v=vpp8iO4 ... bQ&index=1
Motospun
Posts: 102
Joined: Thu Aug 23, 2018 6:44 pm
Team: 1TF
Location: Communist Canada

Re: AI Learning / Training

Post by Motospun »

I've messed around with it myself and I found the best way to teach the bots was to give yourself the same restrictions they have. So basically set your shifting to automatic, don't lean forward or back, set all your suspension settings to default, and don't lean too fast/too hard into turns. I'm not sure what the default stability setting is but you could set that to default as well so you can only lean as quickly as the ai. I've done little testing so I'm not entirely sure but I believe with the --learn feature if you have an inside and an outside on every corner of your track and you run 2 laps with --learn on and you take all the insides one lap and all the outsides the next then the bots will have 2 choices for each turn. If you go in the editor and select a timing gate there are 2 buttons on the bottom left that say "Set AI Switch Point" and "Clear AI Switch Point" I believe if a timing gate is set as a switch point then at that timing gate the ai will switch to a random different line. So for example, if your first lap with --learn was all insides and the 2nd all outside and the bot is taking all insides then hits a switch point he will switch to your lap 2 lap and go outside.
Wahlamt
Posts: 7799
Joined: Mon Sep 13, 2010 3:15 pm
Team: MLG Compton
Location: Sweden
Contact:

Re: AI Learning / Training

Post by Wahlamt »

Donut MX wrote: Fri Jan 14, 2022 6:08 am this storken vid is all i could find -
I based mine off of Poidog's guide :lol: Glad to see your face back around 8)
jlv
Site Admin
Posts: 14348
Joined: Fri Nov 02, 2007 5:39 am
Team: No Frills Racing
Contact:

Re: AI Learning / Training

Post by jlv »

They're likely to have trouble at switch points so you only want switch points where it's smooth and easy. I usually put them at the corner entry. They're also prone to over adjust in some places and under adjust in others so you have to take easy lines when you train them to make sure they can actually hit them.
Josh Vanderhoof
Sole Proprietor
jlv@mxsimulator.com
If you email, put "MX Simulator" in the subject to make sure it gets through my spam filter.
Post Reply