You could probably work for a video game company. Apparently, the challenging part of designing driving games is designing AI for cars, especially in games with lots of terrain elements, for example, grand theft auto 4. We often see the non-player characters doing some very silly things making for decreased realism.
I want to see a lot more work like this (I'd love to participate too) but with more "elbow room", like harder puzzles, bigger generations, more of them, and faster - all ONE massive algorithm constantly updated to solve all kinds of problems. Maybe even a Spore-like setup. I want to see how few constraints we can remove, and see just what AI can do if we let it go for millions of generation.
Hey, I am bringing your project to real life! At the beginning of this month I was offered the posibility of providing a 2 wheel robot with 3 Ultra Sound sensors (front, left, right) to go from one point to another in a room. It must avoid Dynamicaly moving obstacles and must reach the other end in time efficient, distance effiecient or power efficient manner.
I was involved in a project using 4 completely autonomous robots with scoops to transport barrels through an area with obstacles. The robots were each restricted to their own zones and had to leave the barrels in specified drop off points.
I notice it pre-emptively anticipates thenext waypoint. Is its intent to get around the course as quiclky as posibile? Also, how many iterations is this?
The objective to get around the track as fast as possible. This particular controller was evolved for 100 generations with a population of 100 and a standard evolution strategy; the controllers are based on neural networks being fed rangefinder wall sensors, speed and angle to the next way point. There are more details in the papers on my website:
I have been researching alot and have found many alternatives to do this and to make this work. Finally I saw your video and couldnt believe the coincidence.
of course, even on the fastest cluster running a genetic algorithm for every npc in the game would take ages.
ross817 3 years ago
You could probably work for a video game company. Apparently, the challenging part of designing driving games is designing AI for cars, especially in games with lots of terrain elements, for example, grand theft auto 4. We often see the non-player characters doing some very silly things making for decreased realism.
DeimosSaturn 3 years ago
I want to see a lot more work like this (I'd love to participate too) but with more "elbow room", like harder puzzles, bigger generations, more of them, and faster - all ONE massive algorithm constantly updated to solve all kinds of problems. Maybe even a Spore-like setup. I want to see how few constraints we can remove, and see just what AI can do if we let it go for millions of generation.
GBart 4 years ago 2
Brilliant!
SexyMelon 3 years ago
and maybe let the cars themselves evolve. add another wheel, get a little faster, develop better brakes, become more aerodynamic.
yaifibar 2 years ago
Great job, well done!!!
mpeniak 4 years ago 3
Cool stuff, lets use some Genetic Programming with multi-objective fitness functions to evolve interesting driving behaviours....
perastikosssss 4 years ago
By the way, saw your interest in Borge, Im from Argentina!!
AlecDickinson2007 4 years ago
Hey, I am bringing your project to real life! At the beginning of this month I was offered the posibility of providing a 2 wheel robot with 3 Ultra Sound sensors (front, left, right) to go from one point to another in a room. It must avoid Dynamicaly moving obstacles and must reach the other end in time efficient, distance effiecient or power efficient manner.
AlecDickinson2007 4 years ago
I was involved in a project using 4 completely autonomous robots with scoops to transport barrels through an area with obstacles. The robots were each restricted to their own zones and had to leave the barrels in specified drop off points.
neotropic9 4 years ago
I notice it pre-emptively anticipates thenext waypoint. Is its intent to get around the course as quiclky as posibile? Also, how many iterations is this?
billamu 4 years ago
The objective to get around the track as fast as possible. This particular controller was evolved for 100 generations with a population of 100 and a standard evolution strategy; the controllers are based on neural networks being fed rangefinder wall sensors, speed and angle to the next way point. There are more details in the papers on my website:
julian dot togelius dot com
togelius 4 years ago
I have been researching alot and have found many alternatives to do this and to make this work. Finally I saw your video and couldnt believe the coincidence.
Well, hope we could exchange results some time!
My email is alecdickinson at hotmail dot com.
See ya!
AlecDickinson2007 4 years ago
Hi,
do you use Neural Network to evolve the behaviour of the car ?
Do you use some physics library to simulate ?
I made a work on evolutionary algorithms to evolve virtual creatures (see my videos)
felipeportavales 5 years ago
Nice! Five stars!
pelagicdivision 5 years ago
hmm i wonder if they can find, with more tweaking, the fastest turning line on this map or possibly a real track
Wiz4Shiz 5 years ago
The lines are walls. This particular video was taken using Xvid on Suse/KDE. More videos to follow soon, where the car(s) bounce off walls etc.
togelius 5 years ago
hey, what's that desktop system you're using? are the lines part of the background picture, or some weird desktop application?
damnsrm 5 years ago