Neural Network Cars and Genetic Algorithms (1/2)

First attempt on teaching neural networks to drive a car. The architecture of each neural network consists on three layers with five, six and four neurons respectively and the evolution is done using uniform cross-breeding and random mutation.

Part 2:

Background Music by Trevor Lentz:

[kehtabpeg]

I wish cars could speed up at the cost of turning radius and lap time would be counted instead of just progress

[MaximeFAYEMaxouMask]

I'm actually impress how well it does. Normally for this kind of situation you'd use reinforcement learning with "q learning" but your method of mix of nn + genetic algorithm seems to work very well

[CardZoneMax]

Big fan of using neural networks! You should try adding a competitive element like racing to see how that impacts their risk taking

[sportingpatria]

Great performance, interesting game, keep it up bro!

[chirag]

This can be even more amazing if the game engine can be configured to make the cars collide each other. The neural network can soon learn not only to stay on track but to also avoid other cars

[MrSquareart]

Very very nice video!!! Found your video on reddit post

[bonniewhy]

This is so amazing. Those cars are so cuuute. Haha. Go cars go!

[artbyact]

Woah very interesting experimental game you have here. Thumbs up!

[lordty]

Honestly this is so amazing, well explained concept !

[SherwoodBotsford]

Interesting. Next step. Avoid other cars.

Get them to the point they can complete an arbitrary track.
Now select for two things: Speed in doing the track (You will have to add some physics to car motion) And Collision avoidance with other cars. Breed a set of avoiders, and a set of speeders, then hybridize them. This is how real genetics work. Select for one trait at a time then hybridize.

[khalidogilbee]

Very interesting concept, explained very well

[dartme]

Wow, there are so many ways to improve this! 1: Instead of the lines keeping their orientation with the car, maybe like 15 lines coming out from all directions and they stay north, east, south, etc. (not rotate with the car) and one line that comes out the front of the car. Also, of course, try to get the fastest time. Maybe you will find cars doing proper chicane after a while! Also, they tend to prefer right turns (and staying by the right side of the track) I think because the first turn in the course is a right turn, so that was over-trained.

Excellent idea, great execution, and compelling video!

[shaks_]

This is some very interesting stuff, keep it up

[ravinchowdhury]

Awesome! Are you planning on making the code public?

[francescobertini]

This is really beautiful if you could share the code I would love to try more complex cases... Congrats

[Omnicia]

Why not just do selection based on track distance traveled (rather than manually)