Self Car
- 4 Devlogs
- 14 Total hours
This is a program that allows to train cars, so with limited perception of the track, cars would be ablo to ride it using neural learning!
Ship #1
Pending review
I made a neural network where cars try to ride fast
The goal is to tech AI to ride this track as fast aspossible using only three sensors as a data.
Challenging part was optimization and simulation. In a matter of fact making neural network was easy, making a correct point system and simulation was hard.
I am proud of graph and saving system, it looks so slick, it add up ‘spice’ to the project.
By pressing Try project there will be a short video guide on how to run this project locally. Its really simple - only takes 4 commands to run if you have python installed.
3h 51m 49s logged
SAVE BEST ONE
Now you can save models and choose from the 14 best ones. They are saved in models folder and you can even send them to your friends (send whole folder). You can also load saved models with “L” and save with “S”, to choose between different models use arrow keys.
Now DIF_MODE - if you would turn on this option two segments on inner circle will stick out so now its harder to ride a car.
2h 24m 16s logged
They are smart now
Neural network is really good now. It is a race car. Changes:
- I added graph to visualize best score
- I removed deltaTime - time between frames, so simulations doesnt depends on it
- Improved vizualisation
- In neural network I removed direction as parametr so now car
learns how to driveinstead of remembering track
3h 39m 2s logged
ITS ALIVE
A lot of changes has been made - firs of all is reward system - all these dim green lines are checkpoints, for crossing them neural network get a reward.
Then of course I implementer AI itself. Now it spawn a 100 cars and chooses the best each generation.
Also cars receive a benefit when they ride each lap faster. So riding faster is also a plus.
I increased size of cars to make them a bit more realistic.
There is a video of already trained generations which shows what a generation looks like. They all disappear because I set so max amount of laps they can go is three, so new generation is created, because old one wouldnt crash.
There are a lot of thing to add still, like better physics, model saving and training for variety of tracks not only for one.
4h 20m 3s logged
Base for machine learnign
In this devlog I prepared everything to start teaching machines how to ride a car - the car was prepared for them. What has been done:
- Track generation - trach is 4 mathematic functions that create 1000 line segments in total to serve as walls.
- Optimized ray logic - car will have only 3 sensors (green) that will tell distance till wall
- Car class itself - for optimization i wont have 1000s of car object but in the class there will be stored vectorized information about all cars operating with numpy
- Visual part - I will render only one car, but under the hood there will be much more cars training.
