Also, changining the track from one interation to another I don't think it helps. The training should be done on the same track until at least one of the cars is able to complete it. Than it can be changed.
Sometimes some of the cars get stuck and the round doesn't finish, without having to click the "End round" button. There should be some mechanism that detects that cars are stuck and end the round automatically.
I actually have had inactivity detection implemented at some point, but I've resigned of this in favor for user-defined stop condition. But you are right, I will bring that back.
It's actually expected that in some cases it might take many generations to learn, as the population is rather small. In other attempts they can do this in like 5 generations, so please, tray to restart the whole process.
Keeping one track untill at least one car is able to finish it is not as good idea as it seems due to ovefritting. The will simply memorize all the moves depending on some insigificant factors and they won't be albe to finish any other track. I have tested that :D
0.5 mutation rate is way too much..
On the other hand 0.5 is not so large, when you look at the mutation procedure. If a speciman is going to be mutated, only 20% of it's weights are changed and only within a range [-0.1,...,0.1]. But fill free to experiment :)
https://news.ycombinator.com/item?id=12397375
For research purposes :)
While I agree that exposing the NN to diverse tracks each iteration asures a thorough learning, I was kind of curios how susceptible the algorithm is to exploit irelevant particularities of a track instead of finding the general rule.
I wish the creator will add the posiblity to chose if the track should be changed each iteration or only after at least one car completed it.