I wonder how simulations like this transfer to the real world. For example if you want to win in the real world you just start driving like a maniac. People will very quickly get out of dodge.

Spoilers: I tried to encode this idea by setting the inputs to be just tall enough to learn the safety system and wide enough to cover all lanes with no forward visibility. This learns an aggressive lane switching algorithm that lets you hit around ~70mph though not for the reason I thought. Also needed a few more hidden units/layers to make it work.