I've been messing around with their DeepTraffic simulation and reading their documentation about it at http://selfdrivingcars.mit.edu/deeptraffic/. Can anyone provide a link to a solution for the simulation (preferably with an explanation)?
The goal is to reach 65mph, which isn't too difficult of a task. The only parameters that need to be changed to reach that goal are the learning inputs (the area around the car), and the network configuration. I found that having some buffer on the sides and front are helpful in recognizing the conditions for passing a slower car. The size of the hidden layer should also be big enough to take into account the different kind of situations that can happen in the simulation.
Making it on the leaderboard takes a bit more effort. I'm struggling to figure out the insight that takes me over the 70mph mark. I've toyed with the input parameters, types of hidden layers, the weighted random moves, and learning size. It's been frustrating, and has taken me down a deep rabbit hole about reinforcement learning.
If there are any tips for getting past the 'good enough' solution, I would love to hear them.
Exciting stuff! Thank you for publicly releasing this. Deep learning and self-driving cars are exciting spaces, and will definitely see more activity in the future.
http://selfdrivingcars.mit.edu/