This is a loose follow up to the deeptraffic project I’ve been playing with.
I decided to attack the problem initially by working each aspect and watching the network train. This devolved into to figuring out which parameters I thought I could poke and move towards my goal of making the leaderboard. The spoiler first I guess, I’m currently at 75.13 mph, still shy of the now 75.60 mph 10th place entry.
I guess I’ll have keep my micro-farm running when I can find
free AWS cycles.
I was able to use the “headless” code that is running the deeptraffic challenge, and with only a few change and some bash scripting to help organize a hyper parameter search I was running.
This sounds pretty lofty,
but really I manually
tweaked most things. The
random searching I focused
on learning rate and
finding the correct number
of iterations.
I feel like there is still so much to learn here, it’s really fun.