v6-reInventorBust.md

Iteration 6 - reInvent or Bust

Ok, after much trial and error, I'm going to try an alternative strategy. I'm going to assume that there will be a gauntlet of tracks at reInvent which will require either a nice variety of models or one pretty universal model. If I start with a basic model and just train a little over each track, maybe the neural net will develop to handle a variety of scenarios.

I trained across the basic Oval, reInvent track, Bowtie, Empire, Shanghai and Cumulo for 1 hour each, cloning for each iteration. Adjusted the batch size to 128, switched to Huber loss after Empire. Left all other hyperparameters the same.

Results

You'll just have to wait until re:Invent to see.... :)

UPDATE: I managed to miss out on making to the "Round of 16" by just a few tenths of a second. You can see the whole sordid story here: DeepRacer: At the Apex

Reward Function

def reward_function(params):

    reward = 0.001

    if params["all_wheels_on_track"]:
        reward += 1
    if abs(params["steering_angle"]) < 5:
        reward += 1
   
    reward += ( params["speed"] / 8 )
   
    return float(reward)