This is an AWS DeepRacer model, trained on AWS Console for the AWS DeepRacer Competition
The reward function is a basic template that can produce results on any track with 3 to 5 hours of training.
//Reward Function def reward_function(params):
if params['all_wheels_on_track']:
reward = params['progress']
else:
reward = 0.00001 //some very small value
return float(reward)
At initial training with above reward function, I trained with default hyperparameters But as the model got better, increasing the number of experience episodes between each policy-updating iteration produced positive results
The action space of this model Steering angle:: 30 degrees Steering angle granularity:: 5 Max Speed:: 6 m/s Speed graularity:: 3
Results:: after 4 hours of transfer learning anchieved 00:00:22.678 on re:invent 2018 track