RL models learn to make the best short-term decisions while optimizing for a long-term goal. For example, with DeepRacer the short-term decisions are the angle at which the car is steering and the speed at which the car drives at a single moment. The long-term goal is to complete an entire race-track as fast as possible. When having to make a short-term decision, the RL model only has information about its current position and the direct surroundings. However, it does not have global information like the layout of the entire race track to make the decision.
AWS provides a DeepRacer simulator, powered by AWS SageMaker, which can be used to train virtual cars based on Reinforcement Learning. Our Data Science team uses the AWS DeepRacer simulator to train and test different RL models or virtual cars. The question remains: how well do the learnings from the virtual cars perform in the real world? To answer that question, we transfer our learned model from the simulation environment to the AWS DeepRacer car and try it on our track. The AWS DeepRacer car is a fully autonomous 1/18th scale race car where you can upload your own custom RL model to race it around a track. It uses cameras to view the track while the RL model controls throttle and steering. The car shows how a model trained in a simulated environment can be transferred to the real-world.
Next month our Data Science Guild is starting with a training session in our office leading up to a race day. Teams will compete with their virtual self-driving car. Follow us to check the progress and find out who wins this race.
Mobiquity Europe - LinkedIn
Mobiquity Europe - Twitter
Leave your details below and we'll be in touch soon.
