AWS RoboMaker
Developer Guide

Self-Driving using Reinforcement Learning

In this section, you teach a racecar to drive in a simulation through reinforcement learning using Coach Reinforcement Learning Library. View the reward metrics in Amazon CloudWatch Metrics to explore how the machine learning model improves over time. Customize your reward function to improve the machine learning algorithm used for training.

Before you use AWS RoboMaker for the first time, complete the tasks in Create an Account. Then, in the AWS RoboMaker console, launch the Self-Driving using Reinforcement Learning sample application.

View Training Metrics

The Self-Driving using Reinforcement Learning sample application uses AWS RoboMaker cloud extensions to write training metrics to Amazon CloudWatch.

To view robot training metrics

  1. Open the CloudWatch console at https://console.aws.amazon.com/cloudwatch/.

  2. In the CloudWatch console, select Metrics.

  3. On the Metrics page, in the All metrics tab, select AWSRoboMakerSimulation.

  4. Select the metric named DeepRacerRewardPerEpisode.

  5. Hover over the graph to see values for that moment. This graph plots the total reward received by the robot in each trial/episode. An increase in reward with time indicates that the race car is improving its performance on the race track.

View Trained Machine Learning Models

When a simulation job runs, a frozen TensorFlow graph with weights is written to an Amazon S3 bucket. This file can be deployed to a DeepRacer.

To view the TensorFlow data

  1. Open the Amazon Simple Storage Service console at https://console.aws.amazon.com/s3/. Open the bucket name beginning with awsrobomakerdeepracer.

  2. In the bucket, open the folder named model-store/model, and then select model.pb.