AWS DeepRacer - Advanced Model Training Using AWS DeepRacer Console
The guide walks you through how to optimize your model by using an advanced version of the reward function and fine-tuning your model’s hyperparameters.
Model tuning with the advanced reward function
In previous, I created your first trained model using the default settings in console.
https://www.truongbui.com/2020/12/model-training-using-the-aws-deepracer-console.html
To set up your next model, follow these steps.
1. In your AWS account, go to the AWS Management Console.
2. Choose us-east-1 region at the top right corner of the Regions dropdown menu.
3. From the top left of the console, click Services, type DeepRacer in the search box, and select AWS DeepRacer. That will open the AWS DeepRacer console.
4. From the AWS DeepRacer console, click Reinforcement Learning.
5. Select the model you built in previous (Roadstour Model).
6. From the Action dropdown, select Clone.
7. Name the model and add a description.
8. In the Environment simulation section, select the same track used in your first model.
9. Under Action space, familiarize yourself with these settings and accept the defaults.
10. In the Reward function section, select the Advanced function dropdown arrow and then select the first Insert code button. This code block optimizes for turning and requires less time to train. You can also scroll down a little to select the second Insert code button. The code for this function focuses on optimizing for speed. It also requires less training time.
11. In the Algorithm settings section, accept the default settings.
12. For Stop conditions, choose a max time of 60 minutes and click Start training. The new training will initiate in 5 to 6 minutes. You then must wait for the training job to complete before proceeding. If you choose a max time of 60 minutes, it will take up to 60 minutes for this training job to complete. It is complete when the status reads Ready.
Tip: One effective way to train your model is to use shorter training session followed by adjustments and then more training. You may want to compare this type of approach to longer training jobs exceeding several hours.
13. Now, select the name of the model you just trained.
14. Click Start new evaluation.
15. Select the track (this should be the same as the track you trained on).
16. Set the number of trials to 3 and click Start evaluation. The evaluation results will update in approximately 5 minutes.
Note: While the evaluation job is running, watch the live results in the simulator. What’s your car doing that seems to be working well? Is it staying on the center line? Is it moving at a good speed? What’s your car doing that isn’t working well? How’s it doing around corners, for instance?
Tips and tricks
• Try to increase your speed setting and stabilize the training.
• Use progress and speed variables to reward higher speeds and faster lap times.
• Keep as close to the center line as possible.
Additional Resources
• For more information about AWS DeepRacer, see https://aws.amazon.com/deepracer/.
• To troubleshoot and collaborate about the AWS DeepRacer, see https://forums.aws.amazon.com/forum.jspa?forumID=318.
Comments