Shift your ML model into overdrive with AWS DeepRacer analysis tools
summary
Explains concrete ways to shave seconds off DeepRacer laps: reward on progress over steps (steps ≈ 15 Hz, episodes = laps), use waypoint-conditioned logic to change behavior mid-track, and trim action space to avoid wasting training on punished turns. Recommends NumPy/SciPy/Shapely for geometry, warns about overfitting and when to lower learning rate if training oscillates, and shows a log analysis notebook (runs in SageMaker) that produces heat maps, per-action reward maps and 900+ attention images for debugging. League now has Open and Pro divisions; Pro requires top 10% monthly.
AI-written summary · about
chapters (23)
- 0:00Intro
- 2:26Open Division
- 3:04Reminder of a reward function
- 3:44Reward function parameters
- 4:01Useful input parameter: Steps
- 4:48Comparison and sample code
- 5:16Visualize your reward function
- 6:22Useful input parameter: Waypoints
- 7:08Using waypoints to reward different behavior
- 8:27Geometry and trigonometry
- 10:47Do you need more training time?
- 12:01Why are training & evaluation lines separate?
- 13:21Is training not improving?
- 13:59Are you overfitting?
- 15:43More training with adjusted hyperparameters?
- 17:08Example log
- 18:11Log Analysis notebook
- 18:26Visualize the track and waypoints
- 20:07Analyze the reward distribution for your reward function
- 21:14Histogram for action distribution for each turn
- 23:12What is the model looking at?
- 26:31Recap and wrap-up
- 27:04Learn ML with AWS Training and Certification
related talks
28mShift your ML model into overdrive with AWS DeepRacer analysis tools
30mShift your ML model into overdrive with AWS DeepRacer analysis tools
30mPush Your ML Model to Top Speed (AWS DeepRacer)
27mGet rolling with machine learning on AWS DeepRacer
37mAWS DeepRacer: Train, Evaluate, and Tune
30mStart your ML journey with AWS DeepRacer for public sector
24mGet Rolling on Machine Learning
42mFrom Zero to Machine Learning in 45 Minutes