Amazon Web Services DeepRacer
GOALS:
Learn, Build and Compete in the AWS DeepRacer Reinforcement Learning League - Finish Top 100
TECHNOLOGY:
Python, AWS, PPO Algorithms, Selenium
ACCOMPLISHMENTS:
Top 10 in North America at 2020 Amazon reInvent Finals
32nd out of Thousands of Racers (94th Percentile), $400 USD Prize + $35 AWS Credits
DESCRIPTION:
My AWS Deep Racer Progress in May 2020
- 10 Days of Learning + Setup
- 20 Days of Trial and Error Training
Goal:
- Finish at least top 100 out all racers
Result:
Summit Time Trial - 32/341 - Top 9%
Summit Head 2 Head - 37/157 - Top 23%
Virtual Time Trial - 79/1291 - Top 6%
Virtual Object Avoidance - 39/134 - Top 29%
Virtual Head 2 Head - 50/202 - Top 25%
Success :)
***
Model Description:
1. Barcelona Model
- Attempt to use a complex functions and guide the car to follow a drawn line based off of Fernando
Alonso's F1 drive through in 2018 (https://www.youtube.com/watch?v=e5uArfwC4Zk).
- Too complex, convergence was too slow, car wasn't really learning
- Performed decently, was my leading Barcelona model until All You Model (#5)
2. Slow Complex Model
- Similar Attempt as Model 1, without pathfinding
- Surpassed by Pathfinder Model (#3)
3. Pathfinder Model
- Relying only on a given path and speed
- Surpassed by Speed Model (#4)
4. Speed Model
- Rewarded completion and staying left of track, less distance to cover on the left, much simpler
- Surpassed by All You Model (#5)
5. All You Model
- Simplest Model, Don't go off track, finish the lap, fast
- Top finishes, could've been improved with more training time