rdot-lee / deepracer

AWS DeepRacer Reward Function

I participated in the "Invent 2018" contest of DeepRacer Virtual Circuit.

It is a reward function at that time. The lap time was about 8.464 seconds after repeated training.

Environment simulation

• re:Invent 2018 Training

Action space

Action space	Value
Maximum steering angle	30°
Maximum speed	2.4 m/s
0°	2.4 m/s
14°	1.6 m/s
20°	1.6 m/s
21°	1.7 m/s
25°	1.7 m/s
30°	1 m/s
-10°	1.9 m/s
-20°	1.5 m/s
-30°	1.6 m/s

Hyperparameters

Hyperparameters	Value
Gradient Descent Batch Size	32
Entropy	0.01
Discount Factor	0.999
Loss Type	Huber
Learning Rate	0.00006
No# Experience Episodes between each policy-updating iteration	20
No# of Epochs	3

Stop conditions

• 30 or 60 mins

Reference

• https://docs.aws.amazon.com/zh_tw/deepracer/latest/developerguide/deepracer-reward-function-input.html
• https://forums.aws.amazon.com/forum.jspa?forumID=318
• https://github.com/aws-samples/aws-deepracer-workshops
• https://github.com/sasasavic82/deeracer-reward
• https://qiita.com/kai_kou/items/8a45c687baca8c9465f6
• https://qiita.com/Alt_Shift_N/items/2c37fbb26d739b7f3046