Apply Reinforcement Learning in AWS DeepRacer
Mai, Nhan (2022)
Mai, Nhan
2022
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202205179708
https://urn.fi/URN:NBN:fi:amk-202205179708
Tiivistelmä
Reinforcement learning is a machine learning algorithm that has the potential to aid in the development of an AGI system. Among the various types of machine learning algorithms, RL is unique in that it explores the environment without prior knowledge and chooses the appropriate action while the others focus on handling the data.
AWS DeepRacer is a self-driving 1/18th size race car designed to simulate real-world conditions while testing RL models on a physical track. The project aims to gain a better understanding of RL, the mathematics underlying it, and to observe it in action by deploying the trained model in Amazon's DeepRacer automobile. [1].
To fine-tune the model, performance indicators such as the average reward per episode and cumulative reward were investigated. To gain a better understanding of the distribution of action spaces, Amazon's log analysis capabilities were used. Any wasted action was deleted for effective training based on the log analysis data. The model is uploaded as soon as the training was finished to test it in the race track.
The results may be utilized as general suggestions for training models and enhancing RL using AWS DeepRacer. By using the strategies described in the thesis, it is possible to develop more robust and stable models.
AWS DeepRacer is a self-driving 1/18th size race car designed to simulate real-world conditions while testing RL models on a physical track. The project aims to gain a better understanding of RL, the mathematics underlying it, and to observe it in action by deploying the trained model in Amazon's DeepRacer automobile. [1].
To fine-tune the model, performance indicators such as the average reward per episode and cumulative reward were investigated. To gain a better understanding of the distribution of action spaces, Amazon's log analysis capabilities were used. Any wasted action was deleted for effective training based on the log analysis data. The model is uploaded as soon as the training was finished to test it in the race track.
The results may be utilized as general suggestions for training models and enhancing RL using AWS DeepRacer. By using the strategies described in the thesis, it is possible to develop more robust and stable models.