Reinforcement learning with AWS DeepRacer
Rönnberg, Oskar (2021)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2021060213665
https://urn.fi/URN:NBN:fi:amk-2021060213665
Tiivistelmä
The main purpose of this project was to take a look at reinforcement learning with AWS DeepRacer by providing good policies to use when training reinforcement learning models. Another goal was to analyse model performance and how to tune the performance of a model to achieve a better model.
AWS DeepRacer is a 1/18th scale autonomous vehicle which is taught to drive by itself with reinforcement learning along various tracks. Currently DeepRacers are trained for three types of objectives: time trials, object avoidance and head-to-head racing. In this project DeepRacer was trained for time trials and object avoidance. DeepRacer was trained to drive within a simulation created by AWS RoboMaker, and its neural network was updated within AWS SageMaker. DeepRacer could either be driven in simulation or on a physical track.
In this project, estimates were created for a required amount of training time for a model. In addition, estimates for the initial training time for a model were created. Moreover, the thesis discusses how significantly agent parameters affect model performance; which approaches work the best in reward functions; how changing hyperparameters affects the model and its performance; how to evaluate model performance from log files; and how to improve the quality of training and model performance by doing log analysis.
The results can be used as general guidelines for model training and improvement in reinforcement learning with AWS DeepRacer. Following the policies recommended in the thesis, better and more stable models can be achieved.
AWS DeepRacer is a 1/18th scale autonomous vehicle which is taught to drive by itself with reinforcement learning along various tracks. Currently DeepRacers are trained for three types of objectives: time trials, object avoidance and head-to-head racing. In this project DeepRacer was trained for time trials and object avoidance. DeepRacer was trained to drive within a simulation created by AWS RoboMaker, and its neural network was updated within AWS SageMaker. DeepRacer could either be driven in simulation or on a physical track.
In this project, estimates were created for a required amount of training time for a model. In addition, estimates for the initial training time for a model were created. Moreover, the thesis discusses how significantly agent parameters affect model performance; which approaches work the best in reward functions; how changing hyperparameters affects the model and its performance; how to evaluate model performance from log files; and how to improve the quality of training and model performance by doing log analysis.
The results can be used as general guidelines for model training and improvement in reinforcement learning with AWS DeepRacer. Following the policies recommended in the thesis, better and more stable models can be achieved.