The main objective of this project is to optimize the performance of our Bing DeepRacers by designing an efficient reinforcement learning (RL) algorithm. RL models are developed and trained on the Amazon AWS virtual platform. By tweaking actions, hyperparameters and a reward function, it is possible to control how your model will learn as it practices on a virtual racetrack. When training is complete, the model can be evaluated on either a physical or virtual racetrack where it attempts to complete the course as fast as it can. From there, the model can be updated as necessary to improve its performance. The initial goal of this project is to create a RL model that can compete in AWS DeepRacer leagues to achieve the fastest possible time. From there the stretch goal is to create a machine learning model that can simulate a self parking car with the potential ability to avoid obstacles.

The major components of the DeepRacer system are the reward function, hyperparameters and the action space. These are the items that are tweaked and tested in order to improve your RL model. From the system diagram, we can see how these different components tie into each other in general. By using the camera, the DeepRacer will be able to update the parameters input to the reward function. From here the model will earn its reward based on the current state and continue to train. At the same time, the hyperparameters set how the model will improve/learn as it trains. These influence how often feedback is given and how the DeepRacer will make decisions based on the outcome of the reward function. Once it decides to make an action, it will choose one of the actions from the model's action space and execute it.

The software of the project (the RL model) is written, trained, and implemented through the AWS Virtual platform. Once this is completed and the model has been evaluated adequately, it starts the physical testing of the project using an actual DeepRacer on a track built specifically for this project. The DeepRacer is the component that loads trained models and uses them to follow the track.

Once there was a model that passed its virtual evaluations, the physical track was built in a team member’s basement. It is a fairly simple oval loop, with one right turn thrown in to confuse the models being tested.

Before the models can be tested, the DeepRacer must first be calibrated to ensure that the steering, motors, camera, and braking are all functioning properly. Once this is complete, the machine learning models are uploaded to the DeepRacer for testing.