University project work for master course "Project Deep Learning für selbstfahrendes Kraftfahrzeug"

In this project we used imitation learning to train an agent for solving the OpenAI Gym CarRacing environment. The agent was based on a custom convolutional network which was trained on manually generated samples from the environment. The base model was later compressed with different pruning approaches and the fine-tuned pruned networks were tested in the environment. We consistently achieved comparable results to the base model even when parameter count was reduced by 80%. Our best model achieved an average test reward of 854 (with a standard deviation of 52) over 20 episodes in the environment.

Ulm University Winter Semester 20/21