In this course project, our objective was to use a reinforcement learning algorithm to teach an agent how to play Pong from raw images. The task included the application of the concepts learned during the course and the exploration of state-of-the-art techniques in the field.
We could reuse external sources (including code) and adapt them to the needs of our environment Wimblepong, designed after OpenAI Gym's Pong environment, in order to provide us with an easy to use version of the game.
We were also allowed to preprocess the raw images, e.g., by color transformations, stacking multiple frames together, or by a supervised learning model, but extracting the positions of elements (e.g., by color) was not allowed.