ABOUT THE PROJECT
This project explored the viability of using reinforcement learning to play Atari games using only pixel data as input. Whereas Google’s DeepMind used a convolutional neural network as the front end for feature extraction, we experimented with less complex feature extraction methods. I implemented edge-based feature extraction and a DBScan clustering algorithm that dynamically labeled extracted features. These features were then used as the input to a Q-Learning algorithm, which learned optimal actions for multiple atari games. What we learned: for simple games, you can learn to do okay; as the games get more complex deep learning methods simply outperform traditional methods.
Developed For
This project was developed for CS221 and CS229 at Stanford University