Design and implementation of an Artificial intelligence system for chrome offline game

Description

Abstract In this project, implemented both feature-extraction based algorithm and an end-to-end reinforcement learning method to learn to control Chrome offline dinosaur game directly from high-dimensional game screen input. Results show that compared with the pixel features based algorithms, deep reinforcement learning is more powerful and effective. It leverages the high-dimensional sensory input directly and avoids potential errors in feature extraction. The result reveal that the approaches significantly out-perform even the experienced human player in the game. I implemented the extraction of pixel-based features from the game screenshot and the MLP algorithm. I also focus on the implementation of Q-learning framework, and the comparison between hands coded online learning methods and deep reinforcement learning.  Finally, propose a special training methods to tackle class imbalance problems caused by increase in game velocity. After training, our Deep-Q AI is able to outperform human experts.