Deep Reinforcement Learning: An introduction through video games

Details

Supervisors

Van Oirbeek, Robin ; Pennisi, Andrea

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

An increasing number of technologies involving artificial intelligence (AI) are currently being developed all over the world. These technologies could lead to enormous breakthroughs in their sectors. These AI developments are largely made possible thanks to deep learning techniques that use neural networks in their algorithm. In this master thesis, we discover the Deep Reinforcement Learning through three algorithms: Q-Learning, one of the more basic algorithms, Deep Q-Network, an evolution of the Q-Learning using neural networks, and the Proximal Policy Optimization, currently one of the most efficient RL algorithms. To compare these algorithms, we use two games implemented in a Python package called Gym developed by OpenAI. The first game is basic and often used in the literature, Mountain Car. The second is the classical Super Mario Bros. For all these algorithms, we have designed several models with various hyperparameter values and compared their performance by looking at the cumulative average reward produced by each game. For the Mario environment, the models were also compared by their ability to end the first stage.