ECE 6474 – Deep Reinforcement Learning (3C)

Course Description

Reinforcement learning for Markov decision processes. Deep neural networks for reinforcement learning. Value-based reinforcement algorithms with deep Q-networks. Policy gradient methods for discrete and continuous actions. Actor critic methods and advantage actor critic methods. Distributional reinforcement learning. Exploration and exploitation. Model-based reinforcement learning.

Why take this course?

Reinforcement Learning allows machines and software agents to automatically determine the best course of behavior within a set context. Reinforcement Learning is not perfect, and often has issues in dealing with more complex tasks. This is where deep reinforcement learning (DRL) comes in. The combination of deep learning with reinforcement learning has led to breakthrough applications like AlphaGo beating a world champion in the strategy game Go, and self-driving cars performing visual recognition and decision making at a human-performance level. Through this course students will be given the theoretical understanding of Deep Reinforcement Learning as they build their own deep reinforcement agents and teach them how to accomplish complex tasks in environment-agent interactions. The direct motive of this course is to help students gain experience in the study, design and implementation of DRL algorithms/techniques for machine learning applications.

Learning Objectives

• Describe the theoretical concepts underlying deep reinforcement learning.
• Analyze the problems inherent in reinforcement learning with deep Q-learning methods.
• Analyze the problems inherent in reinforcement learning with policy gradients methods.
• Analyze the problems inherent in reinforcement learning with actor-critic methods.
• Design deep reinforcement learning systems with application to real world problems.
• Apply the theory of exploration-exploitation in deep reinforcement learning.