Curriculum
  /  
Descriptions
MLDS 490: Reinforcement Learning


VIEW ALL COURSE TIMES AND SESSIONS

Description

Recent breakthroughs of artificial intelligence in games such as Go and poker, and advancements in robotics, autonomous cars, manufacturing scheduling are driven by reinforcement learning where an agent learns in real time based on an ever-changing environment and receiving a reward after an action is taken. Combined with deep learning, deep reinforcement learning is the current state-of-the-art.

This course covers the modeling concepts and the underlying recursive functions behind reinforcement learning. A substantial part of the material is devoted to modern algorithms for solving the underlying model including deep Q-networks, policy gradient, Monte-Carlo tree search, and actor-critic algorithms. Through homework assignments and projects, students implement and evaluate these algorithms in Python.

Learning Objectives

At the end of the course students should be able to:

  1. Given a business problem identify if reinforcement learning is needed to solve it or a different technique
  2. Recognize states, the role of an agent, its actions, and environment for a given reinforcement learning problem
  3. Select the most appropriate reinforcement learning algorithm for a problem at hand
  4. Implement the overall solution in Python

Topics

Models (value function, Q-factor, optimality equations); Basic algorithms (Value iteration, Policy iteration, Q-learning, TD-learning); State-of-the-art with deep learning (Deep Q-network, policy gradient, actor-critic algorithm, Monte-carlo tree search, PPO)