In this advanced course on deep reinforcement learning, you will learn how to implement policy gradient, actor critic, deep deterministic policy gradient (DDPG), twin delayed deep deterministic policy gradient (TD3), and soft actor critic (SAC) algorithms in a variety of challenging environments from the Open AI gym. There will be a strong focus on dealing with environments with continuous action spaces, which is of particular interest for those looking to do research into robotic control with deep reinforcement learning.
Rather than being a course that spoon feeds the student, here you are going to learn to read deep reinforcement learning research papers on your own, and implement them from scratch. You will learn a repeatable framework for quickly implementing the algorithms in advanced research papers. Mastering the content in this course will be a quantum leap in your capabilities as an artificial intelligence engineer, and will put you in a league of your own among students who are reliant on others to break down complex ideas for them.
Fear not, if it’s been a while since your last reinforcement learning course, we will begin with a briskly paced review of core topics.
The course begins with a practical review of the fundamentals of reinforcement learning, including topics such as:
The Bellman Equation
Markov Decision Processes
Monte Carlo Prediction
Monte Carlo Control
Temporal Difference Prediction TD(0)
Temporal Difference Control with Q Learning
And moves straight into coding up our first agent: a blackjack playing artificial intelligence. From there we will progress to teaching an agent to balance the cart pole using Q learning.
After mastering the fundamentals, the pace quickens, and we move straight into an introduction to policy gradient methods. We cover the REINFORCE algorithm, and use it to teach an artificial intelligence to land on the moon in the lunar lander environment from the Open AI gym. Next we progress to coding up the one step actor critic algorithm, to again beat the lunar lander.
With the fundamentals out of the way, we move on to our harder projects: implementing deep reinforcement learning research papers. We will start with Deep Deterministic Policy Gradients (DDPG), which is an algorithm for teaching robots to excel at a variety of continuous control tasks. DDPG combines many of the advances of Deep Q Learning with traditional actor critic methods to achieve state of the art results in environments with continuous action spaces.
Next, we implement a state of the art artificial intelligence algorithm: Twin Delayed Deep Deterministic Policy Gradients (TD3). This algorithm sets a new benchmark for performance in continuous robotic control tasks, and we will demonstrate world class performance in the Bipedal Walker environment from the Open AI gym. TD3 is based on the DDPG algorithm, but addresses a number of approximation issues that result in poor performance in DDPG and other actor critic algorithms.
Finally, we will implement the soft actor critic algorithm (SAC). SAC approaches deep reinforcement learning from a totally different angle: by considering entropy maximization, rather than score maximization, as a viable objective. This results in increased exploration by our agent, and world class performance in a number of important Open AI Gym environments.
By the end of the course, you will know the answers to the following fundamental questions in Actor-Critic methods:
Why should we bother with actor critic methods when deep Q learning is so successful?
Can the advances in deep Q learning be used in other fields of reinforcement learning?
How can we solve the explore-exploit dilemma with a deterministic policy?
How do we get and deal with overestimation bias in actor-critic methods?
How do we deal with the inherent approximation errors in deep neural networks?
This course is for the highly motivated and advanced student. To succeed, you must have prior course work in all the following topics:
College level calculus
The pace of the course is brisk and the topics are at the cutting edge of deep reinforcement learning research, but the payoff is that you will come out knowing how to read research papers and turn them into functional code as quickly as possible. You’ll never have to rely on dodgy medium blog posts again.
This is a simple two state system with two actions. You have to use this information to calculate the state transition probabilities. Please see the handout from lecture 4 for related material.