Pál Zsámboki: Artificial Intelligence: Tree Search and Deep Reinforcement Learning Algorithms
16.00-16.45: Introductory talk: How to Code AI Algorithms
AI without a working code is like sheet music with nobody to play it.
This is why in this introductory talk, I will not only show you AI
algorithms but I'll aim to give as many pointers as possible for getting
started with coding them. First, I will introduce a couple of webpages
that I recommend for learning to code. Then, we will discuss how puzzles
can be solved by writing a fast simulator and implementing the Breadth
First Search algorithm in Rust programming language which is very
suitable for this purpose. Finally, we will take a look at a simple
Reinforcement Learning task that can be tackled by implementing the
REINFORCE algorithm using the Pytorch machine learning framework.
17.00-18.00: Main lecture
To introduce different aspects of AI, I will focus on a select few
algorithms that you can successfully implement yourself/ves and run on
an average modern laptop. To get a taste of tree search methods, we will
explore writing a program to play a board game using the Monte Carlo
Tree Search algorithm, and optimizing its hyperparameters with a Genetic
Algorithm. Then, we will talk about solving problems with continuous
observation spaces via the Deep Reinforcement Learning algorithms
Rainbow and TD3, both using quantile regression and the Sunrise ensemble
learning framework. I will conclude the session by introducing the big
guns: AlphaGo Zero / Expert Iteration, and transformers.