This workshop will give an introduction to the mathematical aspects of reinforcement learning (RL), its current methods, applications, and general scope. It will also make use of simulators and interactive applets to introduce aspects of RL in a hands-on way.
June 9, 2021 to June 11, 2021
Time:
11:15 am
Facebook Live
Reinforcement learning (RL) addresses problems of sequential decision making and stochastic control and is strongly connected to dynamic programming and Markov decision processes. In the last decades, it has gained importance and has become a major field of study in machine learning and artificial intelligence. Researchers from a variety of scientific fields that reach from cognitive sciences, neurology and psychology to computer science, physics, and mathematics, have developed algorithms and techniques with impressive applications as well as mathematical foundations.
Reinforcement learning is based on the simple idea of learning by trial and error while interacting with an environment. At each step, the agent performs an action and receives a reward depending on the starting state, the action, and the environment. The agent learns to choose actions that maximize the sum of all rewards in the long run. The resulting choice of action for each state is called a policy. Finding optimal policies is the primary objective of reinforcement learning.
The workshop is a joint venture with IMAGINARY, associated partner of Mathematisches Forschungsinstitut Oberwolfach, an institute of the Leibniz Association, Germany.
Date: June 9 - 11, 2021
Time: 11:15 am - 1:15 pm
Venue: Online
To register please click here.
Workshop Details:
The sessions will be conducted over three days, with two hours a day of online lecture followed by 90 minutes of tutorials/interactive sessions, affixed with further take-home exercises.
Speakers:
Jonathan Shock (University of Cape Town)
Andreas Matt (Imaginary, Berlin)
Content:
RL - introduction and basics
History of reinforcement learning: psychology, dynamic programming
Markov Decision Processes
Rewards, Returns
Value Functions, Action-Values, Policies
Bellman Equations
Value Iteration
Policy Iteration
Grid world example with Value Iteration
RL - model free methods
TD-learning
Q-Learning: Tabular
Eligibility Traces
Deep Q learning
Double DQN
Actor Critic Methods
Policy methods - the reinforce algorithm (click here)
ML and RL - applications, communication, and ethics
This would be more an applied, less mathematical, more 'playful' module.
