Master
2022/2023
Mathematical Foundations of Reinforcement learning
Type:
Elective course (Math of Machine Learning)
Area of studies:
Applied Mathematics and Informatics
Delivered by:
Department of Complex System Modelling Technologies
Where:
Faculty of Computer Science
When:
2 year, 2 module
Mode of studies:
offline
Open to:
students of one campus
Master’s programme:
Math of Machine Learning
Language:
English
ECTS credits:
6
Contact hours:
32
Course Syllabus
Abstract
Reinforcement Learning is a fascinating area located on the intersection of approximation techniques, optimal control, statistics and machine learning. The main problem sounds as follows: ”For some agent in some (possibly adaptive) environment, how to learn a way to make decisions to live ”optimally” by know- ing only some scalar reward obtained after taking the action?” It can be argued that this is essentially an optimal control problem... Yes and no. Yes – because the goal is to learn control function for the agent, which would tell what to do in certain state of the world. No – because classic optimal control usually deals with known model of the environment (transi- tion probabilities, stochastic differential equations,..). Reinforcement Learning is concerned with what to do if such model is unavailable and only some general assumptions can be made about its function. There are several decent courses on Reinforcement Learning existing, most of them are practical: in the sense that they introduce many algorithms and ideas of solutions for certain practical problems. The other side of the story, mathematical explanations of why the methods actually work, is mostly skipped. Our course is aimed at closing this gap and focusing mainly on the mathematics behind Reinforcement Learning.".