Probabilistic Decision Making
This theme focuses on decision-making as a crucial skill for intelligent agents, addressing the challenges posed by uncertainty in real-world problems using probabilistic models. Agents must learn, reason, and interpret observations to make effective decisions. The Probabilistic Decision Making theme explores the decision-theoretic approach, combining probabilistic reasoning, causal reasoning, and utility theory to determine optimal actions for rational agents. It covers algorithmic techniques such as model-based planning under uncertainty and reinforcement learning.
Year 1 |
|||
Quarter 1 |
Quarter 2 |
Quarter 3 |
Quarter 4 |
Data management and Engineering | Software Engineering and Testing for AI Systems | Responsible Data Science and AI Engineering | Research course |
Machine and Deep Learning | Theme 1 | Theme 1 | Theme 1 |
Probabilistic AI and Reasoning | Theme 2 | Theme 2 | Theme 2 |
Credits: each course in a theme is 5EC, so each theme is 15EC.
Students choose 2 themes, each of which has 3 courses in the 2nd, 3rd and 4th quarters of the 1st year. For this theme, you will take the following courses:
Q2 - Probabilistic Models and Inference
This course will focus on advanced probabilistic inference and modelling techniques, complementing the core 'Probabilistic AI & reasoning' course, and causality. On the inference side, the topics include representation of uncertainty (Bayesian networks, undirected graphical models, probabilistic programs); exact inference algorithms for Bayesian networks; approximate inference algorithms (importance sampling, metropolis-hastings, particle filtering, variational inference); implementation strategies for probabilistic programs; learning for probabilistic inference; discrete probabilistic programming; types of probabilistic queries; algorithms for causal inference. On the modelling side, it includes a spectrum of common probabilistic models that are often used in practice.
Q3 - Sequential Decision Making
This course will deepen your knowledge of AI decision making. Topics that will be discussed in this course are: how the goals of an agent can be represented properly, algorithms that use the Markov Decision Process (MDP) framework to compute (sub)optimal decisions, via search and/or learning and extensions of the basic MDP framework that are relevant for real-world problems such as partial information, multiple objectives and acting in multiagent systems and techniques to improve the safety and robustness of the decisions of agents. You will apply this theoretical knowledge in lab assignments, where you need to design and programme an algorithmic solution for sequential decision-making problems.
Q4 - Deep Reinforcement Learning
This course will cover the breadth of modern model-free RL methods, discuss their limitations, and introduce a variety of current research topics, such as: deep learning methodology and architectures, stabilization of approximated value estimation, modern actor-critic methods, planning as inference, exploration with deep networks, offline reinforcement learning, deep multi-agent reinforcement learning and multi-task and meta learning.