Online and Reinforcement Learning (OReL)

In the classical machine learning data are collected and analysed offline and it is assumed that new data come from the same distribution as the data that the algorithm was trained on. If not, all the theoretical guarantees become void and the empirical performance may deteriorate dramatically. But what if we want to design an algorithm for playing chess? The opponent is not going to sample the moves from a fixed distribution.

Online and reinforcement learning break out of the static realm and move into the realm of perpetual cycle of getting new information, analysing it, and executing actions based on the updated estimation of reality. We consider agents (computer programs, robots, living beings) learning based on interactions with (real or simulated) environments. Examples include problems like repeated investment in the stock market, spam filtering, online advertising, online routing, medical treatments, games, and robotics. It allows to model a much richer range of problems, including problems with limited feedback, problems with delayed feedback, and even adversarial problems, where the environment deliberately acts against the algorithm (as, for example, in chess or spam filtering). At the same time it stimulates the development of fascinating mathematical tools for developing and analyzing algorithms for these problems.

In the course we will cover:

• The notion of regret: the evaluation measure, which replaces generalization error in offline learning and makes it possible to define and analyse learning in adversarial environments
• Various forms of feedback, including full-information and limited [bandit] feedback
   

We will introduce the following basic online learning settings, algorithms, and their analysis:

• Follow the Leader algorithm
• Prediction with expert advice: the Hedge / Exponential Weights algorithm
• Stochastic and adversarial multiarmed bandits: UCB1 and EXP3 algorithm
• Contextual bandits: EXP4 algorithm

And the following basic reinforcement learning settings, algorithms, and their analysis:

• Markov Decision Processes (MDPs)
• Monte Carlo Methods for reinforcement learning
• Dynamic programming for reinforcement learning
• Temporal Difference Learning (e.g., Q-Learning)
• Reinforcement learning using function approximators (e.g., Deep Q-Learning)
• Online reinforcement learning: average-reward and discounted settings
   

We will also cover a few advanced topics. The selection of advanced topics will depend on the lecturers and will be announced on Absalon.

The students will learn tools for theoretical analysis of most of the algorithms studied at the course and implement them in Python.

The course will bring the students up to a level sufficient for writing a master thesis in the domain of online and reinforcement learning.

WARNING: If you have not taken DIKU's Machine Learning A course, please, carefully check the "Recommended Academic Qualifications" box below. Machine Learning courses given at other places do not necessarily prepare you well for this course, because DIKU's machine learning courses have a stronger theoretical component than average machine learning courses offered elsewhere. It is not advised taking the course if you do not meet the academic qualifications.