Prediction, Learning, and Games

Online Machine Learning

This course was taught by Prof. Yen-Huan Li . The course studies the online learning problem and its extensions from several aspects. The topics include Blackwell approachability, PAC-Bayes analysis, probability forecasting with the logarithmic loss, online portfolio selection, learning with expert advice, and aggregating algorithms (from course website).

The course’s three assignments and the final project all include extensive literature reviews. For hw0, we first did a preliminary literature review in this field to understand the topics that are currently being studied in online machine learning. hw1 mainly focuses on Blackwell approachability and calibrated forcasting problems, in order to understand how Blackwell understood and solved the prediction problem without any statistical assumptions. In hw2, we analyze the performance guarantees of different regrets to capture model dynamics.

In the term project, we studied how to solve the non-convex problem under the framework of online learning. This type of problem usually introduces the oracle model in the optimization theory, and uses this as a basis to improve the performance guarantee of the algorithm. In addition, we also understand the relationship between online learning and statistical learning through literature review (Hazan and Koren, (2016)), and how to leverage each other’s strengths.

The biggest gain and extended thinking of this course is the improvement of the understanding of the literature, and how to apply this type of knowledge to economic problems. Issues such as non-parametric bayesian, model averaging, model selection, etc. are often mentioned in the class, and econometrics also often discusses such issues. In addition, after this course, I also discovered the interesting aspects of machine learning theory. For example, I recently read an article about omnipredictors for constrained optimization (Hu et al. (2023)). Both the design idea and the problem framework are very interesting. I look forward to becoming more familiar with the literature and technology in this area in the future.