Bayesian Machine Learning I
We consider optimization problems whose objective functions are unknown and hence have to be learned from data. Such problems are pervasive in science and industry, e.g., when
- designing prototypes in engineering,
- automated tuning of machine learning algorithms, e.g., in deep learning,
- optimizing control policies in robotics,
- developing pharmaceutical drugs, and many more.
Bayesian optimization methods are popular in the machine learning community due to their high sample-efficiency and have become a key technique in the area of "automatic machine learning". We introduce a general framework in which to understand and formulate such optimal learning problems, and provide a survey of problems, methods, and theoretical results.