P. I. Frazier, "A Tutorial on Bayesian Optimization", 2018, https://arxiv.org/abs/1807.02811 (TAKEN)
The paper provides a review of Bayesian Optimization, a method for finding the global optimum of a black-box function.
Tasks:
-
Give a summary of the overall strategy employed by Bayesian optimization.
-
In particular give an intuitive description of the various acquisition functions to guide the algorithm.
-
There are many well-known test functions, such as Goldstein-Price and Hartmann 6. Choose say 2 or 3 to compare methods in an experiment you will design, analyse, and write up.
- There are two aspects (factors) of interest to me when comparing "methods".
The first factor is the implementation: compare a python implementation such as BoTorch with the R library DiceOptim. The second factor is the acquisition function: try several as time allows,
taking account of what is available in the two implementations.
- You will have to choose a metric to assess the effectiveness of a method on a particular test problem.
- This is a statistical experiment. It is important to use the fundamental principles of statistical design and analysis.
- Feel free to limit any aspect that is causing you difficulty, with an explanation.
Return to the faculty list