Multi-task Bayesian optimization

[nabenabe0928]

Multi-task Bayesian optimization

Main points

• BO using MT-GP
• there are three types of searches:
  1. Just using MT-GP with EI (straightforward and the information from source tasks will assist the precision of the surrogate)
  2. Acquisition function for k-fold cross validation, but I actually did not get the point. (Anyways, it does not look that useful)
  3. information gain per cost acquisition function. The point is that we would like to explore unseen regions with less cost, so if there are some related cheap tasks to the target task, we would like to explore the regions with the related cheap tasks. But because the information gain for the target task is obviously higher when we search the target task, we divide the information gain by the cost (e.g. runtime) function so that we can explore unseen regions with the cheap related tasks. (I think the method is quite similar to FABOLAS? and it could be cost-aware in a way)

[nabenabe0928]

Performance over time

A bunch of self-made benchmarks.