feat: Adding several acquisition functions #459
Description
Adding several acquisition functions
Currently we have Thompson Sampling and Probability of Improvement as acquisition functions. It would be ideal to include
- Expected Improvement (EI).
- Upper Confidence Bound (UCB).
- Log-Probability of Improvement (logPI).
- Log-Expected Improvement (logEI).
Describe Preferred Solution
These acquisition functions could naturally fit inside the decision making module. GPJax already has the core ingredients for e.g. acquisition function maximization using L-BFGS, and BO loops using utility-driven decision making.
Additional Context
Here are some relevant references we could use for implementing these acquisition functions:
- Taking the human out of the loop: a review of BO, by Shahriari et al., in which we can find formulas for EI and UCB.
- Unexpected Improvements to Expected Improvement for Bayesian Optimization, by Ament et al. The authors describe a numerically estable way of computing logEI.
If the feature request is approved, would you be willing to submit a PR?
Yes, I am glad to contribute implementations of e.g. UCB and logEI.