Have we looked at Thompson sampling efficient multiobjective optimization?

[SRFU-NN]

https://link.springer.com/article/10.1007/s10898-018-0609-2

An alternative to the multi-objective sampling we have now, and to Expected HyperVolume.

As I understand it:

1. Draw sample functions for each objective
2. Estimate the Pareto front from the sample functions (e.g. with NSGA-II)
3. For points on the estimated Pareto front, find the hypervolume increase
4. Pick the point on the estimated Pareto front with the highest hypervolume increase.

Since it deals in sample functions, it takes uncertainty into account.

It is less computationally heavy that expected hypervolume.

It automates finding the reference point, which is a problem with expected hypervolume.

The paper suggests using Kriging believer to sample multiple points (though they don't call it that).