EGO and infill criteria

Development
1

Hi all,

Sometimes ago, we had a presentation from @l-brevault and @m-balesdent on benchmarking various infill criteria for bayesian optimization. This rich benchmark made a comparison between 6 criteria that are defined hereafter

– Expected Improvement (EI) [J. Močkus 1975, D. Jones et al. 1998] → the one used currently
– Lower Confidence Bound (LCB) [Cox et al 1992]
– Probability of Improvement (PI) [H. Kushner, 1964]
– Watson and Barnes 2nd (WB2) [A. Watson et al. 1995]
– Scaled Watson and Barnes 2nd (WB2S) [Bartoli et al. 2019]
– Thompson sampling [W. Thompson, 1933]
– Max-value Entropy Search (MES) [Wang et al. 2017]

The benchmark has been done on different use-cases of various complexity, with different dimension and computations budget.
– Ackley : 2D
– Branin : 2D
– Three Hump Camel : 2D
– Colville : 4D
– Trid : 5D
– Styblinski-Tang : 6D
– Schwefell : 10D

Today the expeced improvement is hard-coded in the API. Thus the suggestion is to allow having more infill criteria and potentially having a user-defined one!

What should be an infill criterion ?

An infill criterion (today) is a specific evaluation that has the () operator But we should pay attention that the infill is updated during the optimization process

Suggestion

  • Implement an InfillCriterion class that inherits from EvaluationImplementation
  • InfillCriterion ctor → InfillCriterion(krigingResult, optimalValue, isMinimization, hasNoise)
  • The class should allow to update both optimalValue and KrigingResult via setters
  • Implement current ones (ExpectedImprovement, AugmentedExpectedImprovement that inherit from InfillCriterion)
  • Implement the ones identified from the benchmark

what do you think ?

2

sounds ok, inheriting Evaluation makes senses since we want to make the function to the user

3

It would be great to find them in OT. And your implémentation makes sense. It would be also a good thing to add classical acquisition criteria for the estimation of level sets, quantiles and probabilities. And perhaps with parallel acquisition option. We developed a few things about active learning in gemseo-mlearning, with examples based on the OT Kriging. Do not hesitate to have a look and see if the design can help you. Trieste can also be a source of inspiration.

4

Thanks for the inputs, I will have a look at your designs in gemseo.
I was aware about Trieste and the various infill. The problem today is more in terms of usability.

I will try to prepare something for the next technical committee