Bayesian optimization: bad performance examples

Antonio Sala, UPV

Difficulty: ***** ,       Relevance: ,      Duration: 17:17

Summary:

This video continues video [boloop2EN] with additional examples. The aforementioned video is reviewed in the first two minutes, and then the following examples are discussed:

• Example with one local minimum and the global minimum at an extreme of the search range (works OK)

• Example with “overexploitation” that proposes quite unreasonable samples in its final phase.

• Example with prior confidence interval too small: it gets stuck at a local minimum

• Example with too small a bandwidth of the prior (too large correlation distance): exploration is very bad, blindsided, and ends at a point that is not even a local minimum

• Example with too large prior bandwidth (too small correlation distance): it only extrapolates to a very short distance and, therefore, progress is very slow and, before approaching the true optimal zone, it explores in an almost regular mesh (at the distance where correlation vanishes).

With all these examples, we seek to depict a complete image of the advantages and disadvantages of Bayesian optimization, and the importance of a good choice of the prior... or, if the prior is not so good, detect overexploitation and low probability of the samples (given the prior) to adjust the (hyper)parameters of said prior as samples are acquired. These issues are not considered here.

*Link to my [ whole collection] of videos in English. Link to larger [ Colección completa] in Spanish.