Raghuram Shankar
← Bayesian Optimization

Part 5 of 8

The Maths III: Inside the SOTA Methods

The Papers & Sources part tells the story of the modern methods; this section shows their actual mathematics. Each of the five ideas below changes one specific piece of the machinery you already know — a prior, a search region, an acquisition weight, or the surrogate itself. Nothing here requires more than Maths I and II.

  1. The √d fix: why vanilla GPs drown, and the prior that saves them
  2. SAASBO: sparsity as a prior
  3. TuRBO: the trust-region algebra
  4. πBO: folding expert beliefs into the acquisition
  5. PFNs: surrogates without fitting
  6. Quiz: 5 questions