Skip to content

πBO (prior slot)

Cite: Hvarfner et al., ICLR 20221.

Paper Code

Compile a belief from context into a factorized distribution, then fold it into the acquisition:

\[\alpha_\pi(x) = \alpha(x) \cdot \pi(x)^{\beta / (t+1)}\]

The exponent decays so a wrong prior fades as data arrives.

lenz set-belief --state ./state.json --prior '{
  "lr": {"dist": "lognormal", "mu": -7.0, "sigma": 1.0},
  "dropout": {"dist": "beta", "a": 2, "b": 8}
}' --decay-beta 10

Supported dist values: uniform, normal (mu, sigma), lognormal (mu, sigma), beta (a, b on the range scaled to [0,1]), categorical (probs map). Name a numeric distribution — do not pass adjectives such as "aggressive".

Clear a prior:

lenz set-belief --state ./state.json --prior '{}' --clear

  1. Hvarfner, Stoll, Souza, Lindauer, Hutter, Nardi. πBO: Augmenting Acquisition Functions with User Beliefs for Bayesian Optimization. ICLR 2022. ↩