CMLA

doc: chaining_ucb

Compute the chaining UCB given the canonical distance matrix between point of X

Syntax

 ucb = chaining_ucb(D2, s2, u)
 ucb = chaining_ucb(..., 'Name',Value)

Arguments

  • D2 matrix (n, n) of canonical squared distance
  • s2 matrix (n, 1) of posterior variance
  • u scalar for negative log probability

Name-Value Pair Arguments

  • step scalar > 1 power to use for the geometric decay of $\epsilon$
  • cdfinv function to upper bound the tail probabilities e.g. @(u,d) d.*sqrt(2*u)

Ouputs

  • ucb vector (n, 1) such that $P[\sup_{x^\star}f(x^\star)-f(x)-\mu(x^\star)+\mu(x^\star) > ucb(x)+cst] < \exp(-u)$

See also

gpucb | chaining_tree