Tobias Ryden,  Lund

Title: Non-standard acceptance functions and cooling schedules in simulated annealing

Abstract

In standard simulated annealing the probability of
accepting a proposal where the objective function
is worse than at the current position, decays exponentially
in the difference of the function values. In addition, the
inverse temperature increases at logarithmic speed.
In this talk we will describe more general algorithms
where the acceptance function is not the exponential
function, and the cooling schedule is not logarithmic.
Working on a compact set in R^d we will show that such
algorithms can obtain a wide range of convergence rates,
depending on the cooling schedule, including geometric
ones. We will also discuss the situation when the
objective function cannot be computed exactly, but
approximated with arbitrary precision.