Annealing schedule from population dynamics

TL;DR

This paper proposes a dynamic annealing schedule for population-based optimization algorithms that adaptively adjusts mutation rates based on a statistical mechanics framework, removing the need for manual parameter tuning.

Contribution

It introduces a novel adaptive mutation rate schedule derived from population dynamics and statistical mechanics, eliminating the free parameter of mutation rate.

Findings

Mutation rate adapts to maximize expected rewards.

The schedule improves optimization efficiency.

Mutation rate is no longer a free parameter.

Abstract

We introduce a dynamical annealing schedule for population-based optimization algorithms with mutation. On the basis of a statistical mechanics formulation of the population dynamics, the mutation rate adapts to a value maximizing expected rewards at each time step. Thereby, the mutation rate is eliminated as a free parameter from the algorithm.