Optimal Sampling for Simulated Annealing Under Noise

This paper proposes a simulated annealing variant for optimization problems in which the solution quality can only be estimated by sampling from a random distribution. The aim is to find the solution with the best expected performance, as, e.g., is typical for problems where solutions are evaluated using a stochastic simulation. Assuming Gaussian noise with known standard deviation, we derive a fully sequential sampling procedure and decision rule. The procedure starts with a single sample of the value of a proposed move to a neighboring solution and then continues to draw more samples until it is able to make a decision to accept or reject the move. Under constraints of equilibrium detailed balance at each draw, we find a decoupling between the acceptance criterion and the choice of the rejection criterion. We derive a universally optimal acceptance criterion in the sense of maximizing the acceptance probability per sample and thus the efficiency of the optimization process. We show that the choice of the move rejection criterion depends on expectations of possible alternative moves and propose a simple and practical (albeit more empirical) solution that preserves detailed balance. An empirical evaluation shows that the resulting approach is indeed more efficient than several previously proposed simulated annealing variants.