On Step Width Adaptation in Simulated Annealing for Continuous Parameter Optimisation

摘要

Simulated annealing is a general optimisation algorithm, based on hill-climbing. As in hill-climbing, new candidate solutions are selected from the 'neighbourhood' of the current solution. For continuous parameter optimisation, it is practically impossible to choose direct neighbours, because of the vast number of points in the search space. In this case, it is necessary to choose new candidate solutions from a wider neighbourhood, i.e. from some distance of the current solution, for performance reasons. The right choice of this distance is often crucial for the success of the algorithm, especially in real-world application where the number of fitness evaluations is limited. This paper explains how in such a case the use of a variable radius of this neighbourhood, refereed to as maximum step width, can increase the over-all performance of simulated annealing. A real-world example demonstrates the increased performance of the modified algorithm.