Genetic Algorithms and Simulated Annealing: A Marriage Proposal

Cenetic Algorithms (CA) and Slmulated Annealing (SA) have emerged as the leadlng methodologies for search and optimization problems in high dlmensional spaces. Previous attempts at hybrkllzing these two algorithms have been cumbersome and requlred major changes to both. In this paper we propose a simple scheme of using Simulated-Annealing Mutation (SAM) and Recombination (SAR) as operators In a standard GA envlronment. The operators use the SA stochastic acceptance function internally to limit adverse moves. This is shown to solve two key problem In CA optimization: populations can be kept small, and hlllcllmbing in the later phase of the search Is facilitated. The Implementation of this algorithm withln an existlng CA envlronment is shown to be trivlal, allowing the system to operate as pure SA (or Iterated SA), pure CA, or In various hybrid modes. Performance of the algorithm is tested on various large-scale applicatlons, includlng DeJong's functions, a 100-city travellng-salesman problem, and the optlmlzation of weights In a feed-forward Neural Network. The hybrid algorithm Is seen to improve on pure CA In two ways: better solutions for a given number of evaluations, and more consistency over many runs.