The behavior of simulated annealing in stochastic optimization

In this thesis we examine the performance of simulated annealing (SA) on various response surfaces. The main goals of the study are to evaluate the effectiveness of SA for stochastic optimization, develop modifications to SA in an attempt to improve its performance, and to evaluate whether artificially adding noise to a deterministic response surface might improve the performance of SA. SA is applied to several different response surfaces with different levels of complexity. We first experiment with two basic approaches of computing the performance measure for stochastic surfaces, constant sample size and variable sample size. We found that the constant sample size performed best. At the same time we also show that artificially adding noise may improve the performance of SA on more complex deterministic response surfaces. We develop a hybrid version of SA in which the genetic algorithm is embedded within SA. The effectiveness of the hybrid approach is not conclusive and needs further investigation. Finally, we conclude with a brief discussion on the strengths and weaknesses of the proposed method and an outline of future directions.