Local optimization using simulated annealing

The authors present a tool for general purpose local optimization. Randomized combinatorial search heuristics such as simulated annealing (SA) are an effective means of exploring function space to find regions of high performance. Their ability to optimize functions once regions of high performance are found is limited. To improve, and accelerate, their local optimization capabilities, a perturbation operator that selectively perturbs the solution string was created. During the initial phases of optimization, the significance of each bit position is monitored. As the search continues, the probability of perturbations is shifted away from the most significant bits to the least. As the strength of techniques such as SA is achieved through their ability to sample a diverse set of hyperplanes, the perturbation probability is never allowed to decrease to zero for any bit position. This technique was tested on twelve problems, including problems in which the significance of each bit position was nonstatic through the search progression. The local optimization tool was also tested on K. DeJong´s five-problem test set (1975)