Recently, there have been much research of robust and powerful optimization methods for solving large and difficult combinatorial problems. As a result, several effective stochastic search algorithms have shown up in the literature, for example, simulated annealing, tabu search, and evolutionary algorithms.
Evolutionary algorithms inspired by the natural process of evolution are attracting attentions for dealing with global optimization. The main feature of evolutionary algorithms is the maintenance of a set of solutions that are searched in parallel and the adoption of perturbation mechanism analogous to biological operators. Accordingly, they can be easy parallelizable compared to other methods. There have been a lot of empirical evidences indicating that evolutionary algorithms are good optimization methods, resulting in rapid enlargement of application areas.
However, most of the applications have been performed without considering how effective evolutionary algorithms are in solving the problems compared to other search algorithms.
This thesis consists of two parts. In the first part, we investigate the features of evolutionary algorithms and examine the efficacy of them by carefully controlled empirical comparisons with simulated annealing. As problem size and ruggedness of the landscape increase, the contribution of crossover to evolutionary search becomes less useful and the evolutionary search becomes less efficient than simulated annealing.
The second part of this thesis deals with two possible hybridization techniques to enhance performance of canonical evolutionary algorithms. The first hybrid incorporates independent sophisticated heuristic local search to evolutionary algorithms so as to overcome problems of evolutionary algorithms identified in the first part of this thesis. Experimental results demonstrate that the hybrid is capable of clearly outperforming canonical evolutionary algorithms and is better than simulated annealing for some cla...