Variable Neighborhood Descent with Self-Adaptive Neighborhood-Ordering

In Variable Neighborhood Descent (VND) it is often difficult to decide upon the ordering in which a different types of neighborhoods are considered. This arrangement typically strongly affects the quality of finally obtained solution as well as computation time. We present a VND variant which orders the neighborhoods dynamically in a self-adaptive way during the optimization process. Each neighborhood structure has associated a rating which is updated according to observed success probabilities and required times for evaluation. In this way, more effective neighborhood structures come to the fore and will be applied more frequently. An experimental comparison to a classical VND with a static neighborhood ordering is performed on the generalized edge biconnected network design problem. Results indicate that the selfadaptive VND requires substantially less time for finding solutions of comparable quality.