Fitness Landscapes of Buffer Allocation Problem in Production Lines and Genetic Algorithms Performance

We study structural properties of the buffer allocation problem from the fitness landscape perspective. We consider manufacturing flow lines with series-parallel network structure. The machines are supposed to be unreliable, their time to failure and repair time are exponentially distributed. We carry out computational experiments with local search and genetic algorithms in order to evaluate the fitness landscape properties of previously published instances and their modifications. We show that in many problem instances, several clusters of local optima can be identified. Besides that, the so-called 'massif central' or 'big valley' structure of the fitness landscape is present only partially.