Project Scheduling: The Impact of Instance Structure on Heuristic Performance

Many meta-heuristic approaches have been suggested for or applied to the Resource Constrained Project Scheduling Problem (RCPSP). The existence of a number of highly accessible standard benchmark sets has promoted a research focus on finding anything that improves average solution quality, without investigating what effect is responsible for the improvement, or what is responsible for holding us back. This work focuses instead on understanding the original constructive Schedule Generation heuristics and their interaction with a well known but poorly understood post-processing step called Forward-Backward Improvement that is known to almost always improve any generated RCPSP schedule. We follow an empirical investigation methodology by first observing the effect of FBI on a large generated testset. Based on these observations we explain why FBI works by means of hypotheses on its operation. These hypotheses generate predictions that we subsequently successfully test in a second round of experiments. In the process we are able to propose a novel priority rule heuristic based on the principles of FBI. We find that this new rule outperforms the current best priority rule heuristic.