Robust and Flexible Scheduling with Evolutionary Computation

Over the last ten years, there have been numerous applications of evolutionary al-gorithms to a variety of scheduling problems. Like most other research on heuris-tic scheduling, the primary aim of the research has been on deterministic formula-tions of the problems. This is in contrast to real world scheduling problems whichare usually not deterministic. Usually at the time the schedule is made some in-formation about the problem and processing environment is available, but thisinformation is uncertain and likely to change during schedule execution. Changesfrequently encountered in scheduling environments include machine breakdowns,uncertain processing times, workers getting sick, materials being delayed and theappearance of new jobs. These possible environmental changes mean that a sched-ule which was optimal for the information available at the time of scheduling canend up being highly suboptimal when it is implemented and subjected to the un-certainty of the real world. For this reason it is very important to find methodscapable of creating robust schedules (schedules expected to perform well after aminimal amount of modification when the environment changes) or flexible sched-ules (schedules expected to perform well after some degree of modification whenthe environment changes).This thesis presents two fundamentally different approaches for schedulingjob shops facing machine breakdowns. The first method is called neighbourhoodbased robustness and is based on an idea of minimising the cost of a neighbour-hood of schedules. The scheduling algorithm attempts to find a small set of sched-ules with an acceptable level of performance. The approach is demonstrated tosignificantly improve the robustness and flexibility of the schedules while at thesame time producing schedules with a low implementation cost if no breakdownoccurs. The method is compared to a state of the art method for stochastic sche-duling and concluded to have the same level of performance, but a wider area ofapplicability. The current implementation of the method is based on an evolution-ary algorithm, but since the real contribution of the method is a new performancemeasure, other implementations could be based on tabu search, simulated anneal-ing or other powerful “blind” optimisation heuristics.The other method for stochastic scheduling uses the idea of coevolution to