Explaining a Staff Rostering Genetic Algorithm using Sensitivity Analysis and Trajectory Analysis.

In the field of Explainable AI, population-based search metaheuristics are of growing interest as they become more widely used in critical applications. The ability to relate key information regarding algorithm behaviour and drivers of solution quality to an end-user is vital. This paper investigates a novel method of explanatory feature extraction based on analysis of the search trajectory and compares the results to those of sensitivity analysis using "Weighted Ranked Biased Overlap". We apply these techniques to search trajectories generated by a genetic algorithm as it solves a staff rostering problem. We show that there is a significant overlap between these two explainability methods when identifying subsets of rostered workers whose allocations are responsible for large portions of fitness change in an optimization run. Both methods identify similar patterns in sensitivity, but our method also draws out additional information. As the search progresses, the techniques reveal how individual workers increase or decrease in the influence on the overall rostering solution's quality. Our method also helps identify workers with a lower impact on overall solution fitness and at what stage in the search these individuals can be considered highly flexible in their roster assignment.