A reinforcement learning hyper-heuristic in multi-objective optimization with application to structural damage identification

Multi-objective optimization allows satisfying multiple decision criteria concurrently, and generally yields multiple solutions. It has the potential to be applied to structural damage identification applications which are oftentimes under-determined. How to achieve high-quality solutions in terms of accuracy, diversity, and completeness is a challenging research subject. The solution techniques and parametric selections are believed to be problem specific. In this research, we formulate a reinforcement learning hyper-heuristic scheme to work coherently with the single-point search algorithm MOSA/R (Multi-Objective Simulated Annealing Algorithm based on Re-seed). The four low-level heuristics proposed can meet various optimization requirements adaptively and autonomously using the domination amount, crowding distance, and hypervolume calculations. The new approach exhibits improved and more robust performance than AMOSA, NSGA-II, and MOEA/D when applied to benchmark test cases. It is then applied to an active damage interrogation scheme for structural damage identification where solution diversity/completeness and accuracy are critically important. Results show that this approach can successfully include the true damage scenario in the solution set identified. The outcome of this research can potentially be extended to a variety of applications.