Learning from evolutionary optimization by retracing search paths
Evolutionary search algorithms are used routinely to find optimal solutions for multi-parameter problems, such as complex pulse shapes in coherent control experiments. The algorithms are based on evolving a set of trial solutions iteratively until an optimum is reached, at which point the experiment ends. We have extended this approach by recording the best solution in each iteration and subsequently applying these to a modified system. By studying the shape of the learning curves in different systems, features of the fitness landscape are revealed that aid in deriving the underlying control mechanisms. We illustrate our method with two examples.