Measuring mobility and the performance of global search algorithms

The global search properties of heuristic search algorithms are not well understood. In this paper, we introduce a new metric, mobility, that quantifies the dispersion of local optima visited during a search. This allows us to explore two questions: How disperse are the local optima visited during a search? How does mobility relate to algorithm performance? We compare local search with two evolutionary algorithms, CHC and CMA-ES, on a set of non-separable, non-symmetric, multi-modal test functions. Given our mobility metric, we show that algorithms visiting more disperse local optima tend to be better optimizers.