Large-Scale Expensive Optimization with a Switching Strategy

Some optimization problems in scientific research, such as the robustness for Internet of Things and neural architecture search, are large-scale decision space expensive objective evaluation. In order to get a good solution limited budget optimization, random grouping strategy is adopted divide problem into some low-dimensional sub-problems. A surrogate model then trained each sub-problem using different strategies select training data adaptively. After that, dynamic infill criterion proposed corresponding models currently used surrogate-assisted optimization. Furthermore, an escape mechanism keep diversity population. The performance method evaluated on CEC'2013 benchmark functions. Experimental results show that algorithm has better solving problems.