Towards Transferring Algorithm Configurations Across Problems

Abstract

Automatic approaches for algorithm configuration and design have received significant attention in the last years, thanks to both the potential to obtain high performing algorithms, and the ease for algorithm designers and practitioners. One limitation of current methods is the need to repeat the task for every new scenario encountered. We show how the observation of problem-independent features of the solution landscape can enable the use of past experiments to infer good configurations for unseen scenarios, both in case of new instances and new problems. As a proof of concept, we report preliminary experiments obtained when configuring a metaheuristic with two parameters.