2018/03/26 by Daniel Karapetyan, Karapetyan, Daniel, Andrew J. Parkes +3 · 1 citation
Computer Science · #Data Mining Algorithms and Applications #Algorithms and Data Compression #Machine Learning and Algorithms
paper · pdf · doi:10.48550/arxiv.1803.09785
One way to speed up the algorithm configuration task is to use short runs\ninstead of long runs as much as possible, but without discarding the\nconfigurations that eventually do well on the long runs. We consider the\nproblem of selecting the top performing configurations of the Conditional\nMarkov Chain Search (CMCS), a general algorithm schema that includes, for\nexamples, VNS. We investigate how the structure of performance on short tests\nlinks with those on long tests, showing that significant differences arise\nbetween test domains. We propose a "performance envelope" method to exploit the\nlinks; that learns when runs should be terminated, but that automatically\nadapts to the domain.\n