Transferring Information Across Restarts in MIP

Timo Berthold, Gregor Hendel, Domenico Salvagnin

Abstract: Restarting a solver gives us the chance to learn from things that went good or bad in the search until the restart point. The benefits of restarts are often justified with being able to employ different, better strategies and explore different, more promising parts of the search space. In that light, it is an interesting question to evaluate whether carrying over detected structures and collected statistics across a restart benefits the subsequent search, or even counteracts the anticipated diversification from the previous, unsuccessful search.