Transferring Information Across Restarts in MIP

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. In this paper, we will discuss four different types of global information that can potentially be re-used after a restart of a mixed-integer programming (MIP) solver, present technical details of how to carry them through a represolve after a restart, and show how such an information transfer can help to speed up the state-of-the-art commercial MIP solver FICO Xpress by 7% on the instances where a restart is performed.