Online learning of variable ordering heuristics for constraint optimisation problems

Abstract Solvers for constraint optimisation problems exploit variable and value ordering heuristics. Numerous expert-designed heuristics exist, while recent research learns novel, customised from past problem instances. This article addresses unseen which no historical data is available. We propose one-shot learning of customised, instance-specific To do so, we introduce the concept deep , a data-driven approach to learn extended versions given heuristic online. First, instance, an initial online probing phase collects data, function learned. The learned can look ahead arbitrarily-many levels in search tree instead ‘shallow’ localised lookahead classical A restart-based strategy allows multiple models be acquired exploited solver’s optimisation. demonstrate based on smallest, anti first-fail, maximum regret Results instances MiniZinc benchmark suite show that solve 20% more improving overall runtime Open Stacks Evilshop problems.