Bayesian Hyperparameter Optimization with ALASPO

Abstract

The ALASPO framework applies Large Neighbourhood Search (LNS) to Answer Set Programming (ASP), which has proven to be effective in tackling a number of different optimization problems. To control the behaviour of the LNS, ALASPO offers many configuration options that in combination with all solving and grounding parameters of the underlying ASP solver result in a vast configuration space. Choosing a best from it poses a challenge, which we address by integrating the SMAC3 framework for Bayesian Optimization into ALASPO in order to enable the automatic tuning of ALASPO configurations. In addition, we cluster the provided problem instances based on their characteristics, allowing one to obtain multiple optimal configurations, one for each cluster of similar instances. After tuning, the result can be used for future ALASPO runs on unseen instances to automatically select and apply an appropriate configuration based on the characteristics of the given instance. We evaluate our methodology on a set of four benchmark problems and show that tuning can yield significant gains over using the default configuration for some of them, while also demonstrating its limitations in case of insufficient opportunities for advancement.