Short Portfolio Training for CSP Solving

TL;DR

This paper introduces a simple, efficient portfolio approach for CSP solving that involves short-term training on specific instance sets, improving solver selection performance.

Contribution

It presents a novel short training method for portfolio systems using k-nearest neighbors, tailored for specific CSP instance sets, with demonstrated effectiveness.

Findings

k-nearest neighbors outperforms other machine learning methods

short training improves solver selection accuracy

approach generalizes well to SAT domain

Abstract

Many different approaches for solving Constraint Satisfaction Problems (CSPs) and related Constraint Optimization Problems (COPs) exist. However, there is no single solver (nor approach) that performs well on all classes of problems and many portfolio approaches for selecting a suitable solver based on simple syntactic features of the input CSP instance have been developed. In this paper we first present a simple portfolio method for CSP based on k-nearest neighbors method. Then, we propose a new way of using portfolio systems --- training them shortly in the exploitation time, specifically for the set of instances to be solved and using them on that set. Thorough evaluation has been performed and has shown that the approach yields good results. We evaluated several machine learning techniques for our portfolio. Due to its simplicity and efficiency, the selected k-nearest neighbors method is especially suited for our short training approach and it also yields the best results among the tested methods. We also confirm that our approach yields good results on SAT domain.