Meta-learning to select the best metaheuristic for the maxsat problem

Abstract

Several metaheuristics can be considered for solving a given optimization problem. Unfortunately none of them is better on all instances. Selecting a priori the best metaheuristic for a given instance is a difficult task which can be addressed using meta-learning. In this work, we propose a method to recommend, for a MaxSAT instance, the best metaheuristic among three: Genetic Algorithm (GA), Bee Swarm Optimization (BSO) and Greedy Randomized Adaptive Search Procedure (GRASP). Basically, a learning model is trained to induce associations between MaxSAT instances’ characteristics and metaheuristics’ performances. The built model is able to select the best metaheuristic for a new MaxSAT instance. We experiment different learning algorithms on different instances from several benchmarks. Experimental results show that the best metaheuristic is selected with a prediction rate exceeding 80% regardless the learning algorithm. They also prove the effect of instances used in training on the model performance.