Meta-learning and meta-optimization

Meta-learning is a method of improving results of algorithm by learning from metafeatures which describe problem instances and from results produced by various algorithms on these instances. In this project we tried to apply this idea, which was already proved to be useful in machine learning, to combinatorial optimization. We have developed a general software tool called SEAGE to extract meta-features and tune parameters of heuristics for several combinatorial problems. We have also developed a specialized tool Pheromone for extraction of meta-features for Traveling Salesman Problem (TSP) [1], which serves as a benchmark for combinatorial optimization. Source codes and database with extracted metafeatures and results of optimization are available online as open source. We show preliminary results of meta-learning on results for several algorithms and large set of TSP instances. These results indicate that meta-optimization can improve quality of solution by automatic algorithm recommendation.