Searching For Relevant Features To Classify Crew Pairing Problems

Abstract

Machine learning (ML) is an emerging technology. Jeppesen, a leader of commercial optimization products in the airline industry, has started exploring ML methods to facilitate optimization algorithm development. This thesis investigates one of the company’s products, the crew pairing optimizer. The optimizer can use different algorithms to solve crew pairing problems, and the thesis looks into what features of a pairing problem influence algorithm selection, i.e. the best choice of algorithm for a problem, based on the performance of different algorithms. With little prior knowledge about features of pairing problems and their relation with algorithm performance, using ML, the thesis first generates over twenty features, and then uses different feature selection methods to find the most informative feature subsets. Each feature subset is then fed into multiple classifiers to test its robustness. Besides ML, the thesis also includes statistical analysis as a comparison. The thesis has some interesting findings, including a subset of features that might influence algorithm performance. However, none of the methods used can find a feature subset to accurately classify the pairing problems by the best performing algorithm. The thesis discusses possible reasons for the results. It also lists what to consider before applying ML to real-world problems.