Gelişigüzel düğüm dağılımlarına sahip GSP örneklerinin en iyi tur uzunluğunu tahminlemek için sinir ağı tahminleyicileri

To achieve operational efficiency in logistics, we need to solve complex routing problems. Due to their complexity, these problems are often solved sequentially, i.e., using cluster-first route-second (CFRS) type frameworks. However, such two-phase frameworks generally suffer from sub-optimality arising from the first phase. To mitigate this sub-optimality, information about optimal tour lengths of potential clusters can be exploited first, thereby transforming this two-phase approach into a less myopic solution framework. In that aspect, a quick and highly accurate Traveling Salesperson Problem (TSP) tour length estimator can be utilized for searching high-quality clusters. Motivated by this, we propose novel and computationally efficient neural network-based optimal TSP tour length estimators. Our approach uses an entirely new feature set consisting of node level, instance level, and solution level features by combining the power of artificial neural networks and theoretical knowledge in the routing domain. This data and knowledge hybridization enables us to achieve predictions with less than 0.7 percent deviation (on average) from the optimality. Unlike previous studies, we design and use new instances mimicking real-life logistics networks and morphologies. These instance characteristics introduce a substantial computational cost, making our instances harder to solve. To cope with these pathologies, we devise a new and efficient way of finding lower bounds and partial solutions to TSP later to be used as solution-level predictors. We also conduct a computational study where we produce up to 100 times lower prediction error on out-of-distribution test instances. Finally, we develop an enumeration-like mechanism by incorporating proposed machine learning models and metaheuristics to solve massive-scale rout- ing problems efficiently. We significantly outperform the state-of-the-art solver in terms of solution time and quality, demonstrating the potential of our models and the proposed method.