Towards Interpretable Routing Policy: A Two Stage Multi-Objective Genetic Programming Approach with Feature Selection for Uncertain Capacitated Arc Routing Problem

The Uncertain Capacitated Arc Routing Problem (UCARP) is the dynamic and stochastic counterpart of the well-known Capacitated Arc Routing Problem (CARP). UCARP has a wide range of real-world applications. One of the main challenge in UCARP is to handle the uncertain environment effectively. Routing policy-based approaches are promising technique for solving UCARP as they can respond to the uncertain environment in the real time. However, manually designing effective routing policies is time consuming and heavily replies on domain knowledge. Genetic Programming Hyper-heuristic (GPHH) has been successfully applied to UCARP to automatically evolve effective routing policies. However, the evolved routing policies are usually hard to interpret. In this paper, we aim to improve the potential interpretability of the GP-evolved routing policies by considering both program size and number of distinguished features. To this end, we propose a Two Stage Multi-Objective Genetic Programming Hyper Heuristic approach with Feature Selection (TSFSMOGP). We compared TSFSMOGP with the state-of-the-art single-objective GPHH, a two-stage GPHH with feature selection and a two-stage Multi-Objective GP. The experimental results showed that TSFSMOGP can evolve effective, compact, and thus potentially interpretable routing policies.