A deep reinforcement learning approach for solving the Traveling Salesman Problem with Drone

Reinforcement learning has recently shown promise in quality solutions many combinatorial optimization problems. In particular, the attention-based encoder-decoder models show high effectiveness on various routing problems, including Traveling Salesman Problem (TSP). Unfortunately, they perform poorly for TSP with Drone (TSP-D), requiring a heterogeneous fleet of vehicles coordination—a truck and drone. TSP-D, two are moving tandem may need to wait at node other vehicle join. State-less decoder fails make such coordination between vehicles. We propose hybrid model that uses an attention encoder Long Short-Term Memory (LSTM) network decoder, which decoder’s hidden state can represent sequence actions made. empirically demonstrate improves upon purely both solution computational efficiency. Our experiments min-max Capacitated Vehicle Routing (mmCVRP) also confirm is more suitable coordinated multiple than model. The proposed demonstrates comparable results as operations research baseline methods.