Improving the ALNS Algorithm for PDPTW Problem using a DDQN Agent

In response to practical challenges in food delivery logistics, we introduce a novel algorithm to address the Pickup and Delivery Problem with Time Windows (PDPTW). Our method employs a Double Deep Q-Network (DDQN)-based Large Neighborhood Search (LNS) strategy, which builds on the Adaptive Large Neighborhood Search (ALNS) framework, tailored specifically for the PDPTW. To mitigate the issues of slow convergence and instability observed in the ALNS algorithm, we have refined the state representation of the DDQN agent by integrating information on the current solution, historical data, and the time usage patterns of operators. We evaluated our algorithm using publicly available benchmark instances. The empirical results indicate that our approach not only surpasses the ALNS algorithm in terms of macro-level solution quality and stability but also demonstrates superior performance at the micro-level by achieving faster convergence and enhanced efficiency in operator usage strategy.