Advanced multi-agent pathfinding and optimization system for logistic networks

A system for managing logistics of package transportation from pickup to delivery, comprising:

• a processor and memory coupled to the processor, the processor executing code that implements—
• a database of vendors who pick up packages from their sources and deliver them to their destinations;
• a network represented by a graph where vertices represent hubs and facilities, and the connections with their characteristics are represented by edges;
• a geocoding engine (GE) that converts the package sources and the package destinations into a set of addresses represented by geospatial coordinates;
• a hub assignment (HA) engine that matches each package to its first hub after pickup and its last hub before delivery, as well as corresponding vendors from the database of vendors, for the pick-up and the delivery;
• a VRP (Vehicle Routing Problem) engine that assigns packages to vendors' vehicles and plans an optimal pickup or delivery sequence for each vehicle;
• a Reinforcement Learning (RL) agent trained by
• i. executing sequential stages of expanding graph scope from central hub-centric subgraphs to a full network, with nodes selected based on centrality ranking;
• ii. training the RL agent sequentially across each stage using Q-learning with LightGBM (Light Gradient-Boosting Machine)-based function approximation, wherein each completed stage initializes a subsequent stage;
• v. applying N-Secretary stopping criteria to intermediate training rounds by selecting a baseline performance and terminating the training when the RL agent's performance probabilistically exceeds the baseline performance;
• and
• vi. deploying the RL agent into a registry with warm-start capability for continued training;
• inputting a set of package sources and package destinations;
• a Multi-Modal Engine (MME) that uses the trained RL agent from the registry to attempt an optimal path for the packages from the first hub to the last hub when no deadlock occurs;
• wherein, when the RL agent is deadlocked, the trained RL agent is complemented by a time-skewed A* algorithm to find a valid path and resolves the deadlock by computing an alternate sub-path around the deadlock to a feasible hub, incrementally replanning only a portion of the optimal path that bypasses the deadlock, thereby avoiding a full recomputation of the optimal path
• wherein the RL agent takes into account path distance between the package sources and the package destinations, total time that a package is in a vehicle and time inside a hub, and cost of moving the package from the first hub to the last hub;
• and
• a dispatch module that dispatches the vehicles of the vendors to pick up and deliver the packages based on the optimal path.