Multi-Agent Training-free Urban Food Delivery System using Resilient UMST Network

A team of researchers including Md Nahid Hasan, Vishwam Tiwari, and Aditya Challa has introduced a novel algorithm called the Union of Minimum Spanning Trees (UMST) to revolutionize urban food delivery networks. The system addresses a core logistics challenge: balancing efficiency with resilience. Traditional fully connected graphs are computationally impossible at city scale, while single Minimum Spanning Trees (MSTs) are efficient but fragile to disruptions like road closures. UMST solves this by generating multiple MSTs through randomized edge selection and uniting them, creating a sparse network that still maintains multiple alternative routes between key delivery hubs.

This training-free, multi-agent approach delivers remarkable performance without the computational overhead of machine learning models. In tests across multiple U.S. cities, UMST networks used 20 to 40 times fewer edges than a fully connected graph, drastically reducing complexity. The system enabled substantial order bundling, with 75-83% of orders participating in shared deliveries. Crucially, it matched or outperformed sophisticated learning-based baselines like MADDPG and Graph Neural Networks, achieving 88-96% delivery success rates and reducing travel distance by 44-53%.

The operational advantages are significant. By forgoing the need for training data and model tuning, UMST executes routing decisions 30 times faster than its AI-powered counterparts. This speed and simplicity, combined with its interpretable network structure, make it a highly practical solution. As the global Online Food Delivery market heads toward a projected $500 billion by 2030, the research provides a scalable, resilient, and efficient algorithmic foundation for the next generation of urban logistics platforms, promising faster deliveries and lower operational costs.