The exponential growth of e-commerce and rapid delivery services has made last-mile delivery one of the most significant and complex operational challenges for contemporary urban mobility. Managing fleets of couriers across a given area requires balancing multiple needs: ensuring timely deliveries within agreed-upon time windows, reducing the impact of traffic, and ensuring fair and sustainable workloads for operators. However, while traditional mathematical optimization techniques struggle to adapt to unforeseen events and unstructured instructions, pure generative artificial intelligence models offer great interpretive capabilities but lack an understanding of network topology and strict adherence to physical constraints. The goal of this thesis is to overcome this dichotomy by designing a hybrid architecture that combines the modeling of complex networks using typed hypergraphs with agent-based orchestration guided by large language models (LLMs). The work defines an approach structured around two decision-making horizons, integrating morning strategic planning with reactive management of afternoon requests as they stream in. The system is ultimately validated using real operational data, demonstrating its ability to improve service punctuality, minimize total travel distances, balance the workload among operators, and ensure stable and efficient responses in daily operations.
La crescita esponenziale dell'e-commerce e dei servizi di consegna rapida ha reso la logistica dell'ultimo miglio (last-mile delivery) una delle sfide operative più rilevanti e complesse per la mobilità urbana contemporanea. Gestire flotte di corrieri sul territorio richiede di armonizzare molteplici esigenze: garantire consegne tempestive entro finestre orarie concordate, ridurre l'impatto del traffico e assicurare carichi di lavoro equi e sostenibili per gli operatori. Tuttavia, mentre le tecniche classiche di ottimizzazione matematica faticano ad adattarsi a imprevisti e a istruzioni non strutturate, i modelli di Intelligenza Artificiale generativa pura offrono grandi capacità interpretative ma difettano di comprensione della topologia della rete e rispetto rigoroso dei vincoli fisici. L'obiettivo di questa tesi è superare tale dicotomia progettando un'architettura ibrida che unisce la modellazione delle reti complesse tramite ipergrafi tipizzati con l'orchestrazione agentica guidata da modelli linguistici di grandi dimensioni (Large Language Models). Il lavoro definisce un approccio articolato su due orizzonti decisionali, integrando la pianificazione strategica mattutina con la gestione reattiva delle richieste pomeridiane in streaming. Il sistema viene infine validato su dati operativi reali, dimostrando la capacità di migliorare la puntualità del servizio, contenere le percorrenze complessive, equilibrare il carico tra gli operatori e garantire risposte stabili ed efficienti nell'operatività quotidiana.
Pianificazione della consegna dell’ultimo miglio basata su ipergrafi tipizzati con orchestrazione semantica tramite LLM
MANDORLINI, MATTIA
2025/2026
Abstract
The exponential growth of e-commerce and rapid delivery services has made last-mile delivery one of the most significant and complex operational challenges for contemporary urban mobility. Managing fleets of couriers across a given area requires balancing multiple needs: ensuring timely deliveries within agreed-upon time windows, reducing the impact of traffic, and ensuring fair and sustainable workloads for operators. However, while traditional mathematical optimization techniques struggle to adapt to unforeseen events and unstructured instructions, pure generative artificial intelligence models offer great interpretive capabilities but lack an understanding of network topology and strict adherence to physical constraints. The goal of this thesis is to overcome this dichotomy by designing a hybrid architecture that combines the modeling of complex networks using typed hypergraphs with agent-based orchestration guided by large language models (LLMs). The work defines an approach structured around two decision-making horizons, integrating morning strategic planning with reactive management of afternoon requests as they stream in. The system is ultimately validated using real operational data, demonstrating its ability to improve service punctuality, minimize total travel distances, balance the workload among operators, and ensure stable and efficient responses in daily operations.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi-Mandorlini.pdf
accesso aperto
Dimensione
986.35 kB
Formato
Adobe PDF
|
986.35 kB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/29212