Intra-hospital transportation problem: A two-phase bayesian deep reinforcement learning framework

The rapid growth in intra-hospital transportation demand, driven by increasing patient flows and operational complexity, has placed critical pressure on hospital logistics systems. Transportation requests, ranging from time-sensitive patient transportation to the delivery of medical supplies and equipment, must be fulfilled with high reliability while managing human resource constraints and emerging automation technologies. This thesis addresses the Real-Time Intra-Hospital Transportation Problem (RIHTP), aiming to optimize the assignment and routing of both human porters and service robots under dynamically arriving and heterogeneous transportation requests. We first develop a detailed problem formulation that integrates recharging requirements for robots, compatibility constraints, workload balancing, and urgency-sensitive delay penalties. For the static setting, the problem is modeled as a Dial-a-Ride Problem (DARP) and solved using a Mixed Integer Linear Programming (MILP) formulation to benchmark solution quality. For the dynamic setting, we propose a novel two-phase solution methodology: Phase 1 employs a Bayesian Deep Reinforcement Learning framework to assign requests in real time under uncertainty, and Phase 2 applies a Best Insertion heuristic followed by Adaptive Variable Neighborhood Search (AVNS) for route construction and consolidation. Using a suite of realistically simulated hospital scenarios, our computational experiments demonstrate that the proposed Bayesian-based assignment policy outperforms existing baseline policies in the heterogeneous environment of hospital transportation, achieving a 9.26% improvement over the current hospital policy. Furthermore, the proposed two-phase solution approach yields a 10.88% reduction in total objective value compared to the conventional one-phase reoptimization method widely adopted in the literature, while simultaneously reducing average runtime from 954.97 seconds to 299.56 seconds. Furthermore, experiments investigating other operational and tactical decisions highlight the substantial benefits of integrating service robots into the transportation system: the average workload of human porters decreases by approximately 15%, delay penalties are reduced by up to 85%, and overall objective function values drop by 13%. These findings collectively underscore the efficacy and scalability of our proposed framework in enhancing responsiveness, equity, and efficiency within real-time hospital logistics. The proposed methods are shown to be effective across diverse operational conditions, including varying request arrival rates, transporter fleet configurations, and robot eligibility scenarios. Additionally, the framework captures the complex interplay between assignment policies, route optimization, and automation integration, offering practical insights into transporter coordination and workload management. In summary, this thesis contributes scalable, cost-efficient, and policy-aware solutions for real-time intra-hospital transportation, providing actionable guidance for smart hospital logistics, automation planning, and the future integration of learning-based decision support systems in healthcare environments.