There is a growing need for adaptive operational control to manage stochastic sample arrivals and strict turnaround deadlines in molecular diagnostics systems. Conventional static dispatching rules, however, often struggle to accommodate sudden surges in urgent samples, leading to bottleneck propagation across serial diagnostic processes. The primary objective of this study is to develop a multi-agent reinforcement learning framework for dynamic scheduling within a discrete event simulation environment. In this framework, independent deep Q-network agents at six serial diagnostic stages dynamically select from three dispatching rules―first-in-first-out, urgent-first, and earliest due date. These decisions are based on continuous-state observations of queue congestion, urgency mix, and remaining deadline, with hyperparameters determined through random search. The simulation environment was preliminarily validated by comparing Gantt-chart-based field schedules with simulation outcomes using a paired t-test. The proposed model was evaluated against a rule-based baseline, a genetic algorithm policy, and a tabular Q-learning model under three urgency-ratio regimes. During overload scenarios, the proposed framework demonstrated a statistically significant improvement in the high-urgency on-time rate compared to the rule-based baseline, while simultaneously reducing average waiting times. The genetic algorithm converged to a static, rule-equivalent mapping, whereas the critical-urgency on-time rate remained consistent across all four models. Overall, these results imply that continuous-state multi-agent reinforcement learning provides superior adaptive responsiveness in high-load regimes. Future research should focus on extending this framework to facility-level dispatching to address load imbalance among parallel machines.

1. 서 론
2. 관련연구
    2.1 동적 스케줄링과 강화학습 기반 최적화
    2.2 다중 에이전트 강화학습
3. 문제정의
    3.1 시스템 및 동적 환경 정의
    3.2 MDP 모델
4. 방법론
    4.1 이산사건 시뮬레이션 모델
    4.2 MDP 정의
    4.3 다중 에이전트 Deep Q-Network (DQN)
5. 실험
    5.1 비교 정책 정의
    5.2 실험 환경
    5.3 실험 결과
    5.4 실험 결과 고찰
6. 결 론
References

• Hong-Mo Yang(Department of Smart Production and Management Engineering, Hanbat National University) | 양홍모 (국립한밭대학교 스마트시스템경영공학과)
• Moonsoo Shin(Department of Industrial and Management Engineering, Hanbat National University) | 신문수 (국립한밭대학교 산업경영공학과) Corresponding author