Reconfigurable Manufacturing Systems aim to provide adaptability and responsiveness to changing production requirements. Within this context, the Facility Layout Problems play a central role, as layout decisions directly affect transportation cost, material flow efficiency, and operational performance. This thesis investigates the use of Reinforcement Learning and Deep Reinforcement Learning for solving static, stochastic, and dynamic Facility Layout Problems, including both weak and strong dynamic forms. A unified modelling and experimental framework is developed, integrating layout generation, Autonomous Mobile Robot trajectory construction, and discrete-event simulation. Static Facility Layout Problems are first analysed to study representation effects, scalability, and algorithmic performance. Several value-based and policy-based Deep Reinforcement Learning methods are evaluated and compared with Simulated Annealing. The analysis is then extended to stochastic and dynamic environments through a novel Multi-task Deep Reinforcement Learning framework, termed Episode-level Task Switching with Shared Replay Buffer. The proposed approach enables a single network to learn across multiple transportation demands while preserving learning stability and task diversity. This approach provides a structured mechanism for handling demand variability within a unified reinforcement learning model and offers a novel reinforcement learning perspective on layout optimization in reconfigurable manufacturing contexts.
