A conceptual model for provider production scheduling in a manufacturing-as-a-service systems using deep reinforcement learning
[EN] Manufacturing as a Service (MaaS) is an emerging paradigm in which a Provider exposes manufacturing capabilities such as CNC machines, additive manufacturing systems, and other production assets as on-demand services to multiple Consumers. In this setting, Provider side scheduling becomes a cri...
| Autores: | , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2026 |
| País: | España |
| Institución: | Universitat Politècnica de València (UPV) |
| Repositorio: | RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia |
| Idioma: | inglés |
| OAI Identifier: | oai:riunet.upv.es:10251/232247 |
| Acceso en línea: | https://riunet.upv.es/handle/10251/232247 |
| Access Level: | acceso abierto |
| Palabra clave: | Manufacturing-as-a-Service Dynamic Scheduling Deep Reinforcement Learning Unrelated parallel machine scheduling |
| Sumario: | [EN] Manufacturing as a Service (MaaS) is an emerging paradigm in which a Provider exposes manufacturing capabilities such as CNC machines, additive manufacturing systems, and other production assets as on-demand services to multiple Consumers. In this setting, Provider side scheduling becomes a critical aspect, as orders arrive dynamically and must be allocated to heterogeneous resources while meeting contractual constraints. This paper proposes an implementable conceptual framework for Provider production scheduling in MaaS, formulating the problem as a Markov Decision Process and enabling Deep Reinforcement Learning based decision making. The proposed Provider Planner models resource allocation as an unrelated parallel machine scheduling problem, where decisions are taken at discrete event-driven decision epochs (e.g., order arrivals and resource releases). The framework explicitly specifies the observation space, action space, and discrete-event transition logic, incorporating practical features such as resource availability and efficiency, operating cost rates, setup states across product families, batch-size constraints, order time windows (earliest/latest start), due dates, and delay penalties. A multi-objective reward formulation is defined to jointly minimise tardiness and associated penalties, overall makespan, and total production cost computed from resource uptimes. The resulting model provides a structured basis for developing and evaluating adaptive scheduling policies for MaaS Providers under demand variability and complex operational constraints. |
|---|