Framework for Reinforcement Learning Production Control for One-Piece Flow Modular Concrete Structure Production with Incomplete Information

Cement accounts for a large proportion of global CO2 emissions, which can be reduced by prefabricating light weight modular concrete structures. For such modular construction methods with high accuracy requirements and rapid construction the separation of planning and production is futile due to occurring deviations in geometry, time, process and material. The processes must be connected in order to be able to compensate for deviations. For this purpose, a reinforcement learning agent can be used, which dynamically assigns modular concrete structures to a curing chamber and their processing time. In contrast to most approaches for RL in the literature, the considered problem is located at the interface between production control and process control. Thus, this paper presents a framework that outlines the use of RL in this problem domain by describing the simulation model, the actions of the RL agent as well as a possible state space and reward function.