OPTIMIZATION AND EVALUATION OF A NEURAL NETWORK BASED POLICY FOR REAL-TIME CONTROL OF CONSTRUCTION FACTORY PROCESSES

This paper focuses on the development, optimization, and evaluation of an intelligent real-time control system for the fabrication of precast reinforced concrete components. The study addresses the unique challenges associated with real-time control in the construction manufacturing industry, including high customization, uncertain work demand, and limited stockpiling opportunities. A production system model is built based on a real construction manufacturing factory to simulate real-world precast reinforced concrete component fabrication, and acts as the basis for the development and validation of the control system. A review of alternative decision-making techniques is presented to identify the most suitable for the control of construction manufacturing factories. Ultimately, an artificial neural network approach trained using a reinforcement learning strategy is selected as a promising technique for effective real-time control. The controller is developed and validated, and its performance is optimized using sensitivity analysis, which takes into account both the structure of the artificial neural network and the parameters of the reinforcement learning algorithm. The ANN-based control policy is applied to the sequencing of precast reinforced concrete component production, while a rule-of-thumb policy is used as a benchmark for comparison. The study demonstrates that the optimized ANN-based control policy significantly outperforms the standard rule-of-thumb policy. The paper concludes by providing suggestions for further advancement of the ANN-based approach and potential avenues to increase the control policy's scope of application in construction manufacturing.