A Reinforcement Learning Neural Network Approach to Online Optimization of Construction Manufacturing Processes

Abstract

Precast component manufacturing has the benefits of high-level industrialization, being friendly to the environment, and quick installation. It helps solve the labor shortage and results in high construction efficiency. Due to the unique characteristics of construction manufacturing, for instance, the uncertainty of the component’s design, randomness in job arrival time, and lack of opportunity to stockpile precast components, prefabricated components are typically made-to-order. Online scheduling is required to minimize the delivery delay in real time.

This research study is concerned with the development of an effective method of dynamic control of precast factory production. The objectives are (1) to establish the ability of the reinforcement learning method for improving delivery performance relative to the benchmark techniques, namely, the random selection policy and the least remaining contingency policy; and (2) to determine the optimal structure for the neural network and the optimal reward length for training.

The well trained reinforcement learning agent would control the sequencing of the precast component job orders online in the simulated environment. A construction manufacturing system model would be built with certain prefabrication processes included. The prefabrication datasets are found from the literature and adjusted to be used for neural network training and production simulation.

The neural network’s structure and the reward length implemented in the training procedure are selected through sensitivity analyses. The well-trained reinforcement learning agent’s online sequencing performance is better than the least remaining contingency policy in the simulated prefabrication production test, benchmarked against the random selection policy. Suggestions for future work are also discussed from different perspectives. This research contributes to the application of the reinforcement learning agent in the potential smart construction manufacturing system to optimize construction project planning.