Improving Delivery Performance of Construction Manufacturing Using Machine Learning
Keywords:Construction Manufacturing, Deep Artificial Neural Networks, Delivery Performance, Precast Reinforced Concrete Components, Reinforcement Learning
This paper is concerned with the development, testing, and optimization of a machine learning method for controlling the production of precast reinforced concrete components. A discussion is given identifying the unique challenges associated with achieving production efficiency in the construction industry, namely: uncertain and sporadic demand for work; high customization of the design of components; a need to produce work to order; and little prospect for stockpiling work. This is followed by a review of the methods available to tackle this problem, which can be divided into search-based techniques (such as heuristics) and experience-based techniques (such as artificial neural networks). A model of an actual factory for producing precast reinforced concrete components is then described, to be used in the development and testing of the controller. A reinforcement learning strategy is proposed for training a deep artificial neural network to act as the control policy for this factory. The ability of this policy to learn is evaluated, and its performance is compared to that of a rule-of-thumb and a random policy for a series of testing production runs. The reinforcement learning method developed an effective and reliable policy that significantly outperformed the rule-of-thumb and random policies. An additional series of experiments were undertaken to further optimize the performance of the method, ranging the number of input variables presented to the policy. The paper concludes with an indication of proposed future research designed to further improve performance and to extend the scope of application of the method.
How to Cite
Copyright (c) 2023 Ian Flood, Xiaoyan Zhou
This work is licensed under a Creative Commons Attribution 4.0 International License.
All manuscripts published in JSimE is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0). Submission of a manuscript to JSimE assumes the acceptance of this license.