Authors:
Ian Flood
1
and
Paris D. L. Flood
2
Affiliations:
1
Rinker School, University of Florida, Gainesville, FL 32611, U.S.A.
;
2
Dept. of Computer Science and Technology, University of Cambridge, Cambridge, U.K.
Keyword(s):
Construction Manufacturing, Construction Simulation, Decision Agents, Deep Artificial Neural Networks, Precast Reinforced Concrete Production, Reinforcement Learning, Rule-of-thumb Policies.
Abstract:
This paper is concerned with the development and evaluation of a reinforcement learning approach to the control of factory based construction operations. The unique challenges associated with controlling construction work is first discussed: uneven and uncertain demand, high customization, the need to fabricate work to order, and a lack of opportunity to stockpile work. This is followed by a review of computational approaches to this problem, specifically those based on heuristics and machine learning. A description is then given of a model of a factory for producing precast reinforced concrete components, and a proposed reinforcement learning strategy for training a neural network based agent to control this system. The performance of this agent is compared to that of rule-of-thumb and random policies for a series of protracted simulation production runs. The reinforcement learning method was found to be promising, outperforming the two competing strategies for much of the time. Thi
s is significant given that there is high potential for improvement of the method. The paper concludes with an indication of areas of proposed future research.
(More)