Addressing adjacency constraints in rectangular floor plans using Monte-Carlo Tree Search

• Published in: Automation in construction 2020-07, Vol.115, p.103187, Article 103187
• Main Authors: Shi, Feng, Soman, Ranjith K., Han, Ji, Whyte, Jennifer K.
• Format: Article
• Language: English
• Subjects: Algorithm; Computer simulation; Constraints; Evolutionary algorithms; Floor plan generation; Floorplans; Floors; Generative design; Highly-dense adjacency and non-adjacency constraint; Monte Carlo simulation; Off-policy Monte-Carlo tree search; Reinforcement learning; Search algorithms
• ISSN: 0926-5805; 1872-7891
• DOI: 10.1016/j.autcon.2020.103187

Manually laying out the floor plan for buildings with highly-dense adjacency constraints at the early design stage is a labour-intensive problem. In recent decades, computer-based conventional search algorithms and evolutionary methods have been successfully developed to automatically generate various types of floor plans. However, there is relatively limited work focusing on problems with highly-dense adjacency constraints common in large scale floor plans such as hospitals and schools. This paper proposes an algorithm to generate the early-stage design of floor plans with highly-dense adjacency and non-adjacency constraints using reinforcement learning based on off-policy Monte-Carlo Tree Search. The results show the advantages of the proposed algorithm for the targeted problem of highly-dense adjacency constrained floor plan generation, which is more time-efficient, more lightweight to implement, and having a larger capacity than other approaches such as Evolution strategy and traditional on-policy search. •The proposed algorithm efficiently addresses adjacency constraints for floor plan.•Both adjacency and non-adjacency constraints are tackled.•The algorithm is based on Monte-Carlo Tree Search based reinforcement learning.•The proposed algorithm is time-efficient, lightweight, and scalable.