Addressing adjacency constraints in rectangular floor plans using Monte-Carlo Tree Search
Introduction
Laying out a floor plan is one of the key tasks in architecture design. It involves making decisions on the design and layout of all the rooms, usually in a 2D space, to satisfy various geometric and topological constraints. Conventionally, this has been a manual trial and error drawing process, where different pieces are adjusted, rearranged and reconfigured repetitively until a suitable floor layout that satisfies the various requirements eventually emerges [1]. This iterative manual process requires a significant amount of human labour and time, and becomes ever less possible as the size and complexity of the design problem increases. Due to the iterative and repetitive nature of this problem, automated computational techniques have replaced the manual design process and become the main approach for generating floor plans [2].
Many computational algorithms including heuristic search, mixed-integer programming have been successfully developed to generate satisfactory floor plans [3]. Especially, the evolutionary methods which have dominated this field in the last decade can generate a variety of layouts. However, the adjacency constraints tackled by most of these approaches are small-in-scale, and more importantly sparse-in-density, where the number of rooms is within 10 and the number of constraints is usually equally around (or at least no more than twice) the number of rooms. For example, Camozzato et al. [4] proposed a procedural method to generate a floor plan of 8 rooms with only 1 adjacency constraint. In [5], the authors illustrate a rectangular dissection method through an example of only 4 rooms with 3 adjacency constraints. Case study [6] tackles totally 9 adjacency constraints within 9 rooms, so the number of adjacency constraints is still no more than the number of rooms. Therefore, these approaches become inefficient with increased scale and density due to their limited scalability. For example, Rodrigues et al. [7] have applied the evolutionary methods to generate floor plans for a hotel up to 30 rooms, however the total number of adjacency constraints is only 34 and therefore still leads to a sparse adjacency matrix. Also, their case is not to generate a rectangular floor plan, therefore rooms can be placed in a more creative way with flexible boundaries. Finally, their algorithm had a runtime of 52 min on a 4GHz 8-core computer with multi-threading, which is not expensive when considering all kinds of granular constraints that were tackled in the original work. However, in case to address adjacency constraints only in initial floor plan, it may become not worth to apply the same approach. In addition to being limited to the small-scale and sparse-density of the adjacency constraints, this work hasn't considered the non-adjacency constraints. This paper tried to address the limitation of existing algorithms to handle high density topological adjacency and also non-adjacency constraints.
Topological adjacency constraint is one of the most important requirements during floor plan generation process, which defines the adjacency conditions between any pair of rooms. The complexity of topological adjacency constraints can be represented in terms of three factors: scale, density, and type of constraints. The first factor, the scale of constraints refers to the number of rooms nroom to place in a floor. Rooms can be any enclosed space. The larger number of the rooms we have, the larger scale of the adjacency constraints we need to tackle. The second factor, the density of constraints refers to the ratio between the number of constraints and the number of rooms nconstraints/nrooms. For example, in residential floor plans, the adjacency constraints are often small-in-scale and sparse-in-density where there are only a limited of total rooms, and the number of constraints is roughly equal to or even less than the number of rooms. Whereas in other complex scenarios such as hotel and school planning, the problem usually have high-dimensional and dense adjacency constraints with a larger number of rooms to locate, and the constraint density may be much higher. The third factor, the type of adjacency constraint refers to adjacency constraints and non-adjacency constraints that need to be tackled while generating the floor layout. Adjacency constraint is very common in most types of floor plan design, which requires two rooms to be next to each other. Non-adjacency constraint which requires that two rooms must not be adjacent, though less common, is also necessary for some practical problems. For example, in a hospital floor plan, some rooms are not only required to be adjacent to other rooms for convenient circulating reasons, but also required to be non-adjacent to some other rooms for isolation and infection control.
This paper proposes an efficient and lightweight algorithm which focuses on tackling highly dense adjacency constraint matrix, and taking into consideration both the adjacency and non-adjacency constraints. It uses off-policy Monte-Carlo Tree Search (MCTS) based reinforcement learning algorithm to solve this problem. The rest of this paper is divided into four sections. Section 2 gives a brief review of the related computational approaches on floor plan layout design. Section 3 first introduces the MCTS method and the problem definition, and then presents the proposed off-policy MCTS for solving the floor plan problem with highly-dense adjacency and non-adjacency constraints. Section 4 demonstrates two practical case studies to evaluate the capabilities of the proposed algorithm. Limitation and future work are discussed in Section 5. Finally, conclusions are drawn in Section 6.
Section snippets
State of the art: solving floor plan generation problem
Since the 1970s, researchers have developed computer-based approaches and algorithms for architecture design as detailed in the remaining part of this section. These approaches can be categorised into three main groups: conventional search methods, theoretical and mathematical proofs, and most recently the evolutionary approaches.
Modelling floor plan generation problem using off-policy Monte-Carlo tree search based reinforcement learning
In this section, the background of traditional MCTS is firstly introduced, and then give a formal definition of the floor plan problem with highly-dense constraints of both adjacency and non-adjacency types. Finally, off-policy MCTS is proposed to solve this problem.
Evaluation
The proposed algorithm is evaluated from two perspectives: time efficiency, and capability. The first case study aims to evaluate the time efficiency of the proposed algorithm in solving adjacency constraints. The proposed algorithm is compared with the Evolution Strategy by using the floor plan problem proposed in [27]. In the second case study, the aim is to validate the capability of the proposed algorithm for solving the problem with highly-dense adjacency constraints, where the proposed
Orthogonal polygon boundary and multi-story buildings
As presented above, this paper only shows how this algorithm can be applied to solve rectangular floor plan where both rooms and building boundary are in rectangular shape. However, we argue here that the proposed algorithm can also be similarly used for orthogonal polygons boundary. By following the rules in Section 3.3, the algorithm starts from most top-left point to place the next room, where “top” has higher priority than “left”, which means that when placing next room, we first look at
Conclusions
Inspired by the recent advanced searching and planning algorithms applied in AlphaGo, we propose a novel off-policy Monte-Carlo Tree Search to tackle the complex highly-dense adjacency and non-adjacency constrained floor plan problem in a time efficient and scalable manner. The proposed algorithm updates the state-value function to the max value of the historical total rewards it has ever seen instead of the average of the historical rewards in traditional on-policy MCTS. Two case studies are
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
The authors thank the Lloyd's Register Foundation, UK for funding this research through Data-Centric Engineering Programme at The Alan Turing Institute. The second author’s PhD research is co-funded by Bentley systems, UK and through a Skempton Scholarship from the Department of Civil and Environmental Engineering, Imperial College London.
References (32)
- et al.
An evolutionary strategy enhanced with a local search technique for the space allocation problem in architecture, part 1: methodology
Comput. Aided Des.
(2013) Automated facilities layout: past, present and future
Autom. Constr.
(2000)- et al.
Genetic search for solving construction site-level unequal-area facility layout problems
Autom. Constr.
(2000) - et al.
Clustering of architectural floor plans: a comparison of shape representations
Autom. Constr.
(2017) - et al.
An approach to the multi-level space allocation problem in architecture using a hybrid evolutionary technique
Autom. Constr.
(2013) - et al.
Customization and generation of floor plans based on graph transformations
Autom. Constr.
(2018) - et al.
From the generation of layouts to the production of construction documents: an application in the customization of apartment plans
Autom. Constr.
(2018) - et al.
Optimal space planning in practice
Comput. Aided Des.
(1981) - et al.
Alternative approaches to solve the multi-floor facility layout problem
J. Manuf. Syst.
(1997) Enumerating generic rectangular floor plans
Autom. Constr.
(2018)
EvoArch: an evolutionary algorithm for architectural layout design
Comput. Aided Des.
Evolving design genes in space layout planning problems
Artif. Intell. Eng.
An evolutionary strategy enhanced with a local search technique for the space allocation problem in architecture, part 2: validation and performance tests
Comput. Aided Des.
An evolutionary approach for 3D architectural space layout design exploration
Autom. Constr.
Procedural floor plan generation from building sketches
Vis. Comput.
Representation and generation of rectangular dissections
Cited by (26)
Architectural layout generation using a graph-constrained conditional Generative Adversarial Network (GAN)
2023, Automation in ConstructionArchitectural spatial layout planning using artificial intelligence
2023, Automation in ConstructionAutomated building layout generation using deep learning and graph algorithms
2023, Automation in ConstructionAutomated generation of floorplans with non-rectangular rooms
2023, Graphical ModelsA sequential space syntax approach for healthcare facility layout design
2023, Computers and Industrial EngineeringGaps and requirements for applying automatic architectural design to building renovation
2023, Automation in Construction