AI Tool Optimizes Homestay Layouts for Older Guests Using Floor Plans

A new computational framework can take an ordinary architectural floor plan of a small tourist accommodation and automatically generate renovation layouts that are easier for older adults to move through, safer to use at night, and simpler to navigate. The method, published in Computational Urban Science, links a deep-learning image parser, a spatial-analysis technique known as space syntax, and a hybrid optimization algorithm into a single closed loop that the authors call SSO. Rather than relying on the experience-based layouts that dominate most homestay renovations today, the framework turns age-friendly design concerns into numbers an algorithm can minimize.

The scope of the work is deliberately narrow and worth stating precisely. The researchers address interior accommodation spaces designed or renovated for older adult guests in short-term lodging scenarios, such as guest rooms, bathrooms, entryways, dining areas, and the circulation paths connecting them. This is not urban-scale planning and not the design of institutional elderly-care facilities. The authors also caution that their method does not directly measure long-term elderly user experience. Instead, it optimizes spatial proxy indicators related to accessibility, fall-risk exposure, visual recognizability, and wayfinding complexity, which are intended to support layout decisions before any human-centered validation takes place.

Technically, the pipeline begins with raster floor plans exported from Building Information Modeling software. A lightweight SegNet semantic segmentation model, trained with cross-entropy loss and an Adam optimizer over up to 200 epochs with early stopping, parses these images at the pixel level into walls, doors, corridors, rooms, bathrooms, furniture, and obstacles. Depthwise separable convolutions shrink the parameter count so the model can run on modest hardware, an important consideration for design iteration on mobile devices. The segmentation output then feeds a syntactic network built from an adjacency matrix of spatially connected nodes, from which the team computes connectivity, total and mean depth, and choice, a measure of how often a node lies on shortest paths between other locations.

A Visibility Graph Analysis layer supplements the topology. Sampling points are placed along corridor centerlines, entrance grids, and areas in front of key facilities, and a visibility matrix records whether unobstructed sightlines exist between pairs of points, accounting for walls, partitions, and furniture. The authors are explicit that these visibility indicators are spatial proxies rather than direct measurements of fall events or psychological security. Low visual connectivity does not inherently cause unsafe behavior, but it can flag locations where destinations, exits, restrooms, or caregivers lack visual accessibility, and the results are meant to be read together with accessibility constraints and later human-centered review.

The optimization stage defines six decision variables: door clear width, corridor clear width, turning radius, functional adjacency coding, a dry-wet separation boundary for bathrooms, and the positions of furniture and other obstacles. Three objective functions are minimized simultaneously: accessible path cost over key origin-destination pairs such as guest room to bathroom, weighted by the number of turns; a risk cost combining queueing-like congestion at intersections, slippery bathroom exposure, and collision proxies; and a cognitive or wayfinding cost derived from mean depth, connectivity, and obstructed visibility. Hard constraints enforce normative thresholds for widths and turning radii within the feasible domain set by load-bearing walls and functional integrity. A supervised multilayer perceptron, trained on expert preference annotations averaged across multiple assessors, adds a learned comprehensive cost score to reduce manual weight setting.

To solve this multi-objective problem, the team paired the genetic algorithm NSGA-II with Simulated Annealing. NSGA-II handles population initialization, non-dominated sorting, crowding distance, crossover, and mutation for global exploration, while the annealing criterion perturbs key design parameters locally and occasionally accepts worse solutions to escape local optima. The output is a Pareto front of trade-off solutions plus three representative recommendations: a balanced Knee solution, a safety-oriented minimum-risk solution, and a cost-oriented minimum-accessibility-cost solution, letting renovators weight preferences differently for ambulatory guests, wheelchair users, or people with low vision.

The experiments used an in-house dataset of BIM-exported floor plans split by layout typology rather than randomly, so that identical geometric patterns could not leak between training and testing sets. In segmentation comparisons, the best configuration, P1b, reached a mean Intersection over Union of 0.8883, with precision above 0.91 and recall above 0.90, outperforming an ablated variant and a U-Net baseline that plateaued near 0.86 and 0.83 respectively. Its F1 score stabilized around 0.93 after roughly 100 epochs, and distribution analysis showed its per-sample accuracy was both higher and more concentrated than the baselines.

On the optimization side, the hybrid NSGA-II-SA solver converged to a final comprehensive cost of 0.18, compared with 0.30 to 0.42 for ablated and baseline configurations, and achieved the highest hypervolume of 0.72 and the lowest inverted generational distance of 0.22, with differences against conventional algorithms reaching statistical significance. Its solutions clustered in the lower-left region of the accessibility-risk trade-off plot, at accessibility costs of roughly 0.20 to 0.50 and risk costs of 0.10 to 0.32. The extra sophistication did carry overhead: full optimization runs took about 56 to 60 minutes per scheme against 5 minutes for a simplified baseline, although per-generation time stayed in the 1.2 to 1.6 second range. Among the generated layouts, the cost-oriented scheme achieved weighted shortest-path costs as low as 34.79 on the entrance-to-guest-room route while reducing turns and bottlenecks, and compliance with net-width, turning-radius, slope, and wet-dry separation standards approached ideal levels.

Because the indicators are proxies, the authors supplemented the computational results with a walk-through evaluation by a five-expert panel including two accessibility design specialists, a senior care space designer, an interior renovation practitioner, and an architectural researcher. On a five-point scale covering passage safety, nighttime toilet convenience, visual recognition, furniture collision avoidance, and implementation feasibility, all three optimized schemes outperformed the manual renovation scheme. The safety-oriented variant scored highest on traffic safety and collision avoidance, the cost-oriented variant on nighttime toilet convenience, and the balanced Knee scheme received the highest overall preference score.

The researchers acknowledge real limitations. The dataset concentrates on particular apartment layout styles, the visibility analysis uses static two-dimensional occlusion modeling that ignores eye-level height, seated wheelchair sightlines, and obstruction by furniture of meaningful height, and the presentation of solution sets relied on two-dimensional projections of a few representative schemes. The framework, they stress, optimizes computable spatial indicators rather than proving long-term age-friendly experience, so its outputs should be treated as decision-support recommendations to be checked against local barrier-free standards, expert review, and feedback from older users before implementation in real projects. Within those bounds, the study demonstrates that automatic parsing, syntactic evaluation, and multi-objective optimization can be fused into one workflow that turns a floor plan into ranked, interpretable renovation options.

Subject of Research: A computational framework integrating SegNet semantic segmentation, space syntax, and hybrid multi-objective optimization to generate age-friendly homestay renovation layouts.

Article Title: Optimization design of age-friendly homestay spaces integrating Space syntax and SegNet

Article References: Dai, Y. (2026). Optimization design of age-friendly homestay spaces integrating Space syntax and SegNet. Computational Urban Science, 6(1), Article 53. https://doi.org/10.1007/s43762-026-00285-2

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00285-2

Keywords: age-friendly design, homestay, SegNet, space syntax, Visibility Graph Analysis, NSGA-II, Simulated Annealing, semantic segmentation, multi-objective optimization, floor plans, accessibility, older adults