Illuminating Spaces: Deep Reinforcement Learning and Laser-Wall Partitioning for Architectural Layout Generation

Benjamin Dillenburger; Reza kakooee

REVIEW 4 major objections 6 minor 14 references

Illuminating Spaces: Deep Reinforcement Learning and Laser-Wall Partitioning for Architectural Layout Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A deep reinforcement learning agent, acting on laser-wall partitions in the SpaceLayoutGym simulator, can generate 4-to-9-room layouts whose areas and aspect ratios stay within 5% of the target and that satisfy 70 of 72 required…

desk verdict Genuinely new laser-wall partitioning for RL-based layout generation, but the quantitative claims are under-supported without code, per-seed stats, and a reachability analysis. read the letter →

arxiv 2502.04407 v1 pith:DJJHC3R7 submitted 2025-02-06 cs.LG cs.AI

Reza Kakooee , Benjamin Dillenburger This is my paper · ORCID

classification cs.LGcs.AI

keywords deepreinforcementlearningspacelayoutdesignlaser-wallpartitioningdynamicplanningproximalpolicyoptimizationarchitecturalfloorplangenerationadjacencyconstraintsGym

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that space layout design can be recast as a reinforcement learning problem, provided the partition method is expressive enough for an agent to explore. It introduces 'laser-wall' partitioning, in which each wall is a hard base plus imaginary light beams that cut space until they meet other walls or beams, producing rooms without pixel-level rendering. On top of this, the paper builds SpaceLayoutGym, a simulator where a Proximal Policy Optimization agent starts from a random wall configuration and refines it through discrete moves and rotations. Across six design scenarios with 4 to 9 rooms, the trained agent produces layouts whose room areas and aspect ratios are within 5% of the targets and that satisfy 70 of the 72 required adjacencies. If the results hold, the method offers an architecturally intuitive, procedural alternative to image-based generative layout tools.

What carries the argument

The load-bearing object is the laser-wall: a hard base wall made of two connected segments plus imaginary light beams emitted from its ends. The beams travel until they hit the plan boundary, another base wall, or another beam, and the infiltration rate decides whether a beam can cut through a beam it meets. This converts a wall placement into a complete partition in one shot, and the discrete action space (8 directional moves, 90-degree rotations, two transformation modes) turns layout refinement into a Markov decision process whose state is an RGB image of the current plan. The reward function balances hard constraints, geometric closeness, and adjacency bonuses, and Proximal Policy Optimization trains the policy.

What would settle it

Train the same PPO setup on a scenario whose only feasible layout requires a partition shape not in the predefined wall library, such as a T-junction or a non-orthogonal intersection that the infiltration rules forbid, and check whether the agent can still meet the reward thresholds; a systematic failure in such cases would show the successes depend on favorable wall choices rather than on the laser-wall mechanism itself.

Watch

Extended reading notes

Core claim

The central claim is that a deep-RL agent, acting on laser-wall partitions through dynamic planning, can generate diverse and functional architectural layouts that simultaneously satisfy geometric constraints (room areas, aspect ratios) and topological requirements (adjacencies to living room and facade). The evidence is six design scenarios: averaged area and aspect-ratio errors under 5%, and 70 of 72 required connections established. The paper attributes this to the laser-wall representation, which bridges vector-based and pixel-based partitioning, and to the on-light/off-light transformation scheme that lets walls be moved and rotated while keeping room identity coherent.

Load-bearing premise

The laser-wall library and the 8-directional, 90-degree-rotation action space can transform any initial random wall configuration into any layout the reward function demands; the paper does not analyze this reachability.

Editorial extensions

If this is right

Architects can use a trained agent to explore diverse functional layouts for briefs of 4 to 9 rooms, with room areas and aspect ratios within 5% of the requested values.
Angled walls achieve adjacency requirements more reliably than straight-only walls, indicating that layout methods should allow non-rectangular partitions.
Because the agent works from an image state and discrete transformations, the same setup can be re-run for new scenarios without redefining the environment, only the reward and scenario table.
Both on-light and off-light transformations support successful training, so the laser-wall partitioning is robust to the choice of transformation mechanics.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The MDP framing is domain-neutral, so the same laser-wall plus PPO recipe could be applied to facility layout, urban block layout, or other partition problems, provided the wall library is extended accordingly.
A quantitative comparison against pixel-based generative models on coverage, constraint satisfaction, and wall length would test whether the claimed 'intuitive' procedural representation actually yields better architectural outcomes, something the paper does not yet provide.
The 2-of-72 missed adjacencies may reflect a bound of the current action space; extending it with diagonal or multi-segment walls could close the gap.
A reachability analysis of the laser-wall action space would turn the empirical success into a statement about what layouts are generatable and would predict scenario difficulty before training.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Reza et al. introduce laser-wall partitioning, where walls emit light beams that partition space, and combine it with PPO in a gym-style simulator. The core idea is new relative to the pixel-based GAN/diffusion work and the earlier RL-for-rectangular-dissection work, and the on-light/off-light transformation distinction is a useful engineering contribution. The qualitative layouts look plausible, and the open-source simulator is a plus.

What I can't verify is the quantitative story. The reward function is described only as a list of penalty categories; no weights, no terminal thresholds, no PPO hyperparameters. The claims 'less than 5% area error' and '70 of 72 adjacencies' are reported as aggregate numbers with no per-seed variance or error bars, and no baseline comparison against any existing floor-plan generator. Since area, aspect ratio, and adjacency targets are encoded in the reward, the reported successes are partly the optimizer telling you it optimized what it was told to optimize; that's not meaningless, but it's weaker than the abstract suggests.

The more load-bearing gap is reachability. Dynamic planning starts from a random wall configuration and the only actions are translating and rotating existing walls. There is no add/remove/resize. So every state lies in the translation/rotation orbit of the initial wall set. The paper never says how the initial walls are sampled, how many walls there are relative to the target room count, or whether the action space can actually reach layouts that meet the constraints. If the orbit can't reach a target for some scenario, no policy can succeed, and the reported successes would be an artifact of initialization. That is not a fatal flaw by itself—the method clearly does find good layouts in the shown cases—but it means the headline claims are under-supported.

My recommendation: send it to peer review, but only with major revisions. The authors should release the code with a commit hash, specify the reward function and hyperparameters exactly, report per-seed statistics, and add at least one baseline (even a simple optimizer or a previous floor-plan generator). They should also either analyze reachability of the action space or state clearly that the method is a local refinement of a given wall set rather than a complete generator. The idea is worth engaging with; the evidence as presented is not.

Referee Report

4 major / 6 minor

Summary. This paper proposes "laser-wall" partitioning, a space-composition representation in which walls consist of a hard base segment plus emitted light beams that partition a plan, and uses it to frame architectural space layout design as an MDP. The authors build SpaceLayoutGym, an OpenAI-Gym-compatible simulator, and train PPO agents with two variants of "dynamic planning" (on-light and off-light wall transformations) on six design scenarios of 4 to 9 rooms. The paper reports that trained agents produce layouts whose room areas and aspect ratios differ from desired values by less than 5% on average, and that 70 of 72 required adjacencies are satisfied across all scenarios; it also compares straight versus angled walls and visualizes the action trajectory of one agent. One-shot planning is mentioned but deferred to previous work, and the experiments focus exclusively on dynamic planning.

Significance. If the reported results survive closer scrutiny, the paper's main contribution is a procedural, interpretable RL formulation of layout synthesis, and the laser-wall representation is a genuinely distinct middle ground between vector-based and pixel-based partitioning. The open-sourced simulator (claimed in Section 4) and the explicit disclosure of failed adjacencies (two missing connections) are commendable. The trajectory analysis in Figure 6 gives a rare view of how a trained agent refines a layout, which is valuable for designer-facing tools. However, the significance is currently capped by the absence of quantitative per-scenario results, seed-level statistics, and any comparison to the prior generative (e.g., House-GAN++, HouseDiffusion) or one-shot-policy baselines the paper criticizes; the strength of the central claim is therefore hard to calibrate.

major comments (4)

[2.5, 2.8] The dynamic-planning action space contains only two decisions—selecting an existing wall and applying one of 8 directional moves or 6 rotation options—so every reachable layout lies in the translation/rotation orbit of the initial random wall configuration described in Section 2.8. If that orbit cannot produce the target areas, aspect ratios, and adjacencies within the acceptance threshold, the task is infeasible for the MDP and any success would be an artifact of initialization or scenario choice rather than evidence for the laser-wall method. The paper neither provides a completeness or reachability argument nor reports success rates over repeated random initializations, so the reader cannot determine whether the layouts in Figures 3 and 5 are typical outcomes or selected positive cases.
[3.1.1] The headline quantitative claims (area and aspect-ratio errors "both less than 5%" and "all but 2 of 72" connections) are presented only as aggregate prose, with no per-scenario table of desired versus achieved areas and aspect ratios, no standard deviations across training seeds, and no failure rate over episodes. Because the reward function in Section 2.6 directly encodes these same geometric and topological targets, the reported numbers partly measure how well the reward was optimized; without the underlying error distributions, the claim that the approach "can generate diverse and functional space layouts" is not quantitatively established.
[2.6] The reward function is specified only as qualitative bullet categories (negative penalties for hard constraints and deviations, non-negative terminal rewards within an unspecified threshold, and an adjacency bonus), and the reward weights, terminal acceptance thresholds, and scaling functions are never given. Section 4 states that SpaceLayoutGym is "open sourced" but provides no repository URL or version identifier, and Section 2.7 gives no PPO hyperparameters or network architecture details; these omissions make the experiments impossible to reproduce or compare against.
[3] No quantitative baseline is provided against any prior method (e.g., House-GAN++, HouseDiffusion, or the authors' own one-shot planning work cited as Kakooee and Dillenburger 2024), even though Section 1 motivates the paper by the limitations of pixel-based generative approaches. As a result, the relative claim of flexibility and exploratory power is supported only by qualitative figure inspection, and the reader cannot assess whether laser-wall dynamic planning improves over existing methods on the same scenarios.

minor comments (6)

[2.6] The sentence "The reward consists of two components: instant reward, and terminal reward. The reward is calculated from two elements: an instant reward and a terminal reward." is duplicated; one occurrence should be removed.
[2.7] There are small language errors: "following by fully connected layers" should read "followed by fully connected layers," and "SDL" in Section 2.3 should be "SLD."
[4] The open-source claim names the project as "SpaceLayoutGym" without a URL or a citation; a working link or DOI is needed for the reproducibility claim to be actionable.
[2.1.1] The infiltration rate is described verbally (beams stop upon encountering other beams, or the rate decreases with distance) but never defined formally; an equation would be needed to implement the fixed versus decreasing variant.
[3.1] The sentence "The agent's ability to find a design solution indicates that the generated layouts closely match the geometric properties... because geometric properties are integral constraints of the optimization problem" is circular, since a "solution" is defined as a layout satisfying those constraints; the claim should instead be supported by the reported error metrics.
[Figures 3 and 5] The figure captions do not say which design scenario each panel corresponds to; adding panel labels referencing Table 1 would make the per-scenario claims checkable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RL results are genuine optimization outcomes, not reductions to the reward definition, and the sole self-citation is not load-bearing.

full rationale

The paper makes no first-principles derivation or prediction in the sense required for circularity. Its central demonstration is that a PPO agent, trained in SpaceLayoutGym with a reward that penalizes geometric and topological deviations, produces layouts that score well on those same geometric and topological criteria. This is an empirical optimization result: maximizing the terminal reward requires nontrivial search over the discrete wall-transformation action space, and the reward definition alone does not guarantee that any policy reaches a satisfying layout. Although the reported figures (area and aspect-ratio errors below 5%, 70 of 72 adjacencies) mirror the terminal-reward criteria, this is objective alignment rather than construction-equivalence. The only self-citation, to Kakooee and Dillenburger (2024) for one-shot planning, is explicitly set aside by the statement 'we focus exclusively on dynamic planning' and is not load-bearing for the dynamic-planning results. The unexamined reachability of the initial random wall configurations is a validity or correctness threat, not a circularity: it concerns whether the action space can express target layouts, which is an assumption about the MDP rather than a conclusion that reduces to its inputs. No step in the manuscript equates a claimed output with an input by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim depends on several unreported constants: reward weights, terminal thresholds, infiltration rates, wall library contents, and training hyperparameters. None of these are given numeric values in the paper, so a reader cannot tell whether the reported 'less than 5%' and '70 of 72' results are robust or tuned per scenario.

free parameters (5)

Reward penalty weights and terminal threshold
Section 2.6 lists instant and terminal reward components qualitatively but does not report the numerical weights or acceptance threshold. These shape all reported success rates.
Infiltration rate
Section 2.1.1 defines fixed and decreasing infiltration rates but gives no values; beam behavior depends on these unspecified constants.
Wall library contents
Section 2.2 refers to a 'predefined wall library'; the set of available base walls and lengths is not listed, so the expressiveness of the action space is unknown.
Maximum transformations per episode
Section 2.8 says episodes terminate when a satisfactory layout is reached or a maximum number of transformations is reached, but the maximum is not stated.
PPO and network hyperparameters
No learning rate, batch size, architecture depth, or training iterations are reported; Section 2.7 only describes convolutional layers and fully connected layers.

assumptions (4)

domain assumption The MDP state can be represented as an RGB image of the current layout, and the transition depends only on the current image and action.
Section 2.4 uses an RGB image as the state space, but no proof is given that this representation is Markov or sufficient for the design criteria.
domain assumption Laser-wall light propagation and infiltration rules always yield a valid partition into subregions that can be assigned to rooms.
Section 2.1.1 describes beams stopping at obstacles, but no consistency proof is given for overlapping beams or ties.
domain assumption The discrete action space (8 directions, 90-degree rotations) can reach a near-optimal layout from any initial random wall configuration.
Section 2.2 defines dynamic planning transformations; reachability and completeness are never analyzed.
domain assumption PPO with the described convolutional architecture converges to a policy that maximizes the reward within the training budget.
Sections 2.7 and 2.8 rely on standard PPO convergence behavior, but no convergence guarantee or seed analysis is provided.

invented entities (2)

Laser-wall (base wall plus emitted light beams)
purpose: Partition the floor plan procedurally; beams define room boundaries via infiltration rates.
It is an algorithmic abstraction, not a physical object; the paper provides no external falsifiable handle, and its architectural validity depends entirely on downstream evaluation.
On-light and off-light wall transformations
purpose: Allow walls to move and rotate while updating the partition; off-light deactivates the moved wall's light, on-light keeps it active.
These are internal environment mechanics, not independently testable entities; they affect agent behavior but are not validated against external criteria.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Illuminating Spaces: Deep Reinforcement Learning and Laser-Wall Partitioning for Architectural Layout Generation." pith.science (2026). https://pith.science/paper/DJJHC3R7

@misc{pith2026250204407,
  author       = {Pith},
  title        = {Pith review of: Illuminating Spaces: Deep Reinforcement Learning and Laser-Wall Partitioning for Architectural Layout Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DJJHC3R7}},
  note         = {Machine review of arXiv:2502.04407}
}

read the original abstract

Space layout design (SLD), occurring in the early stages of the design process, nonetheless influences both the functionality and aesthetics of the ultimate architectural outcome. The complexity of SLD necessitates innovative approaches to efficiently explore vast solution spaces. While image-based generative AI has emerged as a potential solution, they often rely on pixel-based space composition methods that lack intuitive representation of architectural processes. This paper leverages deep Reinforcement Learning (RL), as it offers a procedural approach that intuitively mimics the process of human designers. Effectively using RL for SLD requires an explorative space composing method to generate desirable design solutions. We introduce "laser-wall", a novel space partitioning method that conceptualizes walls as emitters of imaginary light beams to partition spaces. This approach bridges vector-based and pixel-based partitioning methods, offering both flexibility and exploratory power in generating diverse layouts. We present two planning strategies: one-shot planning, which generates entire layouts in a single pass, and dynamic planning, which allows for adaptive refinement by continuously transforming laser-walls. Additionally, we introduce on-light and off-light wall transformations for smooth and fast layout refinement, as well as identity-less and identity-full walls for versatile room assignment. We developed SpaceLayoutGym, an open-source OpenAI Gym compatible simulator for generating and evaluating space layouts. The RL agent processes the input design scenarios and generates solutions following a reward function that balances geometrical and topological requirements. Our results demonstrate that the RL-based laser-wall approach can generate diverse and functional space layouts that satisfy both geometric constraints and topological requirements and is architecturally intuitive.

Figures

Figures reproduced from arXiv: 2502.04407 by the authors.

**Figure 2.** Illustration of off-light and on-light transformations in laser-wall partitioning [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison of layouts generated using straight walls only (top row) and angled walls only (bottom row). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Average rewards and episode lengths over training iterations . [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Comparison of space layouts generated using off-light (top two rows) and on-light (bottom two rows) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Agent’s action trajectory from initial random layout to final optimal layout. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 9 canonical work pages

[8]

Spatial geometric reasoning for room layout estimation via deep reinforcement learning

Liangliang Ren, Yangyang Song, Jiwen Lu, and Jie Zhou. Spatial geometric reasoning for room layout estimation via deep reinforcement learning. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16, pages 550–565. Springer,

work page 2020
[9]

Lawrence Mandow, José-Luis Pérez-de-la Cruz, Ana Belén Rodríguez-Gavilán, and Manuela Ruiz-Montiel

doi:https://doi.org/10.1016/j.aei.2012.12.004. Lawrence Mandow, José-Luis Pérez-de-la Cruz, Ana Belén Rodríguez-Gavilán, and Manuela Ruiz-Montiel. Architec- tural planning with shape grammars and reinforcement learning: Habitability and energy efficiency. Engineering Applications of Artificial Intelligence, 96:103909,

work page doi:10.1016/j.aei.2012.12.004 2012
[10]

Feng Shi, Ranjith K Soman, Ji Han, and Jennifer K Whyte

doi:https://doi.org/10.1016/j.engappai.2020.103909. Feng Shi, Ranjith K Soman, Ji Han, and Jennifer K Whyte. Addressing adjacency constraints in rect- angular floor plans using monte-carlo tree search. Automation in Construction , 115:103187,

work page arXiv 2020
[11]

Matthias Klar, Moritz Glatt, and Jan C Aurich

doi:https://doi.org/10.1016/j.autcon.2020.103187. Matthias Klar, Moritz Glatt, and Jan C Aurich. An implementation of a reinforcement learning based algorithm for factory layout planning. Manufacturing Letters, 30:1–4,

work page arXiv 2020
[12]

Multi-Agent Reinforcement Learning of 3D Furniture Layout Simulation in Indoor Graphics Scenes

Xinhan Di and Pengqian Yu. Multi-agent reinforcement learning of 3d furniture layout simulation in indoor graphics scenes. arXiv preprint arXiv:2102.09137,

work page Pith review arXiv
[13]

Proximal policy optimization algorithms

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347,

arXiv
[1998]

Architectural layout planning using genetic algorithms

Manish K Thakur, Monika Kumari, and Madhabananda Das. Architectural layout planning using genetic algorithms. In 2010 3rd International Conference on Computer Science and Information Technology, volume 4, pages 5–11. IEEE,

work page 2010
[2010]

Benjamin Dillenburger

doi:https://doi.org/10.1145/1882261.1866203. Benjamin Dillenburger. Raumindex. Ein datenbasiertes Entwurfsinstrument. PhD thesis, ETH Zurich,

work page arXiv

Show all 14 references

[2013]

Gplan: Computer-generated dimensioned floorplans for given adjacencies

Krishnendra Shekhawat, Nitant Upasani, Sumit Bisht, and Rahil Jain. Gplan: Computer-generated dimensioned floorplans for given adjacencies. arXiv preprint arXiv:2008.01803,

2008 arXiv
[2017]

doi:https://doi.org/10.48550/arXiv.1707.06347. 11

doi arXiv
[2020]

10 Illuminating Spaces A PREPRINT Dasong Wang and Roland Snooks

URL https://proceedings.neurips.cc/paper_files/paper/ 2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf. 10 Illuminating Spaces A PREPRINT Dasong Wang and Roland Snooks. Artificial intuitions of generative design: An approach based on reinforcement learning. In Proceedings ...

2020
[2021]

Mohammad Amin Shabani, Sepidehsadat Hosseini, and Yasutaka Furukawa

doi:https://doi.org/10.48550/arXiv.2103.02574. Mohammad Amin Shabani, Sepidehsadat Hosseini, and Yasutaka Furukawa. Housediffusion: Vector floorplan genera- tion via a diffusion model with discrete and continuous denoising. In Proceedings of the IEEE/CVF Conference on Computer...

doi arXiv
[2023]

Jonathan Ho, Ajay Jain, and Pieter Abbeel

doi:https://doi.org/10.48550/arXiv.2211.13287. Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6...

doi arXiv
[2024]

House-gan++: Generative adversarial layout refinement networks

Nelson Nauata, Sepidehsadat Hosseini, Kai-Hung Chang, Hang Chu, Chin-Yi Cheng, and Yasutaka Furukawa. House-gan++: Generative adversarial layout refinement networks. arXiv preprint arXiv:2103.02574 ,

arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [8]

Spatial geometric reasoning for room layout estimation via deep reinforcement learning

Liangliang Ren, Yangyang Song, Jiwen Lu, and Jie Zhou. Spatial geometric reasoning for room layout estimation via deep reinforcement learning. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16, pages 550–565. Springer,

work page 2020

[2] [9]

Lawrence Mandow, José-Luis Pérez-de-la Cruz, Ana Belén Rodríguez-Gavilán, and Manuela Ruiz-Montiel

doi:https://doi.org/10.1016/j.aei.2012.12.004. Lawrence Mandow, José-Luis Pérez-de-la Cruz, Ana Belén Rodríguez-Gavilán, and Manuela Ruiz-Montiel. Architec- tural planning with shape grammars and reinforcement learning: Habitability and energy efficiency. Engineering Applications of Artificial Intelligence, 96:103909,

work page doi:10.1016/j.aei.2012.12.004 2012

[3] [10]

Feng Shi, Ranjith K Soman, Ji Han, and Jennifer K Whyte

doi:https://doi.org/10.1016/j.engappai.2020.103909. Feng Shi, Ranjith K Soman, Ji Han, and Jennifer K Whyte. Addressing adjacency constraints in rect- angular floor plans using monte-carlo tree search. Automation in Construction , 115:103187,

work page arXiv 2020

[4] [11]

Matthias Klar, Moritz Glatt, and Jan C Aurich

doi:https://doi.org/10.1016/j.autcon.2020.103187. Matthias Klar, Moritz Glatt, and Jan C Aurich. An implementation of a reinforcement learning based algorithm for factory layout planning. Manufacturing Letters, 30:1–4,

work page arXiv 2020

[5] [12]

Multi-Agent Reinforcement Learning of 3D Furniture Layout Simulation in Indoor Graphics Scenes

Xinhan Di and Pengqian Yu. Multi-agent reinforcement learning of 3d furniture layout simulation in indoor graphics scenes. arXiv preprint arXiv:2102.09137,

work page Pith review arXiv

[6] [13]

Proximal policy optimization algorithms

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347,

arXiv

[7] [1998]

Architectural layout planning using genetic algorithms

Manish K Thakur, Monika Kumari, and Madhabananda Das. Architectural layout planning using genetic algorithms. In 2010 3rd International Conference on Computer Science and Information Technology, volume 4, pages 5–11. IEEE,

work page 2010

[8] [2010]

Benjamin Dillenburger

doi:https://doi.org/10.1145/1882261.1866203. Benjamin Dillenburger. Raumindex. Ein datenbasiertes Entwurfsinstrument. PhD thesis, ETH Zurich,

work page arXiv