A Reinforcement Learning-Based Spatial Optimization Design Method for Landscape Planning

Abstract

This invention proposes a reinforcement learning-based spatial optimization design method for landscape planning. The method includes: acquiring user group feature vectors, a landscape module library, and a spatial area label set; determining a set of candidate layout areas based on the candidate module set, combined with the recommended layout area labels and the spatial area label set, and calculating the corresponding spatial layout score; identifying the optimal area, candidate modules, and the layout area corresponding to the candidate modules based on the highest score of the spatial layout score, to generate an initial state set; collecting user feedback data, combining the initial state set and recommended layout area labels, and generating an optimized state set and policy network through a reinforcement learning model; and generating structured drawings, module annotation diagrams, an evaluation report, and a general deployment data package based on the optimized state set and policy network. This invention significantly improves the effectiveness of health interventions while ensuring the efficiency of landscape design.

Description

Technical Field

[0001] This invention belongs to the field of landscape planning and spatial optimization, and particularly relates to a landscape planning and spatial optimization design method based on reinforcement learning. Background Technology

[0002] Against the backdrop of rapid urban development, residents' demands for their living environment are increasingly evolving from basic safety and functionality to higher levels of health, comfort, and emotional support. As a crucial component of urban green infrastructure, garden spaces not only fulfill ecological functions but are also gradually taking on social functions that promote human physical and mental well-being. In a society characterized by an aging population, children's lack of contact with nature, and a surge in sub-healthy individuals, garden design has been endowed with multiple values: sensory therapy, psychological recovery, and the promotion of social interaction. Modern healthy gardens emphasize personalized experiences, multi-sensory interaction, and adaptability to diverse population groups, requiring spaces that are both functional and humanistic. To improve design efficiency and adaptability, modular garden spaces have become a trend, with gardens being broken down into combinable functional units (such as visual modules, olfactory modules, rest modules, and participation modules) to support rapid deployment and customized design. However, in current design practices, module combinations and spatial layouts largely rely on designer experience or static rules, lacking scientific quantitative feedback mechanisms, and especially lacking the systematic ability to model and optimize the sensory preferences, emotional responses, and behavioral patterns of different population groups.

[0003] While existing spatial optimization techniques (such as heuristic search, rule-based reasoning, and graph theory analysis) can be used for preliminary layout suggestions, their generalization ability and adaptability are clearly insufficient in the open, unstructured, and highly interactive environment of gardens. In recent years, artificial intelligence technology has been gradually applied to the fields of architecture and landscape, such as using machine learning to predict pedestrian density or path preferences. However, these methods are mainly based on static data or historical statistics, lacking dynamic optimization capabilities based on real-time feedback mechanisms. Reinforcement learning, as an optimization algorithm with self-learning and policy evolution capabilities, has demonstrated excellent performance in scenarios such as robot path planning and traffic control, but its application in garden spaces remains largely unexplored. This is partly because the behavior of people in garden settings is highly uncertain and subjective, and partly because the causal mechanism between sensory stimuli and psychological states is complex and difficult to express using traditional modeling methods.

[0004] Therefore, existing methods generally suffer from the following technical defects: (1) They lack a dynamic optimization mechanism that integrates people's sensory preferences and emotional feedback, resulting in the inability to personalize the therapeutic effect of garden spaces; (2) The combination of spatial functional modules lacks scientific computational support and relies on human experience, making it difficult to achieve both efficiency and adaptability; (3) They lack a feedback loop, making it impossible to realize an iterative optimization process of design-use-adjustment. Faced with this challenge, there is an urgent need for a spatial intelligent planning method that can integrate user perception data, behavioral responses, and emotional indicators, and has automatic optimization capabilities, in order to achieve truly people-oriented healthy garden space design. Summary of the Invention

[0005] The purpose of this invention is to propose a spatial optimization design method for garden planning based on reinforcement learning. It is built around the core path of reinforcement learning and human sensory behavior feedback, and establishes a system process that starts from user sensory preferences, integrates human usage behavior and emotional feedback, and continuously iterates and optimizes the spatial layout.

[0006] To achieve the above objectives, this invention provides a spatial optimization design method for landscape planning based on reinforcement learning, the method comprising the following steps:

[0007] S1. Obtain user group feature vector, garden module library and spatial area label set, wherein each module in the garden module library is accompanied by a sensory attribute vector. The sensory attribute vector is matched with the user group feature vector by weighted Euclidean distance calculation to generate a matching degree score corresponding to the user group feature vector.

[0008] Modules with matching scores higher than a preset threshold are selected as candidate module sets, and recommended area labels are assigned based on the module's own design standards or prior uses to generate recommended layout area labels.

[0009] S2. Based on the candidate module set, and in combination with the recommended layout area label and spatial area label set, determine the set of candidate layout areas and calculate the corresponding spatial layout score. The highest score corresponding to the spatial layout score is used to determine the best area, candidate modules, and the layout area corresponding to the candidate modules, so as to generate an initial state set.

[0010] S3. Collect user feedback data, combine the initial state set and the recommended placement area labels, and generate an optimized state set and policy network through a reinforcement learning model. In the optimized state set, each candidate module has a new placement area, but the corresponding sensory attribute vector remains unchanged.

[0011] S4. Generate structured drawings, module annotation diagrams, evaluation reports, and general deployment data packages based on the optimized state set and policy network.

[0012] Furthermore, the user group feature vector is obtained through surveys of user basic information and sensory preferences; the garden module library is obtained based on garden product design catalogs, supplier databases, or preliminary expert annotations; the spatial area label set is composed of functional sub-areas initially divided by the designer based on the project site.

[0013] The sensory attribute vector includes color contrast, plant fragrance index, sound decibel and pitch distribution, interactivity rating, and boundary perception attribute.

[0014] The spatial area label set includes quiet areas, interactive areas, shaded areas, and open areas.

[0015] Further, step S1 specifically includes:

[0016] The sensory attribute vector of each module is matched with the user preference vector using a weighted Euclidean distance to generate a matching score between the module and the target group's preferences. The value range is (0,1], with the closest value being 1.

[0017] Modules with matching scores higher than a preset threshold are selected as candidate module sets. Based on the module's own design standards or prior uses, recommended area labels are assigned through a lookup table function to generate recommended layout area labels.

[0018] Furthermore, S2 specifically includes:

[0019] Each candidate module in the candidate module set is placed in a spatial region that matches its recommended label. A matching function is used to select all regions that satisfy the recommended label of the candidate module in the region as the recommended placement region label to form a set of candidate placement regions.

[0020] For the set of candidate layout areas, a spatial layout scoring function is designed to select the area most suitable for the layout of the module, thus obtaining the optimal area corresponding to each different candidate module;

[0021] Once the optimal region corresponding to each candidate module is generated, it needs to be specifically placed within the current region to generate the layout region corresponding to the candidate module.

[0022] An initial state set is generated by combining the candidate module, the arrangement area corresponding to the candidate module, and the sensory attribute vector corresponding to the candidate module.

[0023] Furthermore, once the optimal region corresponding to each candidate module is generated, it needs to be specifically positioned within the current region to generate the arrangement region corresponding to the candidate module, specifically including:

[0024] The optimal region space corresponding to each candidate module is divided into an N×N uniform grid;

[0025] Exclude meshes that are obscured by obstacles or contain other modules, and generate candidate mesh locations;

[0026] For each candidate grid location, calculate the penalty for matching the candidate module size with the sensory conflict of neighboring modules;

[0027] The first empty space that meets the criteria of non-conflict and high coordination is selected using a greedy algorithm as the initial coordinates of the candidate module, which serves as the placement area.

[0028] Furthermore, S3 specifically includes:

[0029] Collect user behavior feedback data, including trajectory heatmaps, dwell time distribution, and emotion recognition mapping; and discretize the user behavior feedback data into a tensor form aligned with the layout area;

[0030] The initial set of states is used as the current state and input into the policy network to generate an action; wherein the action is a translation operation of a candidate module, and the action of the translation operation is limited by the plot corresponding to the recommended layout area label of the candidate module;

[0031] Perform the actions to generate new states, and design a joint reward function based on new user behavior feedback data to guide policy network updates, thereby generating an optimized set of states and policy network.

[0032] Furthermore, the joint reward function is designed as follows:

[0033]

[0034] in, The emotion consistency score, derived from the mean and standard deviation of the emotion indices within the range surrounding the candidate module in the emotion recognition mapping data, represents the degree of influence of the layout on the collective emotion; Φ t This is a behavior path deviation penalty term, measuring the deviation between the user behavior trajectory in the trajectory heatmap after module changes and the original high-frequency path, in order to ensure that the user behavior guidance logic is not disrupted; Ψ t The module sensory conflict term measures the sensory inconsistency between any two modules in the current layout, and is defined as follows:

[0035]

[0036] Where, d ij Indicates candidate module M i ′ and M j The Euclidean distance of ′, where δ is the critical threshold for sensory interference. For the indicator function, ||A i -Aj ||1 represents the L1 difference between sensory attributes.

[0037] Furthermore, the policy network is implemented using a three-layer fully connected neural network, with each layer having a width of 128 / 64 / 32 and an activation function of ReLU. The final output is a continuous action, which is used to translate the position of the module. During training, the PPO optimization method is used, and a region boundary pruning strategy is set to ensure that the module does not leave the plot to which the recommended layout region label of the candidate module belongs.

[0038] Furthermore, S4 specifically includes:

[0039] The design involves two main sub-tasks: generating structured drawings and module annotation diagrams; and generating an adaptability assessment report and a general deployment data package.

[0040] The steps for generating structured drawings and module annotation diagrams include:

[0041] The layout coordinates of each candidate module are extracted from the optimized state set and mapped to the coordinate system in the design drawing base to generate the final coordinates;

[0042] By combining the names of the candidate modules, the regions corresponding to the recommended layout area labels, and the final coordinates, a module annotation vector is generated.

[0043] The steps for generating the compatibility assessment report and the general deployment data package include:

[0044] Collect user group feature vectors newly input by users, generate a fast layout through the optimized strategy network, and compare the structure with the optimized state set to obtain a layout consistency score, so as to generate an adaptability evaluation report.

[0045] The general deployment data package includes an optimized set of states and module annotation vectors.

[0046] Furthermore, in the step of extracting the layout coordinates of each candidate module from the optimized state set and mapping them to the coordinate system in the design drawing base map to generate the final coordinates, if the position of the candidate module is not within the area represented by the recommended layout area label, the module will be marked on the output map in the form of a design deviation highlight, supporting designers to perform manual correction.

[0047] The beneficial technical effects of the present invention are at least as follows:

[0048] This invention addresses the shortcomings of existing garden space optimization design methods in sensory response modeling, adaptive module combination capabilities, and dynamic optimization mechanisms. It proposes a garden module spatial layout optimization method based on an intelligent learning mechanism. This method is constructed around the core path of reinforcement learning and user sensory behavior feedback, establishing a systematic process that starts from user sensory preferences, integrates user behavior and emotional feedback, and continuously iteratively optimizes the spatial layout.

[0049] First, by constructing a correspondence between population characteristics and sensory preferences, the initial combination recommendation of garden functional modules is realized, solving the problem of lack of personalization in module combination in traditional schemes. Second, multimodal emotional and behavioral data of users are collected during space use to establish a feedback model of space combination → emotional response, providing a quantitative basis for layout adjustment. Finally, by using reinforcement learning technology, a dynamic closed-loop system of perception-feedback-optimization is constructed to continuously optimize the module layout strategy based on feedback data, replacing the traditional linear process of design first and then use. This system significantly improves the health intervention effect while ensuring the efficiency of garden design, and is particularly suitable for community-based, customized, and highly participatory health garden design scenarios, demonstrating significant technological innovation and application promotion value. Attached Figure Description

[0050] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0051] Figure 1 This is a flowchart of the reinforcement learning-based spatial optimization design method for landscape planning according to the present invention. Detailed Implementation

[0052] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0053] like Figure 1 As shown in the embodiment of the present invention, a spatial optimization design method for landscape planning based on reinforcement learning is provided. The method includes:

[0054] S1. Obtain user group feature vector, garden module library and spatial area label set, wherein each module in the garden module library is accompanied by a sensory attribute vector. The sensory attribute vector is matched with the user group feature vector by weighted Euclidean distance calculation to generate a matching degree score corresponding to the user group feature vector.

[0055] Modules with matching scores higher than a preset threshold are selected as candidate module sets, and recommended area labels are assigned based on the module's own design standards or prior uses to generate recommended layout area labels.

[0056] S2. Based on the candidate module set, and in combination with the recommended layout area label and spatial area label set, determine the set of candidate layout areas and calculate the corresponding spatial layout score. The highest score corresponding to the spatial layout score is used to determine the best area, candidate modules, and the layout area corresponding to the candidate modules, so as to generate an initial state set.

[0057] S3. Collect user feedback data, combine the initial state set and the recommended placement area labels, and generate an optimized state set and policy network through a reinforcement learning model. In the optimized state set, each candidate module has a new placement area, but the corresponding sensory attribute vector remains unchanged.

[0058] S4. Generate structured drawings, module annotation diagrams, evaluation reports, and general deployment data packages based on the optimized state set and policy network.

[0059] Specifically, step S1 includes:

[0060] The goal of this step is to construct a reproducible and adjustable matching mechanism between sensory preferences of different groups and garden modules. This mechanism will identify suitable combinations of garden functional modules for specific populations and allocate initial spatial layout areas based on the design characteristics of each module. Garden spaces contain a wide variety of functional modules, including planting modules, pathway modules, interactive installations, seating facilities, sound installations, and shade structures. Different groups (such as the elderly, children, and rehabilitation patients) exhibit significant differences in their sensory responses to these modules. Only by modeling these differences as computable variables can a sensory-driven data foundation be provided for subsequent reinforcement learning-based layout optimization, ultimately achieving a healthy garden space design with physiological and psychological adaptability.

[0061] This step serves as the starting point of the system. The input data comes from three parts, all of which can be collected and standardized using existing methods.

[0062] First, we have the user group feature vector P. P is derived from user basic information and sensory preference surveys, which can be collected through questionnaires, community surveys, and user profiles provided by elderly care institutions. In the pilot design, we use the following five sensory dimensions to build P: visual stimulus sensitivity, olfactory stimulus sensitivity, auditory stimulus comfort threshold, willingness to participate, and psychological safety needs. Each dimension is standardized to the [0,1] interval. For example, for elderly users, the survey questionnaire will include questions such as: Do you prefer a quiet environment? Are you sensitive to strong smells? Do you enjoy participating in garden activities? The quantitative results of each question are mapped to the corresponding dimension in P.

[0063] Secondly, there is the garden module library {M1,M2,...M i M n}, where each module M i Includes a sensory attribute vector A i Its components include color contrast (visual dimension), plant fragrance index (olfactory dimension), sound decibel and pitch distribution (auditory dimension), interactivity rating (participation dimension), and boundary perception attributes (safety dimension). These data come from landscape product design brochures, supplier databases, or preliminary expert annotations, and are all converted to the [0,1] interval.

[0064] The third part is the spatial region label set. The site is initially divided into functional sub-zones by the designer based on the project site. For example, if the left side of the site is shaded by trees, the designer will mark it as a shaded area; the area near the plaza exit will be an interactive area; and the central open lawn can be designated as an open area. These labels are not used for precise coordinate positioning, but rather to provide regional constraints for subsequent layout.

[0065] First, each module M i Sensory attribute vector A i Matching with the user preference vector P is achieved using weighted Euclidean distance:

[0066]

[0067] Among them, s i Represents module M i The matching score with the target audience's preferences ranges from (0,1], with a higher value closer to 1 indicating a better match; d represents the total number of sensory dimensions, which is fixed at 5 in this system; P j This refers to the sensitivity score of the target user group on the j-th sensory dimension, derived from the survey questionnaire results; A i,j It is module M i The attribute value in the j-th sensory dimension comes from the module database; w jThis is the weighting coefficient for this dimension, calculated from the designer's experience or historical data, satisfying ∑w j =1. exp(·) is used to transform distance into matching degree, emphasizing non-linear sensitivity.

[0068] Then, modules with a matching degree higher than the threshold θ are selected as candidate module set C, and recommended region labels are assigned based on the module's own design criteria or prior purpose:

[0069] C = {M i ′∣s i >θ}, L i =module_suggest(M i (2)

[0070] Where θ is the matching threshold, typically set between [0.7, 0.85]; module_suggest(·) is a lookup function that determines the recommended layout area L based on the module's functional suggestions. i For example, water features are generally recommended for open areas, while herb clusters are recommended for shaded areas; L i The range of values ​​is The tags in the text are used for area constraints in subsequent layout steps.

[0071] For example, if a group P = [0.8, 0.2, 0.9, 0.3, 0.7] represents high visual sensitivity, weak olfactory receptivity, and extreme need for quiet, then a certain module M... k The attribute is A k =[0.75,0.3,0.85,0.4,0.6], after weighted matching s k =0.92, satisfying the threshold θ=0.8, then it is included in C, and L k =Quiet zone.

[0072] Understandably, the core innovation of this mechanism lies in mapping sensory preferences to module selection and further introducing a recommended placement area label L. i This allows subsequent spatial optimization to not only consider what information to select, but also the structural priors for placing the modules in which areas. Traditional methods either manually select modules or use random initial layouts, lacking this ability to model the interconnectedness of psychology, senses, and spatial structure.

[0073] Output: C = {M} i '}: A set of functional modules that satisfy user sensory preferences; L = {L} i}: Recommended layout area label set, the recommended layout area labels for each module, for use in subsequent steps.

[0074] Specifically, step S2 includes:

[0075] The initial spatial layout construction guided by sensory preferences in this step plays a crucial role in building the foundational state space for reinforcement learning optimization within the entire landscape planning optimization system. Its goal is to integrate the set of landscape functional modules C = {M} selected in the previous step. i Based on the sensory attribute A of the module i Instead of recommending the placement of area labels L i In a given garden space Ω, initially position the plants and output the state set S0 = {(M i ′,x i ,y i A i This provides a trainable starting point for the next step of reinforcement learning. Unlike general graph optimization or building layout tasks, this patent targets healthy garden design, which has strong sensory-driven characteristics, user emotional adaptability requirements, and significant characteristics of unstructured area labeling of the site. Therefore, this step introduces modeling items unique to the patent scenario in terms of spatial layout logic and evaluation indicators, thus forming a complete and original initial layout method.

[0076] Input: Module set C = {M i ′}: Represents the garden functional modules selected based on the sensory preferences of the population; the sensory attribute vector A of each module. i The dimensions include visual (color brightness), olfactory (fragrance intensity), auditory (sound source type), interactivity, and sense of security, with each dimension ranging from [0,1]. The recommended area label for each module is L. i : indicates M i 'Most suitable for the type of space area to be arranged, and the value belongs to the preset tag set.' Site spatial division information Ω={Ω1,Ω2,...,Ω K The garden site is divided into several functional zones by the designer or a GIS layer, each Ω k correspond A certain tag in the middle.

[0077] Understandably, these input elements collectively define the multidimensional relationships of what a module is suitable for, where it should be placed, and what its sensory attributes are, serving as the basis for constructing the state space in this step.

[0078] First, each module M needs to be... i Place it in its recommended tag L i Matching spatial region Ω k In the middle. We select all that satisfy label(Ω) using a matching function. k ) = L i The region composition can be arranged into a set of candidate regions Ω (i) Each region Ω kIt contains metadata such as boundary definition, deployable area, and existing obstacle information.

[0079] To enhance the overall spatial harmony and human-centered therapeutic orientation of the sensory layout, we designed a spatial layout scoring function Q(M). i ′,Ω k This function is used to select the most suitable area for arranging the module. The scoring function integrates three aspects of innovative evaluation indicators:

[0080]

[0081] in, Sensory coordination score: measures M i ′ and already in Ω k The sensory differences between the modules arranged in the middle are defined as the mean of the Euclidean differences of each sensory dimension attribute. The lower the value, the better the coordination. Sensory diversity compensation factor: used to encourage moderate sensory hierarchy variation within a region and avoid module homogenization. It is calculated as the absolute value of the L1 difference between the module and the sensory center within the region. The region load regularization term penalizes modules placed in regions that are already close to their capacity limits, based on the current region's placement percentage r. k Nonlinear growth calculation improves layout balance; α, β, and λ are hyperparameters used to balance the weights of the three in actual deployment and can be adjusted to adapt to different landscape design requirements (such as healing orientation vs. activity orientation).

[0082] The above function design embodies three levels of scenario innovation: 1) The regional coordination score is constructed based on sensory characteristics, which fully serves the goal of sensory healing; 2) Sensory diversity serves as a proxy for the richness of spatial experience, and the introduction of positive scores is a unique requirement in healthy gardens; 3) The load regularization term is based on the regional health density, which is a necessary mechanism to avoid over-concentration and ensure the walking experience and ventilation.

[0083] Wherein, once for each M i The best area has been selected. (i.e., the region corresponding to the maximum value of Q), we need to pinpoint the location within this region. Specific location (x) i ,y i It is generated by the following process:

[0084] Will The space is divided into an N×N uniform grid;

[0085] Exclude meshes that are obscured by obstacles or contain other modules;

[0086] For each candidate grid position g i=(x,y) calculates the penalty for sensory conflict between the module size matching and the neighboring modules;

[0087] The first empty slot that satisfies both non-conflict and high coordination is selected as M using a greedy algorithm. i The initial coordinates (x') i ,y i ).

[0088] Finally, the system outputs a complete set of initial states:

[0089] S0={(M i ′,x i ,y i A i )∣M i ∈C} (4)

[0090] This set is the starting point of the state in the reinforcement learning system. Each state element contains ternary information of module identifier, spatial location, and sensory attribute, which is the basic observation input for policy selection in reinforcement learning.

[0091] Understandably, the innovation of this step lies not only in considering where the modules are placed, but also in incorporating three soft indicators from garden therapy design—sensory adjustment, spatial density control, and sensory diversity—into the layout function. The Q function integrates harmony, diversity, and healthy space density control, representing a type of humanistic factor that truly exists in garden health scenarios but has not been modeled in the past. Simultaneously, it avoids the problem of initial states in reinforcement learning deviating from reality due to simple module placement, enabling faster convergence and greater spatial feasibility in subsequent learning stages.

[0092] Output: S0 = {(M i ′,x i ,y i A i )}: Serves as the initial state input for reinforcement learning optimization; each M i The layout area and coordinate information (x) of ′ i ,y i ): Used in spatial modeling and graphics rendering systems.

[0093] Specifically, step S3 includes:

[0094] This step, reinforcement learning to optimize the spatial layout (driven by sensory and behavioral feedback), is the core algorithm of this patented solution. Its goal is to achieve the initial layout state S0 = {(M} generated in step two. i ′,x i ,y i A iBased on this, actual user behavior and emotional feedback data are introduced to construct a reinforcement learning optimization model for the sensory therapeutic effects of garden spaces, achieving adaptive optimization of module spatial locations. This process ultimately outputs the optimized spatial layout state S. * And optimal strategy π * This approach not only achieves a rational physical layout of garden spaces but also harmonizes sensory stimulation, behavioral guidance, and psychological therapeutic effects. Compared to traditional spatial planning methods that optimize area utilization or connectivity efficiency, this step constructs a reinforcement learning reward mechanism through sensory attributes and user feedback, representing an innovative approach oriented towards a new planning concept for healthy gardens.

[0095] This step fully utilizes all the structured data output from step two, combined with additional feedback data collected:

[0096] Initial state S0 = {(M i ′,x i ,y i A i )}: where M i ′ is the module identifier, (x i ,y i ) represents the module location, A i The sensory attribute vectors of the module (such as color brightness, sound source intensity, olfactory stimulation, etc.) are determined by step two;

[0097] Recommended layout area label L for each candidate module i Used to limit the spatial range in which modules can be moved, preventing out-of-bounds errors during optimization;

[0098] User behavior feedback dataset Including trajectory heatmap T g (x,y), D stay time distribution t Emotion recognition mapping E(x,y) was collected in an experimental garden space equipped with cameras and infrared sensors, using devices such as dual-camera infrared motion probes, thermal infrared grid arrays, and wristband synchronization systems.

[0099] Behavioral and emotional data were discretized into tensors aligned with a site grid space, the size of which was related to the landscape design site. Figure 1 To.

[0100] Furthermore, the reinforcement learning environment uses the above data as state input, constraints, and reward feedback to guide the optimization of action strategies.

[0101] The optimization model employs a policy gradient method based on an Actor-Critic structure for space optimization. In each iteration, the system sets the current state S... t Input to policy network π(a) t|S t ), generate action a t That is, a certain adjustable module M i Translation operation (Δx) i ,Δy i Actions are limited by the original recommended region L of the module. i The corresponding plot of land. The action is then executed, generating a new state S. t+1 The reward R is calculated by combining comprehensive feedback data from the population with the degree of sensory conflict. t The guidance strategy network is updated.

[0102] To maximize the adaptability of the patent scenario, especially for the three core design requirements of a healthy garden—emotional regulation, path behavior, and sensory coordination—this step constructs a three-term joint reward function:

[0103]

[0104] in, The emotional consistency score is derived from the mean and standard deviation of the emotional data within the module's surrounding area, representing the degree of influence of the layout on collective emotion; Φ t This is a penalty term for behavioral path deviation, measuring the user's behavioral trajectory T after module changes. g The deviation between (x,y) and the original high-frequency path is used to ensure that the user behavior guidance logic is not disrupted; Ψ t The module sensory conflict term measures the sensory inconsistency between any two modules in the current layout, and is defined as follows:

[0105]

[0106] Where, d ij Represents module M i ′ and M j The Euclidean distance is δ, where δ is the critical threshold for sensory interference (e.g., 1.5m). For indicator functions (1 if the modules are too close, 0 otherwise), ||A i -A j ||1 represents the L1 difference between sensory attributes. This encourages systems to avoid placing high-contrast sensory modules in adjacent locations, thus maintaining the sensory harmony of the garden space.

[0107] The above three parts together constitute a composite optimization objective for garden health scenarios, significantly different from traditional reinforcement learning approaches that focus on rewards, path lengths, etc. The most unique design is:

[0108] Introduced in the reward function Derived from real-time emotion recognition tensor data, this is the first time it has been used as an optimization signal in the field of landscape architecture.

[0109] Behavioral path punishment Φ t Maintaining the rationality of spatial structure in a way that directly reflects people's behavioral habits is a sign of respect for real user feedback;

[0110] Inter-module sensory conflict degree Ψ t Using the L1 difference between module attribute tensors as the penalty basis is a spatial semantic modeling of the sensory layout system.

[0111] The policy network is implemented using a three-layer fully connected neural network, with each layer having a width of 128 / 64 / 32 and using ReLU as the activation function. The final output is a continuous action (Δx). i ,Δy i This is used to translate the module's position. During training, the PPO optimization method is used, and a region boundary clipping strategy is set to ensure that the module does not leave the original proposed region L. i The land parcel in which it belongs.

[0112] Each iteration includes: state input → module action generation → action feasibility detection (whether it exceeds limits or overlaps) → layout state update → reward calculation → policy update. The entire optimization process continues until the reward value converges or a maximum number of iterations is set, ultimately yielding the optimized state S. * .

[0113] Output: S * ={(M i ,x i ′,y i ′,A i )}: Optimized spatial state, each module has a new location (x i ′,y i ′), sensory attributes remain unchanged; policy network π * The trained policy network can be quickly deployed to other similar user needs scenarios.

[0114] Specifically, step S4 includes:

[0115] The core objective of this step is to optimize the final layout state S obtained from the previous reinforcement learning step. * ={(M i ,x i ′,y i ′,A i )} and strategy model π * This is further transformed into feasible and deliverable landscape design solutions to assist designers in construction implementation and subsequent iterative optimization.

[0116] This step is divided into two main sub-tasks: (1) generating structured drawings and module annotation diagrams; (2) generating an adaptability assessment report and a general deployment data package.

[0117] Part One: Generation of Structured Drawings. The system starts from S... * Extract each module M i The arrangement coordinates (x') i ′,y i The coordinate system is then mapped to the base map Ω of the design drawing (e.g., CAD coordinates, GIS spatial grid). This transformation is accomplished using a spatial scale ratio κ defined during design, for example, κ = 1 meter / unit grid.

[0118] Each module is represented on the drawing using its prototype graphic template T(M). i ′) represents (standardized planar tiles provided by the module library), which the system embeds into coordinate points and performs region label comparison. If M i The position of ′ is not in the recommended area L i Within the indicated area, the system will highlight the module on the output diagram as a design deviation, allowing designers to perform manual corrections.

[0119] Furthermore, to enhance human-centered interpretability, the system simultaneously generates module annotation vectors Γ. i The format is as follows:

[0120] Γ i =[name(M i ′),A i ,L i ,(x i ′,y i ′)] (7)

[0121] This vector is embedded in the interactive design diagram, allowing designers to access its sensory attribute A by clicking on the module. i Recommended area L i Final coordinates (x) i ′,y i This allows for quick identification of module functions and layout rationality. For example, if M... i ′ is an aromatic herbal group, its A i =[0.2,0.9,0.1] represents low visual stimulation, high olfactory stimulation, and weak sound stimulation, L i =Shaded areas will be highlighted by the system when placed in open areas.

[0122] The second part involves adaptability reporting and deployment data generation. Considering the need for rapid reuse of the design solution across different user scenarios, this system provides a policy network-based π... * The rapid layout simulation function. The user inputs another group's characteristics P.test Then, the system uses π * Generate quick layout and the current layout S * Structural comparison was performed to obtain a layout consistency score ξ:

[0123]

[0124] Where ∈ represents the positional tolerance (e.g., 0.8 meters). This is an indicator function that shows whether the two schemes are consistent in their layout of the same module. This score reflects the generalization ability of the current design scheme to different user groups and can be used as a reference indicator for the universality of the layout.

[0125] In addition, all module layout information {M i ′,x i ′,y i ′,A i} and Γ i Vector information will be exported as structured data packages (such as JSON or XML format) to support automatic generation of BIM models, interface integration with project management systems, or invocation by intelligent construction robots.

[0126] Output: Drawing output Used for design visualization and construction reference; Structural consistency score ξ: used to evaluate the stability of the current design under multi-user requirements; Structured data package Includes S * With Γ i This is provided for use with construction and BIM interfaces.

[0127] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0128] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0129] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0130] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A spatial optimization design method for landscape planning based on reinforcement learning, characterized in that, The method includes the following steps: S1. Obtain user group feature vector, garden module library and spatial area label set, wherein each module in the garden module library is accompanied by a sensory attribute vector. The sensory attribute vector is matched with the user group feature vector by weighted Euclidean distance calculation to generate a matching degree score corresponding to the user group feature vector. Modules with matching scores higher than a preset threshold are selected as candidate module sets, and recommended area labels are assigned based on the module's own design standards or prior uses to generate recommended layout area labels. S2. Based on the candidate module set, and in combination with the recommended layout area label and spatial area label set, determine the set of candidate layout areas and calculate the corresponding spatial layout score. The highest score corresponding to the spatial layout score is used to determine the best area, candidate modules, and the layout area corresponding to the candidate modules, so as to generate an initial state set. S3. Collect user feedback data, combine the initial state set and the recommended placement area labels, and generate an optimized state set and policy network through a reinforcement learning model. In the optimized state set, each candidate module has a new placement area, but the corresponding sensory attribute vector remains unchanged. S4. Generate structured drawings, module annotation diagrams, evaluation reports, and general deployment data packages based on the optimized state set and policy network.

2. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 1, characterized in that, The user group feature vector is obtained through surveys of user basic information and sensory preferences; the garden module library is based on garden product design brochures, supplier databases, or preliminary expert annotations; the spatial area label set is composed of functional sub-areas initially divided by the designer based on the project site. The sensory attribute vector includes color contrast, plant fragrance index, sound decibel and pitch distribution, interactivity rating, and boundary perception attribute. The spatial area label set includes quiet areas, interactive areas, shaded areas, and open areas.

3. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 2, characterized in that, The specific steps of S1 include: The sensory attribute vector of each module is matched with the user group feature vector using a weighted Euclidean distance to generate a matching score between the module and the user group feature vector, with a value range of [value range missing]. The one closest to 1 is the best match; Modules with matching scores higher than a preset threshold are selected as candidate module sets. Based on the module's own design standards or prior uses, recommended area labels are assigned through a lookup table function to generate recommended layout area labels.

4. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 1, characterized in that, S2 specifically includes: Each candidate module in the candidate module set is placed in a spatial region that matches its recommended label. A matching function is used to select all regions that satisfy the recommended label of the candidate module in the region as the recommended placement region label to form a set of candidate placement regions. For the set of candidate layout areas, a spatial layout scoring function is designed to select the area most suitable for the layout of the module, thus obtaining the optimal area corresponding to each different candidate module; Once the optimal region corresponding to each candidate module is generated, it needs to be specifically placed within the current region to generate the layout region corresponding to the candidate module. An initial state set is generated by combining the candidate module, the arrangement area corresponding to the candidate module, and the sensory attribute vector corresponding to the candidate module.

5. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 4, characterized in that, Once the optimal region corresponding to each candidate module is generated, it needs to be specifically positioned within the current region to generate the layout region corresponding to the candidate module, specifically including: The optimal region space corresponding to each candidate module is divided into Uniform grid; Exclude meshes that are obscured by obstacles or contain other modules, and generate candidate mesh locations; For each candidate grid location, calculate the penalty for matching the candidate module size with the sensory conflict of neighboring modules; The first empty space that meets the criteria of non-conflict and high coordination is selected using a greedy algorithm as the initial coordinates of the candidate module, which serves as the placement area.

6. The method for optimizing garden planning space based on reinforcement learning according to claim 1, characterized in that, S3 specifically includes: Collect user behavior feedback data, including trajectory heatmaps, dwell time distribution, and emotion recognition mapping; and discretize the user behavior feedback data into a tensor form aligned with the layout area; The initial set of states is used as the current state and input into the policy network to generate an action; wherein the action is a translation operation of a candidate module, and the action of the translation operation is limited by the plot corresponding to the recommended layout area label of the candidate module; Perform the actions to generate new states, and design a joint reward function based on new user behavior feedback data to guide policy network updates, thereby generating an optimized set of states and policy network.

7. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 6, characterized in that, The joint reward function: ； in, The emotion consistency score is derived from the mean and standard deviation of the emotion in the range around the candidate module in the emotion recognition mapping data, indicating the degree of influence of the layout on the collective emotion. The behavior path deviation penalty measure measures the deviation between the user behavior trajectory in the trajectory heatmap after the module change and the original high-frequency path, in order to keep the user behavior guidance logic from being disrupted. The module sensory conflict term measures the sensory inconsistency between any two modules in the current layout, and is defined as follows: ； in, Indicates candidate module and European distance, The threshold for sensory interference. For indicator functions, The L1 difference is the difference between sensory attributes.

8. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 6, characterized in that, The policy network is implemented using a three-layer fully connected neural network, with each layer having a width of 128 / 64 / 32 and an activation function of ReLU. The final output is a continuous action, which is used to translate the module position. During training, the PPO optimization method is used, and a region boundary pruning strategy is set to ensure that the module does not leave the plot to which the recommended layout region label of the candidate module belongs.

9. The method for optimizing garden planning space based on reinforcement learning according to claim 1, characterized in that, S4 specifically includes: The design involves two main sub-tasks: generating structured drawings and module annotation diagrams; and generating an adaptability assessment report and a general deployment data package. The steps for generating structured drawings and module annotation diagrams include: The layout coordinates of each candidate module are extracted from the optimized state set and mapped to the coordinate system in the design drawing base to generate the final coordinates; By combining the names of the candidate modules, the regions corresponding to the recommended layout area labels, and the final coordinates, a module annotation vector is generated. The steps for generating the compatibility assessment report and the general deployment data package include: Collect user group feature vectors newly input by users, generate a fast layout through an optimized policy network, and compare the structure with the optimized state set to obtain a layout consistency score, so as to generate an adaptability evaluation report. The general deployment data package includes an optimized set of states and module annotation vectors.

10. The reinforcement learning-based spatial optimization design method for landscape planning according to claim 9, characterized in that, In the step of extracting the layout coordinates of each candidate module from the optimized state set and mapping them to the coordinate system in the design drawing base map to generate the final coordinates, if the position of the candidate module is not within the area represented by the recommended layout area label, the module will be marked on the output map with a design deviation highlight, supporting designers to perform manual correction.

Images