Multi-type street network walkability dynamic optimization method based on reinforcement learning
By employing reinforcement learning and multi-source data fusion methods, a dynamic walkability optimization framework is constructed, which solves the problem of insufficient adaptability in traditional street network optimization and achieves efficient and differentiated optimization of various types of street networks and improvement of the pedestrian environment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-04-03
AI Technical Summary
Existing methods for optimizing the walkability of street networks fail to fully consider the spatial structural differences of different types of street networks within cities. They rely on static evaluation indicators and lack dynamic response mechanisms, resulting in insufficient adaptability of optimization strategies and difficulty in achieving precise and differentiated spatial intervention.
A reinforcement learning-based approach is adopted to construct a dynamic optimization method for the walkability of multi-type street networks. By using a two-level optimization action set and multi-source data fusion, combined with a comprehensive evaluation index system of dynamic behavior results and static road network structure, Pareto optimal strategies are generated using the entropy weight-TOPSIS method and multi-objective optimization algorithm to achieve autonomous learning and real-time optimization.
It has achieved efficient and differentiated optimization of various types of street networks, improved the overall quality of the pedestrian environment, shortened the optimization cycle, reduced human intervention, and improved the accuracy and adaptability of pedestrian assessment.
Smart Images
Figure CN121787022A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of urban planning technology, and in particular relates to a dynamic optimization method for the walkability of multi-type street networks based on reinforcement learning. Background Technology
[0002] Walking, as a fundamental mode of daily travel for citizens, not only contributes to promoting physical and mental health and improving environmental friendliness, but also enhances social interaction and community cohesion, making it an important manifestation of sustainable urban development. However, with the increasing motorization of cities, urban street spaces are gradually becoming vehicle-dominated, and the pedestrian environment faces multiple challenges such as a lack of continuity, insufficient safety, and decreased comfort, severely restricting citizens' willingness and experience of walking. Therefore, systematically improving the walkability of street networks can significantly enhance the safety, convenience, and comfort of walking for citizens; for cities, it helps optimize the layout of public spaces, alleviate traffic congestion, and create a more livable and healthy urban environment.
[0003] However, current research and practice in street network walkability optimization suffer from several problems. Firstly, optimization methods often employ experience-based local adjustments or idealized models for unified planning, failing to fully consider the significant differences in spatial structure, connection patterns, and micro-environmental characteristics among different street network types within cities. This results in optimization strategies being disconnected from actual street spaces and lacking adaptability. These methods often treat heterogeneous road network morphologies as homogeneous systems, neglecting the direct impact of different topologies on walking efficiency, comfort, and spatial experience, making precise and differentiated spatial intervention difficult. Secondly, existing methods largely rely on single, static evaluation indicators, limited to one-time static scheme generation. They fail to consider dynamically changing pedestrian flow patterns and real-time environmental feedback within the street network, and also fail to introduce continuous learning and self-adjustment mechanisms, thus hindering the systematic optimization and long-term improvement of various street network types. In other words, traditional street walkability evaluation relies heavily on static structural indicators, reflecting only the physical attributes of the road network while ignoring the dynamic needs of pedestrians in actual movement, such as accessibility, efficiency, and experience. Furthermore, they fail to consider the different adaptations of evaluation indicators to various road network morphologies, such as radial and grid patterns. Traditional street optimization relies on a manual process of "parameter adjustment - simulation verification - iterative modification". It requires manually setting and adjusting parameters, and then verifying the effect through traffic simulation software. Iteration of a single solution can take several days or even weeks. Moreover, when faced with multiple conflicting objectives, manual weighing and selection are required, which can easily lead to inaccurate optimization due to limited experience. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a dynamic optimization method for the walkability of multi-type street networks based on reinforcement learning. This invention can integrate multi-source data, distinguish morphological types, and support dynamic response and autonomous learning, thereby more efficiently improving the overall quality of the urban pedestrian environment.
[0005] Technical solution: To achieve the above objectives, the method for identifying, diagnosing, and optimizing the daily activity space of urban residents as described in this invention includes the following steps: S01. Street data collection and intelligent road network type identification: Obtain street network data of the target city, perform preprocessing, use a pre-trained model to identify the road network characteristics of the target city and classify the morphology. The road network types include radial road network, grid road network, strip road network or free road network. S02. Setting up two-level optimization actions: Two sets of optimization actions are set up for different types of road networks. The two sets of optimization actions include a first-level general structure optimization action set and a second-level type-specific morphology optimization action set. Based on the specific road network morphology type identified in step S01, optimization actions are set for its physical structure characteristics. S03. Comprehensive evaluation of walkability: Construct a dynamic walkability evaluation index system based on behavioral outcomes, conduct random walk simulation experiments, and complete the numerical calculation of dynamic walkability evaluation indexes based on the random walk subject's movement influence index and walk subject index; construct a static adaptive walkability evaluation index system based on road network structure, and calculate it in conjunction with the basic attributes of the road network; use the entropy weight-TOPSIS method to standardize and weight the original data of the two index systems, and generate a comprehensive walkability score. S04. Dynamic optimization of gaitability to generate an optimization scheme: The gaitability score in step S03 is used as the input of the reward function of the reinforcement learning model. Based on the action settings in step S02, the agent executes a combination of actions and calculates the gaitability score in the new state in real time. After multiple rounds of state-action-reward iteration, a Pareto optimal policy set is generated using a multi-objective optimization algorithm, and an optimization scheme is generated. S05. Comparison and interaction of optimization schemes and display of results.
[0006] Optionally, in step S01, obtaining and preprocessing the target city street network data involves writing a Python script using the OSMnx library to obtain a directed graph object of the target city streets. Unnecessary road types are filtered out based on road level attributes, and nodes used only for connection without actual intersection significance are merged. The directed graph object of the streets is then converted into the data object required by the GNN framework PyTorch Geometric to obtain the global vector features required by GNN. The data node features include latitude and longitude coordinates, and the edge features include connection relationships, road length, road level, and direction.
[0007] Optionally, step S01, which utilizes a pre-trained model to identify the road network features of the target city and perform morphological classification, specifically includes the following steps: In the pre-training stage, a smart recognition architecture based on a GNN backbone, global pooling, and a classifier is constructed using self-supervised contrastive learning and Cross-Entropy Loss. Meta-learning optimization is performed on a multi-source road network dataset. A road network morphological feature extractor with zero-shot generalization capability is obtained, enabling the model to autonomously classify the regular connection pattern of "grid", the central clustering pattern of "radial", and the linear direction pattern of "strip". The "free-form" chaotic patterns are encoded into global feature vectors with multi-scale semantic awareness of morphological embeddings. In the classification stage, the target city road network data is standardized and preprocessed according to the graphing rules in the model pre-training stage to ensure the consistency of feature space distribution. The road network data is input into the pre-trained GNN backbone network as a feature extractor, and discriminative graph-level embedding vectors are extracted through global attention pooling layers. The few-sample classifier is quickly fine-tuned based on the embedding vectors. Finally, the optimized classifier is used to output the classification probability distribution of radial, grid, strip, or free-form road networks.
[0008] Optionally, the first-level general structural optimization action set in step S02 includes, but is not limited to: adding / removing road intersections, adjusting intersection connectivity, changing road segment lengths, and adjusting road width, specifically by adjusting the number of lanes to adjust road width; the first-level general structural optimization action set is implemented by defining a discretized action space; specifically, the intersection addition / removal operation is encoded as... Binary action; quantize the connectivity adjustment. The degree change operation; the road segment length and width adjustment is discretized into multiple levels of change step size, and within each optimization step size, the agent selects a combined action from the action space.
[0009] Optionally, the second-level type-specific morphological optimization action set in step S02 includes: for radial road networks, adding / reducing ring roads and optimizing the radial road hierarchy; for grid road networks, introducing diagonal connecting roads and increasing / reducing local grid density; for strip road networks, adding parallel auxiliary roads and optimizing the connection node layout between main roads and parallel roads; for free-form road networks, adding connecting road segments between dead-end road endpoints and adding orthogonal roads in low connectivity areas; the second-level type-specific morphological optimization action set is implemented through an action selector based on an attention mechanism; specifically, the activation weight of the action subset is calculated according to the road network morphological classification probability distribution obtained in step S01, and for radial road networks, the operation of adjusting the number of ring roads and reconstructing the radial road hierarchy is activated; for grid road networks, the operation of optimizing the insertion angle of diagonal connections and adjusting the grid density probability is activated; for strip road networks, the operation of optimizing the spacing between parallel roads and improving the connectivity of main and auxiliary roads is activated; for free-form road networks, the operation of generating the minimum connection path for dead-end roads and repairing the topology of orthogonal roads is activated; then, one-hot encoding is used. Encoding technology converts action parameters into fixed-dimensional binary operation instruction vectors; finally, the Policy Network outputs the action selection probability distribution to intelligently select the optimal operation.
[0010] Optionally, the construction of a dynamic walking evaluation index system based on behavioral outcomes in step S03 refers to constructing a dynamic walking evaluation index system based on the behavioral outcomes of a random walk subject from four levels: accessibility, efficiency, resilience, and experience. Accessibility indicators include route diversity, number of intersection visits, and walking coverage; efficiency indicators include average path tortuosity and shortest path distance; resilience indicators include interruption sensitivity; experience indicators include starting point facility coverage, POI contact frequency, and 15-minute coverage; route diversity refers to the sum of the number of routes from a fixed walking starting point to a fixed public facility and the number of routes from a non-fixed walking starting point to a fixed public facility; the number of intersection visits refers to the number of intersection visits obtained by the random walk subject through numerous iterations; and walking coverage refers to the number of fixed random... The walking activity's steps count indicates the maximum and minimum walking range of the walking activity; average path tortuosity is the ratio of the shortest path distance to a public facility for a random walking activity to the straight-line distance; shortest path distance is the shortest route distance from a fixed walking starting point to a fixed public facility; interruption sensitivity is the rate of decrease in the ratio of the total number of routes and the total distance to the average distance for a random walking activity before and after a 5% road interruption at a fixed starting point and a fixed public facility; starting point facility coverage rate is the number of facilities within a 1km radius centered on the random walking activity's random starting point; POI contact frequency is the average number of times a random walking activity encounters a public facility per unit of steps; 15-minute coverage rate is the proportion of public facilities passed by a random walking activity within 1.2km of a random starting point to the total number of public facilities in the area.
[0011] Optionally, in step S03, the movement impact index and the walk subject index, the random walk subject movement impact factor index includes road segment attributes and intersection attributes, the walk subject index includes movement direction and number of steps; the road segment attributes include street continuity and street length; the intersection attributes include intersection connectivity and intersection elevation; the movement direction setting involves constructing a direction vector matrix at the location of the walk subject, and selecting the next movement position only within the range of the direction vector matrix constructed from the starting point to the ending point; at the same time, the straight-going probability is 0.6, and the left / right turn probability is 0.2; the number of steps is either a fixed number of steps defined in experiments with a fixed radius / fixed movement distance, or a random number of steps with only a fixed starting point and ending point and no specified movement distance.
[0012] Optionally, the static adaptive walkability evaluation index system based on road network structure in step S03 specifically includes road network density, dead-end road density, road network closure, road intersection density, average street length, and street continuity, with the following calculation formulas: , , , , , .
[0013] Optionally, step S03 employs the entropy weight-TOPSIS method, specifically including the following steps: First, based on the number of road network assessment samples and evaluation indicators Construct the initial evaluation matrix To eliminate the influence of different indicator dimensions, the extreme value method is used to standardize the original data of two systems in the matrix: dynamic walking evaluation indicators based on behavioral results and static adaptive walking evaluation indicators based on road network structure. The indicators are then divided into two categories: benefit-type indicators and cost-type indicators. Higher values for benefit-type indicators are better, while lower values for cost-type indicators are better. The standardized matrix is obtained, and the calculation formula is as follows: , in, It is the first The first road network assessment sample in the first Standardized values for each benefit-type indicator It is the first The first road network assessment sample in the first The original values of each benefit-type indicator It is the first The maximum original value of each benefit-type indicator It is the first The minimum original value for each benefit-type indicator; , in, It is the first The first road network assessment sample in the first Standardized values for cost-related indicators It is the first The first road network assessment sample in the first The original values of each cost-type indicator It is the first The maximum original value of each benefit-type indicator It is the first The minimum original value for each benefit-type indicator; Secondly, calculate the proportion of each indicator value in all samples, and calculate the information entropy of each indicator based on the proportion. The smaller the entropy value, the greater the data variation of that indicator, and the more important it is in the evaluation. Finally, derive the objective weight of each indicator from the information entropy. The calculation formula is as follows: , , , , , , in, It is the first The first benefit-type indicator in the first The proportion of each road network assessment sample It is the first The cost-type indicator in the first The proportion of each road network assessment sample It is the first Information entropy of a benefit-type indicator It is the first Information entropy of a cost-type indicator It is a constant. It is the first Entropy weight of a benefit-type indicator It is the first Entropy weight of a cost-type indicator; Subsequently, the Topology-Optimal Solution Ranking (TOPSIS) method was used for comprehensive scoring. The obtained weights were then used to weight the standardized data to form a weighted evaluation matrix. In the weighted evaluation matrix, the optimal values of each indicator are determined to form the positive ideal solution, and the worst values of each indicator are determined to form the negative ideal solution. The Euclidean distance between each evaluation object and the positive and negative ideal solutions is calculated. Finally, the proximity score is calculated based on the relative distance to the positive and negative ideal solutions. The proximity score is used as the comprehensive walkability score; a higher score indicates better walkability. The calculation formula is:
[0014]
[0015]
[0016] in, It is the Euclidean distance from the road network evaluation sample to the positive ideal solution. It is the Euclidean distance from the road network evaluation sample to the negative ideal solution. It is the first The positive ideal solution for each indicator It is the first The negative ideal solution of each indicator It is the first The relative closeness of each road network evaluation sample to the optimal solution is the comprehensive walkability score.
[0017] Optionally, the step S04 of generating an optimization scheme specifically includes the following steps: evaluating all strategies in the library using a multi-objective optimization algorithm, and selecting Pareto optimal strategies that cannot be comprehensively surpassed by other strategies in multiple conflicting traversability evaluation indicators, thereby forming a Pareto optimal strategy set; traversability evaluation indicators include accessibility, efficiency, resilience, experience, and static structure indicators, and the final generated optimization scheme is one or more optimal action sequences selected from the frontier of the Pareto optimal strategy based on actual decision preferences.
[0018] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) This invention integrates multi-source road network data, constructs a two-dimensional walkability evaluation index system of "dynamic behavior results + static road network structure", and introduces machine learning algorithms such as reinforcement learning and random walk model to build a full-process technical framework covering "multi-source road network morphology analysis - walkability quantitative evaluation - optimization scheme generation - effect visualization comparison". This invention can integrate multi-source data, distinguish morphology types and support dynamic response and autonomous learning walkability optimization methods to improve the overall quality of urban pedestrian environment more efficiently. (2) This invention constructs a dynamic evaluation index system covering four dimensions through multi-scenario random walk simulation experiments, and combines it with static structural indicators adapted to the road network morphology to form a complete evaluation dimension of "dynamic behavior + static structure". Finally, the two types of indicators are standardized and integrated through the entropy weight-TOPSIS method, so that the evaluation results not only reflect the rationality of the road network structure, but also fit the real experience of pedestrians, thus solving the core problem of traditional evaluation that "emphasizes structure, neglects behavior, has a single dimension and poor adaptability".
[0019] (3) This invention constructs an “autonomous iterative optimization mechanism” through reinforcement learning, using the comprehensive walkability score as the reward function. The agent can autonomously call two-level actions of “general structure optimization + type-specific optimization” to calculate the walkability score in the new state in real time. It can complete multiple rounds of iteration of “execute action - update road network state - calculate new score - adjust strategy” without human intervention. At the same time, the iteration complexity is reduced by discretizing the action space to ensure efficient progress of the optimization process. Compared with traditional manual trial and error, this mechanism can greatly shorten the optimization cycle, save a lot of manpower investment in manual parameter tuning and repeated simulation, and avoid the subjective bias of human decision-making by autonomously balancing multiple objective conflicts through the algorithm.
[0020] (4) This invention achieves the technical connection between simulation and optimization through a closed-loop design of "real-time simulation - dynamic optimization - multi-scheme output", and generates the optimal and differentiated road network optimization schemes. Relying on the real-time data interaction between the BIM+GIS fusion engine and the reinforcement learning model, after the agent performs optimization actions, the BIM+GIS engine immediately updates the three-dimensional base model of the street network, and the walkability evaluation module synchronously calculates the new score, realizing a lag-free cycle of "action execution - state update - score feedback", ensuring that the optimization rhythm matches the dynamic of the road network. At the same time, this invention selects the Pareto optimal strategy set through a multi-objective optimization algorithm. This set includes comprehensive optimal schemes that cannot be completely surpassed by other strategies, as well as differentiated schemes under specific objectives. Combined with the BIM+GIS fusion engine, the walkability scores of different schemes are rendered in the three-dimensional model in the form of heat maps in real time, supporting multi-terminal interactive comparison, and the scheme can be quickly selected according to the actual guidance. It breaks through the technical and efficiency problems of "disconnect between simulation and optimization" and "single scheme supply" in traditional optimization. Attached Figure Description
[0021] Figure 1 This is a flowchart of the technical method of the present invention; Figure 2 This is a technical roadmap for the dynamic and static comprehensive evaluation of street network walkability in this invention; Figure 3 This is a technical roadmap for optimizing the dynamic iterative process of street network walkability in this invention. Detailed Implementation
[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0023] like Figure 1 As shown, the present invention provides a dynamic optimization method for walkability of multi-type street networks based on reinforcement learning, comprising the following steps: S01. Street data collection and intelligent road network type identification: The target city street network data is obtained using OSM and preprocessed using Python. The global vector features of the GNN model are automatically extracted based on the DGL framework. The target city road network features are identified and morphologically classified using a graph autoencoder pre-trained model built on PyTorch. The road network types include radial road networks, grid road networks, strip road networks, or free-form road networks.
[0024] In step S01, the target city street network data is obtained and preprocessed by writing a Python script using the OSMnx library. This script obtains a directed graph object of the target city streets, filters out unnecessary road types based on road level attributes, and merges nodes that are only used for connection and have no actual intersection significance. The directed graph object of the streets is then converted into the data object required by the GNN framework PyTorchGeometric to obtain the global vector features required by GNN. The data node features include latitude and longitude coordinates, and the edge features include connection relationships, road length, road level, and direction.
[0025] Step S01 utilizes a pre-trained model to identify the road network features of the target city and perform morphological classification. Specifically, it includes the following steps: In the pre-training stage, self-supervised contrastive learning and cross-entropy loss are employed. The Loss architecture constructs an intelligent recognition system based on a GNN backbone, global pooling, and a classifier, and performs meta-learning optimization on a multi-source road network dataset. It obtains a road network morphology feature extractor with zero-shot generalization capability, enabling the model to autonomously encode the regular connection patterns of "grids," the centrally clustered patterns of "radial" networks, the linear directional patterns of "strip" networks, and the chaotic patterns of "free-form" networks into morphological embedding global feature vectors with multi-scale semantic awareness. In the classification stage, the target city road network data is standardized and preprocessed according to the graphing rules established during the model's pre-training stage to ensure consistent feature space distribution. The road network data is input into the pre-trained GNN backbone network, which serves as the feature extractor, and discriminative graph-level embedding vectors are extracted through a global attention pooling layer. The few-shot classifier is then rapidly fine-tuned based on these embedding vectors. Finally, the optimized classifier outputs the classification probability distributions for radial, grid, strip, or free-form road networks.
[0026] like Figure 2As shown, S02. The setting of the two-level optimization actions is based on the road network editing and simulation functions of the traffic simulation software SUMO, and combined with Python scripts to realize automated control and interaction. Two levels of optimization action sets are set for different types of road networks. The first level is a general structure optimization action set, including but not limited to: adding / removing road intersections, adjusting intersection connectivity, changing road segment lengths, adjusting road width, and adjusting road width by increasing or decreasing the number of lanes. According to the specific road network morphology type identified in step S01, optimization actions are set for its physical structure characteristics. The second level is a type-specific morphology optimization action set, including: for radial road networks, adding / removing ring roads and optimizing the radial road hierarchy structure; for grid road networks, introducing diagonal connecting roads and increasing / reducing local grid density; for strip road networks, adding parallel auxiliary roads and optimizing the connection node layout between main roads and parallel roads; for free-form road networks, adding connecting road segments between dead-end road endpoints and adding orthogonal roads in low connectivity areas. The specific parameters for optimizing actions can be adjusted by modifying the .net.xml file of SUMO, or by using the open-source tool netedit for visual assistance.
[0027] In step S02, the general structure optimization action set is implemented by defining a discretized action space. Specifically, the intersection addition / deletion operation is encoded as a {0, 1} binary action; the connectivity adjustment is quantized as a {-1, 0, +1} degree change operation; the road segment length and width adjustment are discretized into multiple levels of change step size, and within each optimization step size, the agent selects a combined action from the action space.
[0028] In step S02, the type-specific morphological optimization action set is implemented through an action selector based on an attention mechanism. Specifically, based on the road network morphology classification probability distribution obtained in step S01, the activation weights of action subsets are calculated. For radial road networks, the operation of adjusting the number of ring roads and reconstructing the radial road level is activated. For grid-shaped road networks, the operation of optimizing the insertion angle of diagonal connections and adjusting the grid density probability is activated. For strip-shaped road networks, the operation of optimizing the spacing between parallel roads and improving the connectivity between main and auxiliary roads is activated. For free-form road networks, the operation of generating the minimum connection path for dead-end roads and repairing the orthogonal road topology is activated. Subsequently, one-hot encoding technology is used to convert the action parameters into fixed-dimensional binary operation instruction vectors. Finally, the action selection probability distribution is output through the policy network to intelligently select the optimal operation.
[0029] S03. Comprehensive Walkability Evaluation: Construct a dynamic walkability evaluation index system based on behavioral outcomes, and conduct random walk simulation experiments using the SUMO simulation platform. Based on the random walk subject's movement impact index and walk subject index, complete the numerical calculation of the dynamic walkability evaluation index. Simultaneously, utilize the GIS platform and open-source libraries such as OSMnx to construct a static adaptive walkability evaluation index system based on the road network structure, and perform calculations in conjunction with the basic attributes of the road network. Use the entropy weight-TOPSIS method to standardize and weight the original data of the two index systems, and generate a comprehensive walkability score.
[0030] Step S03 involves constructing a dynamic walking evaluation index system based on behavioral outcomes. This system is built from four levels: accessibility, efficiency, resilience, and experience. Accessibility indicators include route diversity, number of intersection visits, and walking coverage. Efficiency indicators include average path tortuosity and shortest path distance. Resilience indicators include interruption sensitivity. Experience indicators include starting point facility coverage, POI contact frequency, and 15-minute coverage. Route diversity refers to the sum of the number of routes from a fixed starting point to a fixed public facility and the number of routes from a non-fixed starting point to a fixed public facility. Number of intersection visits refers to the number of intersection visits obtained through numerous iterations by the random walker. Walking coverage refers to the number of intersection visits obtained by setting a fixed random walker... The metrics include: Walk Steps (maximum and minimum walking range); Average Path Twists and Turns (ratio of shortest path distance to straight-line distance); Shortest Path Distance (shortest route distance from a fixed starting point to a fixed public facility); Interruption Sensitivity (rate of decrease in the ratio of total number of routes and total distance to average distance between a fixed starting point and a fixed public facility before and after a 5% road interruption); Starting Point Facility Coverage (number of facilities within a 1km radius of the random starting point); Point of Interest (POI) Contact Frequency (average number of public facilities encountered per unit of steps); and 15-Minute Coverage (proportion of public facilities passed within 1.2km of a random starting point to the total number of public facilities in the area). The correspondence between experiment categories and walkability evaluation metrics based on behavioral outcomes is shown in Table 1: Table 1
[0031] In step S03, the movement impact index and the walk subject index are as follows: the random walk subject movement impact factor index includes road segment attributes and intersection attributes; the walk subject index includes movement direction and number of steps; the road segment attributes include street continuity and street length; the intersection attributes include intersection connectivity and intersection elevation; the movement direction setting involves constructing a direction vector matrix at the location of the walk subject, and selecting the next movement position only within the range of the direction vector matrix constructed from the starting point to the ending point; at the same time, the probability of going straight is 0.6, and the probability of turning left / right is 0.2; the number of steps is either a fixed number of steps defined in experiments with a fixed radius / fixed movement distance, or a random number of steps with only a fixed starting point and ending point and no specified movement distance.
[0032] Step S03 involves a static adaptive walkability evaluation index system based on road network structure, specifically including road network density, dead-end road density, road network closure, road intersection density, average street length, and street continuity. The calculation formulas are as follows: , , , , , .
[0033] Step S03 employs the entropy weight-TOPSIS method, specifically including the following steps: First, based on the number of road network assessment samples and evaluation indicators Construct the initial evaluation matrix To eliminate the influence of different indicator dimensions, the extreme value method is used to standardize the original data of two systems in the matrix: dynamic walking evaluation indicators based on behavioral results and static adaptive walking evaluation indicators based on road network structure. The indicators are then divided into two categories: benefit-type indicators and cost-type indicators. Higher values for benefit-type indicators are better, while lower values for cost-type indicators are better. The standardized matrix is obtained, and the calculation formula is as follows: Formula for benefit-type indicators: , in, It is the first The first road network assessment sample in the first Standardized values for each benefit-type indicator It is the first The first road network assessment sample in the first The original values of each benefit-type indicator It is the first The maximum original value of each benefit-type indicator It is the first The minimum original value for each benefit-type indicator; Cost-based indicator formula: , in, It is the first The first road network assessment sample in the first Standardized values for cost-related indicators It is the first The first road network assessment sample in the first The original values of each cost-type indicator It is the first The maximum original value of each benefit-type indicator It is the first The minimum original value for each benefit-type indicator.
[0034] Secondly, calculate the proportion of each indicator value in all samples, and calculate the information entropy of each indicator based on the proportion. The smaller the entropy value, the greater the data variation of that indicator, and the more important it is in the evaluation. Finally, derive the objective weight of each indicator from the information entropy. The calculation formula is as follows: , , , , , , in, It is the first The first benefit-type indicator in the first The proportion of each road network assessment sample It is the first The cost-type indicator in the first The proportion of each road network assessment sample It is the first Information entropy of a benefit-type indicator It is the first Information entropy of a cost-type indicator It is a constant. It is the first Entropy weight of a benefit-type indicator It is the first Entropy weight of a cost-type indicator.
[0035] Subsequently, the Topology-Optimal Solution Ranking (TOPSIS) method was used for comprehensive scoring. The obtained weights were then used to weight the standardized data to form a weighted evaluation matrix. In the weighted evaluation matrix, the optimal values of each indicator are determined to form the positive ideal solution, and the worst values of each indicator are determined to form the negative ideal solution. The Euclidean distance between each evaluation object and the positive and negative ideal solutions is calculated. Finally, the proximity score is calculated based on the relative distance to the positive and negative ideal solutions. The proximity score is used as the comprehensive walkability score; a higher score indicates better walkability. The calculation formula is: , , , in, It is the Euclidean distance from the road network evaluation sample to the positive ideal solution. It is the Euclidean distance from the road network evaluation sample to the negative ideal solution. It is the first The positive ideal solution for each indicator It is the first The negative ideal solution of each indicator It is the first The relative closeness of each road network evaluation sample to the optimal solution is the comprehensive walkability score.
[0036] like Figure 3 As shown, S04. Walkability dynamic optimization generates an optimization scheme. Based on the reinforcement learning framework, a training environment is constructed. The walkability score in step S03 is used as the input of the reward function of the reinforcement learning model. According to the action settings in step S02, the agent executes action combinations and calculates the walkability score in the new state in real time. After multiple rounds of state-action-reward iteration, a Pareto optimal policy set is generated using a multi-objective optimization algorithm, and finally, an optimization scheme is generated through a Python automated script.
[0037] Step S04, which generates an optimization scheme, specifically includes the following steps: all strategies in the library are evaluated using a multi-objective optimization algorithm, and Pareto optimal strategies that cannot be comprehensively surpassed by other strategies in multiple conflicting traversability evaluation metrics are selected, thus forming a Pareto optimal strategy set. The traversability evaluation metrics include accessibility, efficiency, resilience, experience, and static structure metrics. The final generated optimization scheme is one or more optimal action sequences selected from the frontier of the Pareto optimal strategy based on actual decision preferences.
[0038] S05. Optimization Scheme Comparison and Interaction Module. A 3D base model of the street network is constructed using a BIM+GIS fusion engine. WebGL technology is used to dynamically link walkability score data, rendering score changes in real-time on the 3D model surface in the form of a heatmap. A web front-end framework supports multi-terminal interactive operation, and combined with back-end services and a database, it provides scheme comparison and displays the 3D optimization process and results of the urban street network.
[0039] Step S05, the BIM+GIS fusion engine, specifically includes the following steps: Based on the updated road network morphology data obtained in step S04, by integrating the spatial coordinate system and topographic data of GIS with the parametric component library and geometric expression capabilities of BIM, three-dimensional models of roads and intersections are instantiated in the GIS spatial environment, and detailed attribute information and adjustable parameters are assigned to them. By providing a unified application programming interface (API) and rendering services, it supports the optimization of visual mapping of action sequences, dynamic correlation of walkability score data, and real-time three-dimensional rendering.
Claims
1. A dynamic optimization method for walkability of multi-type street networks based on reinforcement learning, characterized in that, Includes the following steps: S01. Street data collection and intelligent road network type identification: Obtain street network data of the target city, perform preprocessing, use a pre-trained model to identify the road network characteristics of the target city and classify the morphology. The road network types include radial road network, grid road network, strip road network or free road network. S02. Setting up two-level optimization actions: Two sets of optimization actions are set up for different types of road networks. The two sets of optimization actions include a first-level general structure optimization action set and a second-level type-specific morphology optimization action set. Based on the specific road network morphology type identified in step S01, optimization actions are set for its physical structure characteristics. S03. Comprehensive evaluation of walkability: Construct a dynamic walkability evaluation index system based on behavioral outcomes, conduct random walk simulation experiments, and complete the numerical calculation of dynamic walkability evaluation indexes based on the random walk subject's movement influence index and walk subject index; construct a static adaptive walkability evaluation index system based on road network structure, and calculate it in conjunction with the basic attributes of the road network; use the entropy weight-TOPSIS method to standardize and weight the original data of the two index systems, and generate a comprehensive walkability score. S04. Dynamic optimization of gaitability to generate an optimization scheme: The gaitability score in step S03 is used as the input of the reward function of the reinforcement learning model. Based on the action settings in step S02, the agent executes a combination of actions and calculates the gaitability score in the new state in real time. After multiple rounds of state-action-reward iteration, a Pareto optimal policy set is generated using a multi-objective optimization algorithm, and an optimization scheme is generated. S05. Comparison and interaction of optimization schemes and display of results.
2. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 1, characterized in that, In step S01, obtaining and preprocessing the target city street network data involves writing a Python script using the OSMnx library to obtain a directed graph object of the target city streets. Unnecessary road types are filtered out based on road level attributes, and nodes that are only used for connection and have no actual intersection significance are merged. The street directed graph object is then converted into the data object required by the GNN framework PyTorch Geometric to obtain the global vector features required by GNN. The data node features include latitude and longitude coordinates, and the edge features include connectivity, road length, road grade, and direction.
3. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 2, characterized in that, Step S01, which utilizes a pre-trained model to identify the road network features of the target city and perform morphological classification, specifically includes the following steps: In the pre-training stage, self-supervised contrastive learning and cross-entropy loss are employed. The Loss architecture constructs an intelligent recognition system based on a GNN backbone, global pooling, and a classifier, and performs meta-learning optimization on a multi-source road network dataset. It obtains a road network morphology feature extractor with zero-shot generalization capability, enabling the model to autonomously encode the regular connection patterns of "grids," the centrally clustered patterns of "radial" networks, the linear directional patterns of "strip" networks, and the chaotic patterns of "free-form" networks into morphological embedding global feature vectors with multi-scale semantic awareness. In the classification stage, the target city road network data is standardized and preprocessed according to the graphing rules established during the model's pre-training stage to ensure consistent feature space distribution. The road network data is input into the pre-trained GNN backbone network, which serves as the feature extractor, and discriminative graph-level embedding vectors are extracted through a global attention pooling layer. The few-shot classifier is then rapidly fine-tuned based on these embedding vectors. Finally, the optimized classifier outputs the classification probability distributions for radial, grid, strip, or free-form road networks.
4. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 3, characterized in that, The first-level general structural optimization action set in step S02 includes, but is not limited to: adding / removing road intersections, adjusting intersection connectivity, changing road segment length, and adjusting road width by increasing or decreasing the number of lanes; The first-level general structure optimization action set is implemented by defining a discretized action space; specifically, the intersection addition / removal operation is encoded as follows: Binary action; quantize the connectivity adjustment. The degree change operation; the road segment length and width adjustment is discretized into multiple levels of change step size, and within each optimization step size, the agent selects a combined action from the action space.
5. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 4, characterized in that, The second-level type-specific morphological optimization action set in step S02 includes: for radial road networks, adding / reducing ring roads and optimizing the radial road hierarchy; for grid road networks, introducing diagonal connecting roads and increasing / reducing local grid density; for strip road networks, adding parallel auxiliary roads and optimizing the connection node layout between main roads and parallel roads; for free-form road networks, adding connecting road segments between dead-end road endpoints and adding orthogonal roads in low connectivity areas. This second-level type-specific morphological optimization action set is implemented through an action selector based on an attention mechanism. Specifically, based on the road network morphological classification probability distribution obtained in step S01, the activation weight of the action subset is calculated. For radial road networks, the operation of adjusting the number of ring roads and reconstructing the radial road hierarchy is activated. For grid road networks, the operation of optimizing the insertion angle of diagonal connections and adjusting the grid density probability is activated. For strip road networks, the operation of optimizing the spacing between parallel roads and improving the connectivity of main and auxiliary roads is activated. For free-form road networks, the operation of generating the minimum connection path for dead-end roads and repairing the topology of orthogonal roads is activated. Then, one-hot encoding is used. Encoding technology converts action parameters into fixed-dimensional binary operation instruction vectors; finally, the Policy Network outputs the action selection probability distribution to intelligently select the optimal operation.
6. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 5, characterized in that, The construction of a dynamic walking evaluation index system based on behavioral outcomes in step S03 refers to constructing a dynamic walking evaluation index system based on the behavioral outcomes of random walk subjects from four aspects: accessibility, efficiency, resilience, and experience. Accessibility indicators include route diversity, number of intersection visits, and walking coverage. Efficiency indicators include average path tortuosity and shortest path distance; resilience indicators include interruption sensitivity; experience indicators include starting point facility coverage, POI contact frequency, and 15-minute coverage. Route diversity refers to the sum of the number of routes from a fixed starting point to a fixed public facility and the number of routes from non-fixed starting points to a fixed public facility; intersection visit count refers to the number of intersection visits obtained by the random walk subject through numerous iterations; walking coverage refers to the maximum and minimum walking range of the random walk subject by setting a fixed number of steps; average path tortuosity refers to the ratio of the shortest path distance to a public facility to the straight-line distance; shortest path distance refers to the shortest route distance from the fixed starting point to the fixed public facility; interruption sensitivity refers to the rate of decrease in the ratio of the total number of routes and the total distance to the average distance between the fixed starting point and the fixed public facility before and after a 5% road interruption; starting point facility coverage rate refers to the number of facilities within a 1km radius centered on the random starting point of the random walk subject; POI contact frequency refers to the average number of times the random walk subject encounters a public facility per unit of steps; 15-minute coverage rate refers to the proportion of public facilities passed by the random walk subject within 1.2km of the random starting point to the total number of public facilities in the area.
7. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 6, characterized in that, In step S03, the movement impact index and the walk subject index are as follows: the random walk subject movement impact factor index includes road segment attributes and intersection attributes, and the walk subject index includes movement direction and number of steps. The road segment attributes include street continuity and street length; the intersection attributes include intersection connectivity and intersection elevation; the movement direction setting involves constructing a direction vector matrix at the location of the walking subject, and selecting the next movement position only within the range of the direction vector matrix constructed from the starting point to the ending point; at the same time, the probability of going straight is 0.6, and the probability of turning left / right is 0.2; the number of movement steps is either a fixed number of steps defined in experiments with a fixed radius / fixed movement distance, or a random number of steps with only a fixed starting point and ending point and no specified movement distance.
8. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 7, characterized in that, The static adaptive walkability evaluation index system based on road network structure in step S03 specifically includes road network density, dead-end road density, road network closure, road intersection density, average street length, and street continuity, with the following calculation formulas: ， ， ， ， ， 。 9. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 8, characterized in that, Step S03 employs the entropy weight-TOPSIS method, specifically including the following steps: First, based on the number of road network assessment samples and evaluation indicators Construct the initial evaluation matrix: ， To eliminate the influence of different indicator dimensions, the extreme value method is used to standardize the original data of two systems in the matrix: dynamic walking evaluation indicators based on behavioral results and static adaptive walking evaluation indicators based on road network structure. The indicators are then divided into two categories: benefit-type indicators and cost-type indicators. Higher values for benefit-type indicators are better, while lower values for cost-type indicators are better. The standardized matrix is obtained, and the calculation formula is as follows: ， in, It is the first The first road network assessment sample in the first Standardized values for each benefit-type indicator It is the first The first road network assessment sample in the first The original values of each benefit-type indicator It is the first The maximum original value of each benefit-type indicator It is the first The minimum original value for each benefit-type indicator; ， in, It is the first The first road network assessment sample in the first Standardized values for cost-related indicators It is the first The first road network assessment sample in the first The original values of each cost-type indicator It is the first The maximum original value of each benefit-type indicator It is the first The minimum original value for each benefit-type indicator; Secondly, calculate the proportion of each indicator value in all samples, and calculate the information entropy of each indicator based on the proportion. The smaller the entropy value, the greater the data variation of that indicator, and the more important it is in the evaluation. Finally, derive the objective weight of each indicator from the information entropy. The calculation formula is as follows: ， ， ， ， ， ， in, It is the first The first benefit-type indicator in the first The proportion of each road network assessment sample It is the first The cost-type indicator in the first The proportion of each road network assessment sample It is the first Information entropy of a benefit-type indicator It is the first Information entropy of a cost-type indicator It is a constant. It is the first Entropy weight of a benefit-type indicator It is the first Entropy weight of a cost-type indicator; Subsequently, the TOPSIS (Topological Solution Approximation System) method was used for comprehensive scoring. The obtained weights were then used to weight the standardized data to form a weighted evaluation matrix. ， In the weighted evaluation matrix, the optimal value of each indicator is determined to form the positive ideal solution, and the worst value of each indicator is determined to form the negative ideal solution. The Euclidean distance between each evaluation object and the positive and negative ideal solutions is calculated. Finally, the proximity score is calculated based on the relative distance to the positive and negative ideal solutions. The proximity score is used as the comprehensive walkability score, and the higher the score, the better the walkability. The calculation formula is as follows: ， ， ， in, It is the Euclidean distance from the road network evaluation sample to the positive ideal solution. It is the Euclidean distance from the road network evaluation sample to the negative ideal solution. It is the first The positive ideal solution for each indicator It is the first The negative ideal solution of each indicator It is the first The relative closeness of each road network evaluation sample to the optimal solution is the comprehensive walkability score.
10. The method for dynamic optimization of walkability of multi-type street networks based on reinforcement learning according to claim 9, characterized in that, The specific steps for generating the optimization scheme in step S04 include the following: evaluating all strategies in the library using a multi-objective optimization algorithm, and selecting Pareto optimal strategies that cannot be comprehensively surpassed by other strategies in multiple conflicting traversability evaluation indicators, thereby forming a Pareto optimal strategy set; traversability evaluation indicators include accessibility, efficiency, resilience, experience, and static structure indicators, and the final generated optimization scheme is one or more optimal action sequences selected from the frontier of the Pareto optimal strategy based on actual decision preferences.