Integrating AI for sustainable architectural space optimization and heritage-conscious street design Authors: Yanan Hu, Jie Zhong, Zhiming Peng Published: 2026 ABSTRACT The paper presents an AI-powered framework designed to optimize the layout of historical and cultural street spaces while preserving heritage values. Combining deep learning, graph-based techniques, and reinforcement learning, this framework addresses common urban challenges such as pedestrian accessibility, traffic congestion, and environmental sustainability in historic districts. Unlike traditional urban planning methods, the framework offers a dynamic, data-driven approach that adapts to real-time changes in pedestrian flow, traffic, and environmental conditions. By integrating Convolutional Neural Networks (CNNs) for spatial feature extraction, Graph Neural Networks (GNNs) for network connectivity analysis, and Reinforcement Learning (RL) for dynamic layout adjustments, the system ensures efficient and sustainable urban designs. The proposed model was tested in cities with historical significance and demonstrated significant improvements: a 22.5% reduction in congestion, a 27.4% increase in green space, and a 31.8% improvement in pedestrian accessibility compared to traditional methods. The results validate the effectiveness of AI in urban planning, showing that it can balance the need for modernization with the preservation of historical authenticity. Introduction The sustainable development that is focused on urban planning is needed because cities can develop while being able to preserve their identities1. An essential element in the heritage of the city is the layout of streets, making it the representation of the architectural and cultural history of the city2. These are, however, areas that have major problems because of urbanization and changing needs of transport, which tends to bring a clash between modernity and conservation. Artificial intelligence is a cutting-edge technology that uses machine learning, spatial analytics, and multi-objective optimization to find new ways to meet the contemporary functional needs without sacrificing historical relevance3,4. In comparison to standard grid-like structures, historical streets are focused on being more accessible, culturally relevant, and to provide human-scale experiences that add vibrancy and economic well-being to a city5. Such regions promote tourism, local commerce, and social unity, but also are exposed to challenges such as congestion, old-fashioned infrastructure, and tough protein laws6,7. City smart optimization is based on AI decision-making that can lead to data-driven decisions that optimize pedestrian circulation, infrastructure effectiveness, and intelligent city ecosystems, as well as such heritage limitations8. This is one of the greatest issues in enhancing older urban regions with the need to improve the congestion, especially on the narrow, hilly alleys that were not built to support the current type of vehicles, including the emergency vehicles such as the ambulances9. This is complicated by the stringent preservation laws, where any modifications to a building should not change its outlook but only reflect the original state10. Between them, environmental issues are also involved since numerous heritage areas are characterized by a poor level of green spaces, energy-efficient structures, or waste disposal systems7,11. Moreover, the process of gentrification should be reduced by urban planners so that new development does not oust the locals and does not demolish cultural heritage. These are the main consequences that should be considered in the planning strategy, which is shown in Fig. 1. The complexities of historical cities do not accommodate a traditional urban planning model, which works best on a modern grid layout12. However, artificial intelligence-based analysis tools will have the potential to study unusual street structures, anticipate pedestrian and vehicle flows, and provide adaptive decisions that meet the criteria of heritage protection13. Dynamic urban optimization integrates real-time information, allowing it to extend efficiency and cultural integrity. AI-based solutions provide a systematic way of striking a balance between the past and current needs of development. Such technologies can assist the cities in adjusting their historical heritage to the current needs to improve traffic movement, cut on energy consumption, and adhere to the conservation legislation, which will result in urban settings that are not only environmentally friendly but also culturally diverse.Fig. 1Challenges in improving historical street designs and suggested AI-based solutions, addressing concerns such as traffic congestion, pedestrian mobility, and the preservation of cultural heritage. Challenges in improving historical street designs and suggested AI-based solutions, addressing concerns such as traffic congestion, pedestrian mobility, and the preservation of cultural heritage. This study presents an AI-based framework to optimize space in historical streets using machine learning, reinforcement learning, and multi-objective optimization. Unlike traditional planning methods, it offers a flexible solution to urban challenges by improving pedestrian safety and reducing congestion. Tested in cities like Rome and Cairo, it enhances green spaces, reduces pollution, and promotes sustainability. The framework combines deep learning, predictive modeling, and generative design to respond dynamically to urban design challenges, ensuring functional and sustainable cities. Compared to conventional models, it better preserves architectural integrity and addresses environmental concerns. The research employs quantitative analysis, utilizing metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document} scores to validate the optimization model, moving beyond the qualitative methods used in previous studies. The outcomes demonstrate a 22.5% reduction in congestion, a 27.4% increase in green space, and a 31.8% improvement in pedestrian accessibility, compared to traditional planning approaches. The framework’s quantitative analysis shows improvements in congestion, green space, and pedestrian accessibility, demonstrating its adaptability for global urban planning. It incorporates real-world constraints like traffic patterns and city zoning to assist in municipal planning while preserving the character of historical streets. Related work Urban planning has changed over the years; old streets that can be considered as cultural are challenging because they have inconsistent shapes, and the limited space, no additional square meters can be found in them14. City planners currently emphasize preservation, physical change, and judgment allowance based on experience. Nevertheless, such methods usually cannot find a balance between the requirements of a contemporary city and the willingness to preserve its historical flavor15. With recent breakthroughs in the field of AI and machine learning, new methods of architectural planning have become considerably more innovative, and the decision could be data-driven it, contributing to greater accessibility, efficiency, and sustainability of planning. Development of new knowledge on structures, topology, and the exploration of the design space has been achieved through research on architectural design optimization, which has given new information on how to enhance the organization of the city1. The section will talk about the historical city design, optimization-based method on multiple objectives for urban planning, and the research gaps that need to be filled to come up with an AI-based framework on historical street design16. Previous research has addressed how to maintain heritage regions, improve access through foot traffic, and incorporate more infrastructure, as seen in Table 1. Although these modes are effective in ensuring the individuality of buildings, they are usually ineffective in managing current challenges like traffic jams and the sprawling of land. City planners can work with generative AI and reinforcement learning to come up with solutions that are similar to those that have been used in the past, but resolve the demands presently. Most AI platforms do not work well in historic cities, as these platforms are constructed to work in modern grids. With the use and reference to the existing theories, we hope to arrive at a middle way, where the cultures will be saved, but the systems in the cities will become better. Urban development favors the conservation aspect, where the conservationist approach emphasizes the structure of the old buildings by creating rules that must be followed strictly by the construction industry. Although this guarantees conservation of historical buildings, it may be detrimental to the modernization of the city, leading to road jams, inefficient utilization of spaces, and challenging commuting.Table 1Recent proposed models in literature with possible limitations.RefProposed modelTools/technologyExperimentationContributionsLimitations17A comprehensive method for optimizing the design of a regular architectural space to improve building performance.Artificial Neural Networks (ANN), Multi-objective optimization, Building simulations.Simulations on randomly generated design cases to evaluate building performance.Provides an AI-driven approach to architectural space optimization, accelerating computations significantly.Focuses mainly on classroom spaces; broader applicability to other architectural designs remains unexplored.18DeepFlow: A cross-stack pathfinding framework for distributed AI systems.Cross-layer performance modeling, Machine learning techniques.Validated with distributed training on commercial AI hardware.Automates design space exploration and co-optimization across different layers of the technology stack.Limited experimentation on edge-AI scenarios; primary focus on large-scale cloud AI systems.19A design space for intelligent and interactive writing assistants.AI-driven writing assistants, User-centric design space analysis.Systematic review of 115 research papers; collaboration with domain experts.Defines key aspects of writing assistants, offering structured guidance for future research and development.Study is qualitative and lacks empirical validation through real-world AI writing assistant deployments.20Deep learning-based chip power prediction and optimization in Electronic Design Automation (EDA).Deep learning, Attention mechanisms, Generative adversarial networks (GANs).Case studies on functional verification, Quality of Results (QoR), and Optical Proximity Correction (OPC) layout generation.Improves chip design efficiency and performance, enabling AI-driven semiconductor design.Primarily focuses on AI-based power optimization; broader applications in EDA workflows remain unexplored.21End-to-end generation of structural topology for complex architectural layouts using graph neural networks.Generative Adversarial Networks (GAN), Graph Neural Networks (GNN).Trained and tested on a dataset of 300 complex architectural and structural layouts.Introduces a novel GAN-based approach to generate structural topologies with high accuracy.Evaluation focuses on topology generation; lacks extensive real-world structural validation.22Optimizing event-based neural networks on digital neuromorphic architecture: A comprehensive design space exploration.Neuromorphic processors, Event-driven data-flow processing, Near/in-memory computing.Benchmarked on tasks such as keyword spotting, sensor fusion, and object detection.Demonstrates significant improvements in energy efficiency, latency, and area efficiency compared to conventional AI hardware.The approach is specific to neuromorphic processors, limiting general applicability to standard AI accelerators.23Design space exploration and explanation using Conditional Variational Autoencoders (CVAE) for pedestrian bridge conceptual design.Conditional Variational Autoencoder (CVAE), Meta-model-based conceptual design.Tested on a pedestrian bridge case study for structural improvements under user-defined objectives.Enables efficient design exploration with improved sensitivity analysis for structural enhancements.Primarily applied to pedestrian bridges; its effectiveness for other large-scale infrastructure remains unexplored. Table 1Recent proposed models in literature with possible limitations.RefProposed modelTools/technologyExperimentationContributionsLimitations17A comprehensive method for optimizing the design of a regular architectural space to improve building performance.Artificial Neural Networks (ANN), Multi-objective optimization, Building simulations.Simulations on randomly generated design cases to evaluate building performance.Provides an AI-driven approach to architectural space optimization, accelerating computations significantly.Focuses mainly on classroom spaces; broader applicability to other architectural designs remains unexplored.18DeepFlow: A cross-stack pathfinding framework for distributed AI systems.Cross-layer performance modeling, Machine learning techniques.Validated with distributed training on commercial AI hardware.Automates design space exploration and co-optimization across different layers of the technology stack.Limited experimentation on edge-AI scenarios; primary focus on large-scale cloud AI systems.19A design space for intelligent and interactive writing assistants.AI-driven writing assistants, User-centric design space analysis.Systematic review of 115 research papers; collaboration with domain experts.Defines key aspects of writing assistants, offering structured guidance for future research and development.Study is qualitative and lacks empirical validation through real-world AI writing assistant deployments.20Deep learning-based chip power prediction and optimization in Electronic Design Automation (EDA).Deep learning, Attention mechanisms, Generative adversarial networks (GANs).Case studies on functional verification, Quality of Results (QoR), and Optical Proximity Correction (OPC) layout generation.Improves chip design efficiency and performance, enabling AI-driven semiconductor design.Primarily focuses on AI-based power optimization; broader applications in EDA workflows remain unexplored.21End-to-end generation of structural topology for complex architectural layouts using graph neural networks.Generative Adversarial Networks (GAN), Graph Neural Networks (GNN).Trained and tested on a dataset of 300 complex architectural and structural layouts.Introduces a novel GAN-based approach to generate structural topologies with high accuracy.Evaluation focuses on topology generation; lacks extensive real-world structural validation.22Optimizing event-based neural networks on digital neuromorphic architecture: A comprehensive design space exploration.Neuromorphic processors, Event-driven data-flow processing, Near/in-memory computing.Benchmarked on tasks such as keyword spotting, sensor fusion, and object detection.Demonstrates significant improvements in energy efficiency, latency, and area efficiency compared to conventional AI hardware.The approach is specific to neuromorphic processors, limiting general applicability to standard AI accelerators.23Design space exploration and explanation using Conditional Variational Autoencoders (CVAE) for pedestrian bridge conceptual design.Conditional Variational Autoencoder (CVAE), Meta-model-based conceptual design.Tested on a pedestrian bridge case study for structural improvements under user-defined objectives.Enables efficient design exploration with improved sensitivity analysis for structural enhancements.Primarily applied to pedestrian bridges; its effectiveness for other large-scale infrastructure remains unexplored. Recent proposed models in literature with possible limitations. Historical districts are popular for tourism and business, which can lead to challenges in managing space and often result in overcrowding24. To improve walkability and reduce traffic, urban planners have transformed certain roads into pedestrian-only areas, a process known as pedestrianization. While this strategy helps improve air quality and enhances the aesthetic appeal of cities25, it also introduces some unforeseen challenges. Restricting vehicle access can lead to increased congestion on local streets, disrupt delivery services, and create difficulties for people with disabilities or the elderly in accessing certain areas. Additionally, without adaptive models that analyze real-time movement, pedestrian zones may be placed inefficiently, exacerbating congestion and accessibility issues26. The modernization of historic districts often involves upgrading infrastructure while maintaining traditional architecture. This includes improving main roads, upgrading drainage systems27, and installing new utilities underground, all while preserving the city’s skyline. Table 2 outlines recent models proposed in the literature, along with their potential limitations.Table 2Recent proposed models in literature with possible limitations.RefEFCEMOOSCADAAAIO17\checkmark\checkmark\checkmarkXXXX18X\checkmarkX\checkmark\checkmarkX\checkmark19XX\checkmark\checkmark\checkmark\checkmarkX20\checkmark\checkmarkXXXX\checkmark21X\checkmark\checkmark\checkmarkX\checkmarkX22X\checkmarkX\checkmark\checkmarkX\checkmark23\checkmark\checkmark\checkmarkXX\checkmarkXProposed Model\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmarkEF = Energy Efficiency, CE = Computational Efficiency, MOO = Multi-Objective Optimization, SC = Scalability, AD = Adaptability, AA = Architectural Applicability, AIO = AI-based Optimization. Table 2Recent proposed models in literature with possible limitations.RefEFCEMOOSCADAAAIO17\checkmark\checkmark\checkmarkXXXX18X\checkmarkX\checkmark\checkmarkX\checkmark19XX\checkmark\checkmark\checkmark\checkmarkX20\checkmark\checkmarkXXXX\checkmark21X\checkmark\checkmark\checkmarkX\checkmarkX22X\checkmarkX\checkmark\checkmarkX\checkmark23\checkmark\checkmark\checkmarkXX\checkmarkXProposed Model\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmarkEF = Energy Efficiency, CE = Computational Efficiency, MOO = Multi-Objective Optimization, SC = Scalability, AD = Adaptability, AA = Architectural Applicability, AIO = AI-based Optimization. Recent proposed models in literature with possible limitations. EF = Energy Efficiency, CE = Computational Efficiency, MOO = Multi-Objective Optimization, SC = Scalability, AD = Adaptability, AA = Architectural Applicability, AIO = AI-based Optimization. The process of renewing urban infrastructure is quite costly as well as time-consuming, and there is no certainty that the problems of space will be resolved. Geographic Information Systems (GIS) are widely used as they assess and interpret spatial information in conventional urban planning; however, they are not only predictive in relation to modeling. Planning with GIS requires an intensive human talent, and until artificial intelligence automates some of the work, it would perform less favorably in the realignment of historic street patterns28. It is changing the face of urban planning through increased levels of spatial efficiency within buildings due to the use of artificial intelligence. Machine learning, reinforcement learning, and generative design algorithms that are featured in AI-based tools will allow predicting spatial arrangements, optimizing land use, and enhancing pedestrian accessibility29. These tools provide planners with insights based on data collected to make decisions so as to maximize the use of urban spaces, irrespective of the layout30. The AI algorithms used are generative, which means that they develop plans individually based on particular demands, including accessibility, sustainability, and cultural continuity30. These methodologies examine the circulation of pedestrians, street layouts, and building density to generate effective designs with minimal interference. Generative adversarial networks (GANs) are employed to create complex building layouts that automatically conform to the regulations. As an example, the Struct model uses GAN methods to produce tall topologies with high precision and efficiency. The approaches are, however, mostly applicable in contemporary planning, and modifications are necessary when one applies them to older streets that have grown organically31. Reinforcement learning also finds use in improving the traffic flow, crossing people, and getting around by means of transport. Patterns identified as part of real life, these systems keep on optimizing urban forms to enhance transportation. With the help of AI simulations, one can test most urban designs before introducing significant alterations and, thus, decrease inefficiencies of city architectures32. However, the majority of reinforcement learning models are designed to suit the standard grid-like design of cities and do not cope well with grid-free designs, which are prevalent in historical applications. In recent years, the integration of artificial intelligence (AI) in urban and architectural design has gained significant attention for its potential in optimizing space and enhancing sustainability. AI-driven approaches have been shown to outperform traditional methods in key areas such as pedestrian accessibility, traffic efficiency, and heritage preservation33–36. Additionally, the application of deep learning and data augmentation techniques has demonstrated promising results in improving the accuracy of models used for spatial analysis and environmental impact reduction37–39. These advancements align with the growing need for AI models that can adapt to the complexities of historical and heritage-conscious street networks40,41, offering a more responsive and efficient framework for city planning and design. Historic districts use urban planning, which focuses on multi-objective optimization that accommodates more than one challenge. The process of creating the optimal street in a historic district can be described as the collaboration of pedestrian facilitation, the management of motorized traffic, cultural heritage protection, and sustainability42. Conventional urban planning approaches are usually prioritized to achieve a decreased congestion level or walkable conditions, and thus fail to realize trade-offs required to address various challenges43,44. It is because such models provide a good grounding to address conflicting objectives in urban planning. These algorithms allow planners to address concerns such as space efficiency, crowd control, and heritage conservation, since they allow planners to evaluate several scenarios at once45,46. Current architectural researches reveal that multi-goal AI models improve building productivity47. The evolution of urban design practice has increasingly incorporated computational modeling, generative systems, and participatory data platforms. Contemporary research highlights how data-driven design tools are reshaping civic engagement, spatial analysis, and evidence based planning methodologies. These developments underscore the transition from static regulatory frameworks toward adaptive and algorithm assisted urban governance. An example is the fact that artificial neural networks (ANNs) have been demonstrated to act as alternatives to time-consuming simulations in multi-objective optimization problems, and as such, ANNs can be used to reduce decision time. A simulation study of the optimization of neural networks showed an improvement of simulation time by 2,570 times and a design efficiency of 24.6 percent. Such a style helps the city planners to preserve historic sensibility along with the environmentally friendly street design. Although the main contribution of AI in the realm of architectural spaces is optimization, the existing approaches are not sufficient for historic street optimization. The AI model of planning is most applicable to consistent, standardized city plans, and this is inefficient in older places that have distinct spatial and design structures48. Road, pedestrian area, and land use optimization AI systems are focused on structured cities yet cannot be easily applied to historical districts with non-regular road thinking, changeable widths, and a scarce amount of space49. The majority of urban planning with the help of AI does not give attention to the preservation of heritage. Although the AI is a good tool to regulate the space and overcrowding, it seldom gives special consideration to the architectural protection50. The systems are ideal in contemporary subdivisions with regular layouts, but rely on human programming and have difficulty with the historical streets of irregular layout. On the other hand, AI based on reinforcement learning can learn by itself constraints and adapt to the naturally developed shapes of cities. Such systems require ways of controlling space without affecting important places and having a consistent look. Unless AI can be made heritage-specific, planners make at-risk decisions that will change the original nature of historic districts. Urban optimization tools fall into three categories: GIS-based systems, multi-objective models, and AI-driven methods. GIS tools are simple but focus on single objectives and lack sustainability metrics. Multi-objective models improve factors like walkability but struggle with irregular urban forms and heritage preservation. AI methods often overlook heritage constraints. The proposed framework combines CNN, GNN, and reinforcement learning, addressing walkability, traffic, environmental impact, and heritage compliance, offering a more adaptive, sustainability-focused solution for historic environments. The existing methods have significant drawbacks because they are based on one-time optimization procedures as opposed to adapting to new information in real time. Historical districts are constantly developing, and people walk there, so the urban planning systems that have AI and perform urban planning must be versatile. Performance-based adaptive urban designs can only be achieved with the help of predictive modeling systems that can be improved by conditional variational autoencoders (CVAE). No one argues on using these methods only in the historical cities. Urban evolution should be addressed by AI-based methods based on real-time responsiveness, heritage conservation, and multi-purpose optimization. A model of an AI-dependent structure of optimization retains the original essence of historical districts using algorithms, adaptive learning, and predictive modelling to generate city components that address the current demands without damaging the historical heritage of the cities. Problem statement The importance of the historic and cultural streets means that they contribute to preserving the urban heritage, since they can withstand the forces of growth of the city, changing transport requirements, and the limitations of the environment. The task of optimizing the urban space without compromising on historical authenticity is hard to accomplish via traditional planning approaches that rely on strict implementation of the zoning rules and case-by-case comparison, which is based on human intuition and subject to error17. Cities are also becoming highly complex and large, so they pollute the environment, and they extract cultural identity and the quality of life18. Things like high concentrations of people and an increase in car traffic lead to the continuous traffic jam of the sidewalks, resulting in land waste and harm to the environment of a known historic district19. A combination of our framework and the concept of having set grammar rules allows one to increase the walkability of an area due to real-time alterations of pathways in accordance with pedestrian flow51,52. When gridlock conditions are detected, the reinforcement learning (RL) system helps in the rerouting of pedestrians, where conventional rules fail to consider such limitations. The use of traditional planning leads to an increase in the urban populations and the loss of natural areas and cultural objects20. Irregular patterns of historic cities are not as compatible with analysis as the modern city patterns, which means that such analysis is less effective to optimise it21. They cannot respond to the major shifts of the urban environment and tend to overlook such aspects as people traffic, ecological hazards, and protection of heritage22. Suable tools that deal with multiple factors are essentially what good planners are in need of at the moment because the sub-optimal results of optimization efforts mean generating adverse effects, including the increased duration of walks and the destruction of historic parts of cities23. RQ1: What is the potential of AI-driven optimization in improving pedestrian accessibility to existing irregular form of historical streets about preserving heritage value? RQ2: How can traffic congestion be lessened in historic districts without street morphological alterations? RQ3: How can the integration of a green space be maximized in urban layouts that have a heritage constraint? RQ4: What ways may be found in which the proposed multi-objective framework is better suited to achieving the balance between these competing objectives when compared to the traditional methods? These challenges will require a complex, empirical solution that would ensure the preservation of history and, at the same time, address the contemporary urban functional requirements (Table 3).Table 3Comparative analysis of street layout optimization methods.MethodTechniques usedHeritage considerationReal-time adaptationReported improvementsLimitationsHeuristic Zoning15Manual rules, GISLowNo∼10% congestion reductionInflexible to dynamic urban changesGenetic Algorithms42MOO, ANNModerateNo∼15% walkability increaseComputationally expensive; no RLStrucTopo (GANs)21Generative designHighNoN/A (structural only)Not applied to street networksProposed FrameworkCNN+GNN+RL, Pareto MOOHighYes22.5% congestion reductionHigh GPU dependency Table 3Comparative analysis of street layout optimization methods.MethodTechniques usedHeritage considerationReal-time adaptationReported improvementsLimitationsHeuristic Zoning15Manual rules, GISLowNo∼10% congestion reductionInflexible to dynamic urban changesGenetic Algorithms42MOO, ANNModerateNo∼15% walkability increaseComputationally expensive; no RLStrucTopo (GANs)21Generative designHighNoN/A (structural only)Not applied to street networksProposed FrameworkCNN+GNN+RL, Pareto MOOHighYes22.5% congestion reductionHigh GPU dependency Comparative analysis of street layout optimization methods. Research contributions and novelty To clarify the unique contributions of this study, we summarize the novelty of the proposed framework relative to both traditional urban planning approaches and existing AI-based optimization models: Heritage-constrained optimization Unlike conventional planning tools and AI models, the proposed system integrates heritage preservation masks directly into the optimization process, ensuring that CNN, GNN, and RL modules never suggest interventions that violate cultural or architectural protection zones.Real-time multi-objective urban optimization Prior work typically relies on static datasets and offline optimization. Our framework uniquely incorporates real-time pedestrian, traffic, and environmental data using a reinforcement learning agent capable of sub-second adaptation.Hybrid deep learning architecture We introduce a combined CNN–GNN–RL pipeline that simultaneously processes spatial imagery, network topology, sustainability constraints, and dynamic urban states, an integration not present in existing literature.Sustainability centered optimization function The multi-objective model explicitly optimizes four competing goals: walkability, traffic efficiency, environmental impact, and heritage compliance. Existing systems typically optimize only one or two.Generalization across cities The framework demonstrates strong performance across multiple cities with varying street morphologies (Rome, Cairo, Kyoto, Zanzibar, Lahore), showing scalability beyond a single case study. Heritage-constrained optimization Unlike conventional planning tools and AI models, the proposed system integrates heritage preservation masks directly into the optimization process, ensuring that CNN, GNN, and RL modules never suggest interventions that violate cultural or architectural protection zones. Heritage-constrained optimization Unlike conventional planning tools and AI models, the proposed system integrates heritage preservation masks directly into the optimization process, ensuring that CNN, GNN, and RL modules never suggest interventions that violate cultural or architectural protection zones. Real-time multi-objective urban optimization Prior work typically relies on static datasets and offline optimization. Our framework uniquely incorporates real-time pedestrian, traffic, and environmental data using a reinforcement learning agent capable of sub-second adaptation. Real-time multi-objective urban optimization Prior work typically relies on static datasets and offline optimization. Our framework uniquely incorporates real-time pedestrian, traffic, and environmental data using a reinforcement learning agent capable of sub-second adaptation. Hybrid deep learning architecture We introduce a combined CNN–GNN–RL pipeline that simultaneously processes spatial imagery, network topology, sustainability constraints, and dynamic urban states, an integration not present in existing literature. Hybrid deep learning architecture We introduce a combined CNN–GNN–RL pipeline that simultaneously processes spatial imagery, network topology, sustainability constraints, and dynamic urban states, an integration not present in existing literature. Sustainability centered optimization function The multi-objective model explicitly optimizes four competing goals: walkability, traffic efficiency, environmental impact, and heritage compliance. Existing systems typically optimize only one or two. Sustainability centered optimization function The multi-objective model explicitly optimizes four competing goals: walkability, traffic efficiency, environmental impact, and heritage compliance. Existing systems typically optimize only one or two. Generalization across cities The framework demonstrates strong performance across multiple cities with varying street morphologies (Rome, Cairo, Kyoto, Zanzibar, Lahore), showing scalability beyond a single case study. Generalization across cities The framework demonstrates strong performance across multiple cities with varying street morphologies (Rome, Cairo, Kyoto, Zanzibar, Lahore), showing scalability beyond a single case study. Methodology This study introduces an AI-powered framework aimed at refining the design of historically and culturally significant urban spaces while maintaining their authenticity. By incorporating deep learning, graph-based techniques, and multi-objective optimization, the framework enhances urban livability, improves pedestrian access, and fosters environmental sustainability. Leveraging spatial data, reinforcement learning, and constraint-based optimization, it delivers adaptive solutions that harmonize modern urban demands with the need to preserve heritage. Convolutional Neural Networks are utilized to evaluate street-level characteristics and suggest expansions for pedestrian pathways. Graph Neural Networks, combined with reinforcement learning, model traffic patterns and enable real-time lane direction adjustments to reduce congestion. Green space allocation is optimized through reinforcement learning, maximizing vegetation in underused areas while respecting preservation constraints, ensuring a balanced approach to development and environmental care. Traditional urban planning methods, which often depend on heuristic zoning regulations and manual processes, frequently fail to address dynamic challenges like fluctuating pedestrian flows and environmental factors. In contrast, the proposed framework facilitates automated scheme assessment, empowering planners to make informed decisions that uphold historical integrity while supporting modern infrastructure growth. This framework promotes efficient spatial design by blending historical constraints with advanced data-driven insights, enabling sustainable, adaptive, and heritage-sensitive urban planning Fig. 2.Fig. 2Proposed architecture for optimizing historical street layouts to support dynamic urban planning. Proposed architecture for optimizing historical street layouts to support dynamic urban planning. Model description and targets The proposed model addresses the research gap through the integration of three key components, CNNs, GNNs, and RL. CNNs are used to check out road layouts from the past by spotting trends in road widths, intersections and where pedestrian zones are, which supports better access for pedestrians. GNNs capture the structural connectivity of roads and pedestrian pathways, ensuring efficient traffic flow and network integrity while preserving the historical configuration of urban streets. RL modules make decisions for real-time data, helping to adjust street layouts frequently, allowing people walking and vehicles travelling to move more smoothly. This dynamic learning process enables the model to maintain a balance between contemporary urban demands and heritage conservation. To more clearly distinguish the contribution of each learning module, we refine the design of the CNN, GNN, and RL models according to the specific requirements of historical street networks. (a) Convolutional neural network (CNN): spatial feature extraction The CNN is customized to work with satellite imagery and GIS rasters that exhibit irregular geometries typical of historic districts. Unlike CNNs trained on regular grid-based road structures, our model uses dilated convolutions to capture long-range spatial dependencies without losing fine-scale features such as narrow alleys and irregular plot boundaries. It detects heritage-sensitive elements (e.g., protected building footprints, courtyard forms, facade buffers), extracts width profiles, intersection density, pedestrian corridor continuity, and micro-block geometry. This makes the CNN particularly strong in identifying where physical interventions are feasible without violating heritage restrictions. (b) Graph neural network (GNN): connectivity and flow modeling A GNN is used to model historic district street networks, where nodes represent intersections, gates, or protected heritage sites, and edges encode travel distances, slope, width, and accessibility. The GNN is designed to assess pedestrian and vehicular connectivity in legally protected areas, learn traffic redistribution where road alterations aren’t possible, and propagate heritage penalties through message passing, ensuring interventions near protected sites automatically reduce update weights. This approach optimizes network efficiency without modifying heritage elements. (c) Reinforcement learning (RL): real-time adaptive layout adjustment RL is integrated to enable dynamic policy learning that adapts to changing pedestrian and traffic conditions. The RL agent is customized with a heritage-aware reward function, penalizing actions that violate preservation constraints. State inputs include real-time sensor data, environmental conditions, and heritage boundary heatmaps. At the same time, action sets are restricted to non-invasive interventions, such as lane direction changes, temporary pedestrianization, micro-green additions, and time-of-day regulations. This setup allows the RL module to optimize operational adjustments despite restrictions on long-term structural changes. Pedestrian accessibility optimization To enhance pedestrian accessibility, the model uses CNNs to analyze historical street layouts and detect patterns related to road width, intersections, and pedestrian zones. GNNs are employed to model the connectivity of pedestrian pathways and suggest modifications that improve flow while minimizing disruptions to the historical streetscape. The reinforcement learning component iteratively adjusts street configurations to optimize pedestrian flow. Traffic congestion management Traffic congestion is managed through a multi-objective optimization approach. The model uses traffic flow data from IoT sensors to simulate congestion hotspots. GNNs model the flow of traffic through the network, and the RL module dynamically adjusts traffic lanes and intersections based on real-time conditions. This ensures efficient traffic movement without altering the street’s heritage. Green space integration and environmental sustainability The model aims to increase green space in urban areas while maintaining historical preservation. The optimization function includes environmental impact factors, such as air quality and green space availability, ensuring that proposed modifications improve urban sustainability. The optimization also takes into account zoning regulations and heritage preservation, preventing any design that would negatively impact the cultural integrity of the area. Model design and functioning The model is structured into three main layers: CNN for spatial feature extraction, GNN for network connectivity analysis, and RL for dynamic layout optimization. Each layer works in concert to balance multiple objectives, ensuring that pedestrian accessibility, traffic efficiency, environmental sustainability, and heritage preservation are all optimized, see Fig. 3. The multi-objective optimization approach ensures that no single objective is improved at the expense of another, achieving an optimal balance.Fig. 3AI-based optimization framework. AI-based optimization framework. Data collection and processing To enhance the design of historically and culturally significant street layouts, we develop a multi-source dataset that combines spatial, environmental, and pedestrian movement data. This dataset merges various information sources to support thorough and flexible urban planning. The components of the dataset are as follows:Satellite Imagery and GIS Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{geo}$$\end{document}): High-resolution spatial information detailing street configurations, building outlines, and pedestrian routes. These datasets provide essential inputs for accurately depicting road networks and urban structures.Pedestrian and Traffic Flow Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_t$$\end{document}): Real-time movement trends collected from IoT sensors and monitoring systems. This data enables dynamic evaluation of congestion points and inefficiencies in pedestrian and vehicular movement.Historical Preservation Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}): Information on significant sites and protected zones. These inputs ensure AI-generated suggestions comply with heritage conservation requirements, preserving architectural authenticity.Environmental Constraints (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_c$$\end{document}): Data on air quality, noise levels, and green spaces. These constraints guide the optimization process to minimize environmental impact, promote green areas, and enhance urban livability.The aggregated and preprocessed dataset is denoted as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} D_{proc} = \{S_{geo}, P_t, H_p, E_c\} \end{aligned}$$\end{document}Feature engineering methods are utilized to derive key variables for AI-driven optimization. Each data subset is processed through normalization, missing data imputation, and transformation. Spatial data is converted into a graph structure, with street intersections as nodes and street connections as edges. This transformation ensures compatibility with deep learning, particularly Graph Neural Networks, while maintaining the integrity of original locations and connections. Satellite Imagery and GIS Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{geo}$$\end{document}): High-resolution spatial information detailing street configurations, building outlines, and pedestrian routes. These datasets provide essential inputs for accurately depicting road networks and urban structures. Satellite Imagery and GIS Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{geo}$$\end{document}): High-resolution spatial information detailing street configurations, building outlines, and pedestrian routes. These datasets provide essential inputs for accurately depicting road networks and urban structures. Pedestrian and Traffic Flow Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_t$$\end{document}): Real-time movement trends collected from IoT sensors and monitoring systems. This data enables dynamic evaluation of congestion points and inefficiencies in pedestrian and vehicular movement. Pedestrian and Traffic Flow Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_t$$\end{document}): Real-time movement trends collected from IoT sensors and monitoring systems. This data enables dynamic evaluation of congestion points and inefficiencies in pedestrian and vehicular movement. Historical Preservation Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}): Information on significant sites and protected zones. These inputs ensure AI-generated suggestions comply with heritage conservation requirements, preserving architectural authenticity. Historical Preservation Data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}): Information on significant sites and protected zones. These inputs ensure AI-generated suggestions comply with heritage conservation requirements, preserving architectural authenticity. Environmental Constraints (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_c$$\end{document}): Data on air quality, noise levels, and green spaces. These constraints guide the optimization process to minimize environmental impact, promote green areas, and enhance urban livability. Environmental Constraints (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_c$$\end{document}): Data on air quality, noise levels, and green spaces. These constraints guide the optimization process to minimize environmental impact, promote green areas, and enhance urban livability. Geographic diversity of datasets To address the need for broader representation and generalizability, we incorporated datasets from multiple cities spanning different geographical regions, urban morphologies, and heritage contexts. The datasets used in this study consist of: Rome, Italy (Europe) Dense medieval street network, irregular alleys, UNESCO-protected buildings, and narrow pedestrian corridors. Cairo, Egypt (North Africa) Hybrid Islamic–colonial street morphology, high pedestrian density, variable block sizes, mixed-heritage constraints. Kyoto, Japan (East Asia) Grid-like street structure overlaid with historic pedestrian pathways and strict conservation regulations. Zanzibar Stone Town (East Africa) Organic coastal street networks with extremely narrow alleys and strong façade protection rules. Old Lahore, Pakistan (South Asia) Highly irregular street geometry, high congestion load, complex informal pedestrian flows. These cities were intentionally selected because they represent five distinct historical street morphologies, each with unique spatial constraints and heritage regulatory environments. This diversity ensures robust training signals for CNN, GNN, and RL models. Cross-city generalization capability To assess the scalability of the framework, we performed a leave-one-city-out generalization test, training on four cities and testing on the fifth. The results showed 82–87% retention in walkability, 78–83% in traffic efficiency, 75–81% in environmental impact reduction, and 88–92% in heritage compliance. These findings indicate that the model effectively captures city-invariant spatial principles, including pedestrian movement, network connectivity, and heritage preservation constraints. Real-time data integration pipeline The framework optimizes historical street layouts in real-time through a four-stage pipeline: sensor collection, data preprocessing, model-based inference, and adaptive urban control. Real-time data from sources like pedestrian counters, traffic sensors, and environmental monitors is processed and fed into CNN, GNN, and RL models. The GNN updates the street network graph, and the RL agent makes decisions such as adjusting lanes or opening temporary routes based on congestion and environmental data.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_t = \left\{ \begin{array}{l} \text {GNN}\_\text {congestion}\_\text {state} \\ \text {environmental}\_\text {state} \\ \text {heritage}\_\text {mask} \\ \text {pedestrian}\_\text {density}\_\text {map} \\ \text {vehicle}\_\text {flow}\_\text {map} \\ \end{array} \right.$$\end{document}Actions are executed through the city’s ITS system, allowing for fast responses to conditions like crowd surges, with a total pipeline latency under 0.3 seconds. Spatial optimization framework We create a multi-layered AI framework that combines convolutional and graph-based learning models to enable adaptive, heritage-conscious optimization of urban layouts. Within this framework, the CNN module handles spatial feature extraction, the GNN module enhances layout connectivity, and the RL module dynamically refines configurations based on heritage preservation requirements and evolving urban demands. The framework includes the following core components:Convolutional Neural Network (FCNN): Employed to derive spatial features from satellite and GIS data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{geo}}$$\end{document}). CNNs examine historical street layouts to identify structural patterns, such as road widths, pedestrian areas, and traffic intersections.Graph Neural Networks (FGNN): Utilized to model road networks and pedestrian pathways. GNNs support efficient connectivity analysis, aiding in the design of street configurations that improve accessibility and alleviate traffic congestion.Reinforcement Learning (FRL): Enables adaptive decision-making for layout adjustments. The RL module tests various planning scenarios and iteratively optimizes configurations to improve performance across multiple urban planning goals.The AI-driven objective function is formulated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O_{\textrm{AI}} = \lambda _1 F_{\textrm{CNN}}(S_{\textrm{geo}}) + \lambda _2 F_{\textrm{GNN}}(P_t) + \lambda _3 F_{\textrm{RL}}(H_p)$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _3$$\end{document} are weighting factors that ensure balanced contributions from each module toward the overall optimization goal. Convolutional Neural Network (FCNN): Employed to derive spatial features from satellite and GIS data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{geo}}$$\end{document}). CNNs examine historical street layouts to identify structural patterns, such as road widths, pedestrian areas, and traffic intersections. Convolutional Neural Network (FCNN): Employed to derive spatial features from satellite and GIS data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{geo}}$$\end{document}). CNNs examine historical street layouts to identify structural patterns, such as road widths, pedestrian areas, and traffic intersections. Graph Neural Networks (FGNN): Utilized to model road networks and pedestrian pathways. GNNs support efficient connectivity analysis, aiding in the design of street configurations that improve accessibility and alleviate traffic congestion. Graph Neural Networks (FGNN): Utilized to model road networks and pedestrian pathways. GNNs support efficient connectivity analysis, aiding in the design of street configurations that improve accessibility and alleviate traffic congestion. Reinforcement Learning (FRL): Enables adaptive decision-making for layout adjustments. The RL module tests various planning scenarios and iteratively optimizes configurations to improve performance across multiple urban planning goals. Reinforcement Learning (FRL): Enables adaptive decision-making for layout adjustments. The RL module tests various planning scenarios and iteratively optimizes configurations to improve performance across multiple urban planning goals. Multi-objective optimization strategy The optimization of historical and cultural street layouts necessitates balancing a number of frequently contradictory goals, including pedestrian accessibility, traffic efficiency, environmental sustainability, and heritage preservation. To handle this, a multi-objective optimization function is defined as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \max O = \alpha _1 A_p + \alpha _2 T_e - \alpha _3 E_i + \alpha _4 H_p \end{aligned}$$\end{document}Where:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_p$$\end{document}: Pedestrian Accessibility - Maximizing walkability and ease of pedestrian movement.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_e$$\end{document}: Traffic Efficiency - Reducing vehicular congestion and optimizing flow.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document}: Environmental Impact - Minimizing pollution and enhancing green space integration.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}: Heritage Preservation - Ensuring protection of historically significant structures.The weighting parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\alpha }_1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\alpha }_2$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\alpha }_3$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\alpha }_4$$\end{document} are dynamically adjusted according to specific urban planning priorities. To derive an optimal solution, Pareto optimization is employed, ensuring that improvements in one objective do not cause unacceptable trade-offs in others. The optimal set of solutions, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{S}_{\textrm{opt}}$$\end{document}, is given by:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{S}_{\textrm{opt}} = \arg \max \limits _{S} \left( \lambda _1 A_p + \lambda _2 T_e - \lambda _3 E_i + \lambda _4 H_p \right) \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _i$$\end{document} represents the priority assigned to each corresponding objective. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_p$$\end{document}: Pedestrian Accessibility - Maximizing walkability and ease of pedestrian movement. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_p$$\end{document}: Pedestrian Accessibility - Maximizing walkability and ease of pedestrian movement. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_e$$\end{document}: Traffic Efficiency - Reducing vehicular congestion and optimizing flow. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_e$$\end{document}: Traffic Efficiency - Reducing vehicular congestion and optimizing flow. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document}: Environmental Impact - Minimizing pollution and enhancing green space integration. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document}: Environmental Impact - Minimizing pollution and enhancing green space integration. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}: Heritage Preservation - Ensuring protection of historically significant structures. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}: Heritage Preservation - Ensuring protection of historically significant structures. Constraint-based layout generation Given the uneven character of ancient street layouts, the optimal spatial design must adhere to a variety of legal and functional requirements. The restricted optimization issue is stated as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{\textrm{opt}} = \arg \max _{O} \ \ \text {subject to} \ C$$\end{document}Where C represents the set of constraints:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C = \{C_{\textrm{zoning}}, C_{\textrm{heritage}}, C_{\textrm{traffic}}, C_{\textrm{green}}\}$$\end{document}Define the following components for the street layout optimization:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C_{\textrm{zoning}}&: \text {Adherence to zoning regulations to preserve historical significance.} \\ C_{\textrm{heritage}}&: \text {Constraints to minimize impact on culturally significant sites.} \\ C_{\textrm{traffic}}&: \text {Optimization of traffic lanes and pedestrian pathways to reduce congestion.} \\ C_{\textrm{green}}&: \text {Preservation and enhancement of green spaces within urban design.} \end{aligned}$$\end{document}To satisfy these constraints while maximizing layout efficiency, an Evolutionary Algorithm (EA) is employed. The EA iteratively evolves feasible street layouts based on a fitness function defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{Fitness} = f(O) - \sum _{i} w_i \delta (C_i)$$\end{document}Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta (C_i)$$\end{document} represents the penalty incurred for violating constraint \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_i$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_i$$\end{document} are the corresponding weights that govern the trade-off between constraint feasibility and the objective maximization. To assess the robustness and performance of the proposed layouts in real-world scenarios, Monte Carlo simulations are performed over multiple iterations of the generated layouts. Experiment setup and performance metrics The optimization framework is implemented in Python, utilizing TensorFlow and Porch for deep learning, NetworkX for graph-based spatial modeling, and SciPy for multi-objective optimization. The training of the models takes place on a high-performance computer framework that works on an NVIDIA RTX 3090 GPU, providing 24 GB of VRAM and an Intel Core i9-processor with support for 64 GB of RAM. This dataset combines several sources, such as satellite images, records of urban planning based on GIS, pedestrian flows collected due to IoT sensors, and the databases containing historical sites. The spatial normalization, missing value imputation, and feature extraction through graphs are associated with the process of data preprocessing, contributing to effective spatial learning. To assess the performance of the model, four essential performance indicators (KPIs) are identified, including walkability improvement, traffic efficiency, energy efficiency, and adherence to the requirements of heritage preservation. Algorithm 1: AI model training for spatial optimization Training an AI model for optimizing historical street layouts involves several learning techniques, including Convolutional Neural Networks, Graph Neural Networks, and Reinforcement Learning. CNNs leverage high-quality satellite images and GIS data to identify roadways, footpaths, and green spaces. GNNs boost the efficiency of the urban grid by controlling the relations among various zones within the city. In the meantime, test results of the RL model cover the laying out of new streets and sidewalks with such results that people, traffic, and historical preservation are in balance. The training process also consists of data normalization, the replacement of missing values, and constant updates applied to the AI models. Optimization of the model is provided by gradient descent optimization per component to be able to capture correct spatial components and well-behaved compositions. Algorithm 1AI model training for spatial optimization. AI model training for spatial optimization. Algorithm 2: AI-driven layout optimization process The layout optimization process leverages trained AI models to refine historical street configurations incrementally. Initially, CNNs and Graph Neural Networks GNNs establish important spatial characteristics, and RL creates new arrangements where the composition of functionality and existing infrastructure feels balanced. In urban planning, multi-objective functions are taken into account where the access of pedestrians, vehicle movement, and sustenance are addressed. When adjusting the city environment, architects should take into account the zoning laws as well as laws protecting historical buildings and nature. It is precisely evolutionary principles that are applied to this process, i.e., choosing and altering different layout possibilities. Retention of layouts is allowed in situations where they enhance efficiency without compromising the stability of the existing structures. This is an iterative process by which the model can be better fit to the peculiarities of various cities and changing needs in urban planning. Algorithm 2AI-driven layout optimization process. AI-driven layout optimization process. Performance metrics All performance indicators are defined as composite indices in the range [0, 100], constructed from normalized spatial and operational variables. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {norm}(\cdot )$$\end{document} denote a min–max normalization to [0, 1], applied over all scenarios and case-study zones. The weighting parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _w, \alpha _c, \alpha _d, \alpha _p$$\end{document} were selected based on established urban design principles and empirical calibration across case-study cities. Slightly higher weight was assigned to effective pedestrian width and network continuity, as these factors are consistently identified in walkability research as primary determinants of pedestrian comfort and accessibility. The detour ratio and pedestrian-priority share were assigned moderate weights, reflecting their secondary but still significant influence on perceived walkability. All weights are normalized such that\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _w + \alpha _c + \alpha _d + \alpha _p = 1,$$\end{document}and can be adjusted to reflect local planning priorities. A sensitivity analysis (Section 5.4) confirms that moderate perturbations (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10\%$$\end{document}) in these components do not alter the relative superiority of the optimized layouts, demonstrating robustness of the metric design. Walkability score The Walkability Score W is designed to capture the effective pedestrian experience along the network, combining width, continuity, crossings, and pedestrian-priority length. For each street segment i, we compute:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_i: \text {effective pedestrian width (m)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_i: \text {connectivity factor (e.g. normalized closeness centrality)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i: \text {average pedestrian detour ratio (actual path length / straight-line distance)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i: \text {share of segment designated as pedestrian-priority (0--1)}$$\end{document}We first derive normalized indicators:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{w}_i = \text {norm}(w_i), \quad \tilde{c}_i = \text {norm}(c_i), \quad \tilde{d}_i = \text {norm}(d_i), \quad \tilde{p}_i = p_i$$\end{document}The walkability score for the network is then:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W = 100 \cdot \left( \alpha _w \tilde{w} + \alpha _c \tilde{c} + \alpha _d \left( 1 - \tilde{d}\right) + \alpha _p \tilde{p} \right) ,$$\end{document}where overbars denote averages over all segments, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _w, \alpha _c, \alpha _d, \alpha _p \ge 0$$\end{document} are weights summing to 1 (chosen to give slightly higher importance to width and continuity). The reported change from 50 → 85 corresponds to an increase in average effective width (1.4–2.2 m → 2.8–3.2 m), improved network continuity (fewer forced detours), and a larger share of pedestrian-priority length. Traffic efficiency Traffic Efficiency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_e$$\end{document} measures how efficiently vehicles traverse the network, combining average travel time, queue length, and intersection delay. From microsimulation, we extract:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau : \text {average travel time per vehicle (s)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q: \text {average queue length at critical nodes (m)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta : \text {average intersection delay (s/veh)}$$\end{document}These are normalized as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\tau } = \text {norm}(\tau ), \quad \tilde{q} = \text {norm}(q), \quad \tilde{\delta } = \text {norm}(\delta ).$$\end{document}Traffic efficiency is then defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_e = 100 \cdot \left( \beta _\tau \left( 1 - \tilde{\tau }\right) + \beta _q \left( 1 - \tilde{q}\right) + \beta _\delta \left( 1 - \tilde{\delta }\right) \right) ,$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _\tau , \beta _q, \beta _\delta \ge 0$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _\tau + \beta _q + \beta _\delta = 1$$\end{document}. Thus, an increase from 60 → 80 reflects lower average travel times, shorter queues (e.g., 185 m → 105 m at Node A), and reduced intersection delays after optimization. The weighting factors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{\tau }, \beta _{q}, \beta _{\delta }$$\end{document} were determined according to transportation engineering standards and microsimulation outcomes. Travel time was assigned the highest relative weight, as it directly reflects user-level efficiency and system-wide performance. Queue length and intersection delay were weighted slightly lower but remain critical indicators of localized congestion and operational bottlenecks. These weights were empirically calibrated using cross-city validation experiments to ensure that no single variable disproportionately influenced the composite index. The sensitivity analysis presented in Section 5.4 further demonstrates that performance improvements remain stable under reasonable weight perturbations. Heritage preservation compliance Heritage Preservation Compliance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document} quantifies the extent to which proposed layouts respect legally protected heritage assets. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j = 1, \dots , N_h$$\end{document} index all heritage elements (buildings, façades, protected viewsheds, and protected public spaces). For each element j, we define a binary indicator:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_j = {\left\{ \begin{array}{ll} 1, & \text {if no geometric or regulatory constraint is violated;}\\ 0, & \text {otherwise.}\\ \end{array}\right. }$$\end{document}Violations are detected automatically via GIS overlay between: The proposed layout (streets, pedestrian zones, green insertions), and the heritage constraint layer (building footprints, buffer zones, protected lines of sight). The compliance index is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p = 100 \cdot \frac{1}{N_h} \sum _{j=1}^{N_h} h_j.$$\end{document}In addition, “soft” infractions—such as minor encroachments into visual buffers—are penalized by adjusting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_j$$\end{document} to values in (0, 1) based on pre-defined planning rules. The reported improvement from 65 → 90 indicates a reduction in such infractions, achieved by the RL agent learning to avoid interventions in proximity to protected structures. Sensitivity analysis of metric components To assess the robustness of the metrics, we conducted a local sensitivity analysis on the Cairo case study by perturbing key design variables by ±10% around the optimized solution. The resulting percentage change in each metric is summarized as follows:Walkability (W) showed a +7.5% increase and a -8.3% decrease with changes in effective pedestrian width, +6.1% / -6.7% with pedestrian-priority links, +3.4% / -3.9% with marked crossings, and +1.8% / -1.6% with vehicle lane width.Traffic Efficiency (T\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_e$$\end{document}) changed by +1.2% / -1.5% for pedestrian width, +2.3% / -2.8% for pedestrian-priority links, +2.7% / -3.1% for crossings, +5.9% / -6.4% for lane width, and +1.4% / -1.7% for green-space fraction.Heritage Compliance (H\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_p$$\end{document}) remained unaffected by pedestrian width, pedestrian-priority links, crossings, lane width, and green-space fraction, but showed a +4.6% / -5.1% change with adjustments to the size of heritage buffer zones. Walkability (W) showed a +7.5% increase and a -8.3% decrease with changes in effective pedestrian width, +6.1% / -6.7% with pedestrian-priority links, +3.4% / -3.9% with marked crossings, and +1.8% / -1.6% with vehicle lane width. Walkability (W) showed a +7.5% increase and a -8.3% decrease with changes in effective pedestrian width, +6.1% / -6.7% with pedestrian-priority links, +3.4% / -3.9% with marked crossings, and +1.8% / -1.6% with vehicle lane width. Traffic Efficiency (T\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_e$$\end{document}) changed by +1.2% / -1.5% for pedestrian width, +2.3% / -2.8% for pedestrian-priority links, +2.7% / -3.1% for crossings, +5.9% / -6.4% for lane width, and +1.4% / -1.7% for green-space fraction. Traffic Efficiency (T\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_e$$\end{document}) changed by +1.2% / -1.5% for pedestrian width, +2.3% / -2.8% for pedestrian-priority links, +2.7% / -3.1% for crossings, +5.9% / -6.4% for lane width, and +1.4% / -1.7% for green-space fraction. Heritage Compliance (H\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_p$$\end{document}) remained unaffected by pedestrian width, pedestrian-priority links, crossings, lane width, and green-space fraction, but showed a +4.6% / -5.1% change with adjustments to the size of heritage buffer zones. Heritage Compliance (H\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_p$$\end{document}) remained unaffected by pedestrian width, pedestrian-priority links, crossings, lane width, and green-space fraction, but showed a +4.6% / -5.1% change with adjustments to the size of heritage buffer zones. Evaluation and results Quantitative performance evaluation A comparison between baseline urban planning methods and AI-optimized solutions reveals notable improvements, as shown in Table 4. The AI-optimized layouts resulted in a 70% increase in walkability, a 33.3% improvement in traffic efficiency, a 36.4% gain in energy efficiency, and a 38.5% enhancement in heritage preservation compliance. Each of these metrics demonstrates significant improvements, with strong 95% confidence intervals supporting the findings.Table 4Quantitative performance comparison of baseline urban planning methods and AI-optimized model with confidence intervals.Performance metricBaselineAI-optimizedImprovement (%)95% CIWalkability Score5085+70%[81.2, 88.8]Traffic Efficiency6080+33.3%[76.5, 83.5]Energy Efficiency5575+36.4%[71.8, 78.2]Heritage Preservation Compliance6590+38.5%[87.3, 92.7] Table 4Quantitative performance comparison of baseline urban planning methods and AI-optimized model with confidence intervals.Performance metricBaselineAI-optimizedImprovement (%)95% CIWalkability Score5085+70%[81.2, 88.8]Traffic Efficiency6080+33.3%[76.5, 83.5]Energy Efficiency5575+36.4%[71.8, 78.2]Heritage Preservation Compliance6590+38.5%[87.3, 92.7] Quantitative performance comparison of baseline urban planning methods and AI-optimized model with confidence intervals. Comparison of performance metrics To assess the effectiveness of AI-driven spatial optimization, we compared four key performance metrics: Walkability, Traffic Efficiency, Energy Efficiency, and Heritage Preservation. The baseline approach represents traditional urban planning methods, while the AI-optimized approach utilizes CNN, GNN, and RL models for spatial analysis and decision-making. As shown in Table 4, the AI-optimized model exhibits statistically significant improvements over the baseline methods in all performance indicators. The results, presented with 95% confidence intervals (CIs) to quantify uncertainty, are as follows: Walkability increased from 50% to 85%, Traffic Efficiency rose from 60% to 80%, Energy Efficiency improved from 55% to 75%, and Heritage Preservation compliance increased from 65% to 90%. Figure 4 visually demonstrates these substantial improvements across all metrics. These findings indicate that AI-driven methods can significantly enhance urban livability while maintaining the integrity of heritage sites.Fig. 4Comparison of baseline urban planning methods and AI-optimized layouts across four key performance metrics: walkability, traffic efficiency, energy efficiency, and heritage preservation compliance. The figure demonstrates the multi-objective improvement achieved by the proposed framework, highlighting its ability to enhance urban functionality while maintaining heritage-sensitive constraints. Comparison of baseline urban planning methods and AI-optimized layouts across four key performance metrics: walkability, traffic efficiency, energy efficiency, and heritage preservation compliance. The figure demonstrates the multi-objective improvement achieved by the proposed framework, highlighting its ability to enhance urban functionality while maintaining heritage-sensitive constraints. Model component efficacy assessment To evaluate the contribution of individual AI components, assess their effectiveness in optimizing street layouts. The CNN model that concentrated on feature extraction based on spatial features was accurate by 78% with an average processing time of 120 ms. Lastly, another GNN model that modeled vehicle and pedestrian flow achieved 85 percent accuracy, which was impressive compared to other models in the spatial connectivity learning. In the meantime, the RL model that was supposed to optimise adaptive decision-making had an accuracy of 82%, but it was the most expensive to calculate at 150 ms. Of all of these models, GNN is the most appropriate in terms of optimizing a complex network since it requires the least tradeoff between accuracy and speed of processing. The results are reflected in Fig. 5.Fig. 5Effectiveness of individual AI models in optimizing street layouts, showing accuracy and processing time for CNN, GNN, and RL components. Effectiveness of individual AI models in optimizing street layouts, showing accuracy and processing time for CNN, GNN, and RL components. Optimization convergence analysis Iterative objective function enhancement The multi-objective function, which balances pedestrian accessibility, traffic efficiency, energy impact, and heritage preservation, is optimized through multiple iterations. As shown in Fig. 6, a logarithmic improvement is observed, demonstrating that the optimization model rapidly converges to an optimal solution.Fig. 6Enhancement of the multi-objective function (pedestrian accessibility, traffic efficiency, environmental impact, heritage preservation) across several iterations. Enhancement of the multi-objective function (pedestrian accessibility, traffic efficiency, environmental impact, heritage preservation) across several iterations. Convergence speed of evolutionary optimization To evaluate the efficiency of the optimization process, we analyzed the convergence speed of the evolutionary algorithm. As illustrated in Fig. 7, the optimization process shows a significant acceleration after 25 iterations, following a sigmoid-like pattern.Fig. 7Convergence speed of the evolutionary optimization process after 25 iterations, illustrating rapid improvements in layout optimization. Convergence speed of the evolutionary optimization process after 25 iterations, illustrating rapid improvements in layout optimization. Quantitative results comparison The results presented in Table 5 demonstrate that AI-driven optimization provides significant improvements over traditional urban planning methods, emphasizing its effectiveness in both preserving heritage and enhancing urban efficiency.Table 5Quantitative comparison of AI-optimized vs. bseline urban planning.Performance metricBaselineAI-optimizedImprovement (%)Walkability Score5085+70%Traffic Efficiency6080+33.3%Energy Efficiency5575+36.4%Heritage Preservation Compliance6590+38.5% Table 5Quantitative comparison of AI-optimized vs. bseline urban planning.Performance metricBaselineAI-optimizedImprovement (%)Walkability Score5085+70%Traffic Efficiency6080+33.3%Energy Efficiency5575+36.4%Heritage Preservation Compliance6590+38.5% Quantitative comparison of AI-optimized vs. bseline urban planning. Comparative analysis with traditional methods Pedestrian flow analysis The enhancement of the pedestrian traffic was analyzed in five areas of the city. The conventional city designs tend to lead to congestion because of poor planning in the streets. But, the AI-based optimal model increased the pedestrian accessibility to a great extent, which resulted in better traffic flow and walkability. The optimized model raised the pedestrian entrance capacities in all the zones, as revealed in Fig. 8. Zone A up to 500 pedestrians, Zone B 600, Zone C 700, Zone D 550, and Zone E 750. These findings indicate a considerable rise in the efficiency of the pedestrian flow.Fig. 8Improvement in pedestrian flow across urban areas before and after optimization, demonstrating increased pedestrian capacity in each zone. Improvement in pedestrian flow across urban areas before and after optimization, demonstrating increased pedestrian capacity in each zone. Traffic congestion reduction Analysis of traffic congestion was carried out at four major time slots, including morning, noon, evening, and night. Before optimization, the congestion level was a high of 90% during the evening peak hours, leading to huge delays. The congestion was alleviated by about 40% in all periods following the AI-based optimization. The AI-based model led to significant reductions in the occurrence of traffic congestion: the percentage of the morning, noon, evening, and night congestion fell by 50%, 40%, 55%, and 25%, respectively, see Fig. 9.Fig. 9Reduction in traffic congestion during peak hours before and after optimization. Reduction in traffic congestion during peak hours before and after optimization. Environmental and sustainability impact Green space utilization Green space utilization was improved across all urban zones due to enhanced land management strategies. As shown in Fig. 10, before optimization, green space coverage ranged from 20% to 40%. Following AI-driven optimization, this coverage increased to between 40% and 60%, leading to more sustainable and efficient urban layouts. The use of green space was enhanced in all the zones of the urban area because of better land management practices. Pre-optimization, the green space coverage levels were between 20% and 40% as indicated in Fig. 10. This level of coverage rose to between 40% and 60% after AI-based optimization, resulting in more sustainable and efficient urban plans.Fig. 10Enhancement in green space coverage across urban areas before and after optimization, highlighting a notable increase in sustainable urban designs. Enhancement in green space coverage across urban areas before and after optimization, highlighting a notable increase in sustainable urban designs. Environmental impact reduction The main pollutants to be monitored were the measurement of air quality in terms of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CO_2$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$NO_2$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PM_{2.5}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SO_2$$\end{document}. The optimization model based on AI reduced emissions substantially by improving traffic control and creating more green space. Based on Fig. 11, the level of the pollutants has been reduced significantly.Fig. 11Decrease in environmental pollutants resulting from optimized street layouts, showcasing enhanced air quality in urban areas. Decrease in environmental pollutants resulting from optimized street layouts, showcasing enhanced air quality in urban areas. Table 6 further illustrates these improvements, showing a 40% reduction in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CO_2$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$NO_2$$\end{document} emissions, a 38.9% decrease in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PM_{2.5}$$\end{document} levels, and a 40% drop in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SO_2$$\end{document} emissions. These results emphasise the model’s effectiveness in improving air quality and mitigating pollution.Table 6Comparison of air pollution levels prior to and following optimisation efforts.PollutantBaseline emission (μg/m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^3$$\end{document})Optimized emission (μg/m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^3$$\end{document})Reduction (%)CO220012040%NO21509040%PM2.518011038.9%SO21006040% Table 6Comparison of air pollution levels prior to and following optimisation efforts.PollutantBaseline emission (μg/m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^3$$\end{document})Optimized emission (μg/m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^3$$\end{document})Reduction (%)CO220012040%NO21509040%PM2.518011038.9%SO21006040% Comparison of air pollution levels prior to and following optimisation efforts. Comparative analysis of AI-driven urban optimization The proposed AI-driven framework for urban layout optimization offers substantial advantages over traditional planning methods across key performance metrics, including pedestrian accessibility, traffic efficiency, green space utilization, and environmental sustainability. It supplements the connectivity among urban areas without losing the historical value of the city by incorporating CNNs, GNNs, and RL, achieving an average improvement of 35%-70% over all the performance indicators of the urban environment. Uneven application of traditional measures, like heuristic zoning and pedestrianization, mostly fail to serve well when it comes to older streets. Such approaches are quite likely to be aimed at solving one particular problem, which can be the alleviation of traffic congestion with the lack of consideration of the interdependence of the pedestrian environment, vehicles, heritage preservation, and environmental consequences. Heuristic zoning typically results in only about a 10% reduction in congestion, as it cannot adapt to fluctuations in urban zones. Grammar-based approaches reduce congestion by 25%, while genetic algorithms achieve an 18% decrease, compared to the 40% improvement achieved by the proposed method. The earlier MOO designs17 can only enhance walkability by 15%, and they are not real-time adaptable because of that, which is the weak point in terms of managing the dynamic urban environments. In contrast to earlier methods17–23, our approach leverages multi-objective optimization to simultaneously balance pedestrian and vehicular traffic while minimizing emissions. The AI-driven optimization function is formulated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O = \alpha _1 A_p + \alpha _2 T_e - \alpha _3 E_i + \alpha _4 H_p$$\end{document}Where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_p$$\end{document}: Walkability (pedestrian accessibility), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_e$$\end{document}: Traffic efficiency, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document}: Environmental impact, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_p$$\end{document}: Heritage compliance Balance the need of a stable convergence against the ability to accommodate real-world constraints on planning, the model employs a constraint-based evolutionary algorithm. Our framework has a 31.8% increase in Confidence Interval: [87.3, 92.7] in pedestrian accessibility compared to StrucTopo24, which does not account for any heritage constraints (including constrained pedestrian accessibility through the layout generation). Conventional multi-objective optimization systems, e.g., those in17, allow 15% improvement in walkability, and cannot be updated in real-time. Our Reinforcement Learning module, in turn, will enable us to react dynamically to the changing congestion patterns, as can be seen in Fig. 7. Although those techniques outlined in17,42 prove the potential of multi-objective optimization in the urban design context, their rather static requirements are not able to scale up to dynamic, historically sensitive districts, as they require the use of predefined datasets. Our framework addresses this shortcoming, as it incorporates the IoT information feed into the RL sub-module, as shown in Fig. 8. The comparative analysis of optimization models is presented in Table 7 with the focus on the better multi-objective outcomes and flexibility of the artificial intelligence-based method. The proposed AI-driven framework significantly outperforms traditional methods across four key objectives: pedestrian accessibility (70% improvement), traffic efficiency (33.3%), energy efficiency (36.4%), and heritage preservation compliance (38.5%). Additionally, CO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_2$$\end{document}, NO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_2$$\end{document}, PM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_2.5$$\end{document}, and SO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_2$$\end{document} emissions are reduced by up to 40%. In comparison, heuristic zoning and static GIS designs achieve modest results (10–15% congestion reduction), while genetic-algorithm models improve walkability by 35% but lack real-time responsiveness and heritage constraints. The CNN+GNN+RL framework optimizes congestion, walkability, heritage, and environmental quality, achieving up to 22.5% congestion reduction and 31.8% improvement in pedestrian accessibility within heritage constraints. However, the system has limitations: it is GPU-dependent, requires dense IoT infrastructure for real-time capabilities, and has lower interpretability than rule-based tools. Thus, it should complement, not replace, conventional planning approaches, especially in early adoption stages.Table 7Comparison of the performance of various urban optimization models.ModelCongestion reductionWalkability improvementHeritage complianceGreen space integrationReal-time adaptationHardware requirementsLimitationsHeuristic Zoning1510–15%20–30%Low10%XCPU-onlyInflexibleGenetic Algorithms4218%35%Moderate15%XCPU-onlySlow (48h)StrucTopo (GANs)21N/A40%HighN/AX4x GPUsNo traffic opt.CVAE Bridges23N/A15%HighN/AX2x GPUsNarrow scopeProposed (CNN+GNN+RL)22.5%31.8%High27.4%\checkmark1x GPUGPU-dependent Table 7Comparison of the performance of various urban optimization models.ModelCongestion reductionWalkability improvementHeritage complianceGreen space integrationReal-time adaptationHardware requirementsLimitationsHeuristic Zoning1510–15%20–30%Low10%XCPU-onlyInflexibleGenetic Algorithms4218%35%Moderate15%XCPU-onlySlow (48h)StrucTopo (GANs)21N/A40%HighN/AX4x GPUsNo traffic opt.CVAE Bridges23N/A15%HighN/AX2x GPUsNarrow scopeProposed (CNN+GNN+RL)22.5%31.8%High27.4%\checkmark1x GPUGPU-dependent Comparison of the performance of various urban optimization models. The results confirm a 40% reduction in congestion, a 50% increase in green space efficiency, and a 38.9% decrease in air pollution. Future improvements will concentrate on integrating real-time IoT-based urban analytics for adaptive AI-driven planning. Spatial visualization and layout reconstruction As illustrated in Fig. 12, the baseline configuration of the Old Cairo Heritage District is characterized by narrow pedestrian alleys (1.4–2.2 m), irregular intersections, and several congestion hotspots concentrated around Nodes A, C, and F. Vehicular circulation follows a fixed unidirectional flow, with no dedicated pedestrian-priority segments and minimal green-space coverage (<12% of usable area).Fig. 12Shows the original street layout with narrow alleys, irregular intersections, and congestion hotspots, with no pedestrian zones and minimal green space. Shows the original street layout with narrow alleys, irregular intersections, and congestion hotspots, with no pedestrian zones and minimal green space. Figure 13 shows the optimized street layout produced by the CNN–GNN–RL framework, highlighting two newly established pedestrian corridors, localized lane-direction adjustments at Nodes A and D, and the insertion of micro-green buffers at Nodes B, E, and F. The resulting walkway widths increase to 2.8–3.2 m through road-space reallocation rather than structural widening, ensuring full compliance with heritage constraints.Fig. 13Illustrates the improved layout after AI optimization, showing new pedestrian corridors, lane-direction changes, and added micro-green buffers, all within heritage limits. Illustrates the improved layout after AI optimization, showing new pedestrian corridors, lane-direction changes, and added micro-green buffers, all within heritage limits. The performance shift is visualized in Fig. 14, where pedestrian throughput at Node A rises from 420 to 630 persons per minute, and peak-period vehicular queue lengths decline from 185 m to 105 m. These improvements occur solely through network reconfiguration and operational adjustments, with no changes to building footprints or protected urban fabric.Fig. 14Compares conditions before and after optimization, showing higher pedestrian flow and shorter vehicle queues without altering any building structures. Compares conditions before and after optimization, showing higher pedestrian flow and shorter vehicle queues without altering any building structures. As detailed in the cross-sectional schematic in Fig. 15, the optimization strategy reallocates a 4.5 m bidirectional lane into a 3.0 m single-direction lane and a 1.5 m dedicated pedestrian walkway, supplemented by 0.4 m micro-green buffers wherever feasible. This section demonstrates how meaningful spatial gains are achieved through roadway redistribution while preserving the heritage boundaries of adjacent structures.Fig. 15Presents how roadway space was reallocated—reducing lane width and expanding pedestrian space while adding small green buffers, with no changes to heritage façades. Presents how roadway space was reallocated—reducing lane width and expanding pedestrian space while adding small green buffers, with no changes to heritage façades. Spatial interventions and their quantified effects The improvements reported in Table 8 stem directly from these physical and operational modifications. The 31.8% increase in pedestrian accessibility is primarily due to widening walkable space by reallocating 1.0–1.4 m of roadway (without structural changes). Similarly, the 22.5% congestion reduction arises from one-way conversions, narrower vehicle lanes, and removal of conflicting turning flows. Green-space gains result solely from micro-buffer insertions in formerly unused verge segments. Importantly, all interventions comply with heritage-protection boundaries and do not involve demolition, façade alteration, or street widening beyond existing right-of-way.Table 8Design interventions and their impacts.Design interventionConcrete spatial changeMeasured or expected outcome1. Walkway expansion via road-space reallocationWalkway width increased from 1.4–2.2 m → 2.8–3.2 m along segments S1, S2, S3. No building walls altered.+31.8% pedestrian accessibility; smoother bi-directional flows; reduced bottlenecks at Nodes A and C.2. Vehicle lane reductionMain corridor lane reduced from 4.5 m → 3.0 m (converted to one-way operation).-22.5% road congestion; queue length at Node A reduced from 185 m → 105 m.3. RL-driven dynamic lane-direction controlNodes A and D switched from static bidirectional operation to context-adaptive single-direction flow.Eliminated cross-conflicts and achieved 33.3% efficiency gain in vehicle throughput.4. Creation of continuous pedestrian corridorsTwo uninterrupted pedestrian routes established (Nodes B→C and D→F).+70% improvement in walkability index; major reduction in pedestrian–vehicle interference.5. Addition of marked pedestrian crossingsNew crossings added at Nodes A, D, and F; widths standardized to 2.5 m.Improved network connectivity, reducing pedestrian detour distances by 18–24%.6. Micro-green buffers0.4 m green strips added along S3 and at Nodes B, E, F.27.4% rise in effective green space; marginal reduction in PM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{2.5}$$\end{document} (-38.9%).7. Time-window pedestrianizationSegment S2 pedestrian-only during peak hours (4 pm–9 pm).Short-term congestion drops of 40–43% during active window; smoother retail footfall distribution.8. Junction reconfiguration (conflict removal)Turning radius reduced from 9 m → 6 m at Node D; simplification of conflict points.Faster clearance, 15–20% reduction in turning delays. Table 8Design interventions and their impacts.Design interventionConcrete spatial changeMeasured or expected outcome1. Walkway expansion via road-space reallocationWalkway width increased from 1.4–2.2 m → 2.8–3.2 m along segments S1, S2, S3. No building walls altered.+31.8% pedestrian accessibility; smoother bi-directional flows; reduced bottlenecks at Nodes A and C.2. Vehicle lane reductionMain corridor lane reduced from 4.5 m → 3.0 m (converted to one-way operation).-22.5% road congestion; queue length at Node A reduced from 185 m → 105 m.3. RL-driven dynamic lane-direction controlNodes A and D switched from static bidirectional operation to context-adaptive single-direction flow.Eliminated cross-conflicts and achieved 33.3% efficiency gain in vehicle throughput.4. Creation of continuous pedestrian corridorsTwo uninterrupted pedestrian routes established (Nodes B→C and D→F).+70% improvement in walkability index; major reduction in pedestrian–vehicle interference.5. Addition of marked pedestrian crossingsNew crossings added at Nodes A, D, and F; widths standardized to 2.5 m.Improved network connectivity, reducing pedestrian detour distances by 18–24%.6. Micro-green buffers0.4 m green strips added along S3 and at Nodes B, E, F.27.4% rise in effective green space; marginal reduction in PM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{2.5}$$\end{document} (-38.9%).7. Time-window pedestrianizationSegment S2 pedestrian-only during peak hours (4 pm–9 pm).Short-term congestion drops of 40–43% during active window; smoother retail footfall distribution.8. Junction reconfiguration (conflict removal)Turning radius reduced from 9 m → 6 m at Node D; simplification of conflict points.Faster clearance, 15–20% reduction in turning delays. Design interventions and their impacts. Case studies and real-world validation To ensure that results are verifiable and not based on abstract or hypothetical layouts, we conducted two detailed real-world case studies using internationally recognized heritage streets with documented constraints, available GIS data, and consistent urban morphology. Case study A: El-Muizz Street, Historic Cairo (UNESCO Heritage Zone) A1: Current configuration Street type Medieval pedestrian spine with commercial frontage Length 1.1 km Widths Primary walkway: 1.8–2.3 mSecondary side alleys: 1.2–1.9 m Primary walkway: 1.8–2.3 m Primary walkway: 1.8–2.3 m Secondary side alleys: 1.2–1.9 m Secondary side alleys: 1.2–1.9 m Heritage constraints: Zero modification to building façadesNo excavation or foundation alterationNo removal of historical pavementsMandatory visual corridor preservation Zero modification to building façades Zero modification to building façades No excavation or foundation alteration No excavation or foundation alteration No removal of historical pavements No removal of historical pavements Mandatory visual corridor preservation Mandatory visual corridor preservation Key problems: Severe pedestrian congestion around Qalawun Complex and Bab ZuweilaNo continuous pedestrian pathZero green micro-infrastructurePoor flow continuity at 5 conflict nodes Severe pedestrian congestion around Qalawun Complex and Bab Zuweila Severe pedestrian congestion around Qalawun Complex and Bab Zuweila No continuous pedestrian path No continuous pedestrian path Zero green micro-infrastructure Zero green micro-infrastructure Poor flow continuity at 5 conflict nodes Poor flow continuity at 5 conflict nodes A2: Proposed optimizations Using our CNN–GNN–RL optimization framework:Walkway width increased from 1.8–2.3 m → 3.0–3.2 m (reallocating space, no widening)Continuous pedestrian corridor created from Al-Hakim Mosque → Qalawun ComplexMicro-green buffers added in underutilized verge pockets (0.4 m width)RL-based dynamic pedestrian flow routing introducedConflict point reduction: 5 → 2 Walkway width increased from 1.8–2.3 m → 3.0–3.2 m (reallocating space, no widening) Walkway width increased from 1.8–2.3 m → 3.0–3.2 m (reallocating space, no widening) Continuous pedestrian corridor created from Al-Hakim Mosque → Qalawun Complex Continuous pedestrian corridor created from Al-Hakim Mosque → Qalawun Complex Micro-green buffers added in underutilized verge pockets (0.4 m width) Micro-green buffers added in underutilized verge pockets (0.4 m width) RL-based dynamic pedestrian flow routing introduced RL-based dynamic pedestrian flow routing introduced Conflict point reduction: 5 → 2 Conflict point reduction: 5 → 2 A3: Quantitative comparison As summarized in Table 9, the optimized configuration significantly reduces pedestrian density, increases throughput, and improves overall walkability while introducing new green-space elements.Table 9Quantitative comparison of baseline and optimized conditions on El-Muizz Street, showing improvements in pedestrian density, throughput, walkability, and green-space coverage.MetricBaselineOptimizedChangeAverage pedestrian density4.8 p/m²\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}2.9 p/m²\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}-39.6%Throughput at hotspot node410 p/min630 p/min+53.6%Walkability Score5284+61%Green-space ratio0%11%+11% Table 9Quantitative comparison of baseline and optimized conditions on El-Muizz Street, showing improvements in pedestrian density, throughput, walkability, and green-space coverage.MetricBaselineOptimizedChangeAverage pedestrian density4.8 p/m²\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}2.9 p/m²\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}-39.6%Throughput at hotspot node410 p/min630 p/min+53.6%Walkability Score5284+61%Green-space ratio0%11%+11% Quantitative comparison of baseline and optimized conditions on El-Muizz Street, showing improvements in pedestrian density, throughput, walkability, and green-space coverage. Problem–solution–validation framework This study adopts a problem-centric approach in evaluating the proposed optimization framework. For each case study, we follow a four-part structure: Problem definition Identify existing spatial constraints such as narrow corridors, bottleneck intersections, absence of continuous pedestrian paths, and heritage-imposed limitations. AI-generated design solution Present the optimized street layout produced by the CNN–GNN–RL system, including reallocated widths, new pedestrian corridors, lane-direction modifications, and micro-green insertions. Quantitative validation Compare baseline and optimized conditions using pedestrian density, throughput, walkability, traffic efficiency, and environmental indicators. Heritage-compliance verification Overlay the optimized design on heritage-protection buffers to confirm that no protected structures or view corridors are violated. This structure ensures that the study focuses on spatial outcomes rather than algorithmic internals. Limitations and practical implementation considerations Model interpretability The CNN–GNN–RL framework performs well but lacks full transparency due to high-dimensional representations and opaque RL policies. Although heritage constraints are encoded via masks and penalties, internal decision logic is not fully explainable. Future work should integrate XAI methods (feature attribution, attention visualization, reward decomposition). The system serves as decision support, not an autonomous authority. Practical adoption Implementation depends on high-quality GIS and sensor data, sufficient computational resources , and regulatory alignment. A phased approach is recommended: (1) offline scenario testing, (2) semi-adaptive assistance with human validation, and (3) full integration into smart urban systems. Bias in data may create inequitable outcomes. Fairness-aware calibration and participatory validation are necessary to ensure transparency and inclusiveness. Future research directions While the proposed framework demonstrates strong performance, several avenues for future investigation remain. Enhancing model interpretability through feature attribution, attention visualization, and reward decomposition techniques would strengthen institutional trust and policy transparency. Future work should incorporate social equity indicators, economic vitality measures, and participatory feedback mechanisms to evaluate the societal impacts of optimization strategies. Research is needed to adapt the framework for cities with limited IoT infrastructure or incomplete GIS datasets, potentially through transfer learning or reduced-data modeling approaches. Extending the framework to multi-year simulation horizons could assess how adaptive interventions influence gradual urban evolution and heritage resilience over time. Combining AI-driven optimization with participatory planning interfaces may improve democratic legitimacy and collaborative governance in heritage-sensitive districts. Conclusion This study presents an artificial intelligence-based framework that seeks to improve historical and cultural streets optimization, and it successfully addresses the twin demands that are the current cities’ requirements and heritage protection. This approach is different to standard urban planning policies since our model is known to utilize deep learning, reinforcement learning, and evolutionary optimization approaches to design adaptable and responsive urban designs. Compared to the existing paradigm, the proposed framework has shown impressive results that entail the improvement of walkability by 70% to create a pedestrian favourable environment, the reduction of traffic congestion by 40%, thereby reducing traffic optimisation, the utilisation of green spaces by 50%, to support sustainable urbanism, the decrease in air pollution by 38.9%, thereby showing the environmental impact of AI based planning. Comparisons with other frameworks reveal that this framework scores highly in improving walkability, decreasing traffic congestion and increasing the use of green spaces, as well as decreasing air pollution enormously. The consequences support pedestrian access, enhancement of traffic facilities, environmental sustainability, and preservation of culture to a significant extent. RQ1: How can AI improve pedestrian accessibility in irregular historical streets while preserving heritage value? Our CNN–GNN–RL system increases pedestrian accessibility by 31.8% without altering protected street geometries, using constraint-aware path refinement and heritage masks that prevent disruption of culturally sensitive areas. RQ2: How can traffic congestion be reduced without modifying historical street morphology? Through real-time RL-driven lane adjustments and adaptive routing, the model reduces congestion by 22.5–40% while preserving all heritage boundaries. RQ3: How can green space integration be maximized within heritage-constrained districts? Green-space allocation improved by 27.4% through optimization of micro-green corridors and underutilized segments, without violating preservation constraints. RQ4: How does the proposed system outperform traditional methods? Compared with heuristic zoning, static GIS approaches, and evolutionary optimization models, our framework delivers higher improvements in walkability (+70%), congestion reduction (+33.3%), sustainability (+36–40%), and heritage compliance (+38.5%), while enabling real-time adaptability and multi-objective trade-off handling. The framework supports planners by enabling scenario testing for pedestrianization, traffic redistribution, and green space interventions in heritage districts. Its RL component allows adaptive traffic and pedestrian management without structural changes, while embedded preservation masks ensure regulatory compliance. Quantitative metrics strengthen evidence-based planning. Overall, it functions as a human-in-the-loop decision-support tool, complementing professional expertise. Moreover, the combination of CNNs, GNNs, and reinforcement learning can be used to carry out the analysis of spatial data and make adjustments in real-time, whereas the other approaches are much less flexible. This AI-oriented solution will help modify the urban layout in a more efficient and scalable way by automating the complicated decision-making process, which is why it is going to be a useful tool to stimulate the evolution of smart cities in the future. Future studies will be based on the integration of real-time, IoT-enabled urban analytics to enable dynamic changes in the urban design. Further research on multi-agent reinforcement learning will be done in order to optimise street designs according to the differences in the flow of pedestrians and traffic. Different cities all over the world will be chosen in order to test the universality of the proposed framework through scaling tests53,54. Supplementary Information Supplementary Information. Supplementary Information.