Spatial layout optimization algorithm based on AI driving

Abstract

The invention relates to the technical field of spatial layout, and discloses an AI-driven spatial layout optimization algorithm, which comprises the following steps: S1, collecting multi-source data to clean, classify and label the data, and expanding the scale and complexity of a data set by using a data enhancement technology; and S2, training an AI model by using deep learning in combination with a transfer learning technology, so that the AI model can learn a spatial configuration rule from diversified data, and by using a few-sample learning technology, the model can effectively identify innovative and non-traditional design requirements under the condition of data scarcity. Through multi-source data fusion, transfer learning and few-sample learning technologies, the model can learn the law of spatial configuration from diversified data, the recognition and response ability to innovative and non-traditional design requirements is significantly improved, and the system is used for traditional building and spatial layout optimization to adapt to emerging design trends and unique scenes, and has a wide application prospect. And the application range is wider.

Description

Technical Field

[0001] The present invention relates to the field of space layout technology, and specifically to an AI-driven space layout optimization algorithm. Background Art

[0002] With the development of modern architectural design and urban planning, the demand for optimization of spatial layout has become diversified. Traditional spatial layout design usually relies on manual work by experienced designers. This method has many limitations and is difficult to achieve the optimal layout when dealing with complex and changing design requirements. At the same time, traditional methods have problems such as slow response speed and low flexibility when dealing with innovative and non-traditional design requirements, making it difficult to adapt to the ever-changing emerging design trends and unique scenarios.

[0003] Traditional spatial layout optimization methods mostly rely on limited, single-source data to make design decisions. Their methods lack the ability to integrate and utilize diverse data, resulting in low generalization and adaptability of optimization models when facing emerging and non-traditional design requirements.

[0004] Most optimization algorithms are single algorithms or simple combinations, which are prone to fall into local optimality and fail to fully consider multi-objective requirements, resulting in low quality and diversity of generated layout solutions;

[0005] Existing layout optimization tools are insufficient in terms of interaction with designers, and the feedback they provide is not intuitive and transparent, making it difficult to effectively support designers’ creativity and personalized needs.

[0006] In existing technologies, the lack of support for intelligent design assistants and augmented reality / virtual reality limits designers’ understanding and control over optimization algorithms during the design process, hindering the application of innovative solutions and the improvement of trust.

[0007] Therefore, those skilled in the art provide an AI-driven spatial layout optimization algorithm to solve the problems raised in the above background technology. Summary of the invention

[0008] In view of the shortcomings of the prior art, the present invention provides an AI-driven spatial layout optimization algorithm to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: a spatial layout optimization algorithm based on AI drive, comprising the following steps:

[0010] S1. Collect multi-source data to clean, classify and annotate the data, and use data augmentation techniques to expand the size and complexity of the data set;

[0011] S2. Use deep learning combined with transfer learning technology to train AI models so that they can learn the rules of spatial configuration from diverse data, and use few-sample learning technology to enable the model to effectively identify innovative and non-traditional design requirements when data is scarce;

[0012] S3. Input the initial layout plan, use the trained AI model to analyze the layout, identify problems and potential optimization points in function and space usage, and use the generative adversarial network to generate diverse candidate layouts;

[0013] S4. Use a hybrid optimization algorithm to generate multiple alternative optimization schemes, and adjust algorithm parameters through reinforcement learning to perform multi-objective optimization;

[0014] S5. Simulate and evaluate the generated alternatives, use intelligent algorithms for multi-dimensional analysis, and adjust the plans based on historical cases and expert knowledge through heuristic intelligent optimization;

[0015] S6. Introduce intelligent design assistants to interact with designers through natural language processing, understand designers’ intentions and preferences, and provide personalized optimization suggestions. Use augmented reality or virtual reality technology to provide an immersive design experience, allowing designers to view, adjust and experience layout solutions in real time.

[0016] S7. Provide visualization tools to display the algorithm decision-making process and explain the pros and cons of each layout solution, so that designers can deeply understand and control the output of the algorithm and further adjust the optimization solution.

[0017] S8. Based on multiple rounds of interactions and feedback, the optimal layout solution is screened out, adjusted based on the designer’s creativity, and the generated design drawings and implementation guidelines are dynamically tracked and optimized during the actual implementation process;

[0018] S9. During actual use, collect user feedback and space usage data, adjust and improve the model, and continuously train and optimize the AI ​​model based on new data and design trends.

[0019] Preferably, in the step S1, in the data cleaning stage, a missing value filling formula is used for processing, and the missing value filling formula is:

[0020] in, is the missing value after filling, μ j is the mean value of feature j, β k is the linear regression coefficient of feature k on feature j, X ik is the value of feature k of the i-th sample, ∈ ij is the error term;

[0021] In the step S1, for data classification and labeling, a clustering algorithm is used for automatic labeling. The clustering algorithm formula is:

[0022] a. Core object conditions: |{x j ∈D|d(x i ,x j )≤∈}|≥MinPts,

[0023] Among them, x i is a data point, D is a data set, ∈ is a distance threshold, and MinPts is the minimum number of neighborhood data points;

[0024] b. Distance function:

[0025] Among them, x ik and x jk is the data point x i and x j The value in dimension k;

[0026] In the S1 step, data augmentation can be performed using a generative adversarial network in an image dataset. The generative adversarial network includes a generator and a discriminator. The basic formula of the generator network is as follows:

[0027] The generator formula of the generative adversarial network is: G(z;θ g )=tanh(W g ·z+b g ),

[0028] Among them, G(z;θ g ) is the generator output, θ g is the parameter set of the generator, W g is the weight matrix of the generator, z is the random noise vector, b g is the bias of the generator;

[0029] The loss function of the discriminator is:

[0030]

[0031] Among them, D(x (i) ) is the discriminator for the real data x (i) The discriminant probability of G(z (i) ) is the discriminator's response to the generated data z (i) The discriminant probability of , m is the number of samples in the batch.

[0032] Preferably, in the deep learning model of step S2, a convolutional neural network is used to extract spatial features of image data, and the core operations of the convolutional neural network include convolution and pooling;

[0033] Convolution operation formula:

[0034] Among them, S i,j is the convolution output, X is the input matrix, K is the convolution kernel matrix, M and N are the sizes of the convolution kernel;

[0035] The transfer learning in step S2 adapts the parameters of the pre-trained model to the new data set by fine-tuning the parameters of the pre-trained model. The fine-tuning process is achieved by freezing some layers and training the parameters of other layers.

[0036] Loss function fine-tuning formula:

[0037]

[0038] Among them, L(θ) is the loss function after fine-tuning, is the standard loss, θ is the parameter of the model, is the parameter of the pre-trained model, λ is the regularization coefficient, N is the number of samples, and M is the number of parameters.

[0039] Preferably, in the few-sample learning of step S2, the prototype network performs classification by learning class prototypes, and the prototypes of each class are represented by the mean of the samples in the class. The prototype calculation formula is:

[0040]

[0041] Among them, c k is the prototype of category k, S k is the support set belonging to category k, f φ (x i ) is the sample x i The embedding vector of φ generate;

[0042] Category prediction formula:

[0043] Among them, p(y=k|x) is the probability that sample x belongs to category k, (f φ (x),c k′ ) is the embedding of sample x and class prototype c k The distance between them is usually the Euclidean distance:

[0044] d(f φ (x),c k )=∥f φ (x)-c k ∥ 2 ;

[0045] The transfer learning in step S2 is combined with few-shot learning (MAML), which aims to quickly adapt to new tasks with a small amount of training data. The inner loop update formula of MAML is:

[0046] Among them, θ is the initial model parameter, α is the learning rate, In task T i The loss function on the task T is i Updated model parameters;

[0047] MAML's outer loop optimization formula:

[0048] Where T is the set of tasks.

[0049] Preferably, the hybrid optimization algorithm in step S4 utilizes the advantages of genetic algorithm and particle swarm optimization by combining them, while genetic algorithm is good at global search and particle swarm optimization is good at local search;

[0050] The genetic algorithm formula is as follows:

[0051] Fitness calculation: F i =f(X i ),

[0052] Among them, F i is the fitness value of individual i, X i is the gene expression of an individual, f(X i ) is the fitness function;

[0053] Crossover operation: C1 = (P 1,1:k ,P 2,k+1:n ), C2=(P 2,1:k ,P 1,k+1:n ),

[0054] Among them, C1 and C2 are newly generated offspring, P1 and P2 are parents, and k is the crossover point position;

[0055] Mutation operation: X i ′=X i +η·N(0,1),

[0056] Among them, X i ′ is the individual after mutation, η is the mutation rate, and N(0,1) is the standard normal distribution;

[0057] The particle swarm optimization formula is as follows:

[0058] Speed ​​update formula:

[0059] in, is the velocity of particle i at time t+1, ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, is the optimal position of particle i, g t is the global optimal position of the group;

[0060] Position update formula:

[0061] in, is the position of particle i at time t+1.

[0062] Preferably, the reinforcement learning in step S4 is used to adjust the optimization algorithm parameters, and the overall optimization effect is improved by learning the optimal strategy. The deep Q network formula is:

[0063] Q value update formula:

[0064]

[0065] Among them, Q(s t ,a t ) is the state s t Take action a t The Q value, α is the learning rate, r t is the immediate reward, γ is the discount factor;

[0066] Loss function:

[0067]

[0068] Among them, θ is the parameter of the Q network, θ - are the parameters of the target Q network.

[0069] Preferably, the simulated annealing algorithm in step S5 is a heuristic optimization algorithm for solving the global optimal problem and gradually approaches the optimal solution by simulating the physical annealing process;

[0070] State transition probability formula:

[0071]

[0072] Where ΔE=E new -E current is the energy difference between the new state and the current state, T is the current temperature;

[0073] Temperature drop formula: T k+1 =αT k ,

[0074] Among them, T k is the temperature of the kth iteration, and α is the cooling coefficient.

[0075] Preferably, the step S5 is based on heuristic optimization of historical cases and expert knowledge, using a case-based reasoning system to adjust the solution according to historical cases;

[0076] Case similarity calculation formula:

[0077]

[0078] Among them, Sim(C,C′) is the similarity between cases C and C′, a i and a i ′ is the value of case C and C′ on the i-th attribute, w i is the weight of the i-th attribute.

[0079] Preferably, the interpretability model of step S7 is used to explain the output of the machine learning model and measure the contribution of each feature to the prediction result. The interpretability model formula is:

[0080]

[0081] Among them, φ i is the value of feature i, N is the set of all features, S is the feature subset, and v(S) is the contribution of feature subset S to the model output.

[0082] Preferably, the interpretability model of step S7 is used to explain the output of the machine learning model and measure the contribution of each feature to the prediction result. The interpretability model formula is:

[0083]

[0084] Among them, φ i is the value of feature i, N is the set of all features, S is the feature subset, and v(S) is the contribution of feature subset S to the model output.

[0085] The present invention provides an AI-driven spatial layout optimization algorithm, which has the following beneficial effects:

[0086] 1. Through multi-source data fusion, transfer learning and few-sample learning technology, the model of the present invention can learn the rules of spatial configuration from diversified data, significantly improving the ability to identify and respond to innovative and non-traditional design needs. The system is used to optimize traditional buildings and spatial layouts to adapt to emerging design trends and unique scenarios, making its application scope more extensive.

[0087] 2. The present invention adopts a hybrid optimization algorithm and reinforcement learning to achieve multi-objective comprehensive optimization, taking into account multi-dimensional requirements such as space utilization, functional adaptability, and user experience, thereby improving the quality and diversity of optimization solutions, avoiding the limitations of local optimality, and ensuring that the generated solutions are more global and practical, and can better meet complex design requirements.

[0088] 3. The present invention introduces intelligent design assistants and augmented reality / virtual reality technologies, interacts with designers through natural language processing, and provides personalized and intuitive optimization suggestions, allowing designers to view, adjust and experience layout solutions in real time. Its enhanced interactivity and transparency make the design process more intuitive and controllable, promote designers' creativity and trust in innovative solutions, and significantly improve design efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0090] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0091] The present invention is described in detail below in conjunction with the accompanying drawings:

[0092] Example:

[0093] Please refer to the attached Figure 1 , an embodiment of the present invention provides an AI-driven spatial layout optimization algorithm, comprising the following steps: S1, collecting multi-source data to clean, classify and annotate the data, and using data enhancement technology to expand the scale and complexity of the data set;

[0094] S2. Use deep learning combined with transfer learning technology to train AI models so that they can learn the rules of spatial configuration from diverse data, and use few-sample learning technology to enable the model to effectively identify innovative and non-traditional design requirements when data is scarce;

[0095] S3. Input the initial layout plan, use the trained AI model to analyze the layout, identify problems and potential optimization points in function and space usage, and use the generative adversarial network to generate diverse candidate layouts;

[0096] S4. Use a hybrid optimization algorithm to generate multiple alternative optimization schemes, and adjust algorithm parameters through reinforcement learning to perform multi-objective optimization;

[0097] S5. Simulate and evaluate the generated alternatives, use intelligent algorithms for multi-dimensional analysis, and adjust the plans based on historical cases and expert knowledge through heuristic intelligent optimization;

[0098] S6. Introduce intelligent design assistants to interact with designers through natural language processing, understand designers’ intentions and preferences, and provide personalized optimization suggestions. Use augmented reality or virtual reality technology to provide an immersive design experience, allowing designers to view, adjust and experience layout solutions in real time.

[0099] S7. Provide visualization tools to display the algorithm decision-making process and explain the pros and cons of each layout solution, so that designers can deeply understand and control the output of the algorithm and further adjust the optimization solution.

[0100] S8. Based on multiple rounds of interactions and feedback, the optimal layout solution is screened out, adjusted based on the designer’s creativity, and the generated design drawings and implementation guidelines are dynamically tracked and optimized during the actual implementation process;

[0101] S9. During actual use, collect user feedback and space usage data, adjust and improve the model, and continuously train and optimize the AI ​​model based on new data and design trends.

[0102] Benefits of S1: Improve data quality and diversity, expand the scale of data sets, and improve model generalization capabilities; Benefits of S2: Efficiently learn spatial configuration rules, significantly improve the spatial analysis capabilities of the model, reduce training time and data requirements, and improve the adaptability and flexibility of the system; Benefits of S3: The system can generate a variety of candidate layouts to improve the innovation of the design; Benefits of S4: Enhance global search capabilities and local optimization capabilities, and the generated solutions are more global and practical, achieving more accurate multi-objective optimization; Benefits of S5: Save design time and costs; Benefits of S6: Understand the designer's intentions and preferences, provide designers with personalized optimization suggestions, intuitively understand the effects of different design solutions, and improve the flexibility and creativity of the design process; Benefits of S7: Enhance the sense of control over the design process and improve design efficiency and quality; Benefits of S8: Provide specific references and operating procedures for actual construction by generating design drawings and implementation guidelines to ensure that the solution can be smoothly implemented; Benefits of S9: Improve the long-term value of the system and enhance the ability to respond to future changes and uncertainties.

[0103] In step S1, during the data cleaning phase, the missing value filling formula is used for processing. The missing value filling formula is:

[0104] in, is the missing value after filling, μ j is the mean value of feature j, β k is the linear regression coefficient of feature k on feature j, X ik is the value of feature k of the i-th sample, ∈ ij is the error term;

[0105] In step S1, for data classification and labeling, a clustering algorithm is used for automatic labeling. The clustering algorithm formula is:

[0106] a. Core object conditions: |{x j ∈D|d(x i ,x j )≤∈}|≥MinPts,

[0107] Among them, x i is a data point, D is a data set, ∈ is a distance threshold, and MinPts is the minimum number of neighborhood data points;

[0108] b. Distance function:

[0109] Among them, x ik and x jk is the data point x i and x j The value in dimension k;

[0110] In step S1, data augmentation can be performed using a generative adversarial network in an image dataset. The generative adversarial network includes a generator and a discriminator. The following is the basic formula of the generator network:

[0111] The generator formula of the generative adversarial network is: G(z;θ g )=tanh(W g ·z+b g ),

[0112] Among them, G(z;θ g ) is the generator output, θ g is the parameter set of the generator, W g is the weight matrix of the generator, z is the random noise vector, b g is the bias of the generator;

[0113] The loss function of the discriminator is:

[0114]

[0115] Among them, D(x (i) ) is the discriminator for the real data x (i) The discriminant probability of G(z (i) ) is the discriminator's response to the generated data z (i) The discriminant probability of , m is the number of samples in the batch.

[0116] The S1 step significantly improves the quality and effect of the data preprocessing stage. Specifically, the missing value filling formula ensures the integrity and accuracy of the data, the clustering algorithm enhances the structure and classification efficiency of the data through automatic labeling, and the generative adversarial network greatly expands the scale and diversity of the data set through data enhancement. The joint action of the algorithms builds a sound data foundation, providing high-quality, rich and diverse data support for subsequent deep learning model training and optimization algorithm applications, and significantly improving the overall performance and innovation capabilities of the spatial layout optimization system.

[0117] In the deep learning model of step S2, the convolutional neural network is used to extract spatial features of image data. The core operations of the convolutional neural network include convolution and pooling.

[0118] Convolution operation formula:

[0119] Among them, S i,j is the convolution output, X is the input matrix, K is the convolution kernel matrix, M and N are the sizes of the convolution kernel;

[0120] The transfer learning in step S2 adapts the parameters of the pre-trained model to the new dataset by fine-tuning the parameters of the pre-trained model. The fine-tuning process is achieved by freezing some layers and training the parameters of other layers.

[0121] Loss function fine-tuning formula:

[0122]

[0123] Among them, L(θ) is the loss function after fine-tuning, is the standard loss, θ is the parameter of the model, is the parameter of the pre-trained model, λ is the regularization coefficient, N is the number of samples, and M is the number of parameters.

[0124] In step S2, the combined use of convolutional neural networks and transfer learning plays an important role in the training process of deep learning models. The convolution operation extracts spatial features from image data to provide the model with key local pattern recognition capabilities and enhance the model's understanding of image layout. Transfer learning further improves the model's performance on new data sets by fine-tuning the pre-trained model. Fine-tuning the loss function can effectively utilize existing knowledge and can achieve rapid adaptation to new tasks while ensuring model stability.

[0125] The formulas work together to achieve efficient model training driven by data. The convolution operation extracts rich spatial features, and transfer learning enables the model to maintain efficiency and accuracy under diverse and innovative design requirements. Their combination makes the deep learning-based spatial layout optimization algorithm more robust and adaptable, providing a solid foundation for further optimization and innovative design.

[0126] In the few-shot learning of step S2, the prototype network performs classification by learning class prototypes. The prototype of each class is represented by the mean of the samples in the class. The prototype calculation formula is:

[0127]

[0128] Among them, c k is the prototype of category k, S k is the support set belonging to category k, f φ (x i ) is the sample x i The embedding vector of φ generate;

[0129] Category prediction formula:

[0130] Among them, p(y=k|x) is the probability that sample x belongs to category k, (f φ (x),c k′ ) is the embedding of sample x and class prototype c k The distance between them is usually the Euclidean distance:

[0131] d(f φ (x),c k )=∥f φ (x)-c k ∥ 2 ;

[0132] The transfer learning in step S2 is combined with few-shot learning (MAML), which aims to quickly adapt to new tasks with a small amount of training data. The inner loop update formula of MAML is:

[0133] Among them, θ is the initial model parameter, α is the learning rate, In task T i The loss function on the task T is i Updated model parameters;

[0134] MAML's outer loop optimization formula:

[0135] Where T is the set of tasks.

[0136] In step S2, the combination of few-shot learning and transfer learning provides AI models with powerful learning capabilities when data is limited. The prototype network achieves efficient classification by calculating class prototypes, so that the model can quickly adapt to new classification tasks in few-shot scenarios. MAML optimizes the initial model parameters so that the model can quickly generalize and adapt to new tasks in a multi-task environment. The internal and external loop optimization strategy greatly reduces the model's dependence on training data, while improving the model's generalization ability and adaptability.

[0137] The combined use of algorithms builds an efficient, flexible and robust learning system, which enables the model to perform well in various complex and changeable design requirements. By integrating the advantages of deep learning in few-sample learning and transfer learning, the system can extract useful features from limited data and quickly apply them to new task scenarios, providing solid technical support and innovation capabilities for spatial layout optimization algorithms. This method ensures that AI-driven spatial layout optimization algorithms are efficient and accurate in modern applications.

[0138] The hybrid optimization algorithm in step S4 takes advantage of the advantages of genetic algorithm and particle swarm optimization by combining them, while genetic algorithm is good at global search and particle swarm optimization is good at local search;

[0139] The genetic algorithm formula is as follows:

[0140] Fitness calculation: F i =f(X i ),

[0141] Among them, F i is the fitness value of individual i, X i is the gene expression of an individual, f(X i ) is the fitness function;

[0142] Crossover operation: C1 = (P 1,1:k ,P 2,k+1:n ), C2=(P 2,1:k ,P 1,k+1:n ),

[0143] Among them, C1 and C2 are newly generated offspring, P1 and P2 are parents, and k is the crossover point position;

[0144] Mutation operation: X i ′=X i +η·N(0,1),

[0145] Among them, X i ′ is the individual after mutation, η is the mutation rate, and N(0,1) is the standard normal distribution;

[0146] The particle swarm optimization formula is as follows:

[0147] Speed ​​update formula:

[0148] in, is the velocity of particle i at time t+1, ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, is the optimal position of particle i, g t is the global optimal position of the group;

[0149] Position update formula:

[0150] in, is the position of particle i at time t+1.

[0151] In step S4, the hybrid optimization algorithm combines the advantages of genetic algorithm and particle swarm optimization to achieve efficient exploration and optimization of the solution space. Genetic algorithm is good at global search. Through fitness evaluation, crossover and mutation operations, the population can evolve continuously to avoid premature convergence. Particle swarm optimization is good at local search. Through the update of particle speed and position, combined with the best experience of individuals and groups, it can quickly adjust the direction and refine the search path.

[0152] =The combination enables the hybrid optimization algorithm to explore the solution space extensively at the global level to avoid falling into the local optimum, and to make fine adjustments in local areas to improve the quality of the solution and the speed of convergence. Through the synergy of genetic algorithm and particle swarm optimization, the system can show stronger adaptability and efficiency in complex and multi-dimensional spatial layout optimization tasks, significantly improve the overall performance and effect of the optimization process, and ensure that the spatial layout optimization algorithm can provide stable, fast and high-quality solutions when facing complex problems under multi-objective and multi-constraint conditions.

[0153] The reinforcement learning in step S4 is used to adjust the optimization algorithm parameters and improve the overall optimization effect by learning the optimal strategy. The deep Q network formula is:

[0154] Q value update formula:

[0155]

[0156] Among them, Q(s t ,a t ) is the state s t Take action a t The Q value, α is the learning rate, r t is the immediate reward, γ is the discount factor;

[0157] Loss function:

[0158]

[0159] Among them, θ is the parameter of the Q network, θ - are the parameters of the target Q network.

[0160] In step S4, reinforcement learning adjusts the parameters of the optimization algorithm through a deep Q network to effectively improve the overall optimization effect. The Q value update formula is used to learn the value of each state-action pair. Through continuous updating and estimation, the model's strategy is gradually improved so that the optimal action can be taken in each state. The design of the loss function further ensures the accuracy of the Q value and guides the training of the Q network, so that the model can stably and efficiently learn and adapt to complex optimization environments.

[0161] Through the deep Q network, reinforcement learning can effectively adjust the parameters of the optimization algorithm, learn the optimal strategy in the optimization task, and dynamically adapt to the changing environment. It can achieve fast and efficient parameter adjustment in complex spatial layout optimization and improve the overall optimization effect. Through the combination of Q value update and loss function, the model gradually optimizes the strategy in the process of continuous exploration and utilization, and finally achieves the global optimal optimization result.

[0162] The simulated annealing algorithm in step S5 is a heuristic optimization algorithm used to solve the global optimal problem and gradually approach the optimal solution by simulating the physical annealing process;

[0163] State transition probability formula:

[0164]

[0165] Where ΔE=E new -E current is the energy difference between the new state and the current state, T is the current temperature;

[0166] Temperature drop formula: T k+1 =αT k ,

[0167] Among them, T k is the temperature of the kth iteration, and α is the cooling coefficient.

[0168] In step S5, the simulated annealing algorithm approaches the global optimal solution by simulating the principle of the physical annealing process. The state transition probability formula dynamically decides whether to accept the new state by calculating the energy difference between the new state and the state, combined with the current temperature. The mechanism ensures that the algorithm can search for a better solution in the direction of energy reduction, and can jump out of the local optimal solution at the appropriate time to explore a wider solution space;

[0169] The temperature drop formula controls the temperature change of the search process. By gradually lowering the temperature, the algorithm's initial extensive random search gradually shifts to detailed local optimization, helping the algorithm to eventually converge to the global optimal solution. The gradual decrease in temperature reduces the probability of the algorithm accepting a worse solution, prompting the solution to get closer and closer to the optimal state.

[0170] Step S5 is a heuristic optimization based on historical cases and expert knowledge, using a case-based reasoning system to adjust the solution based on historical cases;

[0171] Case similarity calculation formula:

[0172]

[0173] Among them, Sim(C,C′) is the similarity between cases C and C′, a i and a i ′ is the value of case C and C′ on the i-th attribute, w i is the weight of the i-th attribute.

[0174] In step S5, the heuristic optimization based on historical cases and expert knowledge adopts a case reasoning system. By calculating the similarity between cases, the current solution is adjusted by referring to the solutions in historical cases. The case similarity calculation formula is the core tool of the process. By calculating the similarity between the new case and the historical case, the most similar case is identified and its solution is referred to to provide guidance for the current problem.

[0175] This formula uses weighted Euclidean distance to measure the similarity between cases, and adjusts the importance of each attribute through weights, so that the similarity calculation can flexibly adapt to different optimization problems. During the optimization process, the similarity measurement can quantitatively evaluate the relationship between cases, and can highlight the influence of key attributes through weight adjustment, thereby improving the rationality and pertinence of the solution.

[0176] The interpretability model in step S7 is used to explain the output of the machine learning model and measure the contribution of each feature to the prediction result. The interpretability model formula is:

[0177]

[0178] Among them, φ i is the value of feature i, N is the set of all features, S is the feature subset, and v(S) is the contribution of feature subset S to the model output.

[0179] In step S7, the interpretable model provides an intuitive explanation for complex machine learning models by calculating the contribution of each feature to the model output. This interpretability makes the internal decision-making process of the black box model transparent, providing users and decision makers with a basis for trust in the model. By quantifying the impact of each feature on the output, the decision-making process of the model becomes understandable and verifiable, and more detailed adjustments and optimizations can be made based on these contribution values.

[0180] Step S7 visualizes the decision-making process of the algorithm through a decision tree, showing the path and decision rules. The decision tree splitting condition formula is:

[0181] Gini Impurity:

[0182] Among them, P i is the sample proportion of the i-th class;

[0183] Information Gain:

[0184]

[0185] in, is the entropy of set S, S v is the subset of feature A with value v.

[0186] In step S7, the decision tree visualizes the decision process of the algorithm, displays the paths and decision rules, and makes the complex machine learning model more transparent and easy to understand. Gini impurity and information gain are key indicators for selecting split features in the decision tree. Gini impurity measures the degree of mixing of data within a node, while information gain measures the improvement in data purity after splitting. By maximizing information gain and minimizing Gini impurity, the decision tree can efficiently select the optimal split features and gradually establish a refined decision path, thereby improving the classification or regression capabilities of the model.

[0187] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven spatial layout optimization algorithm, characterized in that: The steps include: S1. Collect multi-source data to clean, classify and annotate the data, and use data augmentation techniques to expand the size and complexity of the data set; S2. Use deep learning combined with transfer learning technology to train AI models so that they can learn the rules of spatial configuration from diverse data, and use few-sample learning technology to enable the model to effectively identify innovative and non-traditional design requirements when data is scarce; S3. Input the initial layout plan, use the trained AI model to analyze the layout, identify problems and potential optimization points in function and space usage, and use the generative adversarial network to generate diverse candidate layouts; S4. Use a hybrid optimization algorithm to generate multiple alternative optimization schemes, and adjust algorithm parameters through reinforcement learning to perform multi-objective optimization; S5. Simulate and evaluate the generated alternatives, use intelligent algorithms for multi-dimensional analysis, and adjust the plans based on historical cases and expert knowledge through heuristic intelligent optimization; S6. Introduce intelligent design assistants to interact with designers through natural language processing, understand designers’ intentions and preferences, and provide personalized optimization suggestions. Use augmented reality or virtual reality technology to provide an immersive design experience, allowing designers to view, adjust and experience layout solutions in real time. S7. Provide visualization tools to display the algorithm decision-making process and explain the pros and cons of each layout solution, so that designers can deeply understand and control the output of the algorithm and further adjust the optimization solution. S8. Based on multiple rounds of interactions and feedback, the optimal layout solution is screened out, adjusted based on the designer’s creativity, and the generated design drawings and implementation guidelines are dynamically tracked and optimized during the actual implementation process; S9. During actual use, collect user feedback and space usage data, adjust and improve the model, and continuously train and optimize the AI ​​model based on new data and design trends.

2. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: In the step S1, during the data cleaning phase, the missing value filling formula is used for processing. The missing value filling formula is: in, is the missing value after filling, μ j is the mean value of feature j, β k is the linear regression coefficient of feature k on feature j, X ik is the value of feature k of the i-th sample, ∈ ij is the error term; In the step S1, for data classification and labeling, a clustering algorithm is used for automatic labeling. The clustering algorithm formula is: a. Core object conditions: |{x j ∈D|d(x i ,x j )≤∈}|≥MinPts, Among them, x i is a data point, D is a data set, ∈ is a distance threshold, and MinPts is the minimum number of neighborhood data points; b. Distance function: Among them, x ik and x jk is the data point x i and x j The value in dimension k; In the S1 step, data augmentation can be performed using a generative adversarial network in an image dataset. The generative adversarial network includes a generator and a discriminator. The basic formula of the generator network is as follows: The generator formula of the generative adversarial network is: G(z;θ g )=tanh(W g ·z+b g ), Among them, G(z;θ g ) is the generator output, θ g is the parameter set of the generator, W g is the weight matrix of the generator, z is the random noise vector, b g is the bias of the generator; The loss function of the discriminator is: Among them, D(x (i) ) is the discriminator for the real data x (i) The discriminant probability of G(z (i) ) is the discriminator's response to the generated data z (i) The discriminant probability of , m is the number of samples in the batch.

3. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: In the deep learning model of step S2, a convolutional neural network is used to extract spatial features of image data, and the core operations of the convolutional neural network include convolution and pooling; Convolution operation formula: Among them, S i,j is the convolution output, X is the input matrix, K is the convolution kernel matrix, M and N are the sizes of the convolution kernel; The transfer learning in step S2 adapts the parameters of the pre-trained model to the new data set by fine-tuning the parameters of the pre-trained model. The fine-tuning process is achieved by freezing some layers and training the parameters of other layers. Loss function fine-tuning formula: Among them, L(θ) is the loss function after fine-tuning, l is the standard loss, θ is the parameter of the model, is the parameter of the pre-trained model, λ is the regularization coefficient, N is the number of samples, and M is the number of parameters.

4. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: In the few-sample learning of step S2, the prototype network performs classification by learning class prototypes. The prototype of each class is represented by the mean of the samples in the class. The prototype calculation formula is: Among them, c k is the prototype of category k, S k is the support set belonging to category k, f φ (x i ) is the sample x i The embedding vector of φ generate; Category prediction formula: Among them, p(y=k|x) is the probability that sample x belongs to category k, (f φ (x),c k′ ) is the embedding of sample x and class prototype c k The distance between them is usually the Euclidean distance: d(f φ (x),c k )=∥f φ (x)-c k ∥ 2 ; The transfer learning in step S2 is combined with few-shot learning (MAML), which aims to quickly adapt to new tasks with a small amount of training data. The inner loop update formula of MAML is: Among them, θ is the initial model parameter, α is the learning rate, In task T i The loss function on the task T is i Updated model parameters; MAML's outer loop optimization formula: Where T is the set of tasks.

5. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: The hybrid optimization algorithm in step S4 utilizes the advantages of genetic algorithm and particle swarm optimization by combining them, while genetic algorithm is good at global search and particle swarm optimization is good at local search; The genetic algorithm formula is as follows: Fitness calculation: F i =f(X i ), Among them, F i is the fitness value of individual i, X i is the gene expression of an individual, f(X i ) is the fitness function; Crossover operation: C1 = (P 1,1:k ,P 2,k+1:n ), C2=(P 2,1:k ,P 1,k+1:n ), Among them, C1 and C2 are newly generated offspring, P1 and P2 are parents, and k is the crossover point position; Mutation operation: X i ′=X i +η·N(0,1), Among them, X i ′ is the individual after mutation, η is the mutation rate, and N(0,1) is the standard normal distribution; The particle swarm optimization formula is as follows: Speed ​​update formula: in, is the velocity of particle i at time t+1, ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, is the optimal position of particle i, g t is the global optimal position of the group; Position update formula: in, is the position of particle i at time t+1.

6. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: The reinforcement learning in step S4 is used to adjust the optimization algorithm parameters and improve the overall optimization effect by learning the optimal strategy. The deep Q network formula is: Q value update formula: Among them, Q(s t ,a t ) is the state s t Take action a t The Q value, α is the learning rate, r t is the immediate reward, γ is the discount factor; Loss function: Among them, θ is the parameter of the Q network, θ - are the parameters of the target Q network.

7. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: The simulated annealing algorithm in step S5 is a heuristic optimization algorithm used to solve the global optimal problem and gradually approach the optimal solution by simulating the physical annealing process; State transition probability formula: Where ΔE=E new -E current is the energy difference between the new state and the current state, T is the current temperature; Temperature drop formula: T k+1 =αT k , Among them, T k is the temperature of the kth iteration, and α is the cooling coefficient.

8. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: The S5 step is based on heuristic optimization of historical cases and expert knowledge, using a case-based reasoning system to adjust the solution according to historical cases; Case similarity calculation formula: Among them, Sim(C,C′) is the similarity between cases C and C′, a i and a i ′ is the value of case C and C′ on the i-th attribute, w i is the weight of the i-th attribute.

9. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: The interpretability model in step S7 is used to explain the output of the machine learning model and measure the contribution of each feature to the prediction result. The interpretability model formula is: Among them, φ i is the value of feature i, N is the set of all features, S is the feature subset, and v(S) is the contribution of feature subset S to the model output.

10. The AI-driven spatial layout optimization algorithm according to claim 1, characterized in that: The S7 step visualizes the decision-making process of the algorithm through a decision tree, showing the path and decision rules. The decision tree splitting condition formula is: Gini Impurity: Among them, P i is the sample proportion of the i-th class; Information Gain: in, is the entropy of set S, S v is the subset of feature A with value v.

Images