A residential intelligent agent reward feedback reinforcement learning urban renewal public participation planning method and system
By constructing a resident intelligent agent reward feedback reinforcement learning system, and using a large language model and reinforcement learning strategy network to generate land use adjustment actions, the problem of insufficient resident demand feedback in urban renewal planning was solved, and the planning scheme was optimized in real time and social acceptance was improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG URBAN & RURAL PLANNING & DESIGN INST
- Filing Date
- 2026-05-26
- Publication Date
- 2026-08-04
AI Technical Summary
Existing urban renewal land use planning methods are unable to effectively incorporate residents' subjective needs and personalized feedback, resulting in a disconnect between planning schemes and actual usage requirements. Furthermore, the application of large language models in urban planning has not been deeply integrated with reinforcement learning, making it difficult to drive continuous optimization.
A resident intelligent agent reward feedback reinforcement learning system is constructed. Multi-role resident intelligent agents are generated through a large language model, land use adjustment actions are generated using a reinforcement learning strategy network, and iterative training is performed based on resident satisfaction reward values to form an end-to-end closed loop for public participation in planning.
This enabled real-time feedback of residents' needs and continuous optimization of planning schemes, improving the social acceptance and feasibility of planning schemes and ensuring that diverse residents' demands were incorporated into the planning schemes from the generation stage.
Smart Images
Figure CN122288440B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to a method for public participation planning in urban renewal using resident agent reward feedback reinforcement learning. Background Technology
[0002] Urban renewal land use planning is a core component of sustainable urban development. Its key lies in achieving efficient spatial layout that is "people-centered," ensuring residents' basic needs for public services, ecological greening, and transportation are met. To improve planning efficiency and scientific rigor, deep reinforcement learning has been introduced into the field of urban community spatial planning. This technology, by constructing urban adjacency graph models, transforms land use planning into a sequential decision-making problem on a graph. It also utilizes dual-strategy networks to automate the optimization of land use and road layout, achieving superior results compared to manual planning in objective quantitative indicators such as service accessibility, ecological efficiency, and traffic efficiency.
[0003] However, the aforementioned land use planning methods based on deep reinforcement learning are essentially still decision-making models that rely solely on objective quantitative indicators. Their value networks struggle to incorporate residents' subjective needs and personalized feedback expressed in natural language and other forms. This leads to planning schemes that, while performing well at the indicator level, may be disconnected from residents' actual usage needs, resulting in rework and adjustments after project implementation. Furthermore, existing large language models in urban planning applications are mostly limited to superficial processing stages such as planning text parsing, failing to achieve deep integration with the value networks of reinforcement learning. Consequently, residents' needs cannot be effectively transformed into drivers for continuous planning optimization. Summary of the Invention
[0004] To overcome the shortcomings of the prior art described above, which lacks feedback on residents' needs, this invention provides a method for public participation in urban renewal planning through resident agent reward feedback reinforcement learning.
[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A public participation planning method for urban renewal based on resident intelligent agent reward feedback reinforcement learning, the method comprising: Obtain geospatial data of the city to be planned, divide the geospatial data into multiple types of nodes, and construct edges based on the spatial adjacency relationships between different nodes to obtain an adjacency topology graph model. Based on the public participation corpus of the city to be planned, intelligent agents of residents with different roles are constructed using a large language model; The land use adjustment actions are generated by the reinforcement learning policy network based on the features of the adjacency topology graph model. The land use adjustment actions are input into the resident agent to obtain the evaluation results output by the resident agent, and the resident satisfaction reward value is calculated based on the evaluation results. The reinforcement learning strategy network is iteratively trained with the goal of maximizing the resident satisfaction reward value, and when a preset termination condition is met, the trained reinforcement learning strategy network outputs a land use planning scheme for the city to be planned.
[0006] As a preferred embodiment, the various types of nodes include plot nodes, road segment nodes, and intersection nodes; the spatial adjacency relationships include the adjacency between plot nodes and intersection nodes, the adjacency between plot nodes and road segment nodes, and the adjacency between road segment nodes and intersection nodes.
[0007] As a preferred embodiment, the step of constructing resident intelligent agents corresponding to different roles based on the public participation corpus of the city to be planned using a large language model includes: Extract structured descriptive features corresponding to different role types from public participation corpus; The structured description features are mapped to role embedding vectors through an embedding network; The embedding vector of any role is mapped to a prefix matrix using a multilayer perceptron, and the prefix matrix is added as a prefix vector to each layer of the large language model. The parameters of the multilayer perceptron and the prefix matrix are updated based on the frozen parameters of the large language model to obtain the resident agent of that role type.
[0008] As a preferred embodiment, the training process of the resident agent includes mixing the structured description features of all roles, randomly sampling samples and calculating the loss based on the loss function with their corresponding role embedding vectors, and adjusting the parameters of the resident agent with the goal of minimizing the loss until a preset condition is met.
[0009] As a preferred embodiment, the step of generating land use adjustment actions based on the features of the adjacency topology graph model using a reinforcement learning policy network includes: The spatial feature vector is obtained by passing the adjacency topology graph model through a graph coding network; The population proportion weights corresponding to different roles, the role embedding vectors corresponding to different roles, and the semantic vectors of historical public participation corpus are used as public participation state features. The spatial feature vectors and public participation state features are input into the reinforcement learning policy network to generate the land use adjustment action. The semantic vectors of the historical public participation corpus are obtained by text semantic encoding based on the evaluation results output by the resident agent in the previous iteration.
[0010] As a preferred embodiment, the step of obtaining the evaluation results output by the resident intelligent agent and calculating the resident satisfaction reward value based on the evaluation results includes: The land use adjustment action is converted into natural language description text, and the natural language description text is input into the resident intelligent agent; Obtain the satisfaction score output by any resident agent. Based on the population proportion weight of the role corresponding to the resident agent in the city to be planned, the satisfaction scores output by different resident agents are weighted and summed to obtain the resident satisfaction reward value.
[0011] As a preferred option, the evaluation results output by the resident intelligent agent in the previous iteration include the text opinions corresponding to the demands and suggestions.
[0012] As a preferred embodiment, the step of iteratively training the reinforcement learning policy network with the objective of maximizing the resident satisfaction reward value includes: The policy network is used to output action probabilities based on the public participation state characteristics, and the resident satisfaction reward value is obtained based on the resident agent. The value network is used to output the state value based on the public participation state characteristics; The advantage function is calculated based on the resident satisfaction reward value and state value. The optimization direction is to maximize the objective function, and the parameters of the reinforcement learning policy network are updated synchronously using the advantage function. The process continues until the mean of the resident satisfaction reward value is not lower than the first threshold and the fluctuation range of the resident satisfaction reward value is not higher than the second threshold; or the number of iterations reaches the preset maximum number of steps.
[0013] As a preferred embodiment, the reinforcement learning policy network also filters invalid actions based on an action validity mask when generating the land use adjustment action; The action legality mask is generated based on feature representations extracted from the adjacency topology graph model and preset constraints, and is used to determine candidate positions for facility layout actions between different nodes.
[0014] This application also proposes an urban renewal public participation planning system based on resident agent reward feedback reinforcement learning, applied to the urban renewal public participation planning method using resident agent reward feedback reinforcement learning. The system includes: The graph model construction module is used to acquire geospatial data of the city to be planned, divide the geospatial data into multiple types of nodes, and construct edges based on the spatial adjacency relationships between different nodes to obtain an adjacency topology graph model. The agent construction module is used to construct multi-role intelligent agents based on public participation corpora of the city to be planned using a large language model. The evaluation and reward calculation module is used to generate land use adjustment actions based on feature representations extracted from the adjacency topology graph model through a reinforcement learning policy network, input the land use adjustment actions into the agent, obtain the evaluation results output by the agent, and calculate the resident satisfaction reward value based on the evaluation results. The training and scheme generation module is used to iteratively train the reinforcement learning policy network with the goal of maximizing the resident satisfaction reward value, and obtain the land use planning scheme of the city to be planned when the preset termination condition is met.
[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows: This invention constructs a real-time interactive closed loop between a reinforcement learning policy network and a large language model resident agent. The land use adjustment scheme can be evaluated in real time based on multi-role resident agents. The evaluation results are converted into reward signals and fed back to the reinforcement learning policy network, driving the instant update of network parameters. This allows public participation demands to be embedded in each iteration of planning optimization in real time and continuously, forming an end-to-end closed loop of "scheme generation - public participation - model update". Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the urban renewal public participation planning method using resident intelligent agent reward feedback reinforcement learning in Example 1. Figure 2 This is a system architecture diagram of a public participation planning system for urban renewal based on resident intelligent agent reward feedback reinforcement learning, as shown in Example 3. Detailed Implementation
[0017] In the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.
[0018] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0020] Example 1 This embodiment proposes a public participation planning method for urban renewal using resident intelligent agent reward feedback reinforcement learning, such as... Figure 1 The diagram shown is a flowchart illustrating a public participation planning method for urban renewal using resident intelligent agent reward feedback reinforcement learning, as described in this embodiment.
[0021] A public participation planning method for urban renewal based on resident agent reward feedback reinforcement learning includes the following steps: A public participation planning method for urban renewal based on resident intelligent agent reward feedback reinforcement learning, the method comprising: S1. Obtain geospatial data of the city to be planned, divide the geospatial data into multiple types of nodes, and construct edges based on the spatial adjacency relationships between different nodes to obtain an adjacency topology graph model. S2. Construct resident intelligent agents corresponding to different roles based on the public participation corpus of the city to be planned using a large language model; S3. Generate land use adjustment actions based on the features of the adjacency topology graph model through a reinforcement learning policy network, input the land use adjustment actions into the resident agent, obtain the evaluation results output by the resident agent, and calculate the resident satisfaction reward value based on the evaluation results; S4. The reinforcement learning strategy network is iteratively trained with the goal of maximizing the resident satisfaction reward value, and when the preset termination condition is met, the land use planning scheme of the city to be planned is output using the trained reinforcement learning strategy network.
[0022] In this embodiment, urban geospatial data is first abstracted into an adjacency topology graph model composed of multiple types of nodes and edges, serving as the structured input foundation for the reinforcement learning policy network. Then, a large language model is used to construct resident agents capable of expressing diverse roles and demands, serving as a source of public feedback. Subsequently, the policy network generates land use adjustment actions, which are evaluated by the resident agents, who then provide feedback on resident satisfaction rewards, driving iterative optimization of the policy network. This application addresses the problems of superficial public participation, delayed feedback, and difficulty in quantification in traditional urban planning by embedding a public participation mechanism into the reward feedback loop of reinforcement learning. This allows planning schemes to incorporate diverse resident demands from the generation stage, significantly improving the social acceptability and feasibility of the planning schemes.
[0023] In one optional embodiment, the various types of nodes include land parcel nodes, road segment nodes, and intersection nodes; the spatial adjacency relationships include the adjacency between land parcel nodes and intersection nodes, the adjacency between land parcel nodes and road segment nodes, and the adjacency between road segment nodes and intersection nodes.
[0024] Furthermore, the adjacency topology graph model will be updated as land use adjustment actions occur. The updating steps include: when the land use adjustment action causes changes in the geographic element attributes or subdivision of the geographic elements corresponding to the nodes, splitting or adding the corresponding nodes, and updating the edges and adjacency relationships between the affected nodes.
[0025] In an optional embodiment, the step of constructing resident intelligent agents corresponding to different roles based on the public participation corpus of the city to be planned using a large language model includes: Extract structured descriptive features corresponding to different role types from public participation corpus; The structured description features are mapped to role embedding vectors through an embedding network; The embedding vector of any role is mapped to a prefix matrix using a multilayer perceptron, and the prefix matrix is added as a prefix vector to each layer of the large language model. The parameters of the multilayer perceptron and the prefix matrix are updated while the parameters of the large language model are frozen to obtain the resident agent of that role type. The formula for calculating the prefix vector is as follows:
[0026]
[0027] As an example, the corpus was annotated according to four dimensions: participant type, demand type, demand intensity, and negative feedback type. Participant type: Divided into seven core roles: elderly residents, young working professionals, parents of students, individual business owners, special groups such as the disabled / living alone, property owners, and community organizations. Demand type: Divided into eight core dimensions: education, healthcare, elderly care, commerce, transportation, ecology, living environment, and public safety. Demand intensity: Annotated by high frequency, medium frequency, and low frequency for weight calculation. Negative feedback type: Annotated with explicit opposition, questioning, and suggestions for modification. In addition to the above four dimensions, structured descriptive features were extracted for each role during the annotation process, including: role name, typical age range, core interest keywords, language style features, and decision weight factors. These features constitute the role embedding vector.
[0028] In this embodiment, compared to fine-tuning a large language model with all parameters separately for each type of role, this approach significantly reduces model storage and switching overhead, enabling the system to support different resident roles on the same basic model. At the same time, since the basic model parameters are frozen, forgetting and interference between roles are avoided, ensuring the stability and distinguishability of different resident agents when expressing their respective demands, and providing a high-fidelity and controllable role-based feedback source for subsequent reward value calculation.
[0029] In an optional embodiment, the training process of the resident agent includes mixing the structured description features of all roles, randomly sampling samples and calculating the loss based on a loss function with their corresponding role embedding vectors, and adjusting the parameters of the resident agent with the goal of minimizing the loss until a preset condition is met; the calculation formula of the loss function is as follows:
[0030]
[0031] in, The loss function; The input solution description text; The target text output to the agent includes satisfaction ratings, requests, and suggestions; Character type; A large language model with frozen parameters.
[0032] In this embodiment, by uniformly mixing the structured descriptive features of all roles and then randomly sampling and training, the prefix matrix and the multilayer perceptron can learn the discriminative differences between roles, ensuring the accuracy of each role in generating evaluation text. When adding new participating subject types, only the structured description of that role (name, typical demands, language style, etc.) needs to be provided, and the results can be directly used after computation through the embedding network, without the need to collect a large amount of labeled corpus or retrain. For special needs, a small number of examples can be used for fine-tuning.
[0033] Example 2 This embodiment is an improvement upon the urban renewal public participation planning method proposed in Embodiment 1, which uses resident intelligent agent reward feedback reinforcement learning.
[0034] In an optional embodiment, the step of generating land use adjustment actions based on feature representations extracted from the adjacency topology graph model via a reinforcement learning policy network includes: The spatial feature vector is obtained by passing the adjacency topology graph model through a graph coding network; The population proportion weights corresponding to different roles, the role embedding vectors corresponding to different roles, and the semantic vectors of historical public participation corpus are used as public participation state features. The spatial feature vectors and public participation state features are input into the reinforcement learning policy network to generate the land use adjustment action. The semantic vectors of the historical public participation corpus are obtained by text semantic encoding based on the evaluation results output by the resident agent in the previous iteration.
[0035]
[0036] in: For the first Satisfaction score (0-10 points) of the output of the resident-like intelligent agent; For the first The population weight of each type of resident within the project area, satisfying... ; The value range is [0, 1], where 1 represents complete satisfaction of all characters.
[0037] In this embodiment, the unstructured text opinions output by the resident agent are used as feature inputs to the policy network for the next round of states. This technical solution enables the reinforcement learning model to "remember" the residents' historical feedback and continuously improve in subsequent decisions: when a certain type of demand appears multiple times, the model automatically adjusts the action probability distribution of the policy network, converging towards satisfying that demand. Simultaneously, this mechanism represents a technological leap from "one-time weight setting" to "continuous learning and optimization," significantly improving the model's accuracy in responding to public demands.
[0038] In an optional embodiment, the step of obtaining the evaluation result output by the resident agent and calculating the resident satisfaction reward value based on the evaluation result includes: The land use adjustment action is converted into natural language description text, and the natural language description text is input into the resident intelligent agent; Obtain the satisfaction score output by any resident agent. Based on the population proportion weight of the role corresponding to the resident agent in the city to be planned, the satisfaction scores output by different resident agents are weighted and summed to obtain the resident satisfaction reward value.
[0039] In this embodiment, the satisfaction scores of each role are weighted and summed according to the population proportion, which ensures that the reward value is consistent with the actual population structure of the city to be planned in a statistical sense.
[0040] In an optional embodiment, the evaluation results output by the resident agent in the previous iteration include textual opinions corresponding to the demands and suggestions.
[0041] In an optional embodiment, the step of iteratively training the reinforcement learning policy network with the objective of maximizing the resident satisfaction reward value includes: The policy network is used to output action probabilities based on the public participation state characteristics, and the resident satisfaction reward value is obtained based on the resident agent. The value network is used to output the state value based on the public participation state characteristics; The advantage function is calculated based on the resident satisfaction reward value and state value, with the optimization direction being to maximize the objective function, and the parameters of the reinforcement learning policy network are updated synchronously using the advantage function; The expression for the advantage function is as follows:
[0042] in, For a moment The advantage function represents the degree of advantage of taking an action relative to the average level. For a moment The resident satisfaction reward value; This is a discount factor used to balance the weights of current and future rewards, and its value ranges from [value range missing]. ; The state value output by the value network represents the state value in the value network. The expected cumulative rewards in the future; The city's state in the next moment; The expression for the objective function is as follows:
[0043] in, The pruning objective function optimized for the near-end strategy; These are the parameters of the policy network; This represents the probability ratio between the old and new strategies. The hyperparameters are used to limit the policy update step size. This indicates that for all time steps The mathematical expectation; For the clipping function, the probability ratio is restricted to... Within the range.
[0044] Further optionally, the preset termination condition includes at least one of the following: the average value of the resident satisfaction reward value is not lower than a first threshold and the fluctuation range of the resident satisfaction reward value is not higher than a second threshold; or the number of iteration steps reaches a preset maximum number of steps.
[0045] In this embodiment, by working together with the policy network and the value network, the advantage function is used to estimate the action reward, and the objective function is used to limit the policy update magnitude, which can maintain the stability of policy updates and improve the overall sample efficiency and computational economy.
[0046] In an optional embodiment, the reinforcement learning policy network further filters invalid actions based on an action validity mask when generating the land use adjustment action; The action legality mask is generated based on feature representations extracted from the adjacency topology graph model and preset constraints, and is used to determine candidate positions for facility layout actions between different nodes.
[0047] In this embodiment, by using an action validity mask to constrain and filter the action probability distribution output by the policy network, it is possible to eliminate unreasonable developments or illegal connections between non-adjacent nodes and avoid wasting training computing power on reward evaluation of invalid or illegal actions.
[0048] Example 3 This embodiment proposes a public participation planning system for urban renewal using resident agent reward feedback reinforcement learning, and applies a public participation planning method for urban renewal using resident agent reward feedback reinforcement learning proposed in Embodiment 1. For example... Figure 2The diagram shown is an architecture diagram of an urban renewal public participation planning system based on resident intelligent agent reward feedback reinforcement learning, according to this embodiment.
[0049] A public participation planning system for urban renewal based on resident intelligent agent reward feedback reinforcement learning, the system comprising: The graph model construction module is used to acquire geospatial data of the city to be planned, divide the geospatial data into multiple types of nodes, and construct edges based on the spatial adjacency relationships between different nodes to obtain an adjacency topology graph model. The agent construction module is used to construct multi-role intelligent agents based on public participation corpora of the city to be planned using a large language model. The evaluation and reward calculation module is used to generate land use adjustment actions based on feature representations extracted from the adjacency topology graph model through a reinforcement learning policy network, input the land use adjustment actions into the agent, obtain the evaluation results output by the agent, and calculate the resident satisfaction reward value based on the evaluation results. The training and scheme generation module is used to iteratively train the reinforcement learning policy network with the goal of maximizing the resident satisfaction reward value, and obtain the land use planning scheme of the city to be planned when the preset termination condition is met.
[0050] In different specific implementations, the methods or systems described in this application can be implemented in software, hardware, or a combination thereof. Furthermore, the order of the method steps can be changed, and various elements can be added, reordered, combined, omitted, or modified.
[0051] Obviously, the above embodiments of this application are merely examples for clearly illustrating this application, and are not intended to limit the implementation of this application, nor are they intended to limit this application. For those skilled in the art, other variations or modifications can be made based on the above description. The separate structural / functional modules or units can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. The structure and function of the separate components can be implemented as a combined structure or component. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of the claims of this application.
Claims
1. A method for public participation in urban renewal planning using resident intelligent agent reward feedback reinforcement learning, characterized in that, The method includes: Obtain geospatial data of the city to be planned, divide the geospatial data into multiple types of nodes, and construct edges based on the spatial adjacency relationships between different nodes to obtain an adjacency topology graph model. Based on the public participation corpus of the city to be planned, intelligent agents of residents with different roles are constructed using a large language model; The land use adjustment actions are generated by the reinforcement learning policy network based on the features of the adjacency topology graph model. The land use adjustment actions are input into the resident agent to obtain the evaluation results output by the resident agent, and the resident satisfaction reward value is calculated based on the evaluation results. The reinforcement learning strategy network is iteratively trained with the goal of maximizing the resident satisfaction reward value, and when the preset termination condition is met, the trained reinforcement learning strategy network is used to output the land use planning scheme for the city to be planned. The steps of constructing resident intelligent agents corresponding to different roles based on the public participation corpus of the city to be planned using a large language model include: Extract structured descriptive features corresponding to different role types from public participation corpus; The structured description features are mapped to role embedding vectors through an embedding network; The embedding vector of any role is mapped to a prefix matrix using a multilayer perceptron, and the prefix matrix is added as a prefix vector to each layer of the large language model. The parameters of the multilayer perceptron and the prefix matrix are updated while the parameters of the large language model are frozen to obtain the resident agent of that role type. The formula for calculating the prefix vector is as follows: in, For prefix vectors, It is a multilayer perceptron. Embed vectors for the character; The structured descriptive features of a character include at least the character's name, age, keywords related to interests, and decision-making weight factors. The steps of generating land use adjustment actions based on the features of the adjacency topology graph model using a reinforcement learning policy network include: The spatial feature vector is obtained by passing the adjacency topology graph model through a graph coding network; The population proportion weights corresponding to different roles, the role embedding vectors corresponding to different roles, and the semantic vectors of historical public participation corpus are used as public participation state features. The spatial feature vectors and public participation state features are input into the reinforcement learning policy network to generate the land use adjustment action. The semantic vectors of the historical public participation corpus are obtained by text semantic encoding based on the evaluation results output by the resident agent in the previous iteration.
2. The urban renewal public participation planning method based on resident intelligent agent reward feedback reinforcement learning according to claim 1, characterized in that, The various types of nodes include plot nodes, road segment nodes, and intersection nodes; the spatial adjacency relationships include the adjacency between plot nodes and intersection nodes, the adjacency between plot nodes and road segment nodes, and the adjacency between road segment nodes and intersection nodes.
3. The urban renewal public participation planning method based on resident intelligent agent reward feedback reinforcement learning according to claim 1, characterized in that, The training process of the resident agent includes mixing the structured description features of all roles, randomly sampling them, and calculating the loss based on a loss function with their corresponding role embedding vectors. The parameters of the resident agent are adjusted with the goal of minimizing the loss until a preset condition is met. The formula for calculating the loss function is as follows: in, The loss function; The input solution description text; The target text output to the agent includes satisfaction ratings, requests, and suggestions; Character type; A large language model with frozen parameters.
4. The urban renewal public participation planning method based on resident intelligent agent reward feedback reinforcement learning according to claim 1, characterized in that, The step of obtaining the evaluation results output by the resident intelligent agent and calculating the resident satisfaction reward value based on the evaluation results includes: The land use adjustment action is converted into natural language description text, and the natural language description text is input into the resident intelligent agent; Obtain the satisfaction score output by any resident agent. Based on the population proportion weight of the role corresponding to the resident agent in the city to be planned, the satisfaction scores output by different resident agents are weighted and summed to obtain the resident satisfaction reward value.
5. The urban renewal public participation planning method based on resident intelligent agent reward feedback reinforcement learning according to claim 1, characterized in that, The text opinions corresponding to the satisfaction scores and suggestions output by the resident intelligent agent in the previous iteration.
6. The urban renewal public participation planning method based on resident intelligent agent reward feedback reinforcement learning according to claim 1, characterized in that, The step of iteratively training the reinforcement learning policy network with the objective of maximizing the resident satisfaction reward value includes: The policy network is used to output action probabilities based on the public participation state characteristics, and the resident satisfaction reward value is obtained based on the resident agent. The value network is used to output the state value based on the public participation state characteristics; The advantage function is calculated based on the resident satisfaction reward value and state value. The optimization direction is to maximize the objective function. The parameters of the reinforcement learning policy network are updated synchronously using the advantage function until the mean of the resident satisfaction reward value is not lower than the first threshold and the fluctuation range of the resident satisfaction reward value is not higher than the second threshold; or the number of iteration steps reaches the preset maximum number of steps. The expression for the advantage function is as follows: ； in, For a moment The advantage function represents the degree of advantage of taking an action relative to the average level. For a moment The resident satisfaction reward value; This is a discount factor used to balance the weights of current and future rewards, and its value ranges from [value range missing]. ; The state value output by the value network represents the state value in the value network. The expected cumulative rewards in the future; The city's state in the next moment; The expression for the objective function is as follows: in, The pruning objective function optimized for the near-end strategy; These are the parameters of the policy network; This represents the probability ratio between the old and new strategies. The hyperparameters are used to limit the policy update step size. This indicates that for all time steps The mathematical expectation; For the clipping function, the probability ratio is restricted to... Within the range.
7. The urban renewal public participation planning method based on resident agent reward feedback reinforcement learning according to any one of claims 1 to 6, characterized in that, The reinforcement learning policy network also filters invalid actions based on an action validity mask when generating the land use adjustment action. The action legality mask is generated based on the feature representation extracted from the adjacency topology graph model and preset constraints, and is used to determine the candidate positions of facility layout actions between different nodes.
8. A public participation planning system for urban renewal using resident agent reward feedback reinforcement learning, applied to the public participation planning method for urban renewal using resident agent reward feedback reinforcement learning as described in any one of claims 1 to 7, characterized in that, The system includes: The graph model construction module is used to acquire geospatial data of the city to be planned, divide the geospatial data into multiple types of nodes, and construct edges based on the spatial adjacency relationships between different nodes to obtain an adjacency topology graph model. The agent construction module is used to construct multi-role intelligent agents based on public participation corpora of the city to be planned using a large language model. The evaluation and reward calculation module is used to generate land use adjustment actions based on the features of the adjacency topology graph model through a reinforcement learning policy network, input the land use adjustment actions into the agent, obtain the evaluation results output by the agent, and calculate the resident satisfaction reward value based on the evaluation results. The training and scheme generation module is used to iteratively train the reinforcement learning policy network with the goal of maximizing the resident satisfaction reward value, and obtain the land use planning scheme of the city to be planned when the preset termination condition is met.