Cross-scale urban population movement behavior generation method and model
By using a cross-scale urban population movement behavior generation method, and utilizing a flow-based trajectory generator and a multi-scale discriminator, the problem that existing models cannot fully capture the complexity of individuals and groups is solved, and high-precision urban population movement simulation is achieved.
Patent Information
- Application Number
- CN202510123171.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-26
- Publication Date
- 2025-12-16
- Estimated Expiration
- 2045-01-26
AI Technical Summary
Existing urban population movement models fail to fully capture the heterogeneity and complexity of individual movement, neglecting the interaction between individuals and groups and the influence of the environment, resulting in insufficient realism and practicality of simulation data.
A cross-scale urban population mobility behavior generation method is adopted, which utilizes a flow-based trajectory generator, a multi-scale discriminator, and a collaborative evaluator to generate mobility trajectories that conform to individual and group patterns by learning individual and group mobility patterns. The accuracy of the generated model is improved by multi-scale discrimination and optimization algorithms.
It generates realistic human trajectories that conform to individual movement patterns and group traffic patterns, accurately reflecting the rhythm of urban activities, improving the realism and practicality of simulation data, and enabling a more comprehensive capture of the complexity of urban movement.
Smart Images

Figure CN120046789B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of urban population movement modeling and prediction, and particularly relates to a cross-scale urban population movement behavior generation method and model. BACKGROUND
[0002] Human movement is crucial for the functioning of cities. It connects various socio-economic activities across different areas, while also causing traffic congestion and air pollution, and facilitating the spread of infectious diseases. These dynamic processes reflect two aspects of urban human movement. First, the daily individual movement of each person in time and space, i.e., trajectories, reveals the way each individual lives and consumes services in the city. Second, the large-scale movement flows of the urban population as a whole constitute the daily rhythm of urban activities, which has important implications for individual life and environmental quality.
[0003] Previous studies have shown that the distribution of individual movement waiting time and jump length exhibits a long-tail feature, which suggests the relevance of the Continuous Time Random Walk (CTRW) model. Subsequent studies have found that there is a repeated visit pattern to a small number of high-frequency visited locations in human trajectories, which indicates that in addition to the above scale-free random walk, there is also a periodic visit pattern. This finding inspired the design of the famous Exploration and Preferential Return (EPR) model, which describes human movement as two simple competing mechanisms: exploring new locations and preferring to return to previously visited locations. In contrast to individual movement mainly relying on personal preferences, the anisotropic movement between urban areas is mainly influenced by the trade-off between relative accessibility and transportation cost, which can be well captured by the gravity law. In addition, the distance-frequency power law of population movement further allows for fine prediction of the flow associated with different visit frequencies.
[0004] Although existing physical models and deep learning methods have been accepted in applications, the simplified descriptions they use cannot fully capture the complexity of urban movement. First, empirical data show that individual travel habits have high heterogeneity, such as the long-tail radius distribution of trajectories. However, the universal mechanism in these models (e.g., a single exploration or return process in the EPR model) cannot capture the similarity and diversity of human trajectories. Second, when simulating the impact of the urban environment on population movement, more factors and more complex relationships should be considered in addition to the simple linear relationship between population and distance in the gravity model.
[0005] Therefore, how to consider effective modeling of the movement complexity of large-scale cities to improve the performance of urban movement simulation systems and improve the fidelity and practicality of simulation data is a problem to be solved. SUMMARY
[0006] Therefore, the present application provides a cross-scale urban population movement behavior generation method and model to solve at least one of the above-mentioned problems.
[0007] To achieve the above-mentioned purpose, the present application adopts the following scheme:
[0008] According to a first aspect of the present application, a cross-scale urban population movement behavior generation method is provided, which is executed by a behavior generation model, the generation model comprising a flow-based trajectory generator, a multi-scale discriminator and a collaborative evaluator, the method comprising: obtaining real first individual movement trajectories and corresponding first group movement flows of urban population; the trajectory generator learns individual movement patterns and group flow rules by using the first individual movement trajectories and the first group movement flows, and generates simulated second individual movement trajectories; the second individual movement trajectories are aggregated into simulated second group movement flows; the multi-scale discriminator compares the first individual movement trajectories and the second individual movement trajectories, the first group movement flows and the second group movement flows from individual level and group level respectively, and gives reward signals from individual level and group level based on the comparison results; the collaborative evaluator evaluates the degree of conformity of the second individual movement trajectories to individual movement patterns and the second group movement flows to group flow rules respectively according to the reward signals, and gives an advantage function of a proximal policy optimization algorithm according to the evaluation results; the model parameters of the trajectory generator are updated according to the advantage function and the reward signals.
[0009] As an embodiment of the present application, the behavior generation model in the above-mentioned method comprises a first optimization target and a second optimization target, the first optimization target is used to generate a real movement decision strategy similar to the extracted experience strategy in statistics; the second optimization target is used to minimize the error between the generated flow and the ground truth data;
[0010] The function of the first optimization target is as follows:
[0011]
[0012] The function of the second optimization target is as follows:
[0013]
[0014] Wherein: L dist represents the distance between two distributions; L error represents the loss of prediction result; P data represents the real individual trajectory distribution; F t,data represents the real group flow distribution; (l t |x <t ) represents the given historical movement x <tIn the case that the first individual moves to location l at time t t .
[0015] As an embodiment of the present application, in the above method, the trajectory generator learns individual movement patterns and population flow rules using the first individual movement trajectory and the first population movement flow, and generates a simulated second individual movement trajectory, including: inputting the first individual movement trajectory into a gated recurrent unit network to learn the representation of the individual historical movement trajectory; converting the first population movement flow into a regional-level flow distribution; based on the representation of the individual historical movement trajectory and the regional-level flow distribution, generating the second individual movement trajectory in two stages, where the first stage determines the region, and the second stage selects a specific location within the determined region.
[0016] As an embodiment of the present application, in the above method, the multi-scale discriminator compares the first individual movement trajectory and the second individual movement trajectory, the first population movement flow and the second population movement flow from the individual level and the population level respectively, and gives individual-level and population-level reward signals based on the comparison results, including: the individual-level discriminator in the multi-scale discriminator inputs the first individual movement trajectory and the second individual movement trajectory into a network composed of a gated recurrent unit and a multi-layer perception, and obtains corresponding individual-level reward signals; the population-level discriminator in the multi-scale discriminator compares the first population movement flow and the second population movement flow, and obtains corresponding population-level reward signals by comparing the average error of the two flow matrices.
[0017] As an embodiment of the present application, in the above method, the collaborative evaluator respectively evaluates the degree of conformity of the generated second individual movement trajectory to individual movement patterns, and the generated second population movement flow to population flow rules according to the reward signals, and gives the advantage function of the proximal policy optimization algorithm according to the evaluation results, including: the individual evaluation network in the collaborative evaluator learns the true value function of the predicted state-action, evaluates the matching degree of the generated second individual movement trajectory and the individual movement patterns in the experience data based on the Monte Carlo method, and outputs the individual value; the population evaluation network in the collaborative evaluator evaluates the degree of conformity of the generated second population movement flow to the population flow rules in the experience data, decouples the joint space optimization by decomposing the flow authenticity evaluation value of the whole population into the sum of the population evaluation network of each individual, and outputs the population value; the advantage function of the proximal policy optimization algorithm is given according to the individual value and the population value.
[0018] According to a second aspect of the present application, a cross-scale urban population movement behavior generation model is provided, which comprises a data acquisition unit, an aggregation unit, a parameter updating unit, a flow-based trajectory generator, a multi-scale discriminator and a collaborative evaluator; the data acquisition unit is configured to acquire real first individual movement trajectories and corresponding first group movement flows; the trajectory generator is configured to learn individual movement patterns and group flow rules by using the first individual movement trajectories and the first group movement flows, and generate simulated second individual movement trajectories; the aggregation unit is configured to aggregate the second individual movement trajectories into simulated second group movement flows; the multi-scale discriminator is configured to compare the first individual movement trajectories and the second individual movement trajectories, the first group movement flows and the second group movement flows from individual and group levels respectively, and give reward signals from individual and group levels based on the comparison results; the collaborative evaluator is configured to evaluate the degree of conformity of the second individual movement trajectories to individual movement patterns and the second group movement flows to group flow rules respectively according to the reward signals, and give an advantage function of a proximal policy optimization algorithm according to the evaluation results; and the parameter updating unit is configured to update model parameters of the trajectory generator according to the advantage function and the reward signals.
[0019] As an embodiment of the present application, the behavior generation model comprises a first optimization target and a second optimization target, the first optimization target is configured to generate a real movement decision strategy similar to the extracted experience strategy in statistics, and the second optimization target is configured to minimize the error between the generated flow and the ground truth data.
[0020] The function of the first optimization target is as follows:
[0021]
[0022] The function of the second optimization target is as follows:
[0023]
[0024] Wherein: L dist represents the distance between two distributions; L error represents the loss of the prediction result; P data represents the real individual trajectory distribution; F t,data represents the real group flow distribution; (l t |x <t ) represents the probability of visiting location l <t at time t given the historical movement x t .
[0025] As an embodiment of the present application, the trajectory generator comprises: a trajectory learning module for inputting the first individual trajectory into a gated recurrent unit network to learn a representation of the individual historical trajectory; a flow conversion module for converting the first population movement flow into a regional level flow distribution; and a trajectory generation module for generating the second individual trajectory in two stages based on the representation of the individual historical trajectory and the regional level flow distribution, wherein the first stage determines a region and the second stage selects a specific location within the determined region.
[0026] As an embodiment of the present application, the multi-scale discriminator comprises an individual level discriminator and a population level discriminator: the individual level discriminator is configured to input the first individual trajectory and the second individual trajectory into a network composed of a gated recurrent unit and a multi-layer perceptron respectively, and obtain corresponding individual level reward signals respectively; and the population level discriminator is configured to compare the first population movement flow and the second population movement flow, and obtain a corresponding population level reward signal by comparing the average error of the two flow matrices.
[0027] As an embodiment of the present application, the collaborative evaluator comprises an individual evaluation network, a population evaluation network and an advantage function obtaining module; the individual evaluation network is configured to evaluate the matching degree between the generated second individual trajectory and the individual movement pattern in the experience data based on the Monte Carlo method by learning to predict the true value function of the state-action, and output an individual value; the population evaluation network is configured to evaluate the conformity between the generated second population movement flow and the population flow rule in the experience data, and decouple the joint space optimization by decomposing the flow authenticity evaluation value of the whole population contribution into the sum of each individual population evaluation network, and output a population value; and the advantage function obtaining module is configured to give an advantage function of a proximal policy optimization algorithm according to the individual value and the population value.
[0028] According to a third aspect of the present application, an electronic device is provided, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.
[0029] According to a fourth aspect of the present application, a computer readable storage medium is provided, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the above method.
[0030] From the above technical solutions, the cross-scale urban population movement behavior generation method and model provided by the application overcomes the limitations of the simplified description of the existing model, can more comprehensively capture the complexity of urban movement, including the heterogeneity of individual movement, the interaction between individuals, and the complex relationship between individuals and the environment. It not only can generate real human trajectories that conform to individual movement patterns (including spatial movement rules and time diurnal preference), but also can accurately reflect group flow and accurately present the daily rhythm of urban activities. BRIEF DESCRIPTION OF DRAWINGS
[0031] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, a brief introduction will be given below to the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings. In the drawings:
[0032] Figure 1 is a flowchart of a cross-scale urban population movement behavior generation method provided by an embodiment of the application;
[0033] Figure 2 is a flowchart of a trajectory generator generating simulated second individual movement trajectories provided by an embodiment of the application;
[0034] Figure 3 is a structural diagram of a trajectory generator based on flow provided by an embodiment of the application;
[0035] Figure 4 is a flowchart of a multi-scale discriminator giving individual-level and group-level reward signals provided by an embodiment of the application;
[0036] Figure 5 is a function design diagram of a trajectory discriminator and a flow discriminator provided by an embodiment of the application;
[0037] Figure 6 is a flowchart of a synergistic evaluator giving an advantage function of a proximal policy optimization algorithm provided by an embodiment of the application;
[0038] Figure 7 is a network architecture diagram of an individual evaluation network and a group evaluation network provided by an embodiment of the application;
[0039] Figure 8 is a whole framework diagram of a cross-scale urban population movement behavior generation method provided by an embodiment of the application;
[0040] Figure 9 is a structural diagram of a cross-scale urban population movement behavior generation model provided by an embodiment of the application;
[0041] Figure 10 is a structural schematic diagram of a trajectory generator provided by an embodiment of the present application;
[0042] Figure 11 is a structural schematic diagram of a multi-scale discriminator provided by an embodiment of the present application;
[0043] Figure 12 is a structural schematic diagram of a collaborative evaluator provided by an embodiment of the present application;
[0044] Figure 13 is a system structure schematic block diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION
[0045] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, further detailed descriptions of the embodiments of the present application are given below with reference to the drawings. Here, the illustrative embodiments of the present application and their descriptions are used to explain the present application, but are not used as limitations of the present application.
[0046] As Figure 1 shown is a flow schematic diagram of a cross-scale urban population movement behavior generation method provided by an embodiment of the present application, the method is executed by a behavior generation model, the generation model includes a traffic-based trajectory generator, a multi-scale discriminator and a collaborative evaluator, and the method includes the following steps:
[0047] Step S101: Obtain real first individual movement trajectories and corresponding first group movement traffic of urban population.
[0048] The first individual movement trajectories and the first group movement traffic collected in this step are a small amount of real data. Here, the small amount is relative to the subsequently generated simulated second individual movement trajectories. The individual movement trajectory data may, for example, include mobile phone positioning data, traffic card data or GPS data, etc., and the group movement traffic data may, for example, include road network traffic data, regional population density data, etc. After obtaining these data, they can be cleaned and preprocessed to ensure quality and reliability.
[0049] Step S102: Learn individual movement patterns and group traffic rules by using the first individual movement trajectories and the first group movement traffic, and generate simulated second individual movement trajectories.
[0050] The step is executed by a trajectory generator, which is a parameterized neural network that can learn individual movement patterns (spatial motion rules, diurnal preference in time) and group flow rules, that is, the trajectory generator of the present application not only considers individual preferences (for example, exploration and preference return in the EPR model), but also considers the influence of group flow, so it can overcome the limitations of ignoring individual heterogeneity and group influence in existing models. In addition, the trajectory generator of the present application adopts a hierarchical structure, first selects a region, and then selects a specific location, to solve the problem of scale mismatch.
[0051] Step S103: Aggregating the second individual movement trajectories into simulated second group movement flow.
[0052] The step is to aggregate the generated individual trajectories into group flow data, for example, counting by region or counting by road section, etc. The specific method of aggregation is not limited in the present application, and any existing aggregation method can be used to execute the step.
[0053] Step S104: Comparing the first individual movement trajectory and the second individual movement trajectory, the first group movement flow and the second group movement flow from the individual level and the group level respectively, and giving a reward signal from the individual level and the group level based on the comparison result.
[0054] The step is executed by a multi-scale discriminator, which evaluates the quality of the generated trajectory and gives a reward signal from the individual level (trajectory similarity) and the group level (flow similarity). It makes up for the shortcomings of existing deep learning methods mainly used for simple object modeling, so that the present application can handle complex data of city scale.
[0055] Step S105: According to the reward signal, the degree of conformity of the second individual movement trajectory to the individual movement pattern and the second group movement flow to the group flow rule is evaluated respectively, and the advantage function of the proximal policy optimization algorithm is given according to the evaluation result.
[0056] The step is executed by a collaborative evaluator, which combines the reward signals from the individual and group levels to calculate the advantage function of the PPO algorithm. The advantage function considers individual value (evaluating the matching degree of individual trajectory with experience data based on Monte Carlo method) and group value (evaluating the conformity of group flow with experience data), and is calculated through the advantage function. The core of the step is to establish the connection between individual level travel decision and group level flow, solving the problem of ignoring the interaction between individual and group in existing models.
[0057] Step S106: Updating the model parameters of the trajectory generator according to the advantage function and the reward signal.
[0058] From the above technical solutions, the cross-scale urban population movement behavior generation method provided by the application overcomes the limitations of the simplified description of the existing model, can more comprehensively capture the complexity of urban movement, including the heterogeneity of individual movement, the interaction between individuals, and the complex relationship between individuals and the environment. It not only can generate real human trajectories that conform to individual movement patterns (including spatial motion rules and time diurnal preferences), but also can accurately reflect group flow and accurately present the daily rhythm of urban activities.
[0059] Preferably, the behavior generation model includes a first optimization objective and a second optimization objective, wherein the first optimization objective is used to generate a real movement decision strategy similar to the extracted empirical strategy in statistics; and the second optimization objective is used to minimize the error between the generated flow and the ground truth data.
[0060] In the present application, the input of the behavior generation model is a small number of individual movement trajectories and group movement flow in the city, and the output is the full amount of movement trajectories and group movement flow in the city.
[0061] The function of the first optimization objective of the behavior generation model is as follows formula (1):
[0062]
[0063] The function of the second optimization objective is as follows formula (2):
[0064]
[0065] Wherein: L dist represents the distance between two distributions, for example, can be Kullback-Leibler divergence and Jensen-Shannon divergence; L error represents the loss of the prediction result, for example, can be mean square error, mean absolute error, etc.; P data represents the real individual trajectory distribution; F t,data represents the real group flow distribution; (l t |x <t ) represents the probability of visiting location l <t at time t given the history movement x t .
[0066] Preferably, as shown in Figure 2 , the trajectory generator in the step S102 learns individual movement patterns and group flow rules using the first individual movement trajectory and the first group movement flow, and generates simulated second individual movement trajectories, including:
[0067] Step S1021: input the first individual trajectory into a gated recurrent unit network to learn the representation of the individual historical trajectory.
[0068] Step S1022: convert the first population mobility flow into a regional-level flow distribution.
[0069] Step S1023: generate the second individual trajectory in two stages based on the representation of the individual historical trajectory and the regional-level flow distribution, where the first stage determines the region and the second stage selects a specific location within the determined region.
[0070] Specifically, the structure diagram of the trajectory generator based on flow provided by the embodiment of the present application can be seen from Figure 3 As can be seen from Figure 3 , the individual trajectory is represented by a sequence of locations, while the population flow is described with a coarser spatial resolution, i.e., how many people move from one region to another. To address the scale mismatch, the embodiment of the present application adopts a hierarchical generator, where in the first stage, the individual decides the next region to go to; in the second stage, the individual selects a location belonging to the region. Obviously, the region selection in the first stage is consistent with the spatial resolution of the population flow, i.e., the regional level. Therefore, the embodiment of the present application models the population-to-individual interaction in this stage.
[0071] The first stage and the second stage in the above steps will be described in detail below in combination with Figure 3
[0072] (1) First stage:
[0073] The model first receives historical trajectory data for a period of time as input, Figure 3 The historical trajectory data is processed by a gated recurrent unit (GRU) to extract feature information in the time series and generate a state representation. The output of the GRU is sent to a multi-head attention mechanism module, which includes multiple attention heads (Head 1, Head 2,..., Head k). Each attention head focuses on different aspects of the historical trajectory, capturing information at different time scales and patterns. The output of the multi-head attention mechanism includes an uncertainty variable (var) to represent the degree of uncertainty in the model's prediction of the region selection. "high" and "low" represent high uncertainty and low uncertainty, which reflects the model's ability to express the degree of confidence in the prediction result. Finally, the model outputs a probability distribution representing the likelihood of the individual moving to each region. This probability distribution is calculated based on the output of the multi-head attention mechanism and the uncertainty variable. Figure 3 Two example regions are shown in, where the high and low uncertainty regions are inFigure 3 It has also been embodied.
[0074] (2) Second stage:
[0075] The region information selected in the first stage is converted into vector representation by an embedding layer. Then the region embedding information is concatenated with the state representation generated in the first stage and processed by a GRU again to fuse the region information and the time series information. The output of the GRU is transformed by a multi-layer perceptron (MLP) to extract higher-level features. The output of the MLP is converted into a probability distribution by a Softmax layer, representing the likelihood of the individual moving to each location within the selected region. Figure 3 The next step location and the probability of each location are shown in the figure. The final output is the next step location and the next step action (the next step location and the region selection jointly determine the next step action).
[0076] Preferably, as shown in Figure 4 The multi-scale discriminator compares the first individual trajectory and the second individual trajectory, the first group moving flow and the second group moving flow from the individual level and the group level respectively in the step S104, and gives the reward signal of the individual level and the group level based on the comparison result, which includes:
[0077] Step S1041: The individual level discriminator in the multi-scale discriminator inputs the first individual trajectory and the second individual trajectory into a network composed of a gated recurrent unit and a multi-layer perceptron respectively, and obtains the corresponding individual level reward signal.
[0078] Step S1042: The group level discriminator in the multi-scale discriminator compares the first group moving flow and the second group moving flow, and obtains the corresponding group level reward signal by comparing the average error of the two flow matrices.
[0079] As can be seen from the above, the embodiment of the application defines two-level reward functions to comprehensively evaluate the generation results of the individual level and the group level, and all intermediate steps obtain zero reward. The two reward functions are modeled respectively, wherein the first individual level reward function is modeled by a neural network-based discriminator, and the second group level reward function is directly given. As shown in Figure 5 The design of the two-level functions is shown, and Figure 5It can be seen that the trajectory discriminator and the individual-level discriminator, the flow discriminator and the group-level discriminator are used. Among them: the trajectory discriminator is used to evaluate the authenticity of the generated individual trajectory. It receives real trajectory (Real) and generated trajectory (Fake) as input. The input data first passes through a Sigmoid layer, then passes through a multi-layer perceptron (MLP), and finally the features of the real trajectory and the generated trajectory are spliced (Concat), and finally a probability value is output to judge whether the generated trajectory is real. This part uses a neural network-based discriminator to judge the quality of the generated trajectory by learning the difference between the real trajectory and the generated trajectory. The flow discriminator is used to evaluate the authenticity of the generated group flow. It receives real flow and generated flow as input and calculates the distance (Distance) between them. Figure 5 The distribution of real flow and generated flow is shown in the form of images, and the accuracy of the generated flow is measured by calculating the distance between them. The reward function of this part is directly given without using neural network modeling, but directly calculating some distance measure (such as Euclidean distance or other appropriate distance measure) between real flow and generated flow. The smaller the distance, the closer the generated flow to the real flow, and the higher the reward.
[0080] It can be seen that, Figure 5 The multi-scale discriminator shown evaluates the generated results from individual and group levels respectively through the trajectory discriminator and the flow discriminator, and gives the corresponding reward signal to guide the training of the trajectory generator, and finally improves the authenticity and accuracy of the generated trajectory and group flow. Compared with the prior art, the quality of the generated results is more comprehensively evaluated, and the performance of the model is improved.
[0081] Preferably, as Figure 6 shown, the cooperative evaluator in step S105 respectively evaluates the degree of conformity of the generated second individual moving trajectory to the individual moving mode and the generated second group moving flow to the group flow rule according to the reward signal, and the advantage function of the proximal policy optimization algorithm according to the evaluation result can include:
[0082] Step S1051: The individual evaluation network in the cooperative evaluator learns to predict the true value function of the state-action, evaluates the matching degree of the generated second individual moving trajectory to the individual moving mode in the experience data based on the Monte Carlo method, and outputs the individual value.
[0083] Step S1052: The group evaluation network in the cooperative evaluator evaluates the degree of conformity of the generated second group moving flow to the group flow rule in the experience data, decouples the joint space optimization by decomposing the flow authenticity evaluation value of the whole population contribution into the sum of each individual group evaluation network, and outputs the group value.
[0084] Step S1053: giving the advantage function of the proximal policy optimization algorithm according to the individual value and the group value.
[0085] As shown in Figure 7 , which shows the network architecture of the individual evaluation network and the group evaluation network. The individual evaluation network shares the same state encoder with the trajectory generator, including an embedding layer and a GRU layer, to obtain the state representation, and then the embodiment of the application adopts an MLP to predict the individual state value, outputting the individual value. The group evaluation network is modeled by another MLP network, which receives the concatenation of the state embedding (O N ) and the action embedding (a N ) as input. Specifically, the state embedding is obtained through the network of the individual evaluation network, and the action embedding is obtained through the embedding layer.
[0086] The cross-scale urban population movement behavior generation method proposed by the application establishes the connection between the individual-level travel decision and the population-level aggregated flow. In order to strengthen the representation and learning of the movement model, the application designs the following two kinds of individual and population collaborative mechanisms.
[0087] I. Collaborative movement representation learning
[0088] In the city, people's actions are not independent of each other. The preferences of individuals and collective behaviors will affect the travel decisions of individuals. For example, the trajectory of an individual shows his / her daily routine activities, indicating that future travel behavior is influenced by historical activities. At the same time, individuals will occasionally visit some unfamiliar places recommended by friends, which indicates that there is social interaction between individuals. In order to collaboratively consider individual and population-level factors, the population movement model π θ (a t |o t ) is designed as a collaborative cooperation of two parameterized decision processes:
[0089]
[0090] where and represent two different decision strategies considering individual preferences and population influences, respectively. The first part captures the individual's preferences by learning the movement rules in the historical trajectory X <t , while the second part describes the influence of population flow on social interaction shaped by the urban environment. The collaboration between the two parts is designed to make a discrete choice between them according to the parameterized Bernoulli distribution Bernoulli(u t ), where Learning probability that represents individual's uncertainty in following their own preference. Specifically, the policy can be expressed by the following equation:
[0091]
[0092] If the individual has high uncertainty about the historical visits, it is more likely to follow the population-level decision Instead of personal preference This incorporation of state uncertainty enables each individual to flexibly and effectively utilize the collective decision.
[0093] II. Cooperative movement generation optimization
[0094] In addition to the representation of the population movement model, the present invention further enhances the cooperative learning method at the individual and population levels. Similar to GAIL, the optimization of π θ is guided by an evaluation network V φ , which evaluates the value behind each state-action pair. In order to encourage the policy network to cooperatively capture the travel patterns of individuals and the regularity of population flow, the present invention divides the evaluation network into the above-mentioned individual evaluation network and the group evaluation network, which respectively provide accurate estimates related to the current trajectory and the aggregated flow and The individual evaluation network learns to predict the correct value function of state-action based on whether π θ (a t |o t ) matches the individual movement pattern in the experience data. The goal is to obtain the experience average return through the Monte Carlo method.
[0095] For the evaluation of the authenticity of the flow that contributes to the whole population, it is not feasible to directly optimize the evaluation network, because the dimension of the joint action and observation is high. Therefore, in order to solve this problem, the present invention proposes to decouple the joint space optimization by decomposing the contribution of the whole population into the sum of the group evaluation network of each individual. Formally represented as:
[0096]
[0097] Then, the present invention uses Monte Carlo policy evaluation to optimize the above-mentioned
[0098] Finally, the individual evaluation network and the group evaluation network jointly evaluate the individual strategy, evaluate its ability to capture travel regularity and further impact on population travel, and provide the optimal direction for improving the current movement model, using the advantage function of the proximal policy optimization algorithm (PPO) for model optimization as follows:
[0099]
[0100] wherein η is the weight balancing the synergy of the two evaluators.
[0101] The overall framework diagram of the cross-scale urban population movement behavior generation method can also be seen from Figure 8 For Figure 8 It has been mentioned in the above description, and will not continue to be described here.
[0102] As can be seen from the above, the cross-scale urban population movement behavior generation method provided by the application overcomes the limitations of the simplified description of the existing model, and can more comprehensively capture the complexity of urban movement, including the heterogeneity of individual movement, the interaction between individuals, and the complex relationship between individuals and the environment. It not only generates real human trajectories that conform to individual movement patterns (including spatial motion rules and time diurnal preferences), but also accurately reflects group flow and accurately presents the daily rhythm of urban activities. In addition, the present application enhances the representation and learning ability of the model by establishing the link between individual-level travel decision-making and population-level aggregated flow, thereby improving the performance of the urban movement simulation system and the realism and practicality of the simulation data. Moreover, the present application uses a deep collaborative learning framework based on generative adversarial imitation learning, which realizes high-precision simulation of complex urban movement behavior through the collaborative work of three parameterized neural networks (flow-based trajectory generator, multi-scale movement discriminator, and collaborative evaluator). The quality evaluation provided by the discriminator and the application of the proximal policy optimization algorithm (PPO) further improve the performance of the trajectory generator. Finally, the present application considers multi-scale factors, can model at both the individual trajectory and group flow levels, and links the two through a collaborative mechanism, thereby more comprehensively reflecting the complexity of urban movement.
[0103] As Figure 9 The structure diagram of the cross-scale urban population movement behavior generation model provided by the embodiment of the application is shown, and the behavior generation model comprises a data acquisition unit 910, a flow-based trajectory generator 920, an aggregation unit 930, a multi-scale discriminator 940, a collaborative evaluator 950 and a parameter updating unit 960. Wherein:
[0104] The data acquisition unit 910 is used to acquire the first individual movement trajectory and the corresponding first group movement flow of the urban population.
[0105] The trajectory generator 920 is used to learn individual movement patterns and group flow rules by using the first individual movement trajectory and the first group movement flow, and to generate simulated second individual movement trajectories.
[0106] The aggregation unit 930 is used to aggregate the second individual movement trajectories into simulated second group movement flow.
[0107] The multi-scale discriminator 940 is configured to compare the first individual movement trajectory and the second individual movement trajectory, and the first crowd movement flow and the second crowd movement flow, respectively, from the individual level and the crowd level, and give a reward signal from the individual level and the crowd level based on the comparison results.
[0108] The collaborative evaluator 950 is configured to evaluate the degree of compliance of the second individual movement trajectory to the individual movement pattern, and the degree of compliance of the second crowd movement flow to the crowd flow rule, respectively, according to the reward signal, and give an advantage function of the proximal policy optimization algorithm according to the evaluation results.
[0109] The parameter updating unit 960 is configured to update the model parameters of the trajectory generator according to the advantage function and the reward signal.
[0110] Preferably, the behavior generation model comprises a first optimization target and a second optimization target, the first optimization target is configured to generate a real movement decision strategy similar to the extracted experience strategy in statistics, and the second optimization target is configured to minimize the error between the generated flow and the ground truth data.
[0111] The function of the first optimization target is as follows:
[0112]
[0113] The function of the second optimization target is as follows:
[0114]
[0115] Wherein: L dist represents the distance between two distributions; L error represents the loss of the prediction result; P data represents the real individual trajectory distribution; F t,data represents the real crowd flow distribution; (l t |x <t ) represents the probability of visiting location l <t at time t given the historical movement x t .
[0116] Preferably, as shown in Figure 10 , the trajectory generator 920 comprises:
[0117] A trajectory learning module 921 is configured to input the first individual movement trajectory into a gated recurrent unit network, and learn the representation of the individual historical movement trajectory.
[0118] A flow conversion module 922 is configured to convert the first crowd movement flow into a flow distribution at the regional level.
[0119] a trajectory generation module 923, configured to generate the second individual movement trajectory in two stages based on the representation of the individual historical movement trajectory and the regional level traffic distribution, wherein the first stage determines a region, and the second stage selects a specific location within the determined region.
[0120] Preferably, as shown in Figure 11 The multi-scale discriminator 940 includes an individual level discriminator 941 and a group level discriminator 942.
[0121] The individual level discriminator 941 is configured to input the first individual movement trajectory and the second individual movement trajectory into a network composed of a gated recurrent unit and a multi-layer perception, respectively, and obtain corresponding individual level reward signals.
[0122] The group level discriminator 942 is configured to compare the first group movement traffic and the second group movement traffic, and obtain a corresponding group level reward signal by comparing the average error of the two traffic matrices.
[0123] Preferably, as shown in Figure 12 The cooperative evaluator 950 includes an individual evaluation network 951, a group evaluation network 952, and an advantage function obtaining module 953.
[0124] The individual evaluation network 951 is configured to evaluate the matching degree between the generated second individual movement trajectory and the individual movement pattern in the experience data by learning to predict the true value function of the state-action, and output an individual value based on the Monte Carlo method.
[0125] The group evaluation network 952 is configured to evaluate the conformity between the generated second group movement traffic and the group traffic law in the experience data, and decouple the joint space optimization by decomposing the flow authenticity evaluation value of the whole population contribution into the sum of each individual group evaluation network, and output a group value.
[0126] The advantage function obtaining module 953 is configured to give an advantage function of a proximal policy optimization algorithm according to the individual value and the group value.
[0127] The detailed description of each unit and module can be referred to the corresponding description in the foregoing method embodiments, which will not be repeated here.
[0128] From the above technical solutions, the cross-scale urban population movement behavior generation model provided by the application overcomes the limitation of the simplified description of the existing model, can more comprehensively capture the complexity of urban movement, including the heterogeneity of individual movement, the interaction between individuals, and the complex relationship between individuals and the environment. It can not only generate real human trajectories consistent with individual movement patterns (including spatial movement rules and time diurnal preference), but also accurately reflect group flow and accurately present the daily rhythm of urban activities.
[0129] The embodiment of the application further provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the above method when executing the program.
[0130] The embodiment of the application further provides a computer readable storage medium, which stores a computer program for executing the above method.
[0131] As shown in Figure 13 , the electronic device 600 can also include a communication module 110, an input unit 120, an audio processor 130, a display 160, and a power supply 170. It should be noted that the electronic device 600 does not necessarily include all the components shown in Figure 13 ; in addition, the electronic device 600 can also include components not shown in Figure 13 , which can refer to prior art.
[0132] As shown in Figure 13 , the central processor 100, also known as a controller or operation control, can include a microprocessor or other processor device and / or logic device, which receives input and controls the operation of each component of the electronic device 600.
[0133] The memory 140, for example, can be one or more of a cache, a flash memory, a hard drive, a removable media, a volatile memory, a non-volatile memory or other suitable device. The above information related to failure can be stored, and programs for executing the information can also be stored. The central processor 100 can execute the programs stored in the memory 140 to achieve information storage or processing, etc.
[0134] The input unit 120 provides input to the central processor 100. The input unit 120 is, for example, a key or touch input device. The power supply 170 is used to provide power to the electronic device 600. The display 160 is used to display display objects such as images and text. The display can be, for example, an LCD display, but is not limited thereto.
[0135] The memory 140 can be a solid state memory such as a read only memory (ROM), a random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is turned off, can be selectively erased, and is provided with more data, an example of which is sometimes referred to as an EPROM, etc. The memory 140 can also be some other type of device. The memory 140 includes a buffer memory 141 (sometimes referred to as a buffer). The memory 140 can include an application / function storage 142 for storing application programs and function programs or a flow for executing the operation of the electronic device 600 by the central processing unit 100.
[0136] The memory 140 can also include a data storage 143 for storing data such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. A driver storage 144 of the memory 140 can include various drivers of the electronic device for a communication function and / or for performing other functions of the electronic device such as a messaging application, an address book application, etc.
[0137] The communication module 110 is a transmitter / receiver that transmits and receives signals via the antenna 111. The communication module 110 (transmitter / receiver) is coupled to the central processing unit 100 to provide input signals and receive output signals, which can be the same as in the case of a conventional mobile communication terminal.
[0138] Based on different communication technologies, a plurality of communication modules 110 such as a cellular network module, a Bluetooth module, and / or a wireless local area network module, etc. can be provided in the same electronic device. The communication module 110 (transmitter / receiver) is also coupled to the speaker 131 and the microphone 132 via the audio processor 130 to provide audio output via the speaker 131 and receive audio input from the microphone 132, thereby implementing a general telecommunication function. The audio processor 130 can include any suitable buffer, decoder, amplifier, etc. In addition, the audio processor 130 is also coupled to the central processing unit 100, thereby enabling recording on the local by the microphone 132 and enabling playing of a sound stored on the local by the speaker 131.
[0139] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code thereon.
[0140] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to an embodiment of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and a combination of flows and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for implementing functions specified in the flowchart and / or block diagram block or blocks.
[0141] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for implementing functions specified in the flowchart and / or block diagram block or blocks.
[0142] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a process for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 Figure 1 an apparatus for implementing functions specified in the flowchart and / or block diagram block or blocks.
[0143] The principles and implementation manners of the present application are described in the specific embodiments in the present application, and the above embodiment descriptions are only for helping to understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application ranges; in view of the above, the content of the present description should not be understood as a limitation of the present application.
Claims
1. A method for generating cross-scale urban population movement behavior, characterized in that, The method is executed by a behavior generation model, which includes a first optimization objective and a second optimization objective. The first optimization objective is used to generate a realistic movement decision strategy that is statistically similar to the extracted empirical strategy. The second optimization objective is used to minimize the error between the generated flow and the ground reality data. The function of the first optimization objective is as follows: min π (L dist (P data (l t |x <t ),v(l t |x <t )),L error (F t,data ,F(∏(l t |x <t )))); The function of the second optimization objective is as follows: Where: L dist L represents the distance between two distributions. error P represents the loss from the prediction result. data F represents the distribution of real individual trajectories; t,data Indicates the actual distribution of group traffic; (l t |x <t ) represents a given historical move x <t In the case of accessing location l at time t t The probability of; The method includes: Obtain the real individual movement trajectory of the city's population and the corresponding first group movement flow; Using the first individual's movement trajectory and the first group's movement flow, learn the individual's movement pattern and the group's flow pattern, and generate a simulated second individual's movement trajectory; The second individual's movement trajectory is aggregated into a simulated second group movement flow; The movement trajectories of the first individual and the second individual, and the movement flow of the first group and the movement flow of the second group are compared at the individual level and the group level, respectively, and reward signals are given at the individual level and the group level based on the comparison results. The degree of conformity between the generated second individual movement trajectory and the individual movement pattern, and the second group movement flow and the group flow pattern are evaluated based on the reward signal, and the advantage function of the near-end strategy optimization algorithm is given based on the evaluation results. The trajectory generation parameters of the behavior generation model are updated based on the advantage function and the reward signal; The step of learning individual movement patterns and group flow patterns by utilizing the first individual movement trajectory and the first group movement flow, and generating a simulated second individual movement trajectory, includes: The first individual's movement trajectory is input into a gated recurrent unit network to learn the representation of the individual's historical movement trajectory; Convert the first group's mobile traffic into a regional-level traffic distribution; Based on the representation of the individual's historical movement trajectory and the regional-level traffic distribution, the second individual's movement trajectory is generated in two stages, wherein the first stage determines the region, and the second stage selects a specific location within the determined region.
2. The method for generating cross-scale urban population movement behavior as described in claim 1, characterized in that, The step of comparing the first individual's movement trajectory and the second individual's movement trajectory, and the first group's movement flow and the second group's movement flow at both the individual and group levels, and providing reward signals at both the individual and group levels based on the comparison results, includes: The first individual's movement trajectory and the second individual's movement trajectory are respectively input into a network composed of a gated loop unit and a multilayer perceptron, and corresponding individual-level reward signals are obtained respectively. The mobile traffic of the first group and the mobile traffic of the second group are compared, and the corresponding group-level reward signal is obtained by comparing the average error of the two traffic matrices.
3. The method for generating cross-scale urban population movement behavior as described in claim 1, characterized in that, The process of evaluating the degree of conformity between the generated second individual movement trajectory and the individual movement pattern, and the second group movement flow and the group flow pattern, based on the reward signal, and providing the advantage function of the near-end strategy optimization algorithm based on the evaluation results, includes: By learning the state-action true value function and using the Monte Carlo method, the degree of matching between the generated second individual movement trajectory and the individual movement pattern in the empirical data is evaluated, and the individual value is output. The degree of conformity between the generated second group mobility flow and the group mobility patterns in the empirical data is evaluated. The joint space optimization is decoupled by decomposing the mobility authenticity assessment value of the entire population contribution into the sum of the group evaluation networks of each individual, and the group value is output. The advantage function of the proximal strategy optimization algorithm is given based on the individual value and the group value.
4. A cross-scale urban population movement behavior generation model, characterized in that, The behavior generation model includes a first optimization objective and a second optimization objective. The first optimization objective is used to generate a realistic movement decision strategy that is statistically similar to the extracted empirical strategy. The second optimization objective is used to minimize the error between the generated flow and the actual ground data. The function of the first optimization objective is as follows: min π (L dist (P data (L t |x <t ),π(l t |x <t )),L error (F t,data ,F(∏(l t |x <t )))); The function of the second optimization objective is as follows: Where: L dist L represents the distance between two distributions. error P represents the loss from the prediction result. data F represents the distribution of real individual trajectories; t,data Indicates the actual distribution of group traffic; (l t |x <t ) represents a given historical move x <t In the case of accessing location l at time t t The probability of; The behavior generation model includes: a data acquisition unit, an aggregation unit, a parameter update unit, a traffic-based trajectory generator, a multi-scale discriminator, and a collaborative evaluator; The data acquisition unit is used to acquire the real first individual movement trajectory of the urban population and the corresponding first group movement flow. The trajectory generator is used to learn individual movement patterns and group flow patterns by utilizing the first individual movement trajectory and the first group movement flow, and to generate a simulated second individual movement trajectory. The aggregation unit is used to aggregate the second individual movement trajectory into a simulated second group movement flow; The multi-scale discriminator is used to compare the first individual movement trajectory and the second individual movement trajectory, and the first group movement flow and the second group movement flow from the individual level and the group level, respectively, and to give reward signals at the individual level and the group level based on the comparison results. The collaborative evaluator is used to evaluate the degree of conformity between the generated second individual movement trajectory and the individual movement pattern, and the second group movement flow and the group flow pattern, respectively, based on the reward signal, and to give the advantage function of the near-end strategy optimization algorithm based on the evaluation results. The parameter update unit is used to update the model parameters of the trajectory generator according to the advantage function and the reward signal; The trajectory generator includes: The trajectory learning module is used to input the first individual's movement trajectory into the gated recurrent unit network and learn the representation of the individual's historical movement trajectory; The traffic conversion module is used to convert the mobile traffic of the first group into a regional-level traffic distribution. The trajectory generation module is used to generate the second individual's movement trajectory in two stages based on the representation of the individual's historical movement trajectory and the regional-level traffic distribution. The first stage determines the region, and the second stage selects a specific location within the determined region.
5. The cross-scale urban population movement behavior generation model as described in claim 4, characterized in that, The multi-scale discriminator includes an individual-level discriminator and a group-level discriminator: The individual-level discriminator is used to input the first individual's movement trajectory and the second individual's movement trajectory into a network composed of a gated loop unit and a multilayer perceptron, and obtain the corresponding individual-level reward signals respectively. The group-level discriminator is used to compare the first group mobile traffic and the second group mobile traffic, and obtain the corresponding group-level reward signal by comparing the average error of the two traffic matrices.
6. The cross-scale urban population movement behavior generation model as described in claim 4, characterized in that, The collaborative evaluator includes an individual evaluation network, a group evaluation network, and an advantage function acquisition module; The individual evaluation network is used to learn and predict the state-action true value function, and based on the Monte Carlo method, to evaluate the degree of matching between the generated second individual movement trajectory and the individual movement pattern in the experience data, and output the individual value. The group evaluation network is used to assess the degree of conformity between the generated second group mobility flow and the group flow pattern in the empirical data. It decouples the joint space optimization by decomposing the mobility authenticity assessment value contributed by the entire population into the sum of the group evaluation networks of each individual, and outputs the group value. The advantage function acquisition module is used to provide the advantage function of the near-end strategy optimization algorithm based on the individual value and the group value.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.
Citation Information
Patent Citations
Moving track generation method and device
CN113255951A
Crowd movement modeling method and device based on access location clustering
CN115994313A