Device for generating human movement track based on large language model agent

Abstract

The invention discloses a device for generating a human movement track based on a large language model agent, and the device comprises a personality generation module which is used for extracting personal information from historical data, and recognizing the personal information through the large language model agent, so as to obtain a habitual activity mode of a person; then analyzing personal character characteristics by combining a chain reasoning technology COT and performing verification and evaluation, defining personal habitual activity modes and the verified and evaluated character characteristics as personal characteristics, and using the personal characteristics as priori knowledge of a large language model agent for subsequent analysis; the motivation retrieval module is used for retrieving associated information from the historical data through an enhanced retrieval strategy; the big language model intelligent physical sign inference behavior motivation is guided in combination with the habitual activity mode and the personal characteristic information generated by the personality generation module; and the behavior generation module is used for guiding the large language model agent to generate behavior track data according to the contents generated by the personality generation module and the motivation retrieval module.

Description

Technical Field

[0001] The present invention relates to the field of trajectory generation based on a large language model intelligent agent, and in particular to a method for simulating and generating urban human behavior movement data by using a large language model intelligent agent. Background Art

[0002] Human mobility data has always been an important research topic. As a product of the interaction between humans and physical space and time and space, its influence is ubiquitous in human production activities. Many aspects such as urban transportation planning, scenic spot design, and commercial site selection require human activity data as a reference. However, in real life, because such data involves personal privacy issues, it is difficult to collect a large amount of real and reliable data through effective means. Therefore, designing a low-cost and high-quality human mobility data generation model has been a long-term research goal and pursuit direction.

[0003] Although previous data-driven learning methods such as VAE and GAN provide a variety of solutions for generating individual trajectories, these generated data only imitate real-world data from the perspective of data distribution, rather than from a semantic perspective. Therefore, they are less effective in simulating or explaining activities with significantly different distributions. At the same time, traditional models have certain limitations in understanding the intentions and contexts of human mobility. Especially when understanding human characteristics and needs, traditional models mostly focus on the temporal and spatial meanings behind the behavior, but do not think deeply about the human factors behind the behavior from a human perspective, which limits their effectiveness in generating real and context-rich behavioral data.

[0004] In the past few years, large language models have experienced rapid development. With their powerful logical reasoning ability and excellent role-playing performance, they have brought new research prospects to the field of behavior generation. By guiding the model's thinking process through carefully designed prompt information, its performance in reasoning and analysis tasks can be significantly improved. At the same time, relying on its excellent language generation ability, the model can simulate specific individuals and restore behavioral characteristics, thus making up for the lack of consideration of human factors in the behavior generation process of traditional models. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a device for generating human movement trajectories based on a large language model agent. The device is based on real personal data, combined with personalized generation technology and motivation retrieval technology, to create a large language model LLM agent that matches individual characteristics and has planning, reasoning and memory capabilities. The LLM agent is then guided to simulate the daily activities of urban residents and generate more realistic and effective personal travel data.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A device for generating a human movement trajectory based on a large language model agent, comprising:

[0008] The personality generation module is used to extract personal information from historical data and use the semantic perception ability of the large language model agent to identify personal information and obtain the individual's habitual activity pattern; then, the chain reasoning technology COT is used to analyze the individual's personality characteristics and verify and evaluate them, and the individual's habitual activity pattern and personality characteristics after verification and evaluation are defined as personal characteristics, and the personal characteristics are used as the prior knowledge of the large language model agent for subsequent analysis;

[0009] The motivation retrieval module is used to retrieve relevant information from historical data through enhanced retrieval strategies; and combines the personal characteristics generated by the personality generation module to guide the large language model to infer behavioral motivations through intelligent signs;

[0010] The behavior generation module is used to guide the large language model agent to generate behavior trajectory data based on the content generated by the personality generation module and the motivation retrieval module.

[0011] Furthermore, the workflow of the personality generation module includes:

[0012] (1) Data preprocessing: Extracting personal information from historical data. Personal information includes regular commuting distance, typical start and end times of daily trips, typical start and end locations of daily trips, and the places most frequently visited by individuals.

[0013] (2) Generate individual personality traits: Use personal information to construct prompts to guide the large language model agent to summarize the individual's habitual activity patterns; Based on the individual's historical behavior data and the habitual activity patterns obtained through analysis, the large language model agent further generates corresponding personality traits;

[0014] (3) Verify and evaluate personality traits: Select the current individual’s real m behavioral data from the historical behavioral data, extract a multi-dimensional behavioral feature vector X, and establish a mapping relationship g to map the personality trait N to the prediction of the behavioral feature vector X; based on the given behavioral feature vector X, adjust the personality trait N value to minimize the prediction error; finally, use the adjusted personality trait N and habitual activity pattern as input to drive the large language model agent to output the current individual’s personal characteristics, which include occupational identity, habitual activity pattern and personality trait description, and store the personal characteristics as prior knowledge in the memory module of the large language model agent.

[0015] Furthermore, the personality generation module uses the Big Five personality theory to analyze individual personality traits from five dimensions: openness, conscientiousness, extraversion, agreeableness, and emotional stability.

[0016] Furthermore, the behavioral feature vector X includes the individual’s location diversity index, activity density, functional type ratio, and repeat visit rate.

[0017] In the motivation retrieval module, the neural network model is trained using contrastive learning to retrieve date information from historical data and provide it as input to the large language model to generate personal behavior motivations. The specific steps are as follows:

[0018] First, we define temporal features to capture the differences between different dates, and use a parameterized neural network model to evaluate the similarity between any two dates.

[0019] Secondly, an unsupervised contrastive learning method is used to train the neural network model. After the training is completed, the trained neural network model is used to evaluate the similarity between a given query date and a historical date.

[0020] Thirdly, by retrieving the historical data most similar to the query date, these historical data are input into the large language model agent as prompts to generate a summary of the behavioral motivations corresponding to the individuals related to that period;

[0021] Finally, the scenario-based evaluation method selects the final reasonable behavioral motivation.

[0022] Furthermore, the scenario-based assessment method ensures the rationality of the behavior motivation in the following steps:

[0023] (1) Analyze the trigger points and interference factors of behavioral motivation and construct the corresponding virtual scene: After the large language model agent recognizes and generates the behavioral motivation, it first analyzes the specific trigger points of the behavioral motivation; then, it creates a virtual scene related to the behavioral motivation and describes it in detail in text form;

[0024] (2) Further observe the response of the large language model agent in this virtual scene and configure an environment description containing several interference elements for each behavioral motivation;

[0025] (3) Use behavioral motivations and virtual scenarios containing several interference elements as input to guide the large language model agent to generate new travel plans: Provide the planned behavioral motivations and virtual scenario descriptions to the large language model agent to guide it to regenerate the travel plan; By comparing the differences between the new and old travel plans, evaluate the similarity between the two; the high similarity indicates that the original behavioral motivation is consistent with the actual situation, verifying the reliability of the behavioral motivation.

[0026] Furthermore, when constructing a virtual scene, events or conditions that are closely related to behavioral motivations are introduced to enhance the practical application value of the virtual scene.

[0027] Furthermore, the generated behavior trajectory data can be used for urban traffic planning, commercial site selection and simulation of responding to sudden social situations.

[0028] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0029] 1. Personalization and enhanced semantic richness: By introducing the large language model LLM as a core component, the present invention can generate personalized activity trajectories based on individual characteristics, which not only reflects the true characteristics of individual behavior, but also accurately simulates the motivation of behavior, significantly improving the semantic richness and explanatory power of trajectory data. This overcomes the shortcomings of traditional deep learning methods in terms of the interpretability of results.

[0030] 2. Improved reliability and generalization ability: The motivation retrieval enhancement strategy proposed in this paper can generate more reliable simulation data by extracting multi-dimensional information (such as personal information and behavioral data) from historical data and combining it with the deep understanding ability of LLM. In addition, through virtual scene construction and behavior verification methods, the rationality and generalization ability of generated data in complex situations are ensured, thereby meeting the diverse needs in unforeseen scenarios.

[0031] 3. Enhanced scene adaptation and practicality: This invention constructs a scenario-based virtual scene to analyze and evaluate the reliability of behavioral motivations, and through repeated adjustments, ensures that the model's behavioral output meets actual needs. This scene adaptation capability makes the generated data more valuable in practical applications such as urban traffic planning, commercial site selection, and response to sudden social situations.

[0032] 4. Innovative framework design: This paper constructs a comprehensive framework. Before generating activity trajectories, the personality generation module extracts individual characteristics and verifies personality characteristics. In the motivation retrieval module, the contrastive learning technology is combined to further explore the behavior motivation. Finally, the behavior generation module outputs highly reliable and effective trajectory data. This method not only improves the theoretical framework of existing generation technology, but also significantly improves the practicality of the model.

[0033] 5. Potential for wide application: The generated activity trajectory data has important commercial prospects in the fields of urban traffic planning, commercial site selection, and social simulation. Especially in scenarios such as personalized recommendation systems and urban behavior analysis, the generation framework of the present invention can provide simulation data that is closer to reality, providing strong support for multi-field research and application.

[0034] In summary, the technical solution of the present invention achieves an innovative breakthrough in activity trajectory generation by introducing a large language model intelligent agent and combining personality generation and motivation retrieval strategies. It has significant theoretical value and broad application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the overall structure of the device of the present invention.

[0036] Figure 2 Figure 2. Experimental results obtained based on the device of the present invention and other models. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0038] See Figure 1 This embodiment provides a device for generating a human movement trajectory based on a large language model (LLM) agent, including:

[0039] The personality generation module is used to extract personal information from historical data and use the semantic perception ability of the large language model agent to identify personal information and obtain the individual's habitual activity pattern; then, the chain reasoning technology COT is used to analyze the individual's personality characteristics and verify and evaluate them, and the individual's habitual activity pattern and personality characteristics after verification and evaluation are defined as personal characteristics, and the personal characteristics are used as the prior knowledge of the large language model agent for subsequent analysis;

[0040] The motivation retrieval module is used to retrieve relevant information from historical data through enhanced retrieval strategies; and combines the habitual activity patterns and personal characteristic information generated by the personality generation module to guide the large language model to infer behavioral motivations through intelligent physical signs;

[0041] The behavior generation module is used to guide the large language model agent to generate behavior trajectory data based on the content generated by the personality generation module and the motivation retrieval module.

[0042] Specifically, the workflow of the personality generation module includes:

[0043] (1) Data preprocessing: Extracting personal information from historical data. Personal information includes regular commuting distance, typical start and end times of daily trips, typical start and end locations of daily trips, and the places most frequently visited by individuals.

[0044] (2) Generate individual personality traits: Use personal information to construct prompts to guide the large language model agent to summarize the individual's habitual activity patterns; Based on the individual's historical behavior data and the habitual activity patterns obtained through analysis, the large language model agent further generates corresponding personality traits;

[0045] (3) Verify and evaluate personality traits: Select the current individual’s real m behavioral data from the historical behavioral data, extract a multidimensional behavioral feature vector X, and establish a mapping relationship g to map personality traits N to the prediction of behavioral feature vector X; based on the given behavioral feature vector X, adjust the personality trait N value to minimize the prediction error; finally, use the adjusted personality trait N and habitual activity patterns as input to drive the large language model agent to output the current individual’s personal characteristics, which include occupational identity, habitual activity patterns, and personality trait descriptions, and store the personal characteristics as prior knowledge in the memory module of the large language model agent. The behavioral feature vector X includes the individual’s location diversity index, activity density, functional type ratio, and repeat visit rate.

[0046] This example uses the Big Five personality theory as a basis and requires LLM to comprehensively analyze individual personality traits from five dimensions: openness, conscientiousness, extraversion, agreeableness, and emotional stability.

[0047] Specifically, the motivation retrieval module is as follows: Figure 1 The overall process shown includes three parts: (1) training the neural network model (this implementation uses a multi-layer perceptron model). (2) retrieving similar behavior data from the historical database based on the neural network model. (3) generating behavior motivations based on the obtained data, and selecting the final reasonable behavior motivations using a scenario-based evaluation method.

[0048] The steps for training the neural network model are as follows:

[0049] (101) Construct a training dataset.

[0050] For each individual’s trajectory, similar (positive samples) and dissimilar (negative samples) date pairs are identified through a predefined similarity scoring function.

[0051] The similarity score is calculated by the following formula:

[0052]

[0053] Among them, T d is the total number of time intervals in a day (for example, one time interval is every 10 minutes). da (t) represents the activity trajectory A at time t da Current location. The similarity score is calculated by comparing the number of identical locations in two activity trajectories. Subsequently, the similarity score of the positive pair is the highest, indicating a higher similarity between these trajectories; conversely, the similarity score of the negative pair is lower, indicating a weaker similarity between them.

[0054] (102) Trained by unsupervised contrastive learning. The details are as follows:

[0055] For each date d, generate one positive sample pair and m negative sample pairs, and calculate the similarity score represents the similarity score between date d and its positive sample pair, Represents the similarity score between date d and a negative sample pair

[0056] Input these positive and negative sample pairs into the neural network model, and use the current date as the anchor point to calculate the Euclidean distance between these sample pairs and the anchor point;

[0057] The triplet loss function is used as the contrast loss function to train the neural network model by ensuring that the positive samples are closer to the anchor point and the negative samples are farther away from the anchor point based on the distance difference between the positive and negative samples.

[0058] In this embodiment, the scenario-based evaluation steps in the motivation retrieval module are as follows:

[0059] After training the neural network model, it can be applied to evaluate the similarity between any query date and historical dates. This enables the framework to retrieve the historical data that is most similar to the query date and provide this data to the LLM agent to generate a summary of the behavioral motivations for that period.

[0060] Then analyze the behavioral motivation triggers and interference factors, and construct corresponding scenarios: After identifying and confirming a specific motivation, first analyze the specific trigger points of the behavioral motivation. Subsequently, create a virtual scene related to the behavioral motivation and describe it in detail in text form. When designing a virtual scene, events or conditions closely related to the motivation should be carefully introduced to enhance the practical application value of the scene. For example, if the motivation is "finding a quiet environment to think", you can set a scene where a usually quiet cafe suddenly becomes noisy due to a band performance. Further observe the reaction of the LLM agent in this virtual scene. If the LLM agent chooses to migrate to other quiet environments, this shows that the original motivation does reflect the original intention of the LLM agent. Based on this, an environmental description containing specific interference elements is configured for each travel motivation, providing a solid foundation for in-depth analysis and understanding of the agent's behavior.

[0061] Using behavioral motivations and interference scenarios as inputs, guide the LLM agent to generate new travel plans: Provide the LLM agent with carefully designed behavioral motivations and scenario descriptions to guide it to regenerate travel plans. By comparing the differences between the new and old plans, the similarity between the two is evaluated. A high degree of similarity indicates that the original motivation is highly consistent with the actual situation, thus verifying the reliability of the travel motivation. This process not only tests the applicability of the motivation, but also evaluates the agent's ability to cope with complex situations.

[0062] The working process of the behavior generation module is as follows: first, the results of the personality generation module and the motivation retrieval module are spliced. Then the spliced ​​results are input to the LLM agent, allowing the LLM agent to generate a path.

[0063] The behavior generation module combines the personal characteristics provided by the personality generation module and the behavior motivation inferred by the motivation retrieval module to guide the LLM agent to generate the final behavior activity trajectory, which is expressed as A d =LLM (Personal Profile, Motivation). This method makes full use of the advantages of LLM in human intelligence simulation, taking into account a variety of conditions, and the generated results are not only consistent with the actual situation to the greatest extent, but also have strong explanatory power and good generalization ability. The generated results include not only the final behavioral data, but also the agent's explanation of the day's activity arrangements. By providing explanations, the rationality of the travel trajectory can be effectively evaluated. In addition, based on the powerful understanding and reasoning ability of the LLM agent, the present invention can also simulate travel data under some extreme conditions, such as epidemics or natural disasters that cannot be reproduced in reality. Through this framework, the distribution of behavioral trajectories related to these special situations can be generated, providing an effective simulation tool for dealing with complex and sudden social conditions.

[0064] According to the above implementation steps, the specific experimental results are as follows Figure 2 As shown, this experiment selected the following four key indicators for evaluation: (1) Temporal distance (TI): This indicator is used to measure the time interval between two consecutive locations in the trajectory. (2) Number of daily activities (AN): This indicator is used to count the number of activities per day on an individual trajectory. (3) Daily activity routine distribution (AD): This indicator is used to evaluate the activity distribution of each trajectory. (4) Visit distribution (VD): This indicator is used to analyze the spatial-temporal distribution characteristics of the visited locations within each trajectory. After extracting the above features from the generated and real trajectory data, this study used the Jensen-Shannon divergence (JSD) to quantify the differences between these features. A lower JSD value indicates a smaller difference, which means that the model has a better prediction effect.

[0065] The present invention is not limited to the embodiments described above. The above description of the specific embodiments is intended to describe and illustrate the technical solution of the present invention. The above specific embodiments are merely illustrative and not restrictive. Without departing from the scope of the present invention and the scope of protection of the claims, a person of ordinary skill in the art can also make many forms of specific changes under the guidance of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A device for generating human movement trajectories based on a large language model agent, characterized in that: include: The personality generation module is used to extract personal information from historical data and use the semantic perception ability of the large language model agent to identify personal information and obtain the individual's habitual activity pattern; then, the chain reasoning technology COT is used to analyze the individual's personality characteristics and verify and evaluate them, and the individual's habitual activity pattern and personality characteristics after verification and evaluation are defined as personal characteristics, and the personal characteristics are used as the prior knowledge of the large language model agent for subsequent analysis; The motivation retrieval module is used to retrieve relevant information from historical data through enhanced retrieval strategies; and guide the large language model to infer behavioral motivations through intelligent signs combined with personal characteristics generated by the personality generation module; The behavior generation module is used to guide the large language model agent to generate behavior trajectory data based on the content generated by the personality generation module and the motivation retrieval module.

2. The device for generating human movement trajectories based on a large language model agent according to claim 1, characterized in that: The workflow of the personality generation module includes: (1) Data preprocessing: Extracting personal information from historical data. Personal information includes regular commuting distance, typical start and end times of daily trips, typical start and end locations of daily trips, and the places most frequently visited by individuals. (2) Generate individual personality traits: Use personal information to construct prompts to guide the large language model agent to summarize the individual's habitual activity patterns; Based on the individual's historical behavior data and the habitual activity patterns obtained through analysis, the large language model agent further generates corresponding personality traits; (3) Verify and evaluate personality traits: Select the current individual’s real m behavioral data from the historical behavioral data, extract a multi-dimensional behavioral feature vector X, and establish a mapping relationship g to map the personality trait N to the prediction of the behavioral feature vector X; based on the given behavioral feature vector X, adjust the personality trait N value to minimize the prediction error; finally, use the adjusted personality trait N and habitual activity pattern as input to drive the large language model agent to output the current individual’s personal characteristics, which include occupational identity, habitual activity pattern and personality trait description, and store the personal characteristics as prior knowledge in the memory module of the large language model agent.

3. The device for generating human movement trajectory based on a large language model agent according to claim 1 or 2, characterized in that: The personality generation module uses the Big Five personality theory to analyze individual personality traits from five dimensions: openness, conscientiousness, extraversion, agreeableness, and emotional stability.

4. The device for generating human movement trajectory based on a large language model agent according to claim 2, characterized in that: The behavioral feature vector X includes the individual’s location diversity index, activity density, functional type ratio, and repeat visit rate.

5. The device for generating human movement trajectory based on a large language model agent according to claim 1, characterized in that: In the motivation retrieval module, the neural network model is trained using contrastive learning to retrieve date information from historical data and provide it as input to the large language model to generate personal behavior motivations. The specific steps are as follows: First, we define temporal features to capture the differences between different dates, and use a parameterized neural network model to evaluate the similarity between any two dates. Secondly, an unsupervised contrastive learning method is used to train the neural network model. After the training is completed, the trained neural network model is used to evaluate the similarity between a given query date and a historical date. Thirdly, by retrieving the historical data most similar to the query date, these historical data are input into the large language model agent as prompts to generate a summary of the behavioral motivations corresponding to the individuals related to that period; Finally, the scenario-based evaluation method selects the final reasonable behavioral motivation.

6. The device for generating human movement trajectory based on a large language model agent according to claim 5, characterized in that: The steps of scenario-based assessment method to ensure the rationality of behavioral motivation are as follows: (1) Analyze the trigger points and interference factors of behavioral motivation and build corresponding virtual scenarios: After the large language model agent recognizes and generates behavioral motivation, the specific trigger points of the behavioral motivation are first analyzed; Then, create a fictitious scenario related to the motivation for the behavior and describe it in detail in text form; (2) Further observe the response of the large language model agent in this virtual scene and configure an environment description containing several interference elements for each behavioral motivation; (3) Using behavioral motivations and virtual scenes containing several interference elements as input, guide the large language model agent to generate a new travel plan: Provide the planned behavioral motivations and virtual scene descriptions to the large language model agent to guide it to regenerate the travel plan; By comparing the differences between the new and old travel plans, the similarity between the two is evaluated; the high similarity indicates that the original behavioral motivation is consistent with the actual situation, verifying the reliability of the behavioral motivation.

7. The device for generating human movement trajectory based on a large language model agent according to claim 6, characterized in that: When constructing a virtual scene, events or conditions that are closely related to behavioral motivations are introduced to enhance the practical application value of the virtual scene.

8. The device for generating human movement trajectory based on a large language model agent according to claim 1, characterized in that: The generated behavior trajectory data can be used for urban traffic planning, commercial site selection and simulation of responding to sudden social situations.

Images